PHM Society European Conference

Design Of Digital Twins for In-Service Support and Maintenance

Atuahene Barimah — 2024-06-27

This research aims to examine the challenges in developing Prognostics and Health Management (PHM) analytics for Digital Twin (DT) use cases in industrial applications, with a particular focus on Multi-Component Degradation (MCD) scenarios. A hybrid methodology, integrating physics-informed and data-driven models, is employed, using limited asset degradation data for model development. Preliminary work includes an analysis of the impact of data quality on Fault Detection and Isolation (FDI) algorithm performance, as well as the proposal of a weighted ensemble hybrid approach for assets experiencing MCD scenarios Preliminary results indicate enhanced diagnostics in asset health management through the use of Physics-Informed models for FDI in MCD scenarios with limited prior degradation data. Expected contributions for this research are the development of physics-informed PHM analytics for DT applications in MCD scenarios, adaptive PHM analytics for evolving asset lifecycles in DT applications, and interpretable DT model analytics for PHM in systems facing Multi-Component Degradation.

Development of a Data-driven Condition-Based Maintenance Methodology Framework for an Advanced Jet Trainer

Leonardo Baldo — 2024-06-27

Since their introduction more than 20 years ago, PHM strate- gies for aerospace equipment have gone a long way, enabling operators and Original Equipment Manufacturers (OEM) to monitor their assets, track down abnormal behaviors and plan maintenance action in advance. On the other hand, the tran- sition from PHM strategies using simulated data to solutions utilizing real-life operational data is consistently prone to sig- nificant challenges and demands. This doctoral thesis aims to develop a PHM/CBM framework applied to a Electro-Hydraulic Actuators (EHAs) leveraging real in-service fleet data. In this paper, the first steps of the research project are presented.

Digital Twin Development for Feed Drive Systems Condition Monitoring and Maintenance Planning

Himanshu Gupta — 2024-06-27

Current Prognosis and health management (PHM) technology suffers from challenges such as data availability, system interoperability, scalability, and transferability. In previous years, the PHM field has advanced a lot, but very few studies have been presented in which these challenges are addressed, and hence, PHM solutions are still confined to the lab environment. Digital Twin technology has the potential to address these challenges altogether and can add significant value to the PHM field. This thesis aims to develop an implementable Digital Twin framework for feed drive systems' condition monitoring and maintenance optimization, targeting these prevalent PHM challenges. The proposed framework will employ multiple physics-based models to generate synthetic data for different system states, configurations, and applications, and utilize this data with the help of machine learning to overcome the PHM challenges. The successful address of these challenges will pave the foundation in the direction of generalization of PHM solutions and also enhance the trustworthiness and reliability of PHM solutions.

Generating Realistic Failure Data for Predictive Maintenance: A Simulation and cGAN-based Methodology

Felix Waldhauser — 2024-06-27

Absence of failure data is a common challenge for data-driven predictive maintenance, particularly in the context of new or highly reliable systems. This is especially problematic for system level failure prediction of analog electronics since failure characteristics depend on the actual system layout and thus might change with system upgrades. To address this challenge, this work pursues a novel sim\-u\-la\-tion-as\-sist\-ed failure analysis methodology enabling automated and comprehensive evaluation of system level failure effects and failure detectability. While results obtained from simulations are suitable for comparative studies, they are confined to the simulation environment. To overcome this limitation, failure simulations are combined with generative models to generate realistic representations of missing failure data. Preliminary results demonstrate the capability of conditional generative adversarial networks (cGANs) to generate operational data of healthy systems, which accurately reflects correlations present in the source dataset. The proposed approach, using simulations as an additional source for generative models, not only targets the scarcity of failure data for highly reliable electronic systems but also ensures the adaptability of predictive maintenance algorithms to accommodate future system modifications and upgrades.

Machinery Fault Detection using Advanced Machine Learning Techniques

Dhiraj Neupane — 2024-06-27

Manufacturing industries are expanding rapidly, making it essential to detect early signs of machine faults for safety and productivity. With the extension of machines' runtime due to industrial automation, breakdown risks have increased, leading to economic and productivity consequences and sometimes even causalities. The surge in industrial big data from low-cost sensing technologies has enabled the development of intelligent data-driven Machinery Fault Detection (MFD) systems based on machine learning techniques in recent years. However, most existing methods are based on supervised pattern classification techniques to detect previously known fault types, which have limitations such as lack of generalization across different operational settings, focusing only on specific machinery and/or data types, and considering the identical and independent distribution of training and testing data. Therefore, my PhD research aims to develop a robust MFD framework for practical use by addressing these limitations.I will explore the potential of ensemble learning, unsupervised and semi-supervised anomaly detection, reinforcement learning, transfer learning, and cross-domain adaptation approaches in MFD. My PhD research will contribute to the field of data-driven MFD by proposing novel, effective solutions that can be applied across various manufacturing applications.

Natural Language Processing for Risk, Resilience, and Reliability

Jean Meunier-Pion — 2024-06-27

Natural Language Processing (NLP) has seen a surge in recent years, especially with the introduction of transformer architectures, relying on the now famous self-attention mechanism. Especially, with the rise of Large Language Models (LLM), propelled by the appearance of ChatGPT in 2022, a new hope of extracting relevant information from text has emerged. In the meantime, natural language data have not often been used in risk, resilience, and reliability tasks. However, text data containing reliability-related information, that can be used to monitor health information regarding complex systems, are available in several and diverse shapes. Indeed, text data can either contain theoretical expert knowledge (technical reports, documentation, Failure Modes and Effects Analysis (FMEA)), or in-practice expert knowledge (incident reports, maintenance work orders), or in-practice non-expert knowledge (customer feedback, news articles). Critical infrastructures, such as nuclear powerplants, railway networks, or electrical power grids, are complex systems for which any failure would induce severe consequences affecting many people. Such systems have the advantage of serving many users, thus having many possible text sources from which technical information and past incident data can be mined for anticipating future failures and generating responses to catastrophic scenarios. The goal of this work is to develop methods and apply state-of-the-art NLP techniques to text data relating to critical infrastructures and failures, to (1) mine information from unstructured language data, and (2) structure the extracted information. Preliminary experiments were conducted on customer review data and incident reports, and show promising performance for failure detection from text data with transformers, as well as incident-related information extraction using LLMs.

Prognostics of Remaining Useful Life for Aviation Structures Considering Imperfect Repairs

Mariana Salinas — 2024-06-27

Maintenance plays an important role in fulfilling the goals of
the Prognostics and Health Management (PHM) field. As of
now, no publication has addressed the impact of imperfect
repair actions from the prognostics perspective. Imperfect
repairs introduce complexities, altering system degradation
processes and increasing prediction uncertainties, thereby impacting
the accuracy of Remaining Useful Life (RUL) predictions.
To fill this gap in the literature, the study proposes developing
a robust prognostic model adaptable to post-repair
operations. The prognostic model that will be developed is
stochastic since stochastic models have already proven their
adaptability to unseen test data. However, further development
of such models is needed to deal with data on repaired
systems. In addition to that, the implementation of a Bayesian
Extension allows uncertainty interpretability to be considered
to account for the uncertainty coming from the repair action
itself but also from the different sources of uncertainties that
have not been studied in the field of prognostics.

Trustworthy Machine Learning Operations for Predictive Maintenance Solutions

Kiavash Fathi — 2024-06-27

With the ever-growing capabilities of data acquisition and computational units in industry, development, and deployment of data-driven models (e.g., predictive maintenance solutions) have become more abundant. However, when not trained and maintained properly, these models can be counterproductive as their predictions are not correct, reliable, or interpretable. In addition, unlike conventional software, the issues with such models manifest themselves in reduced productivity and not in forms of traceable software error. Therefore, in this proposal we aim to use model evaluation measures introduced in trustworthy AI operations (TrustAIOps) to trigger re-evaluation of different parts of the data pipeline and the deployed data-driven model given machine learning operations (MLOps) requirements. We argue that by creating an ecosystem capable of monitoring different aspects of a data-driven solution by integrating and managing the implementation concepts in TrustAIOps and MLOps, it is possible to boost the performance of models given the constant changes induced by the specifications of Industry 4.0.

A Comparative Study of Semi-Supervised Anomaly Detection Methods for Machine Fault Detection

Dhiraj Neupane — 2024-06-27

Industrial automation has extended machines’ runtime, thereby raising breakdown risks. Machine breakdowns not only have economic and productivity consequences, but they can also be fatal. Thus, the early detection of fault signs is essential for the safe and uninterrupted operation of machinery and its maintenance. In the last few years, machine learning has been widely used in machine condition monitoring. Most existing approaches rely on supervised learning techniques, which face challenges in real-world scenarios due to the lack of enough labelled fault data. Additionally, models trained on historical fault data might struggle to detect new and unseen faults accurately in the future. Therefore, this research uses semi-supervised Anomaly Detection (AD) techniques to detect abnormal patterns in machines’ vibration signals. As semi-supervised techniques are trained on normal data only, they do not require faulty samples and abnormal patterns are detected based on their deviations from the learned normal pattern. We compared the effectiveness of seven state-of-the-art AD methods, ranging from traditional approaches such as isolation forest and local outlier factor to more recent Deep Learning (DL) approaches based on autoencoders. We evaluated the effectiveness of different feature types extracted from the raw vibration signals, including simple statistical features like kurtosis, mean, peak-to-peak, and more complex representations like the scalogram images. Our study on three public datasets, with unique challenges, shows that the traditional methods based on simple statistical analysis have shown comparable and sometimes superior performance to more complex DL approaches. The use of traditional approaches offers simplicity and lower computational needs. Thus, our study recommends that future researchers start with the traditional approaches first and then jump to DL methods if necessary.

A Computer Vision Deep Learning Tool for Automatic Recognition of Bearing Failure Modes

Stephan Baggerohr — 2024-06-27

We introduce an object detection model specifically designed to identify failure modes in images of bearing components, including the inner ring, outer ring, and rolling elements. The method effectively detects and pinpoints failure features, subsequently determining the associated failure mode within the image. With images sourced from real-world bearing applications, our model can recognize various ISO-failure modes such as surface-initiated fatigue, abrasive wear, adhesive wear, moisture corrosion, fretting corrosion, current leakage erosion, and indents from particles. The proposed model could be used in an assistive tool where failure modes give insights on how to prolong average future bearing life in an asset and therefore reduce related costs and environmental impacts.

A data-driven risk assessment approach for electronic boards used in oil well drilling operations

Delia-Elena Dumitru — 2024-06-27

To assist subject matter experts in investigating electronic failures of drilling tools, an innovative risk assessment approach for oil well drilling operations is developed that relies on synthetic time-series data to emulate environmental factors encountered downhole, explicitly focusing on temperature, shock, and vibration. The approach involves utilizing load cycle counting to extract meaningful features from each environmental channel measured by the drilling tool. The results from experiments with features related to dwell periods (dwell time and dwell damage) and load cycles (cycle means and cycle ranges) show a significant correlation between load cycle features and the risk label. Subsequently, a tree-based machine learning model is trained to label drilling operations based on synthetic data. Several models have been trained initially with comparable results. However, the advantage of using a tree-based model, specifically extra trees, is explainability and the stochastic aspect, which translates into model robustness when applied to real data. Preliminary results from a case study indicate that this new approach is highly effective in categorizing environmental risks associated with drilling operations. This risk assessment method can significantly enhance the decision-making process in investigating electronic board failures by offering reliable decision support.

A Flexible Methodology for Uncertainty-Quantified Monitoring of Abrasive Wear in Heavy Machinery Using Neural Networks and Phenomenology-Based Feature Engineering

Thbate — 2024-06-27

This paper introduces a cutting-edge methodology for the monitoring of abrasive wear, particularly focusing on SAG (Semi-Autogenous Grinding) mills liners. The lack of a regular inspection regime has historically led to opportunistic and thus, irregular wear measurements that are challenging to integrate into machine learning algorithms for condition-based maintenance. The study unveils a virtual sensor designed to estimate the mill liner's remaining thickness, aiming to offer daily updates and assist the maintenance team in determining the optimal timing for liner replacements without the need for halting operations. This approach is positioned as a strategic response to the critical need for efficient maintenance strategies, addressing the inherent challenges in real-world industrial settings where data quality may be poor and operational realities dominate. A significant aspect of this methodology is its emphasis on uncertainty quantification, vital for informed maintenance decision-making. This novel approach has been successfully applied to SAG mills at Minera Los Pelambres, showcasing its potential for broader applications across scenarios characterized by sporadic data collection. The results showcase an error of ±7.4254 mm of remaining thickness on the validation set, demonstrating the effectiveness of the methodology. The key contributions of this work lie in its ability to utilize low-quality data effectively and its low complexity, reducing barriers to implementing predictive health monitoring (PHM) algorithms. The successful implementation highlights the methodology's adaptability and flexibility, marking a significant advancement in the domain of maintenance strategy for the mining industry.

A Gear Health Indicator Based on f-AnoGAN

Hao Wen — 2024-06-27

The development of high-quality health indicators based on Artificial Intelligence (AI) for condition monitoring, reflecting the degradation process and trend, remains a key area of research. Unsupervised deep learning methods, such as deep autoencoders and variational autoencoders, are often employed to establish health indicators for rotating machinery. However, commonly used methods frequently face challenges in controlling and evaluating the quality of learned features that represent this distribution, which subsequently impacts the accuracy of the test data analysis and anomaly detection. Additionally, the empirical nature of threshold setting adds an element of uncertainty to detections.

The research propose a novel approach for constructing gear health indicators and performing anomaly detection using Generative Adversarial Networks (GAN), with a particular emphasis on the f-AnoGAN structure. The research focuses on modeling the distribution of vibration signals acquired from healthy systems using adversarial learning. By comparing test samples against this modeled distribution, the degree of similarity or dissimilarity acts as an indicator of anomalies. Owing to the generative process of the GAN architecture (creating data from randomly sampled low-dimensional noise), GAN-based modeling overcomes the limitation of autoencoders by aiming to reconstruct the continuous distribution of systems in healthy conditions from a limited set of healthy (training) samples. In this way, it offers more generalizability than traditional model learning. Moreover, this study proposes a new method for establishing thresholds based on distribution fitting by the anomaly score of healthy data. The proposed f-AnoGAN-based model and thresholding technique is applied, tested and evaluated in a gear-pitting degradation dataset and result in more accurate and timely fault detection, markedly enhancing the ability to identify subtle faults in systems over traditional methods.

A Hybrid – Machine Learning and Possibilistic – Methdology for Predicting Produced Power Using Wind Turbine SCADA Data

Maneesh Singh — 2024-06-27

During its operational lifetime, a wind turbine is continuously subjected to a number of aggressive environmental and operational conditions, resulting in degradation of its parts. If left unattended, these degraded components will negatively influence its performance and may lead to failure of the wind turbine. In order to mitigate the risk associated with the failure of components, a wind turbine is regularly inspected and maintained.

Currently, there are two commonly used approaches for making maintenance management (inspection and maintenance) plans. Traditional Approach utilises understanding of failure profile of the components for manually developing maintenance plan for the equipment. Condition-Based Approach utilises data collected by condition monitoring of equipment for developing dynamic maintenance plan. SCADA system offers a low-resolution condition-monitoring data that can be used for fault detection, fault diagnosis, fault quantification and fault prognosis and eventually for maintenance planning.

The monitoring data from SCADA system of a wind turbine is often afflicted with variability and uncertainty. The variability in data is the result of continuously changing environmental conditions and uncertainty arises due to imperfections in the recorded data. The uncertainty may be due to many reasons, including, inherent characteristic of sensing devices, drift in calibration with time, deterioration of sensing devices’ sensitivity and response due to environmental attacks, etc.

For handling variability in monitoring data a number of parametric and non-parametric (statistical) predictive models for different aspects of a wind turbine’s structure and operation have been proposed. Depending upon its type – aleatory or epistemic – an uncertainty can be handled in a number of ways. Since, the dynamic nature of wind turbine operation does not allow collection of multiple values under the same condition; hence, uncertainty is mostly epistemic in nature. Possibilistic Approach, based on Fuzzy Set Theory, is especially suitable for handling epistemic uncertainty that may arise due to imprecision or lack of statistical data.

Aim of the ongoing research is to develop a methodology for detecting sub-optimal operation of a wind turbine by comparing Measured Produced Power against Predicted Produced Power. Unfortunately, variability and uncertainty associated with the recorded data make accurate prediction of produced power challenging.

This paper presents methodologies for predicting produced power using SCADA data while simultaneously accounting for variability and uncertainty. The methodologies utilise either parametric (example, power curve) or machine learning (example, XGBoost) models for handling variability; and Possibilistic Approach for handling uncertainty.

A maturity framework for data driven maintenance

Chris Rijsdijk — 2024-06-27

Maintenance decisions range from the simple detection of faults to ultimately predicting future failures and solving the problem. These traditionally human decisions are nowadays increasingly supported by data and the ultimate aim is to make them autonomous. This paper explores the challenges encountered in data driven maintenance, and proposes to consider four aspects in a maturity framework: data / decision maturity, the translation from the real world to data, the computability of decisions (using models) and the causality in the obtained relations. After a discussion of the theoretical concepts involved, the exploration continues by considering a practical fault detection and identification problem. Two approaches, i.e. experience based and model based, are compared and discussed in terms of the four aspects in the maturity framework. It is observed that both approaches yield the same decisions, but still differ in the assignment of causality. This confirms that a maturity assessment not only concerns the type of decision, but should also include the other proposed aspects.

A Novel Approach for Evaluating Datasets Similarities Based on Analytical Hierarchy Process in the Industrial PHM Context

Mohamed Aziz Zaghdoudi — 2024-06-27

In prognostics and health management (PHM), data-driven approaches are crucial for performing prognostics based on historical data, relying on the analysis of extensive datasets to identify patterns and relationships that contribute to predicting or optimizing variables. However, their efficiency is contingent upon the availability of large, high-quality datasets tailored to the specific task at hand.
Yet, real-world applications frequently face challenges as data may not always be readily available due to limitations in data acquisition systems or confidentiality concerns. Paradoxically, the contemporary era witnesses an unprecedented surge in the availability of online databases across various fields. These databases offer a plethora of data that can be harnessed to develop, prototype, and test PHM solutions.
This study endeavors to introduce an innovative approach for assessing the similarity between datasets, specifically tailored for prognostic and health management applications. The objective is to empower the development of PHM solutions for predefined systems without relying on data generated from the system itself, but rather by leveraging analogous datasets.
To quantify the similarity between different datasets, we propose a set of criteria and sub-criteria based on the characteristics of datasets. Subsequently, the analytic hierarchy process (AHP), a well-established multi-criteria decision-making approach, is employed to systematically compare the importance of criteria and sub-criteria for each elementary process within the PHM cycle. This dynamic process considers the varying importance of criteria across different phases, acknowledging that a criterion may not be uniformly significant for all elementary processes. The evaluation of dataset similarity incorporates the proposed criteria and sub-criteria, utilizing a fundamental scale of importance intensity and weights assigned through AHP. This holistic approach yields a comprehensive similarity score, enabling a nuanced understanding of dataset compatibility.
To exemplify the efficiency of our proposed approach, we applied it to a practical case study. The study involves assessing the similarity between a run-to-stop database of mechanical bearings and a set of online databases dedicated to the same application. Our solution facilitated the identification of criteria pertinent to the case study, the determination of criterion weights, and ultimately, the calculation of a similarity score for each database. This process proved instrumental in selecting the most similar database, showcasing the practical utility of our proposed approach in real-world PHM scenarios.

A PHM implementation frame work for MASS (Maritime Autonomous Surface Ships) based on RAM (Reliability Availability Maintainability)

Toby Adam Michael Russell — 2024-06-27

The paper focuses on PHM in the maritime industry, specifically on the maintenance of uncrewed vessels, in contrast to the more commonly discussed navigation. The paper examines the potential challenges of removing the maintenance crew and the potential benefits that can result from this major change in operations.
The removal of the primary maintenance team from a vessel necessitates an increase in monitoring and analysis that can be realised by the techniques of PHM. By looking from the perspective of stakeholders, the challenges and opportunities of PHM implementation become clearer. In comparing the challenges that faced other industries with the maritime industry, roadmaps and proposals can be drawn up for vessel owners. There is a correlation between the phased removal of the engineering crew and the increases in monitoring that is required. Current large vessels that do not carry passengers can operate with UMS (un-manned machinery space) for limited periods. To allow this a specific set of sensors referred to as E0 (Engineers-zero) must be established and maintained. This E0 sensor set forms the basis for what is needed to allow UMS for longer periods of time. The critical equipment, as deemed by class societies, is monitored by E0. Acquiring the data from the E0 sensor set and performing PHM analysis on the data allows remote engineers to accurately determine the current and future state of critical equipment. This equipment list needs to be expanded. Causality based risk modelling is employed to establish a data driven critical equipment list and minimum sensor set to cover the maximum amount of failure modes. This builds on the current required E0 sensor set.
With a conventional maintenance system onboard a vessel the crew are doing a lot of the sensing. The crew act as intermediaries between various systems, taking data from one system to help diagnose another system, making a change to one system to help improve another system. The maintenance crew must balance the interfaces of each system so that a harmony or equilibrium can be achieved. This balancing act is part of what makes a PHM study on a vessel so interesting. Many systems onboard a vessel have a sole purpose to support the crew. With the removal of the crew these support systems can also be removed, simplifying the overall engineering of the vessel.
The methodology that has been used to assess the above points is to create a framework for the design and deployment of PHM to marine assets. The framework relates to RAM (Reliability, Availability, Maintainability) and considers stakeholder points of view and their inputs’ implications. In developing the framework, the stakeholder group is realised. The framework compares the ‘As-Is’ conventional method against the proposed PHM framework. The conclusions are that the E0 philosophy can be expanded upon to facilitate the integration of PHM. Also, the paper concludes that a PHM deployment framework gives the maritime industry a basis for using this modern technique for machinery health. Lastly, the paper shows that PHM is a vital element to uncrewed vessels.

A Physics-Inspired and Data-Driven Approach for Temperature-Based Condition Monitoring

Giacomo Garegnani — 2024-06-27

System overheating is a common problem in electric equipment, as degradation of contacts lead to an increase in Ohmic resistance and increased thermal losses. Temperature measurements are widely employed to monitor a device's health status, to estimate its remaining useful life, and to inform maintenance strategies. An issue that is special to electrical distribution networks is the varying heating power, which is in turn due to changes in the current. This results in varying temperatures, which in addition can often be delayed compared to the currents. Simple threshold-based diagnostics approaches are therefore not reliable for detecting faults in contacts. It is common to analyze the thermal behavior of electric devices using thermal networks, for both design and diagnostic purposes. Unfortunately, identifying the parameters of thermal networks from measured temperature data is a challenging problem, mainly due to identifiability issues and to numerical instabilities in parameter estimation. We propose an alternative data-driven strategy to compute the state-of-health of electrical devices which does not resort to thermal networks. Our approach consists in informing physics-based reduced models of the thermal response with sensor data. We show that our method is linked to the thermal network approach but avoids the full identification of the system, leading to better stability as well as less computational effort in the determination of its parameters. Rigorous testing with synthetic and experimental data confirms the efficacy of our methodology.

A Practical Example of Applying Machine Learning to a Real Turbofan Engine Issue: NEOP

Zdenek Hrncir — 2024-06-27

There are high expectations for the use of Machine Learning algorithms in Engine Health Management, but the practical application for use with turbofan engines is often hindered by small sample sizes and noisy data. This paper discusses a case in which Machine Learning techniques were combined with domain expertise to develop a classifier called Non-seal Erratic Oil Pressure (NEOP). This classifier is used as an engineering tool to support manual review of engines flagged with Honeywell’s OPX (Oil Pressure Transducer) algorithm. The purpose of the classifier is to assist a human in analyzing engine trend data from the HTF7000 turbofan engine, when the OPX algorithm identifies an engine with erratic oil pressure. The NEOP history provides an additional data source when deciding if aft sump maintenance is needed to replace a worn carbon seal, or if the erratic signal is associated with some other cause. The OPX algorithm has enabled the prevention and avoidance of costly unscheduled engine failures resulting in millions of dollars in documented savings, and the NEOP algorithm helps to ensure that the conclusions from the OPX process continue to result in the appropriate engines being identified for maintenance inspection and corrective action.

A Review of Prognostics and Health Management in Wind Turbine Components

Jokin Cuesta — 2024-06-27

Wind turbines (WTs) play an essential role in renewable energy generation, and ensuring their reliable operation is essential for sustainable energy production and reduction of levelized cost of energy. In this context, the field of prognostics and health management (PHM) is a powerful tool to predict and assess the health status of WT components, thereby enabling timely maintenance and reducing downtime. The study begins with an overview of WT components studied, including the blades, gearbox, generator, and bearings, and their common failure modes. For each component, various remaining useful life (RUL) estimation methods are explored, categorizing them into physics-based, data-driven, and hybrid methods. Despite the potential benefits, the application of PHM strategies in WTs is currently limited. Although PHM strategies have been present for years, their development in WTs remains a challenge. These key challenges are presented, including uncertainty management, integrating physical knowledge into models, variable operational conditions, data issues and system complexity.

A rolling bearing state evaluation method based on deep learning combined with Wiener process

Yuntian Ta — 2024-06-27

As a key component of rotating parts, rolling bearings largely determine the operation safety of equipment. However, in practical applications, because the degradation trajectory of rolling bearings cannot be truly characterized, the existing model cannot accurately describe the degradation trajectory of rolling bearings, resulting in the running state of rolling bearings cannot be directly evaluated. Therefore, a method of rolling bearing state assessment based on deep learning combined with Wiener process is proposed in this paper. Firstly, a deep network model is constructed by deep learning to mine the degradation information of rolling bearings. Secondly, the mined degradation information is fused, and then the degradation indicator used to characterize the degraded trajectory of the rolling bearing is constructed. Then, based on Wiener process, the degradation model of rolling bearing is established to describe the degradation mode of rolling bearing. Finally, the constructed degradation indicator is input into the established degradation model to predict its RUL, and then the running state of the rolling bearing is evaluated.

A Semi-supervised Fault Diagnosis Method Based on Graph Convolution for Few-shot Fault Diagnosis

Yuyan Li — 2024-06-27

In practical bearing fault diagnosis, labeled fault data are difficult to obtain, and limited samples will lead to training overfitting. To address the above problems, a semi-supervised fault diagnosis method based on graph convolution is proposed. Firstly, the KNN graph construction method based on Euclidean distance (ED-KNN) is used to achieve label propagation. Then, a graph convolutional network framework based on dot product attention mechanism (GPGAT) was constructed to enhance the weights of high similarity nodes and diagnose bearing faults. The proposed method is validated on a public bearing dataset. The results show that the proposed method can make full use of very few labeled samples for fault diagnosis. Compared with other state-of-the-art methods, the proposed method achieves better diagnosis performance.

A Study on the Equipment Data Collection and Developing Next Generation Integrated PHM System

DEOGHYEON KIM — 2024-06-27

This research presents an integrated PHM system for 2,000 rotating equipment units across press, car body, paint, and assembly lines in Hyundai/Kia factories. The system addresses limitations of individual monitoring systems by consolidating vibration, current, robot AI diagnostics, PLC backup status, and operational data. Vibration monitoring utilizes wired/wireless sensors, server storage, and automated analysis for trend detection and fault diagnosis. PLC data monitoring retrieves motor drive information (current, temperature, frequency, etc.) to predict equipment anomalies.
Robot monitoring integrates with various manufacturers and tracks operational status, motor load, and alarms for maintenance and lifespan management. The PLC backup solution ensures proper backup functionality. The integrated PHM architecture manages data collection, analysis, diagnostics, reporting, and visualization, enabling comprehensive equipment health monitoring and proactive maintenance.

Active learning for gear defect detection in gearboxes

Wenzhi Liao — 2024-06-27

Condition monitoring of gears in gearboxes is crucial to ensure performance and minimizing downtime in many industrial applications including wind turbines and automotive. Monitoring techniques using indirect measurements (i.e. accelerometers, microphones, acoustic emission sensors and encoders, etc.) face challenges, including the defect interpretation and characterization. Vision-based gear condition monitoring, as a direct method to observe gear defects, has the capability to give a precise indication of the starting point of a potential surface failure, but suffers from the image annotations (to train a reliable vision model for automatic defect detection of gears). In this paper, we propose an active learning framework for vision-based condition monitoring, to reduce the human annotation effort by only labelling the most informative examples. In particular, we first train a deep learning model on limited training dataset (annotated randomly) to detect pitting defects. To select which samples have the highest priority to be annotated, we compute the model's uncertainty on all remaining unlabeled examples. Bayesian active learning by disagreement is exploited to estimate the uncertainty of the unlabeled samples. We select the samples with the highest values of uncertainty to be annotated first. Experimental results from defect detection of gears in gearboxes show that with less than 6 times image annotations, we can achieve similar performances.

Advancing Durability Testing in Automotive Component through Prognostics and Health Management (PHM) Integration

Jinwoo Song — 2024-06-27

In automotive Powered Door Systems (PDS), the emergence of grinding and clicking noise over time is a common failure mode. This issue typically arises from design or assembly inconsistencies and intensifies due to wear or increased clearance at its component, becoming noticeable to passengers, and causing discomfort. Numerous automotive manufacturers conduct comprehensive durability tests to tackle such issues during the development. Conventional durability tests, however, rely on the manual effort such as visual and auditory inspection at regular intervals, hence, is subjective and inefficient. This study introduces a novel method by the prognostics and health management (PHM) approach to detect anomaly and assess its severity of the noise during the durability test of the PDS, which may improve the reliability of noise detection and reduces the test time by early termination using prognosis capability. The results demonstrate the potential, paving the way for its broader application across various domains to advance testing processes and reliability.

An Experiment on Anomaly Detection for Fault Vibration Signals Using Autoencoder-Based N-Segmentation Algorithm

YongKwan Lee — 2024-06-27

Most manufacturing facilities driven by motors generate vibration and noise representing critical symptoms against facility malfunctioning conditions in the manufacturing industry. Due to the difficulty of obtaining abnormal data from facilities in manufacturing sites, many prior researchers who have studied predicting facility faults have adopted unsupervised learning-based anomaly detection approaches. Although these approaches have a strength requiring only data on from facility normal behaviors, it is not clear that the anomalies detected by an anomaly detection model are due to the real component faults. Also, the model performance is likely to change according to the diverse abnormal conditions of the given facility. In this paper, we took an experiment with a fault vibration simulator to measure the anomaly detection performance of a one-dimensional convolutional autoencoder model with different fault conditions. In the experiment, we used four different abnormal conditions: imbalance, misalignment, looseness, and bearing faults, which are the most frequently occurring facility component failures from the rotating machineries. Data were gathered from the simulator with the IEPE(Integrated Electronics Piezo-Electric) type sensor. We proposed the N-Segmentation algorithm that performs anomaly detection in segmented frequency region according to corresponding component faults for better anomaly detection performance. In conclusion, the proposed algorithm showed about 15 times better anomaly detection rate than not applying it.

Analytical Modeling of Health Indices for Prognostics and Health Management

Dersin Pierre — 2024-06-27

Understanding the current health condition of complex systems and their temporal evolution is an important step in prognostics and health management (PHM). However, when managing a fleet of complex systems, variations arising from manufacturing, environmental factors, mission profiles, and maintenance practices result in diverse health index (HI) trajectories. Therefore, in PHM, it is essential not only to identify common fleet-wide trends but also to account for individual asset-level variations when inferring HIs.

While several data-driven approaches exist for inferring individual asset-level HIs from unsupervised run-to-failure degradation data, little research has been devoted to deriving analytical probabilistic representations of HIs that encompass both fleet-level trends and individual asset-level fluctuations. This paper aims to bridge this gap by addressing the research question of how to obtain an analytical representation of probability distributions for the time to reach intermediate degradation levels, using run-to-failure data or incomplete degradation trajectories from a fleet of complex systems.

In this work, it is assumed that suitable, asset-specific HI curves have been inferred through a fusion of deep learning techniques and prior expert knowledge of degradation physics . Given this context, we derive an analytical probabilistic description of the health index (HI) that reflects both fleet-wide trends and asset-specific conditions in the cases of Gamma or Weibull time-to-failure (TTF) distributions. Our approach involves defining HIs with a power law function, enabling the modeling of TTF and time to reach intermediate degradation levels. Moreover, we also detail the procedure for estimating the power law exponent from field data through regression analysis and conduct a sensitivity analysis regarding this exponent.

To illustrate our methodology, we present two case studies based on the N-CMAPPS dataset of turbofan engines and Li-ion batteries, validating the aforementioned assumptions and illustrating our methodology steps

Anomaly Detection of a Cooling Water Pump of a Power Plant Based on its Virtual Digital Twin Constructed with Deep Learning Techniques

Miguel A. Sanz-Bobi — 2024-06-27

This paper aims to explore the use of recent approaches of deep learning techniques for anomaly detection of potential failure modes in a cooling water pump working in a gas-combined cycle in a power plant. Two different deep learning techniques have been tested: neural networks and reinforcement learning. Two virtual digital twins were developed with each family of deep learning techniques, able to simulate the behavior of the cooling water pump in the absence of pump failure modes. Each virtual digital twin consists of several models for predicting the expected evolution of significant behavior variables when no anomalies exist. Examples of these variables are bearing temperatures or vibrations in different pump locations. All the data used comes from the SCADA system. The main features and hyperparameters in the virtual digital twins are presented, and demonstration examples are included.

Applying Prognostics and Health Management to Optimize Safety and Sustainability at the First Adaptive High-Rise Structure

Dshamil Efinger — 2024-06-27

Prognostics and Health Management (PHM) offers the potential to increase the acceptance of adaptive structures and to operate them in an optimal way. With suitable design and proper operation, adaptive high-rise structures enable significant increases in sustainability and service life extensions compared to passive high-rise buildings. The control loop for PHM provides a systematic overview of the contents related to PHM and their sequence. However, a framework is required for application to a complex adaptive system. Such a framework is presented in this paper. The framework is divided into the areas of system analysis and modeling as well as the PHM solution. A systematic approach is used to analyze the system and create the basis for full integration of all functional domains. This is then used in modeling to develop an adapted model structure. Finally, the PHM solution looks at the details of the approaches for diagnosis, prognosis, and health management.

Automated Fault Diagnosis Using Maximal Overlap Discret Wavelet Packet Transform and Principal Components Analysis

Fawzi Gougam — 2024-06-27

Bearings and gears are components most susceptible to failure in electromechanical systems, especially rotating machines. Therefore, fault detection becomes a crucial step, as well as fault diagnosis. Over decades, substantial progress in this field has been observed and numerous methods are now proposed for feature extraction from monitoring data. Among these data, vibration signals are most used. However, in the presence of non-Gaussian noise, most conventional methods may be inefficient. In this paper, a hybrid methodology is proposed to address this potential issue. The proposed methodology uses a combination of the Maximal Overlap Discrete Wavelet Packet Transform (MODWPT) and Principal Component Analysis (PCA) techniques. First, the MODWPT technique decomposes the vibration signal with uniform frequency bandwidth, facilitating effective signal processing and introducing diversity for enhanced time-frequency signals. Then, to identify significant patterns and characteristics related to faults, PCA is used for 3D dimensional representation of system health state by capturing the variance in the extracted features. Subsequently, a self-organizing map (SOM) is used for system state classification for diagnostics. This technique is applied to open-access test bench data containing vibration signals with non-Gaussian noise.

Bayesian Networks for Remaining Useful Life Prediction

Erik Hostens — 2024-06-27

Remaining useful life (RUL) prediction is a critical task in the field of condition-based maintenance. It is important to perform RUL prediction in a statistical sound way. However, it is not straightforward to properly combine multiple information sources about an asset, such as available statistics, measurements, derived features, and prior knowledge in the form of mathematical models and relations, including their uncertainties. Bayesian networks (BNs) are a means of graphically representing all statistical information in a comprehensible way and allow for correctly combining all information. BNs allow for inference in all directions, thereby not merely providing a RUL prediction with explicit uncertainty, but select the most informative features, diagnose which degradation mechanism is manifest if multiple mechanisms exist, provide decision support in the form of optimal condition-based maintenance points when combined with a cost model. BNs also explicitly quantify the model uncertainty arising from the scarcity of the training data. We illustrate these benefits on two realworld industrial examples: solenoids and bearings. We also provide a method to correctly include the effect of changing operating conditions.

Characterizing Damage in Wind Turbine Mooring Using a Data-Driven Predictor Model within a Particle Filtering Estimation Framework

Rohit Kumar — 2024-06-27

Floating Offshore Wind Turbines (FOWT) represent a promising solution to renewable energy challenges, yet effective maintenance remains critical for cost management. Traditional machine learning (ML) approaches for detecting FOWT damage often rely on extensive real-world data, which can be impractical and economically unfeasible. Alternatively, stochastic filtering-based time-domain approaches leverage physical understanding through dynamic models, typically finite element models. However, these methods are hindered by excessive simulation calls within the recursive filtering frameworks. This study proposes a novel filtering-based approach that replaces the computationally intensive process model with a Deep Neural Network (DNN) surrogate, addressing the aforementioned limitations. The proposed approach utilizes synthetic data generated from the high-fidelity calibrated OpenFAST model of FOWT dynamics to train a DNN toward learning the dynamic evolution of the FOWT conditioned on the current health state. By offering a computationally efficient representation of system dynamics conditioned on health state, this approach allows for real-time damage detection and interpretable information on damage severity within a stochastic inverse estimation framework, specifically employing Particle Filtering in this study. This approach contrasts with traditional black-box ML-based methods, which typically struggle to provide interpretable information on damage characteristics. Extensive numerical investigations on damaged FOWT mooring lines demonstrate this approach's practical applicability and superiority over traditional ML-based methods. Eventually, integrating explainable ML models within the filtering framework induces promptness in detection without sacrificing transparency.

Comparison among Machine Learning Models Applied in Lithiumion Battery Internal Short Circuit Detection

ZiHong Zhang — 2024-06-27

The world is experimenting a decarbonization process, mainly through lithium-ion-based solutions. Nonetheless, catastrophic events have negatively affected the social acceptance of lithium-ion-based solutions. One of the most interesting projects regarding catastrophic event prevention is the internal short-circuit detection. This paper proposes to detect it using different machine-learning algorithms such as random forest and combination of random forest with neural network-based algorithms through time-instant classification and historical feature classification. The hyper-parameters have been optimized through grid-search. The selected algorithms have been trained thanks to synthetically generated data using a first-order electrical equivalent circuit model. The performance of the generated models has been verified and compared thanks to testing and validation data sets taken from the synthetically generated data. Afterward, the most accurate internal short circuit detection algorithm was selected and validated through laboratory-level data. The selected cell in this study is SLPB526495HE, a pouch cell of 3.7Ah. The generated data are time series of voltage and current, which are the variables that will be available in a real application. The results demonstrate an accuracy above 90% in detecting an internal short circuit in the most interesting cases. The validation with laboratory data has shown that an accuracy of 90% can be achieved. This paper provides learned lessons on the process of developing the internal short circuit detection machine-learning model, highlighting the potential they possess to detect accurately internal short circuits.

Continuous Test-time Domain Adaptation for Efficient Fault Detection under Evolving Operating Conditions

HAN SUN — 2024-06-27

Fault detection is crucial in industrial systems to prevent failures and optimize performance by distinguishing abnormal from normal operating conditions. Data-driven methods have been gaining popularity for fault detection tasks as the amount of condition monitoring data from complex industrial systems increases. Despite these advances, early fault detection remains a challenge under real-world scenarios. The high variability of operating conditions and environments makes it difficult to collect comprehensive training datasets that can represent all possible operating conditions, especially in the early stages of system operation. Furthermore, these variations often evolve over time, potentially leading to entirely new data distributions in the future that were previously unseen. These challenges prevent direct knowledge transfer across different units and over time, leading to the distribution gap between training and testing data and inducing performance degradation of those methods in real-world scenarios. To overcome this, our work introduces a novel approach for continuous test-time domain adaptation. This enables early-stage robust anomaly detection by addressing domain shifts and limited data representativeness issues. We propose a Test-time domain Adaptation Anomaly Detection (TAAD) framework that separates input variables into system parameters and measurements, employing two domain adaptation modules to independently adapt to each input category. This method allows for effective adaptation to evolving operating conditions and is particularly beneficial in systems with scarce data. Our approach, tested on a real-world pump monitoring dataset, shows significant improvements over existing domain adaptation methods in fault detection, demonstrating enhanced accuracy and reliability.

Contrastive Metric Learning Loss-Enhanced Multi-Layer Perceptron for Sequentially Appearing Clusters in Acoustic Emission Data Streams

Oualid Laiadi — 2024-06-27

Conventional structural health monitoring methods for interpreting unlabeled acoustic emission (AE) data typically rely on generic clustering approaches. This work introduces a novel approach for analyzing sequential and temporal acoustic emission (AE) data streams by enhancing a Multi-Layer Perceptron (MLP) with a contrastive metric learning loss function (MLP-CMLL)and Time Series K-means (TSKmeans) clustering. This dual approach, MLP-CMLL with TSKmeans, is crafted to refine cluster differentiation significantly. This method is designed to optimize cluster differentiation, bringing similar acoustic patterns closer and distancing divergent ones, thereby improving the MLP's ability to classify acoustic events over time. Addressing the limitations of traditional clustering algorithms in handling the temporal dynamics of AE data, MLP-CMLL with TSKmeans approach provides deeper insights into cluster formation and evolution. It promises enhanced monitoring and predictive maintenance capabilities in engineering applications by capturing the complex dynamics of AE data more effectively, offering a significant advancement in the field of structural health monitoring. Through experimental validation, we apply this method to characterize the loosening phenomenon in bolted structures under vibrations. Comparative analysis with two standard clustering methods using raw streaming data from three experimental campaigns demonstrates that our proposed method not only delivers valuable qualitative information concerning the timeline of clusters but also showcases superior performance in terms of cluster characterization.

Counterfactual Explanation for Auto-Encoder Based Time-Series Anomaly Detection

Abishek Srinivasan — 2024-06-27

The complexity of modern electro-mechanical systems require the development of sophisticated diagnostic methods like anomaly detection capable of detecting deviations. Conventional anomaly detection approaches like signal processing and statistical modelling often struggle to effectively handle the intricacies of complex systems, particularly when dealing with multi-variate signals. In contrast, neural network-based anomaly detection methods, especially Auto-Encoders, have emerged as a compelling alternative, demonstrating remarkable performance. However, Auto-Encoders exhibit inherent opaqueness in their decision-making processes, hindering their practical implementation at scale. Addressing this opacity is essential for enhancing the interpretability and trustworthiness of anomaly detection models. In this work, we address this challenge by employing a feature selector to select features and counterfactual explanations to give a context to the model output. We tested this approach on the SKAB benchmark dataset and an industrial time-series dataset. The gradient based counterfactual explanation approach was evaluated via validity, sparsity and distance measures. Our experimental findings illustrate that our proposed counterfactual approach can offer meaningful and valuable insights into the model decision-making process, by explaining fewer signals compared to conventional approaches. These insights enhance the trustworthiness and interpretability of anomaly detection models.

Damage Detection using Machine Learning for PHM in Gearbox Applications

Lisa Binanzer — 2024-06-27

Early damage detection in gearbox applications enables the implementation of Prognostics and Health Management (PHM). On the one hand, the earliest possible damage detection provides a precise in-sight into the state of health of a gearbox. In addition, early damage detection offers the possibility to slow down the damage progress and extend the remaining useful life (RUL) by intervening in the operating state at an early damage stage. The main contribution of this work is that existing Machine Learning tools are applied to the challenge of very early damage detection in gearboxes. Thus, the need for complex physically based data evaluation is avoided. The aim of this investigation is a comparison of two different machine learning approaches. To investigate the detection possibilities test bench experiments were conducted with a single stage spur gearbox. For a comprehensive investigation, i.e. to detect damage under different operating conditions, the test runs are carried out at different damage sizes, speeds and torques. Based on the recorded vibration data, the damage detection is examined. Two machine learning approaches of anomaly detection are considered: An encoding approach and a loss approach. The same sparse autoencoder is developed for both approaches Both machine learning approaches are able to detect even the smallest damage of about 0.5 % in most operating states. The loss approach allows the different damage stages to be recognized much more clearly than the encoding approach. The comparison of the different approaches provides valuable insights for the further development of robust damage detection algorithms.

Data Scarcity in Fault Detection for Solar Tracking Systems: the Power of Physics-Informed Artificial Intelligence

Mila Francesca Lüscher — 2024-06-27

Combining physical and domain knowledge in artificial intelligence (AI) models has been gaining attention in various fields and applications.
Applications in machine prognostics and health management (PHM) are natural candidates for such hybrid approaches. In particular, they can be efficiently exploited for high fidelity anomaly detection in technical and industrial systems. A natural way for hybridization is using physical models to generate representative data for the training of AI models. Depending on the level of domain knowledge availability, data augmentation can compensate for scarcity of real data from the field. This is particularly attractive for anomaly detection tasks, in which data from the abnormal regimes is limited by definition. On top of this inherent data limitation, many real-world systems suffer from data limitations even within the normal regimes.
In this paper we propose a physics-informed deep learning algorithm for fault detection in grid scale photovoltaic power plants. We focus on a common data scarce scenario that emerges from a low asset monitoring granularity: instead of monitoring the power production of each solar string, the power output is monitored only at combiner-box or even inverter level (monitoring a large number of strings with a single sensor). As a result, the signatures of single local faults can become very subtle and challenging to detect. We show that in this case a physics-informed AI approach significantly outperforms the alternative of a purely data-driven anomaly detection model. This enables high fidelity fault detection in larger solar power plants, that are often limited in the granularity of their condition monitoring data.

Data-Driven Prognostics with Multi-Layer Perceptron Particle Filter: a Cross-Industry Exploration

Francesco Cancelliere — 2024-06-27

The integration of particle or Kalman filters with machine learning tools like support vector machines, Gaussian processes, or neural networks has seen extensive exploration in the context of prognostic and health management, particularly in model-based applications. This paper focuses on the Multi-Layer Perceptron Particle Filter (MLP-PF), a data-driven approach that harnesses the non-linearity of MLP to describe degradation trajectories without relying on a physical model. The Bayesian nature of the particle filter is utilized to update MLP parameters, providing flexibility to the method and accommodating unexpected changes in the degradation behavior. To showcase the versatility of MLP-PF, this work demonstrates its seamless integration into diverse use cases, such as lithium-ion battery analysis, virtual health monitoring for turbofans, and the assessment of fatigue crack growth. We illustrate how it effortlessly accommodates various contexts through slight parameter modifications. Adjustment includes variation in the number of neurons or layers in the MLP, threshold adjustments, initial training refinements and the adaptation of the process noise. Addressing different degradation processes across these applications, MLP-PF proves its adaptability and utility in various contexts. These findings highlight the method’s versatility in adapting to diverse use cases and its potential as a robust prognostic tool across various industries. MLP-PF offers a practical and efficient means of estimating remaining useful life and predicting degradation in complex systems, with implications for advancing prognostic tools in diverse applications.

Data-Driven Remaining Useful Life Estimation Approach for Neutron Generators in Multifunction Logging-While-Drilling Service

Karolina Sobczak-Oramus — 2024-06-27

This paper introduces a data-driven approach for estimating the remaining useful life of the neutron generator component in logging-while-drilling tools. The approach builds on identification of the incipient failure modes of the neutron generator and constructing a health indicator that serves as a statistical representation of the component’s deterioration over time. Afterwards, a K-nearest neighbors algorithm is trained to establish the relationship between the extracted health indicator values and the corresponding remaining useful life. The effectiveness of the presented approach is verified through the utilization of real-world data gathered from oil well drilling operations. The study is part of a long term project aimed at developing a digital fleet management system for drilling tools.

Defect Data Augmentation Method for Robust Image-based Product Inspection

Youngwoon Choi — 2024-06-27

In this paper, we develop a model for detecting defects in fabric products based on an object segmentation algorithm, including a novel image data augmentation method to enhance the robustness. First, a vision-based inspection system is established to collect image data of the fabric products. The three types of fabric defects, such as a hole, a stain, and a dyeing defect, are considered. To enhance defect detection accuracy and robustness, a novel image data augmentation method, referred to as the defect-area cut-mix, is proposed. In this method, the shapes that are the same as each defect are extracted using the masks, and then they are added to non-defective fabric images. Second, an ensemble process is implemented by combining the results of two models, one with high sensitivity in defect diagnosis and the other with lower sensitivity. The results demonstrated that the model trained on the augmented dataset exhibits improved metrics such as intersection over union and classification accuracy in defect detection on the test dataset.

Detection of Abnormal Conditions in Electro-Mechanical Actuators by Physics-Informed Long Short-term Memory Networks

Chenyang Lai — 2024-06-27

Electro-Mechanical Actuators (EMAs) are projected to revolutionize the flight control actuator paradigm, potentially replacing hydraulic-powered systems in the future. Consequently, the functioning of EMAs is destined to become critical for the safe and reliable operation of aircraft. Abnormal conditions of the mechanical components of EMAs can lead to their failure. The objective of this work is to develop a method for the early detection of abnormal conditions of the components of EMAs. The proposed method is based on a signal reconstruction model that estimates the motor position of EMA as expected in normal conditions of its components. Then, the presence of an abnormal condition is identified when the difference between the motor position and its reconstructed position in normal conditions exceeds a preset threshold. The signal reconstruction model employs a Physics-Informed Long Short-Term Memory network (PILSTM), whose architecture combines a physics-informed cell for the solution of the differential equations governing the EMA operation, and a data-driven Long Short-Term Memory (LSTM) cell which receives in input the output of the physics-informed cell and reconstructs the position expected in normal conditions. The proposed method is applied to data simulated by a high-fidelity model of EMAs. The results show that PILSTM is able to provide accurate, physics-consistent estimates of the motor position of EMA in normal conditions and the missed and false detection alarms are lower than those of other state-of-the-art methods.

Development of a Feature Extraction Methodology for Prognostic Tasks of Aerospace Structures and Systems

Antonio Orru — 2024-06-27

The performance of prognostic models used for prognostic health management (PHM) applications heavily depend on the quality of features extracted from raw sensor data. Traditionally, feature extraction criteria such as monotonicity, prognosability, and trendability are selected intuitively. However, this intuitive selection may not be optimal.
This research introduces an innovative approach to transform raw data into 'high-scoring' data without the need for predefined feature extraction criteria. Our methodology involves generating a set of synthetic high-scoring time series. These synthetic time series, resembling the length and amplitude of target features, are created through Monte Carlo sampling (MCS) of a user-defined hidden semi-markov model (HSMM). We pair these synthetic time series with raw data/features from the signals and use them as targets to train a convolutional neural network (CNN). As a result, the trained CNN can convert input features into high-scoring ones, irrespective of their initial characteristics. So, this study provides the following contribution to PHM frameworks: it transforms raw data/features into high-scoring ones without relying on predefined criteria, rather on stochastically generated labels that resemble the nature of the degradation processes. It is worth noting, that the proposed FE technique is independent of the prognostic model that will be utilised, thus making it applicable to the established prognostic models.
We demonstrate and validate the effectiveness of this approach using acoustic emission (AE) sensor data for remaining useful life (RUL) estimation of open-hole CFRP specimens. We compare prognostic performance using cumulative AE features with their transformations via our proposed framework. The transformed features exhibit superior prognostic performance, underscoring the value of our innovative feature extraction framework.

Development of a PHM system for electrically actuated brakes of a smallpassenger aircraft

Andrea De Martin — 2024-06-27

The evolution towards “more electric” aircraft has seen a decisive push in the last decade, due to the growing environmental concerns and the development of new market segments (Urban Air Mobility). Such push interested both the propulsion components and the aircraft systems, with the latter seeing a progressive trend in replacing the traditional solutions based on hydraulic power with electrical or electromechanical devices. Electro-mechanical brakes, or E-Brakes hereby onwards, would present several advantages over their hydraulic counterparts, mainly related to the avoidance of leakage issues and the simplification of the system architecture. Moreover, although it is expected a weight increase of the brake, the elimination of the hydraulic lanes would still come with an overall weight reduction. Despite these advantages, it remains a new, relatively unproven technology within the civil aviation field. Within this context, the development of PHM solutions would align with the need for an on-line monitoring of a relatively unproven component. This paper deals with the preliminary stages of the development of such PHM system for the E-Brake of a future executive class aircraft, iterating on previously published material and presenting a particle filtering approach based on a new degradation model and data provided through a revised high-fidelity model. The paper opens with the introduction to the research project and the technological demonstrator, positioning the performed work within the available literature. PHM activities, performed on simulated data-set are then presented and the preliminary results discussed.

Development of Anomaly Detection Technology Applicable to Various Equipment Groups in Smart Factory

KIWON PARK — 2024-06-27

This study delves into the creation of anomaly detection technology applicable to a range of equipment groups within smart factories. This advanced technology uses high-performance MEMS vibration sensors, edge CMS devices, and PHM platforms to tackle issues such as data imbalance, learning model limitations, complex equipment operating patterns, and real-time processing. It also addresses central server concentration, data cycling problem, various equipment classification, and algorithm operation problems that can arise when implementing systems in the field. Using AI-based vibration detection algorithms, data can be collected at high sampling rates and analyzed in real-time through edge computing, minimizing latency and mitigating server capacity issues compared to cloud-based analytics. The system continually monitors and learns standard performance data from equipment to provide practical solutions that minimize equipment failures and downtimes. The results of this study are impressive, as it has successfully developed anomaly detection framework and PHM systems that are expected to enhance the efficiency and sustainability of smart factories. Furthermore, the study aims to showcase and improve the effectiveness of predictive maintenance in both domestic and international automotive factory production lines. This revolutionary technology will be a key component in smart and software-defined factories and help companies achieve intelligent automation.

Development of Fault Diagnosis Model based on Semi-supervised Autoencoder

Yongjae Jeon — 2024-06-27

The maintenance paradigm based on PHM (Prognostics and Health Management) technology, utilizing big data to predict process conditions through manufacturing intelligence, is rising. However, in most industries, there is lack of accurate labeling of sensor data, posing challenges in data utilization due to the significant cost of labeling tasks. Consequently, recent research has focused on semi-supervised learning methodologies, which are applicable in label-absent scenarios. Especially, there is a growing emphasis on semi-supervised autoencoder, which learns both labeled and unlabeled data simultaneously. Also, there is a demand for the development of fault diagnosis models for essential components, such as bearings in most mechanical systems. Vibrational data is actively being integrated with artificial intelligence for application in bearing fault diagnosis frameworks. Nonetheless, diagnosing the condition of bearings inside machine systems, especially within the machine tool spindle, remains challenging, as the labeling of collected data causes significant costs. Therefore, this paper aims to develop a fault diagnosis model for unlabeled bearings in machine tool spindle using a semi-supervised autoencoder. Initially, a monitoring system of bearing simulator that imitates a machine tool spindle bearing was constructed, and vibration signals from both normal and fault bearings were collected based on this system. Subsequently, a semi-supervised autoencoder model was developed to construct a fault diagnosis model using labeled simulator data and unlabeled machine tool spindle bearing data. To evaluate the model, additional data of normal and fault bearings in machine tool spindle were collected, and the performance of the model was compared with a conventional fault diagnosis model based on 1D-CNN.

DiffPhysiNet: A Bearing Diagnostic Framework Based on Physics-Driven Diffusion Network for Unseen Working Conditions

Zhinbin Guo — 2024-06-27

Fault diagnosis is essential to ensure bearing safety in industrial applications. Many existing diagnostic methods require large scales of data from a full range of working conditions. However, the structure and working conditions differences between machines lead to significant variation in data distribution, making it difficult to diagnostic with unseen samples. To handle this situation, an unknown condition diagnosis Framework (UCDF) based on physics-driven diffusion network (DiffPhysiNet) is proposed, effectively integrating the generation capability of the diffusion model and learning from the working conditional encoding (WCE). Specifically, signals under limited working conditions are gradually convert to noise through a forward noising process. Then, DiffPhysiNet reconstructs signals from the noise by a reverse denoising process. In addition, a physics-driven UNet (Physi-UNet) structure is designed to extract WCE for noise level prediction during the reverse process. Moreover, an Unsupervised Clustering Filter (UCFilter) is constructed to select signals with high quality after generation. Signals under unknown working condition can be generated with certain WCE. Ultimately, extensive experiments on two bearing datasets (SDUST and PU) validate the effectiveness of our method compared with the state-of-the-art baselines and the ablution test confirms the significant role of Physi-UNet and UCFilter.

Domain Adaptation for Fault Detection in Civil Nuclear Plants

Henry Wood — 2024-06-27

Recent domain adaptation approaches have been shown to generalise well between distant data domains achieving high performance in machine fault detection through time series classification. An interesting aspect of this transfer-learning inspired approach, is that the algorithm need not be exposed to fault data from the target domain during training. This promotes the application of these methods to environments in which fault data is unfeasible to obtain, such as the detection of loss-of-coolant accidents (LOCA) in nuclear power plants (NPPs). A LOCA is a failure mode of a nuclear reactor in which coolant is lost due to a physical break in the primary coolant circuit. If undetected, or not managed effectively, a LOCA can result in reactor core damage. Three high-fidelity physics based models were created with divergent behaviour that represent different data domains. The first model is used to generate source domain data by simulating labelled training data under both nominal and LOCA conditions. The second and third models act as surrogates of real plants and are used to generate target domain data, i.e. to simulate nominal data for training and LOCA condition data for validation. Several deep-learning feature encoders (with varying levels of connectivity) were applied to this LOCA detection problem. Among these, a ’Baseline’ encoder was used to quantify the improvement that domain adaptation techniques make to LOCA detection performance under large domain divergences. Classification accuracy for each model is explored within the context of LOCA break size and location within each plant model. The proposed method for LOCA detection demonstrates how the dependence upon sparse accident-specific data can be alleviated through the use of domain adaptation. Detection capability of the LOCA condition is maintained even when no data examples are available in the target domain.

Domain Adaptation via Simulation Parameter and Data Perturbation for Predictive Maintenance

Kiavash Fathi — 2024-06-27

Conventional data-driven predictive maintenance (PdM) solutions learn from samples of run-to-failures (R2F) to estimate the remaining useful life of an asset. In practice, such samples are scarce or completely missing. Simulation models can be oftentimes used to generate R2F samples as a replacement. However, due to the complexity of the assets, creating realistic simulation models is tedious, or even impossible. Thus generated R2F data cannot be used to create reliable PdM models as they are highly sensitive to noises in the sensors or small deviations in system working condition. To address this, we present a new concept of simulation data generation based on supervised domain adaptation for a regression problem where the remaining useful life (RUL) or the health index (HI) of the system is predicted. Apart from input and output domain shift, given the changes in the dominant failing component and its degradation process, the function mapping sensor readings to RUL and/or HI is also prone to changes and thus is a random process itself. Therefore, we aim to generate R2F training data from different working conditions and possible failure types using parameter randomization in the simulation model. By sampling from various configurations within simulation model's parameter space, we ensure that the trained data-driven PdM model's performance is not impacted by the initial conditions and/or the changes in the degradation of the system's condition indicators. Our results indicate that the model is robust to signal reading manipulation and showcases a more spread-out feature importance across a wider range of sensor readings for making predictions. We also demonstrate its applicability on the real-world factory physical system whilst our models were mainly trained using generated data.

Dynamic Modeling of Distributed Wear-Like Faults in Spur Gears: Simplified Approach with Experimental Validation

Lior Bachar — 2024-06-27

Dynamic models of gears are recognized for offering a promising platform for gaining a profound understanding of the dynamic response, particularly the vibration signature. Wear is considered among the most common and concerning fault mechanisms in gears, yet its recognition and subsequent diagnosis remain challenging. In this study, we introduce an existing dynamic model of spur gear vibrations and extend its validation for distributed wear-like faults. The novelty of this work lies in addressing the complexities associated with modeling distributed faults using simplified yet sophisticated approaches. These involve variance among defected teeth, calculation of time-variant gear mesh stiffness, and consideration of the forces induced by the fault. The model is validated through pioneering controlled experiments, analyzing dozens of degrading distributed wear-like faults. This comparison verifies our capability to generate realistic simulations of vibration signals from worn gears. To bridge the discrepancy between the induced and simulated faults, the model first constructs the healthy profile of the inspected gear, incorporating manufacturing errors and tooth modifications. Subsequently, meticulous photography enables the replication of faults in the model with a high resemblance to the experiment. The results demonstrate a strong correlation between measured and simulated signals, as verified through trend analysis of features extracted from synchronous average signals in both the cycle and order domains. This study lays the groundwork for in-depth investigation into the physics of gear wear, paving the way for potential applications such as fault severity estimation and intelligent fault diagnosis in future studies.

Enhanced Diagnostics Empowered by Improved Mechanical Vibration Component Extraction in Nonstationary Regimes

Fadi Karkafi — 2024-06-27

When analyzing vibration and sound signals from rotating machinery, accurately tracking individual orders is crucial for diagnostic and prognostic objectives. These orders correspond to sinusoidal components, also known as deterministic signals, whose amplitude and phase are modulated in response to the angular speed of the machine. The extraction of these components leads to a more comprehensive approach to differential diagnostics. When the machine operates under varying conditions, consistently tracking the orders becomes challenging, particularly in nonstationary regimes with very fast variations. Typically, this issue is addressed using common techniques such as Vold-Kalman filter (VKF), where the bandwidth of the selective filter is adjusted to handle the speed variations. However, in the presence of high-speed fluctuations, manual adjustment of these weights becomes difficult to balance the compromise between achieving accurate tracking by effectively filtering around the speed variations, and maintaining a low estimation bias by reducing noisy errors. To overcome this constraint, the proposed methodology is driven by the need to integrate speed fluctuations into an optimal solution using VKF. This adaptation involves the consideration of angular acceleration profiles within the innovation process. In this context, the bandwidths are automatically adjusted to their optimal values according to the machine’s regime. Optimality is achieved by crafting a model dependent on the order signal-to-noise ratio (SNR) and the auto-regression coefficient. This optimization allows for a practical adjustment tailored to the distinctive characteristics of each order. A comprehensive analysis of the resulting model transfer function reveals crucial insights into the impact of the given order SNR and the speed fluctuations. Subsequently, the methodology undergoes performance assessment through simulations and synthetic cases, showcasing its viability and effectiveness across various regimes. Notably, its practical application is highlighted in envelope-based bearing diagnosis, during operations characterized by variable-speed conditions, thus underlining its promise in real-world applications.

Enhancing Data-driven Vibration-based Machinery Fault Diagnosis Generalization Under Varied Conditions by Removing Domain-Specific Information Utilizing Sparse Representation

David Latil — 2024-06-27

This paper introduces a novel approach to machinery fault diagnosis, addressing the challenge of domain generalization in diverse industrial environments. Traditional methods often struggle with domain shift and the scarcity of balanced, la- beled datasets, limiting their effectiveness across varied oper- ational conditions. The proposed method leverages the abun- dance of healthy machinery signals as a reference for extract- ing domain-specific information. By doing so, it removes the domain-related variances from the observation signals, focus- ing on the intrinsic characteristics of faults. The methodol- ogy is validated with a case study, demonstrating enhanced diagnosis accuracy and generalization capabilities in unseen domains. This research contributes to the field of vibration- based intelligent fault diagnosis by providing a robust solu- tion to a long-standing problem in machine condition moni- toring.

Enhancing gearbox condition monitoring using randomized singular value decomposition and K-nearest neighbor

Adel Afia — 2024-06-27

Efficient gear and bearing diagnosis has become a critical requirement across diverse industrial applications precisely due to their complex design and exposure to difficult operating conditions, which predispose them to frequent failure. Early fault identification remains problematic, as defects are commonly obscured by extensive background noise. Moreover, the exponential increases in gearbox data further complicate the defect classification process, confusing even the most sophisticated algorithms and significantly making the procedure time consuming. Singular Value Decomposition (SVD) has proved to be highly efficient in signal denoising, stability preservation, and feature extraction reliably under varying conditions, filtering out non-linear signals to reconstruct relevant features only. However, its considerable computation time necessitates exploring alternatives like Randomized SVD (RSVD) to mitigate processing time while maintaining classification accuracy. In this work, an intelligent algorithm for gear and bearing fault diagnosis is developed, incorporating Maximal Overlap Discrete Wavelet Packet Transform (MODWPT) and Time-Domain Features for feature extraction. RSVD is employed for signal denoising and feature reconstruction, while K-Nearest Neighbor (KNN) for feature classification. The combined techniques ensure enhanced diagnostic accuracy, addressing critical challenges in industrial maintenance and performance optimization.

Enhancing Lithium-ion Battery Safety: Analysis and Detection of Internal Short Circuit basing on an Electrochemical-Thermal Modeling

YIQI JIA — 2024-06-27

As the main cause of thermal runaway, the prompt identification of Internal Short Circuit (ISC) occurrences in lithium-ion batteries (LIBs) has emerged as a critical priority for ensuring battery safety. To address this critical need, for a comprehensive understanding of ISC behaviors, an electrochemical-thermal-ISC coupled model has been developed in this work to simulate battery performance across various ISC levels. This model is also utilized to validate the efficacy and robustness of the advanced detection approach proposed. By integrating both thermal and electrical aspects using the Pseudo Two-Dimensional (P2D) and Energy Balance Equation (EBE), our model serves as an efficient surrogate for ISC experiments. Key ISC indicators have been analyzed and integrated into the proposed ISC detection algorithm to enhance its effectiveness. The algorithm utilizes an Equivalent Circuit Model (ECM)-based approach for estimating ISC resistance. This research not only advances our understanding of ISC dynamics but also establishes a robust framework for the timely and reliable detection of ISCs. These advancements significantly enhance the overall safety and reliability of LIBs in electric vehicles (EVs).

Enhancing Lithium-Ion Battery State-of-Charge Estimation Across Battery Types via Unsupervised Domain Adaptation

Mohammad Badfar — 2024-06-27

Accurate estimation of the state-of-charge (SOC) in lithium-ion batteries (LIBs) is paramount for the safe operation of battery management systems. Despite the effectiveness of existing SOC estimation methods, their generalization across different battery chemistries and operating conditions remains challenging. Current data-driven approaches necessitate extensive data collection for each battery chemistry and operating condition, leading to a costly and time-consuming process. Hence, there is a critical need to enhance the generalization and adaptability of SOC estimators. In this paper, we propose a novel SOC estimation method based on Regression-based Unsupervised Domain Adaptation. We evaluate the performance of this method in cross-battery and cross-temperature SOC estimation scenarios. Additionally, we conduct a comparative analysis with a widely-used classification-based unsupervised domain adaptation approach. Our findings demonstrate the superiority of the regression-based unsupervised domain adaptation method in achieving accurate SOC estimation for batteries.

Exploring a Knowledge-Based Approach for Predictive Maintenance of Aircraft Engines: Studying Fault Propagation through Spatial and Topological Component Relationships

Meriem HAFSI — 2024-06-27

Predictive maintenance has become a highly favored application in Industry 4.0, particularly in complex systems with requirements for reliability, robustness, and performance. Aircraft engines are among these systems, and several studies have been conducted to try to estimate their remaining lifespan. The C-MAPSS dataset provided by NASA has greatly served the scientific community, and several works based on physical models and data-driven approaches have been proposed. However, several limitations related to data quality or data availability of failures persist, and integrating domain knowledge can help address these challenges. In this article, we are currently implementing a new approach based on knowledge coupled with qualitative spatial reasoning to study the propagation of faults within system components until complete shutdown. Region Connection Calculus (RCC8) formal model will be used to describe the component relationships, drawing inspiration from the C-MAPSS dataset.

False alarm reduction in railway track quality inspections using machine learning

Isidro Durazo-Cardenas — 2024-06-27

Track quality geometry measurements are crucial for the railways’ timely maintenance. Regular measurements prevent train delays, passenger discomfort and incidents. However, current fault diagnosis or parameter deviation relies on simple threshold comparison of multiple laser scanners, linear variable differential transformer (LVDT) and camera measurements. Data threshold exceedances enact maintenance actions automatically. However, issues such as measurement error, and sensor failure can result in false positives. Broad localisation resolution prevents trending/ inferencing by comparison with healthy data baseline at the same position over periodic inspections.

False alarms can result in costly ineffective interventions, are hazardous and impact the network availability.

This paper proposes a novel methodology based on convolutional neural network (CNN) technique for detecting and classifying track geometry fault severity automatically. The proposed methodology comprises an automatic flow of data for quality assessment whereby outliers, missing values and misalignment are detected, restored and where appropriate curated. Improved, “clean” datasets were then analysed using a pretrained CNN model. The method was compared with a suite of machine learning algorithms for diagnosis including k-nearest neighbour, support vector machines (SVM), and random forest (RF).

The analysis results of a real track geometry dataset showed that track quality parameters including twist, cant, gauge, and alignment could be effectively diagnosed with an accuracy rate of 97.80% (CNN model). This result represents a remarkable improvement of 38% in comparison with the traditional threshold-based diagnosis. The benefits of this research are not only associated with maintenance intervention cost savings. It also helps prevent unnecessary train speed restrictions arising from misdiagnosis.

Fault Diagnosis of Multiple Components in Complex Mechanical System Using Remote Sensor

Jeongmin Oh — 2024-06-27

This study proposes an approach to monitor multiple components in complex mechanical systems using a single, externally placed remote sensor. In automobiles and petrochemical plants, where numerous components (e.g., powertrain, bearing, and gear), sensor placement is often compromised by cost and installation environment constraints, resulting in sensing the components far from the regions of interest. To address this challenge, this paper proposes an Operational Transfer Path Analysis (OTPA)-based approach that derives the transfer functions between the vibration excitation source and the measurement point (i.e., receiver). The model for OTPA enables the reverse estimation of the excitation source’s signal from the receiver. Subsequently, the estimated (i.e., synthesized) source signal is fed into a diagnostic model to identify system faults. The OTPA and diagnostic models are constructed using neural network architectures, enabling better adaptation to operational conditions and system-induced nonlinearities. The proposed approach is validated from case studies using hydraulic piston pumps in construction vehicles and next-generation electric vehicles.

Fault Prediction and Estimation of Automotive LiDAR Signals Using Transfer Learning-Based Domain Generalization

Sanghoon Lee — 2024-06-27

Autonomous vehicles (AVs) are undergoing level 4 technology development and should have a system that can be operated without driver’s intervention, so that it must be possible to diagnose failures and predict life cycle themselves. In this study, we propose a technology to estimate signal changes and sensor faults through transfer learning-based domain generalization (TLDG) using limited actual vehicle test information from LiDAR for autonomous vehicles. Because autonomous vehicles operate in various climate/weather conditions over the world, their mechanical, electrical and electronic components must also have stable performance in all environmental conditions. However, an electronic device, especially laser diode (LD), which is one of core components of LiDAR, shows various degradation aspects depending on environmental conditions. We acquired multivariate LiDAR performance data under various environmental conditions through an actual vehicle test driving of about 2,000 km in summer and winter, and based on this, we create the LiDAR fault diagnosis and performance prediction model generalized to the domain under various environmental conditions. Fault prediction and estimation model created through summer and winter data in the environment domain will also adapt to other environmental conditions such as spring and fall. To develop highly accurate performance estimation models under various environmental conditions based on limited data, it is very important to extract correlations and characteristics between data, including environmental conditions. We employ the data augmentation techniques to solve the problem of lack of training data and apply bi-directional Bayesian transfer learning to generalize data and models under uncertainty. To prove the effectiveness of the present study, the data from actual vehicle tests conducted at different temperatures will be divided into train data and test data, and the validity of the generalized degradation performance estimation model will be statistically validated. The proposed domain generalization method, i.e., TLDG can be utilized to estimate signal changes and sensor faults in LiDAR under unexperienced environmental conditions such as weather changes, and even freezing and hot regions.

From Prediction to Prescription: Large Language Model Agent for Context-Aware Maintenance Decision Support

HAOXUAN DENG — 2024-06-27

Predictive analytics with machine learning approaches has widely penetrated and shown great success in system health management over the decade. However, how to convert the prediction to an actionable plan for maintenance is still far from mature. This study investigates how to narrow the gap
between predictive outcomes and prescriptive descriptions for system maintenance using an agentic approach based on the large language model (LLM). Additionally, with the retrieval-augmented generation (RAG) technique and tool usage capability, the LLM can be context-aware when making decisions in maintenance strategy proposals considering predictions from machine learning. In this way, the proposed method can push forward the boundary of current machine-learning methods from a predictor to an advisor for decision-making workload offload. For verification, a case study on linear actuator fault diagnosis is conducted with the GPT-4 model. The result demonstrates that the proposed method can perform fault detection without extra training or fine-tuning with comparable performance to baseline methods and deliver more informatic diagnosis analysis and suggestions. This research can shed light on the application of large language models in the construction of versatile and flexible artificial intelligence agents for maintenance tasks.

Fully Automated Diagnostics of Induction Motor Drives in Offshore Wind Turbine Pitch Systems using Extended Park Vector Transform and Convolutional Neural Network

Manuel Sathyajith Mathew — 2024-06-27

Due to their location and related complexities, the offshore wind farms (OWF) have higher downtimes and operation and maintenance (O&M) costs compared to their onshore counterparts. Condition monitoring could help in bringing down the O&M costs of OWFs. The pitch system is one of the components most prone to failure. This paper details an approach for enhanced diagnosis of the electric pitch systems especially focusing on the induction motor drives (IMD) in wind turbines. The proposed method uses an extended Park vector approach (EPVA) in conjunction with a convolutional neural network (CNN) to accurately classify the condition of an IMD and localize the faults. The method is validated on data collected from a laboratory setup. The advantage of the proposed approach is that the condition of the IMD can accurately be classified, and faults localized in operating conditions with varying load and frequency without any additional information on the instantaneous operating speed, frequency, or load on the motor drives. This results in a non-invasive diagnostic approach incurring least additional expenses to implement.

Graph Neural Networks for Electric and Hydraulic Data Fusion to Enhance Short-term Forecasting of Pumped-storage Hydroelectricity

Raffael Theiler — 2024-06-27

Pumped-storage hydropower plants (PSH) actively participate in grid power-frequency control and therefore often operate under dynamic conditions, which results in rapidly varying system states. Predicting these dynamically changing states is essential for comprehending the underlying sensor and machine conditions. This understanding aids in detecting anomalies and faults, ensuring the reliable operation of the connected power grid, and in identifying faulty and miscalibrated sensors. PSH are complex, highly interconnected systems encompassing electrical and hydraulic subsystems, each characterized by their respective underlying networks that can individually be represented as graph. To take advantage of this relational inductive bias, graph neural networks (GNNs) have been separately applied to state forecasting tasks in the individual subsystems, but without considering their interdependencies. In PSH, however, these subsystems depend on the same control input, making their operations highly interdependent and interconnected. Consequently, hydraulic and electrical sensor data should be fused across PSH subsystems to improve state forecasting accuracy. This approach has not been explored in GNN literature yet because many available PSH graphs are limited to their respective subsystem boundaries, which makes the method unsuitable to be applied directly. In this work, we introduce the application of spectral-temporal graph neural networks, which leverage self-attention mechanisms to concurrently capture and learn meaningful subsystem interdependencies and the dynamic patterns observed in electric and hydraulic sensors. Our method effectively fuses data from the PSH’s subsystems by operating on a unified, system-wide graph, learned directly from the data, This approach leads to demonstrably improved state forecasting performance and enhanced generalizability.

Health-aware Control for Health Management of Lithium-ion Battery in a V2G Scenario

Monica Spinola Felix — 2024-06-27

In response to the urgent need to combat climate change and reduce greenhouse gas emissions, the transition towards renewable energy sources such as solar and wind power is indispensable. However, the intermittent nature of these sources poses significant challenges to the stability of power grids. Battery Energy Storage Systems (BESS) offer a viable solution, and there is potential for Electric Vehicles (EVs) to serve as energy reservoirs, thereby bolstering grid stability through Vehicle-to-Grid (V2G) technology. While V2G holds promise, concerns persist regarding the longevity of batteries, particularly with the additional demand from charging and discharging cycles. To address these concerns, this study introduces a health-aware control strategy for V2G service scenarios. By employing feedback control mechanisms to adjust degradation rates, the strategy aims to effectively manage battery aging. Simulation outcomes of a V2G scenario with random input sources illustrate the efficacy of this proposed approach, demonstrating its potential applicability in practical settings where battery health needs to be managed. In summary, this research contributes to the advancement of health-aware strategies for an interconnected grid where electric vehicles participate as energy sources, with a primary focus on optimizing battery health management while fulfilling grid demands. Future efforts will concentrate on refining optimization strategies and integrating control methodologies with state estimators to ensure the performance of the approach on embedded battery health management systems.

Human-Centric PHM in the Era of Industry 5.0

Parul Khanna — 2024-06-27

The maintenance industry is undergoing a major transformation as it embraces the shift towards Industry 5.0. The focus of Industry 5.0 is on the integration of human intelligence with advanced technologies. It emphasizes interaction and collaboration between humans and machines and aims to combine the strengths of both. The efficiency of prognostics and health management (PHM) for maintenance in industrial contexts can be enhanced by improving this human-machine interaction and collaboration. This paper investigates the human-centric aspects, with a focus on PHM systems for facilitating the enablement of Industry 5.0 in maintenance. Acknowledging human as an active participant, this study explores their integral role in designing and developing PHM systems. The data collection for this study has been based on available literature, active and passive observations, and unstructured interviews and discussions with experienced industry professionals. As a result of the analysis of collected data, this study identifies and highlights potential areas for research and exploration. The research in these areas can advance the understanding and application of human-centric PHM strategies within Industry 5.0 in maintenance contexts. This is expected to improve the resilience and sustainability aspects of the industrial ecosystem and facilitate the shift towards Industry 5.0.

Hybrid AI-Subject Matter Expert Solution for Evaluating the Health Index of Oil Distribution Transformers

Augustin Cathignol — 2024-06-27

Reliability of oil distribution transformers is paramount, ensuring a stable flow of electricity and shielding from potential fire hazards. The internal insulation system of these transformers utilizes a combination of oil and paper. As the oil circulates through the active part of the system, it collects gaseous and physical traces of existing or past defects or degradations, providing a holistic view of the transformer's health, and allowing for early detection of problems and predictive maintenance. While various and mainly data-driven methods have been developed to calculate a transformer health index from oil samples, they lack accuracy due to limited data. This paper proposes a novel hybrid approach that leverages both Artificial Intelligence and Subject Matter Expertise to enhance the health estimation of oil distribution transformers. Our methodology utilizes a substantial dataset exceeding 65,600 analyzed oil samples, coupled with the valuable knowledge of domain experts. This combined approach achieves an accuracy exceeding 95%, suitable for real-world industrial applications. Furthermore, we introduce a risk management feature that strengthens the ability to identify transformers at high risk of failure. Notably, the health index estimation is implemented as a semi-automatic process, retaining the "expert in the loop" principle for managing critical and ambiguous cases.

Hybrid Prognostics for Aircraft Fuel System: An Approach to Forecasting the Future

Shuai Fu — 2024-06-27

The copious volumes of data collected incessantly from diverse systems present challenges in interpreting the data to predict system failures. The majority of large organizations employ highly trained experts who specialize in using advanced maintenance equipment and have specific certification in predictive maintenance (PdM). Prognostics and health management (PHM) connects research on deterioration models to strategies for PdM. Prognostics refer to the process of estimating the time until failure and the associated risk for one or more current and potential failure modes. Prognostics aim to provide guidance by alerting to imminent failures and predicting the remaining useful life (RUL). This eventually leads to improved availability, dependability, and safety, while also reducing maintenance costs. This research work diverges from existing established prognostic methodologies by emphasising the use of hybrid prognostics to predict the future performance of an aircraft system, especially the point in which the system will cease to operate as intended, often referred to as its time to failure. We have developed a new method that combines a physics-based model with the physics of failure (PoF) and a multiple-layered hyper-tangent-infused data-driven approach. The results are useful. The authors retrieved datasets for analysis using a laboratory aircraft fuel system and simulation model. Consequently, the comparative results demonstrate that the proposed hybrid prognostic approach properly estimates the RUL and demonstrates strong application, availability, and precision.

Influence of Reducing the Load Level of Mission Profiles on the Remaining Useful Life of a TO220 Analyzed with a Surrogate Model

Tobias Daniel Horn — 2024-06-27

A methodology for replacing finite element simulations with a fast-calculating surrogate model for fault tolerance in operating systems is presented. The study focuses on the TO220 rectifier system and explores methods to detect impending failures and calculate the resulting necessary load reduction. The finite element simulation model is described, highlighting the die attach as the relevant connection for failure. A surrogate model is developed using long-short-term-memory models to predict temperature and in-elastic strain. The surrogate model significantly reduces simulation time, allowing for the adjustment of load based on the system's current state of health. The rainflow counting algorithm is applied to calculate the number of cycles to failure, and the Palmgren-Miner linear damage accumulation relation is used to determine the damage and state-of-health. The dependency of the change in lifetime due to variations in scaling factor is evaluated and the results show that load reduction increases the lifetime of the system.

Integrating Network Theory and SHAP Analysis for Enhanced RUL Prediction in Aeronautics

Yazan Alomari — 2024-06-27

The prediction of Remaining Useful Life (RUL) in aerospace engines is a challenge due to the complexity of these systems and the often-opaque nature of machine learning models. This opaqueness complicates the usability of predictions in scenarios where transparency is crucial for safety and operational decision-making. Our research introduces the machine learning framework that significantly improves both the interpretability and accuracy of RUL predictions. This framework incorporates SHapley Additive exPlanations (SHAP) with a surrogate model and Network Theory to clarify the decision-making processes in complex predictive models and enhance the understanding of the hidden pattern of features interaction. We developed a Feature Interaction Network (FIN) that uses SHAP values for node sizing and SHAP interaction values for edge weighting, offering detailed insights into the interdependencies among features that affect RUL predictions. Our approach was tested across 44 engines, showing RMSE values between 2 and 17 and NASA Scores from 0.2 to 1.5, indicating an increase in prediction accuracy. Furthermore, regarding interpretability the application of our FIN, revealed significant interactions among corrective speed and critical temperature points key factors in engine efficiency and performance.

Integration of Condition Information in UAV Swarm Management to increase System Availability in dynamic Environments

Lorenz Dingeldein — 2024-06-27

The approach of prognostics and health management (PHM) focuses on the real-time health assessment of a system under its actual operating condition and even extending this by the prediction of the future state based on up-to-date system information. This pursues the aim to derive more advanced maintenance or asset deployment strategies in order to keep the operation of the system safe and reliable. In this context, the outcome of a PHM system is often used as a decision support. For a high fidelity system where the actual state is considered at every timestep and a decision is executed immediately based up on this information, Reinforcement Learning (RL) becomes a tool to find an optimized solution. Therefore the paper presents a methodology that integrates health and operational data into a RL approach in order to derive immediate operational strategies for lower degradation and higher safety and reliability. The approach is evaluated on the basis of a swarm of unmanned aerial vehicles (UAVs) that performs a complete-area path-coverage (CAPC) mission. It can be shown that the integration of health information as well as environmental data describing dynamic operating conditions lead to lower degradation and result in more reliable operations of the swarm while achieving a more flexible mission performance compared to pre-divided swarm-missions. Varying states are also taken into account, which emphasises this approach to be a highly dynamic PHM system application.

Labeling Algorithm for Outer-Race Faults in Bearings Based on Load Signal

Tal Bublil — 2024-06-27

Rolling element bearings are essential components for the proper functioning of many types of rotating equipment. Diagnosing faults in bearings has traditionally been done using signal processing techniques inspired by physics, wherein acceleration signals are analyzed using time-frequency analysis methods. To study the effect of bearing damage on acceleration signals, experiments are typically performed aiming for a natural propagation of a spall. However, the extent of spall severity during the test remains uncertain. It is possible to disassemble and reassemble the bearing for visual inspection. Nevertheless, previous studies observed that the vibration signal would drastically change if this operation was conducted repeatedly, impacting the identification of trends in the acceleration signal. The objective of this study is to provide a method which can assist with labeling the spall size in endurance tests without the necessity of disassembling and reassembling the test rig. To address this issue, a new algorithm, based on the load cell signal was developed to assess the spall size using low-speed measurements. This algorithm enables the identification of the circumferential angle at which the rolling element interacts with the spall and is only carrying a partial load. The algorithm has been validated through visual inspections conducted during the experiment. This algorithm makes it possible to estimate the spall size without the need for visual inspection in subsequent experiments. A labeled endurance test contributes to a better understanding of spall propagation, such as the effect of speed, load, and material properties on the propagation speed. This study demonstrates how the load signal can be used for fault labeling with relatively simple and common techniques. This approach will enable the tackling of advanced and more complex problems in future endeavors, such as fault severity estimation and even prognosis.

Landing Gear Health Assessment: Synergising Flight Data Analysis with Theoretical Prognostics in a Hybrid Assessment Approach

Haroun El Mir — 2024-06-27

This study addresses a critical shortfall in aircraft landing gear (LG) maintenance: the challenge of detecting degradation that necessitates intervention between scheduled maintenance intervals, particularly in the absence of hard landings. To address this issue, we introduce a Performance Degradation Metric (PDM) utilising Flight Data Recorder (FDR) output during the touchdown and initial roll phases of landing. This metric correlates time-series accelerometer data from a Saab 340B aircraft’s onboard sensors with non-linear response dynamic models that predict expected LG travel and reaction profiles across a set of ground contact cycles within a single landing. This facilitates the early detection of deviations from standard LG response behaviour, pinpointing potential performance abnormalities. The initiator of this approach is the Landing Sequence Typology, which systematically decomposes each aircraft landing into successive dynamic periods defined by their representative boundary conditions. What follows is the setting of initial parameters for the ordinary differential equations (ODE)s of motion that determine the orientation and impact responses of the most critical components of the LG assembly. Solving these ODEs with the integration of a non-linear representation of an oleo-pneumatic shock absorber model compliant with CS25 aircraft standards produces anticipated profiles of LG travel based on factors such as aircraft weight and speed at touchdown, which are subsequently cross-referenced with real accelerometer data, enhanced by video footage analysis. This footage is crucial for verifying the sequence of LG touchdowns and corresponding accelerometer outputs, thereby bolstering the precision of our analysis. Upon the conclusion of this study, by facilitating the early identification of LG performance deviations in specific landing scenarios, this diagnostic tool shall enable timely maintenance interventions. This proactive approach not only mitigates the risk of damage escalation to other components but also transitions main LG maintenance practices from reactive to proactive.

Large Language Model-based Chatbot for Improving Human-Centricity in Maintenance Planning and Operations

Linus Kohl — 2024-06-27

The recent advances on utilizing Generative Artificial Intelligence (GenAI) and Knowledge Graphs (KG) enforce a significant paradigm shift in data-driven maintenance management. GenAI and semantic technologies enable comprehensive analysis and exploitation of textual data sets, such as tabular data in maintenance databases, maintenance and inspection reports, and especially machine documentation. Traditional approaches to maintenance planning and execution rely primarily on static, non-adaptive simulation models. These models have inherent limitations in accounting for dynamic environmental changes and effectively responding to unanticipated, ad hoc events.

This paper introduces a maintenance chatbot that enhances planning and operations, offering empathetic support to technicians and engineers, boosting efficiency, decision-making, and on-the-job satisfaction. It optimizes shift scheduling and task allocation by considering technicians' skills, physical stress, and psychological state, thus reducing cognitive stress. The approach ultimately improves human performance and reliability, embodying a human-centricity in the domain of maintenance and health management.

The practical impact of the maintenance chatbot is illustrated through its application in maintenance of railway cooling systems. The presented use case demonstrates the chatbot's potential as a transformative tool in maintenance management. Finally, the paper discusses the theoretical and practical considerations, in particular in the light of regulative frameworks such as EU AI ACT, highlighting the future pathways for complying with responsible AI requirements.

Leveraging Generative and Probabilistic Models for Diagnostics of Cyber-Physical Systems

Alvaro Piedrafita — 2024-06-27

A critical task for system operators is the precise identification of the root causes underlying an error situation. This identification is fundamental in deciding optimal maintenance actions, such as replacing a component versus calibrating it. However, the actual causes of an error are often neither measured nor unique. The measured quantities are the result of complex interactions between different error causes and system variables. Root cause identification in this context becomes a matter of inferring hidden causes from their measurable effects. This challenge is notably pronounced in cyber-physical systems comprising control loops. Control mechanisms, integral to maintaining system performance, introduce a layer of complexity in diagnostics and ultimately complicate the isolation of the underlying causes of errors. To address this challenge, we introduce a two-step approach to derive the hidden causes as a statistical inference task. First, we develop a generative model leveraging existing control software and expert-based insights into the mechanisms of errors, i.e., a simulator of synthetic data given some hidden error causes. Then, we transform the generative model into a probabilistic program on which statistical inference can be executed within a probabilistic programming language framework. This inference effectively estimates the hidden causes given some measured data from the system. Being intrinsically a statistical approach, these inferences come with a confidence interval. We applied this methodology to an industrial printer’s sheet transport belt, operating in a closed-loop configuration. Our approach successfully discerned the contributions of three distinct hidden causes to the belt’s deviation from its intended position. This paper highlights the efficacy of generative modeling followed by a probabilistic programming approach in unraveling complex interactions within cyber-physical systems for optimal maintenance.

LSTM and Transformers based methods for Remaining Useful Life Prediction considering Censored Data

Jean-Pierre NOOT — 2024-06-27

Predictive maintenance deals with the timely replacement of industrial components relatively to their failure. It allows to prevent shutdowns as in reactive maintenance and reduces the costs compared to preventive maintenance. As a consequence, Remaining Useful Life (RUL) prediction of industrial components has become a key challenge for condition based monitoring. In many applications, in particular those for which preventive maintenance is the general rule, the prediction problem ismade harder by the rarity of failing instances. Indeed, the interruption of data acquisition before the occurrence of the event of interest leads to right censored data.

There are few articles in the literature that take that phenomenon into account for RUL prediction, even though it is common in the industrial environment to have a high rate of censored data. The present article proposes a deep-learning approach based on multi-sensor time series which allows to consider censored data during the training of the neural networks. Two methods are proposed, respectively based on the Dual Aspect Self Attention based on Transformer proposed by (Z. Zhang, Song, & Li, 2022) for non-censored data and on a recurrent neural network. Their evaluation on the C-MAPSS benchmark dataset shows, compared to the state-of-the-art RUL prediction methods, no loss in the absence of censoring, and outperformance on censored data.

Maintenance decision-making model for gas turbine engine components

Hongseok Kim — 2024-06-27

When designing gas turbine engine components, the inspection and maintenance (I&M) plan is prepared using the safe life. However, the I&M plan determined using safe life may be costly since all components are replaced at designated life. Therefore, it is important to make maintenance decisions considering the time-dependent deterioration process of gas turbine engine components for a cost-saving I&M plan. In this study, we proposed a maintenance decision-making model for gas turbine engine components based on a partially observed Markov decision process (POMDP). Using dynamic Bayesian networks, a decision-making model integrating a reliability analysis model, and a decision model for I&M planning was constructed. The signal amplitude data resulting from non-destructive inspection according to operation hour was used as partially observed data. The total cost obtained from the proposed model were compared with the results using a fixed I&M plan. The proposed model resulted in more cost-effectiveness I&M planning within affordable risk levels by considering the interaction between risk cost and I&M cost.

Maintenance Strategies for Sewer Pipes with Multi-State Degradation and Deep Reinforcement Learning

Lisandro Arturo Jimenez-Roa — 2024-06-27

Large-scale infrastructure systems are crucial for societal welfare, and their effective management requires strategic forecasting and intervention methods that account for various complexities. Our study addresses two challenges within the Prognostics and Health Management (PHM) framework applied to sewer assets: modeling pipe degradation across severity levels and developing effective maintenance policies. We employ Multi-State Degradation Models (MSDM) to represent the stochastic degradation process in sewer pipes and use Deep Reinforcement Learning (DRL) to devise maintenance strategies. A case study of a Dutch sewer network exemplifies our methodology. Our findings demonstrate the model's effectiveness in generating intelligent, cost-saving maintenance strategies that surpass heuristics. It adapts its management strategy based on the pipe's age, opting for a passive approach for newer pipes and transitioning to active strategies for older ones to prevent failures and reduce costs. This research highlights DRL's potential in optimizing maintenance policies. Future research will aim improve the model by incorporating partial observability, exploring various reinforcement learning algorithms, and extending this methodology to comprehensive infrastructure management.

Model-based Probabilistic Diagnosis in Large Cyberphysical Systems

Peter J.F. Lucas — 2024-06-27

Model-based diagnosis is concerned with diagnosing faults or malfunction of real-world physical or cyberphysical systems using a model of the structure and behavior of the systems. As cyberphysical systems can be extremely large and complex, and the associated computational models will be then equally large and complex, they impose a hard to beat challenge on the computational feasibility of reasoning with such models. When such a model is able to handle the uncertainty associated with diagnostics, giving rise to probabilistic model-based diagnostics, the computational feasibility becomes even harder. This paper: (1) proposes a novel graphical method underlying model-based diagnostics; (2) demonstrates experimentally how a novel, by the authors developed architecture of partitioned positive weighted model counting, is able to handle exact inference to answer a variety of probabilistic queries regarding the health status of a cyberphysical system. Results obtained are well within acceptable time bounds.

MOXAI – Manufacturing Optimization through Model-Agnostic Explainable AI and Data-Driven Process Tuning

Clemens Heistracher — 2024-06-27

Modern manufacturing equipment offers numerous configurable parameters for optimization, yet operators often underutilize them. Recent advancements in machine learning (ML) have introduced data-driven models in industrial settings, integrating key equipment characteristics. This paper evaluates the performance of ML models in classification tasks, revealing nuanced observations. Understanding model decision-making processes in failure detection is crucial, and a guided approach aids in comprehending model failures, although human verification is essential. We introduce MOXAI, a data-driven approach leveraging existing pre-trained ML models to optimize manufacturing machine parameters. MOXAI underscores the significance of explainable artificial intelligence (XAI) in enhancing data-driven process tuning for production optimization and predictive maintenance. MOXAI assists operators in adjusting process settings to mitigate machine failures and production quality degradation, relying on techniques like DiCE for automatic counterfactual generation and LIME to enhance the interpretability of the ML model's decision-making process. Leveraging these two techniques, our research highlights the significance of explaining the model and proposing the recommended parameter setting for improving the process.

NLP-Based Fault Detection Method for Multifunction Logging-While-Drilling Services

Corina Panait — 2024-06-27

This paper presents a Natural Language Processing (NLP) method aimed at detecting faults within field failure reports of drilling tools. It builds on the definition of entities specifically matched to our unique requirements. These entities have been annotated within the dataset under the guidance of a Subject Matter Expert (SME), laying a foundation for our NLP method. By utilizing a model based on bidirectional encoder representations from transformers, the method achieves an F1-score of 88\% in identifying entities and consequently detecting faults within field failure reports. This work is part of a long-term project aiming to construct a failure analysis and resolution system for drilling tools.

Noise-aware AI methods for robust acoustic monitoring of bearings in industrial machines

Kerem Eryilmaz — 2024-06-27

Traditionally, companies have relied on vibration based condition monitoring technologies to implement condition based maintenance strategies. However, these technologies have drawbacks, such as the requirement of contact accelerometers. As an alternative, acoustic condition monitoring is non-invasive and allows for easy deployment. Furthermore, the use of microphones potentially enables the monitoring of multiple components using a single sensor, making the monitoring system scale better with machine or production complexity. However, microphone signals typically show a low signal-to-noise ratio (SNR), impacted by the high level of background noise which is often present in industrial environments. Particularly, the traditional method for monitoring the health condition of rolling element bearings, based on assessing whether the squared envelope spectrum of the bearing signal exceeds a given threshold at the fault frequencies, cause too many false positives when applied directly to microphone signals. It is therefore crucial to develop strategies to increase the robustness of acoustic monitoring methods.

In this paper, we present and evaluate two data-driven strategies to robustly diagnose bearing faults from a microphone signal. Our proposed strategies are noise weighting based on the detection of background noise, and an artificial intelligence (AI) model that uses as input a combination of the traditional bearing fault frequencies and the mel spectrum of the microphone signal. These methods leverage both domain knowledge and data-driven techniques to increase the detection robustness. Our approach is implemented as a model trained and tested on bearing accelerated lifetime tests performed in the Smart Maintenance Lab setup at Flanders Make. Our results show that the use of our proposed strategies leads to significant improvements in diagnostic performance and time to first detection over noise-unaware acoustic monitoring methods.

On the Feasibility of Condition Monitoring of Belt Splices in Belt Conveyor Systems Using IoT Devices*

Henrik Lindström — 2024-06-27

This paper investigates fully automated condition monitoring of belt splices within operational belt conveyor systems, using IoT devices to predict and inform on potential belt breakage or tearing. Such events cause production stops and potentially harm workers. Belt splices are laminated belt connections subject to deterioration during operation and are usually weak spots. The proposed scheme circumvents manual inspection efforts and uses the HOSCHiris DISCOVER IoT device for sensing and data acquisition. Each belt conveyor is equipped with one individual IoT device acquiring the motion signal of the scraper which is used to learn signal patterns of the pulley and the belt to identify both location and deterioration of the individual splices. Deterioration is characterized from an initial healthy condition to a severe condition of the splice to inform on the potential need for action. To assess the feasibility of the scheme, several tests are designed and performed in an industrial belt conveyor system. The results indicate that the scheme can provide valuable insights into the splice condition and its degradation.

Particle Filter Approach for Prognostics Using Exact Static Parameter Estimation and Consistent Prediction

Kai Hencken — 2024-06-27

Particle filters are widely used in model-based prognostics. They estimate the future health state of an asset based on measurement data and an assumed degradation dynamics. Filters are in general applied to estimate only the states given a known dynamics of the process. In model-based prognostics, the dynamics is assumed to be known in an analytical form, but the parameters vary per device and need to be learned from the measurements as well. This is especially important for the calculation of the remaining useful life (RUL), as the prediction of the future evolution is needed.

There are commonly used approaches for this: Augmenting the state space with the parameter, together with assuming them to stay constant or adding an artificial diffusive evolution to them. The Liu–West filter improves on this by modifying the artificial evolution such that mean and standard deviation of the marginal parameter distribution are kept the same. Both approaches require to choose some tuning parameters, which might be difficult in practical applications. In addition, the model parameter is often assumed frozen for the prediction part, leading to an inconsistency. We propose how a modification of the parameter evolution in case of missing measurements can solve this in both cases.

More recently algorithms for combined state estimation and exact parameter estimation have been introduced, especially the Storvik filter, based on the usage of a sufficient statistic. We analyze how this can be applied to overcome difficulties with existing approaches, avoiding the need for tuning parameters. We also extend the Storvik filter in order to deal with time-steps with missing measurements. Two formally equivalent approaches are presented. These are applicable in all cases of missing measurements, coming either from irregular data acquisition, e.g. only during maintenance or inspection, or as part of the prediction step of the RUL calculation.

We study the different methods for two simple models in order to demonstrate potential issues with existing approaches and to explore the stability of the new one based on the Storvik filter. Finally we apply it to a practical application in the area of electrical distribution systems.

PHM for Spacecraft Propulsion Systems: Developing Resilient Models for Real-World Challenges

Takanobu Minami — 2024-06-27

This paper extends the research presented at the Prognostics and Health Management (PHM) Asia-Pacific 2023 Conference Data Challenge, focusing on a more pragmatic approach to spacecraft propulsion system health assessment. While the previous competition saw a variety of solutions, they predominantly relied on the assumption of highly stable initial hydraulic conditions – an idealization seldom met in real-world scenarios. In practical settings, factors such as operational noise, recent operational states, and ambient environmental conditions significantly disrupt this stability, rendering such solutions less feasible. Addressing this gap, our current study introduces a novel diagnostic model capable of valve faults without depending on the initial stable state of hydraulics. This approach marks a significant shift from our previous methodology, which primarily utilized similarity measures and physics-inspired features to classify health states and identify solenoid valve faults in spacecraft propulsion systems. The proposed model in this paper is validated against a diverse set of conditions, emphasizing its robustness and applicability in fluctuating real-world scenarios. Our findings demonstrate that the new model not only effectively diagnoses system health under varied and less controlled conditions but also enhances the practicality of spacecraft health management, offering a more adaptable solution in the face of operational uncertainties.

Probabilistic Uncertainty-Aware Decision Fusion of Neural Network for Bearing Fault Diagnosis

Atabak mostafavi — 2024-06-27

Reliability is a central aspect of machine learning applications, especially in fault diagnosis systems, where only an accurate and reliable diagnosis system is economically justifiable, considering that any false diagnosis would lead to an increase in maintenance costs or a reduction in system efficiency. Recent advances in machine learning (ML) techniques have encouraged condition monitoring researchers to focus their efforts on finding suitable ML-based solutions for system condition assessment. However, to address the reliability issue, it is crucial to consider a larger amount of data measured by heterogeneous sensors on the system together with non-sensor information. The trend of data fusion has already started in other areas of ML application, and many of today's state-of-the-art models benefit from various types of fusion techniques to improve their accuracy. However, traditional classifiers do not provide any information about the prediction uncertainty, and they tend to show falsely high confidence when encountering low-quality data or previously unseen classes. Fusion of different data sources without considering the epistemic or aleatory uncertainty can lead to a deterioration of the result. Bayesian frameworks have traditionally been used to quantify uncertainty of systems; however, only recent advances made it possible to successfully implement Bayesian ML models.

The research methodology was investigated using the MAFAULDA dataset generated by SpectraQuest's Machinery Fault Simulator. This simulator experimentally simulated various bearing conditions, including normal operation and inner and outer ring bearing failures, at variable speeds. The dataset consists of 1951 instances measured using two triaxial accelerometers, a microphone, and a tachometer.

Diagnosis has been done via two multi label 1D Convolutional Neural Networks - each for a selected sensor - and their prediction along with their associated uncertainty quantity has been fused utilizing Bayesian model averaging. The methodology is capable of fusion of various decisions made based on different data sources and generate a unified decision with associated confidence level. Fusion process is uncertainty aware and application of 1D networks reduce the amount of data needed.

Prognosis of Internal Short Circuit Formation in Lithium-Ion Batteries: An Integrated Approach Using Extended Kalman Filter and Regression Model

Lorenzo Brancato — 2024-06-27

The global transition to electric power, aimed at mitigating climate change and addressing fuel shortages, has led to a rising usage of lithium-ion batteries (LIBs) in different fields, notably transportation. Despite their many benefits, LIBs pose a critical safety concern due to the potential for thermal runaway (TR), often triggered by spontaneous internal short circuit (ISC) formation. While extensive research on LIB fault diagnosis and prognosis exists, forecasting ISC formation in batteries remains unexplored. This paper presents a new methodology that combines the extended Kalman filter (EKF) algorithm for real-time estimation of ISC state with an adaptive linear regressor model for forecasting remaining useful life (RUL). This approach is designed for seamless integration into actual battery management systems, offering a computationally efficient solution. Numerical validation of the framework was conducted due to the current lack of experimental data in the literature. The significance of this work lies in its contribution to ISC prognosis, providing a practical solution to enhance battery safety.

Remaining Useful Lifetime Estimation of Bearings Operating under Time-Varying Conditions

Alireza Javanmardi — 2024-06-27

This paper investigates the remaining useful lifetime (RUL) estimation of bearings under dynamic, i.e., time-varying, operating conditions (OC). Unlike conventional studies that assume constant OC in bearing accelerated life tests, we introduce a dataset with time-varying OC during run-to-failure experiments, simulating real-world scenarios. We explore data-driven approaches to identify the transition point from a healthy to an unhealthy state and estimate the RUL. Additionally, we examine strategies for integrating OC information to enhance RUL estimations. These methodologies are evaluated through numerical experiments using various machine learning algorithms.

Residual Selection for Observer-Based Fault Detection and Isolation in a Multi-Engine Propulsion Cluster

Renato Murata — 2024-06-27

For complex systems, the number of residual candidates generated by Structural Analysis could be in the order of tens of thousands, and implementing all candidates is infeasible. This paper addresses the residual generator candidate selection problem from a state-observer perspective. First, the most suitable candidates to derive state-observers are selected based on two criteria related to the state-space form and a low number of equations. Then, a novel algorithm finds the minimal subset of residual generator candidates capable of detecting and isolating all faults. A procedure is introduced to compare the fault sensitivity of the selected candidates. This residual selection method is applied to the multi-engine propulsion cluster of a reusable launcher to illustrate its benefits.

Robust Remaining Useful Life Prediction Using Jacobian Feature Regression-Based Model Adaptation

Prasham Sheth — 2024-06-27

The accurate and robust prediction of remaining useful life (RUL) is critical for enabling the proactive mitigation of fault effects rather than reacting to them. For RUL prediction, one must model nominal and faulty system behaviors and how different faults progress over time. Complex data-driven machine learning (ML) models may capture both nominal and fault progression by updating the model parameters at different stages. As new data are observed, these model parameters can be updated to keep the system model always accurate. However, complete retraining of these models is both data- and computation-intensive and unsuitable for dynamic, fast-changing environments requiring quick recalibration. This calls for efficiently adapting the model to new operating conditions or the system’s current state. One such efficient way to recalibrate model parameters to newly observed data using Jacobian feature regression (JFR) is presented in Forgione, Muni, Piga, and Gallieri (2023), where a recurrent neural network (RNN) models the current behavior of the dynamic system. Then, any subsequent deviation of observed measurements and the RNN model is attributed to an “unacceptable degradation of the nominal model performance.” To update the RNN model, Forgione et al. (2023) propose augmenting the current model with additive correction terms learned by implementing JFR on observed “perturbed system” data. In this paper, we propose an automated online framework to adapt the model efficiently to always reflect the system’s current state and use it for accurate RUL prediction and select JFR as one such adaptation technique. We extend the implementation of JFR-based model adaptation to hybrid models and demonstrate JFR to be more sustainable than the other retraining methods. Finally, we showcase the application of this approach to the oil and gas industry. A testbed that simulates a digital synthetic oilfield is used to show the effectiveness of this adaptation-based RUL prediction technique.

Simulation-based remaining useful life prediction of rolling element bearings under varying operating conditions

Seyed Ali Hosseinli — 2024-06-27

Remaining useful life (RUL) prediction of rolling element bearings is a complex task in the frame of condition monitoring which brings cost benefits to the industry by reducing unexpected downtimes and failures. Data-driven approaches based on deep learning have demonstrated exceptional performance in estimating RUL effectively. Nevertheless, challenges such as data scarcity for model training and varying operating conditions add more complexity to prognostic tasks using these methods. This study proposes a methodology for simulating the vibration signals during the degradation process of bearings in order to mitigate the need for historical data for training the models. Simulations are realized using a phenomenological model whose free parameters are adapted based on real measurements so that the simulated run-to-failure datasets are under the same influence of speed as the real dataset with almost the same degradation rate. The simulated dataset is used for model training. Moreover, the proposed methodology is able to react to the shaft speed and be flexible at the predictions when the speed of the bearing varies. The proposed model can take extra information regarding the operating speed and the sequential ordering of the measurements to be aware of the working conditions and the dynamics of the damage progression. The positive effect of the extra information is shown in the results. Model training is based on an unsupervised domain adaptation approach to reduce domain discrepancy between the simulated and real feature space. The effectiveness of the proposed method is examined according to bearing run-to-failure tests under varying operating conditions.

Soft Ordering 1-D CNN to Estimate the Capacity Factor of Windfarms for Identifying the Age-Related Performance Degradation

Manuel Sathyajith Mathew — 2024-06-27

Wind energy plays a vital role in meeting the sustainable development goals set forth by the United Nations. Performance of wind energy farms degrades gradually with aging. For deriving maximum benefits from these capital-intensive projects, these degradation patten should be analyzed and understood. Variations in the capacity factor over the years could be an indication of the age-related degradation of the wind farms. In this study, we propose a novel data-driven model to estimate the capacity factor of wind farms, which could then be used to estimate its age-related performance decline. For this, a 1-dimensional convolutional neural network (1-D CNN) is developed with a soft ordering mechanism under this study. The model was optimized using Huber loss to counteract the effects of outliers in data. The developed model could perform very well in capturing the underlying dynamics in the data as evidenced by a normalized root mean squared error (NRMSE) of 0.102 and a mean absolute error (MAE) of 0.035 on the test dataset.

State-of-Charge and State-of-Health Estimation for Li-Ion Batteries of Hybrid Electric Vehicles under Deep Degradation

Hyunjoon Lee — 2024-06-27

In recent industry, hybrid vehicles are gaining more recognition as a practical means for future transportation due to the longer distance, reduced charging time, and less charging stations dependency. The batteries in the hybrid vehicles, however, undergo more complex operation of charge depleting and sustaining modes alternately, which may need more accurate battery state estimation. In this study, a model based method is explored for the Li-ion batteries in the hybrid electric vehicles to estimate State-of Charge (SOC) and State-of-Health (SOH) accurately. While there have been widespread studies for this topic in the batteries research, not many are found that have investigated hybrid operation modes. Also the estimations are mostly limited to normal batteries or shallow degradation with the SOH higher than 90%. In this study, an algorithm based on the dual extended Kalman filter (DEKF) and enhanced self-correcting (ESC) model is developed for the simultaneous estimation of the SOC and SOH. Degradation data for plug-in hybrid vehicle (PHEV) are taken for the study, which undergo the deep degradation of 30%. In order to maintain the accuracy such that the root mean square error (RMSE) of the SOC is within 5% over the entire degradation cycles, two practical methods are proposed: First, the SOH is estimated separately during the battery charging, and is used as a constant in the SOC estimation in the discharging cycles. Second, battery modeling is conducted and the parameters are reset in every intermittent cycles at which the SOH is reduced by 10% initially and by 5% thereafter.

Statistical Knowledge Integration into Neural Networks: Novel Neuron Units for Bearing Prognostics

Thomas Pioger — 2024-06-27

Prognostics and Health Management (PHM) is a framework that assesses the health condition of complex engineering assets to ensure proper reliability, availability, and maintenance. PHM can be used to determine how long a machine can function before failure by predicting the Remaining Useful Life (RUL). Neural networks have been used for RUL prediction, but these data-driven models rely solely on data to explicitly integrate knowledge. Recently, authors have proposed physics-informed neural networks (PINNs) to address this limitation. PINNs are neural networks that incorporate expert knowledge and physics in different ways (observational,inductive, and learning bias). Despite their significance, these models tend to be case-dependent and challenging to configure. In this work, we propose statistical neuron units that can be integrated into any neural network. The proposed neuron units extract features from raw data using various statistical functions. Importantly, these modules can be located in different parts of the neural network, and they can be optimized automatically by backpropagating the modules’ weights during training. In a study involving bearing degradation behavior, we compare a classical neural network with our modular
version. Our proposed RUL estimation model outperformed the baseline, with a reduction of 13% in the root mean square
error and a reduction of 7% in the mean absolute error. We also observe an increase of 40% and 21% for the α − λ accuracy metric for an α equal to 0.1 and 0.2 respectively. Our code is available publicly on Github.

SurvLoss: A New Survival Loss Function for Neural Networks to Process Censored Data

Mahmoud Rahat — 2024-06-27

This paper presents SurvLoss, a novel asymmetric partial loss and error calculation function for survival analysis and regression, enabling the inclusion of censored samples. An observation in a dataset for which the complete information regarding an event of interest is not available is called censored. Censored samples are ubiquitous in the industry and play a crucial role in Prognostics and Health Management (PHM) by providing a realistic representation of data, improving the accuracy of analyses, and supporting better decision-making in various industries and the healthcare sector. The proposed approach can effectively equip the conventional regression loss functions such as Mean Absolute Error (MAE), Mean Squared Error (MSE), or Root Mean Squared Error (RMSE) with the ability to process censored samples. This can impact the field hugely by providing a more accessible usage of neural network models in survival analysis. The proposed survival loss incorporates censored samples by penalizing predictions outside the censoring region and skipping them otherwise. Then, it uses weighted averaging to aggregate the loss from censored samples with the loss from event samples. Unlike many other methods in the field, the proposed model distinguishes itself by avoiding superficial assumptions and exclusively relies on the available information, considering the entirety of the data.

We compared the proposed loss function with its baseline on two publicly available datasets. The first dataset, called C-MAPSS, is from NASA Turbofan Jet Engines simulation, and the second is a recently published real-world dataset from SCANIA trucks. The goal of both datasets is to predict the remaining useful life (RUL) of the machines.
The experimental results show that optimization algorithms for training deep neural networks like Adam can effectively utilize the proposed loss function to calculate gradients, update the model's weights, and reduce training and test errors. Moreover, the proposed model outperformed the baseline by taking advantage of the censored samples. The proposed loss function paves the way for the employment of advanced architectures of neural networks with bigger training sizes in survival analysis.

System-level Probabilistic Remaining Useful Life Prognostics and Predictive Inspection Planning for Wind Turbines

Davide Manna — 2024-06-27

Wind energy plays a crucial role in the energy transition. However, it is often seen as an unreliable source of energy, with many production peaks and lows. Some of the drivers of uncertainty in energy production are the unexpected wind turbine (WT) failures and associated unscheduled maintenance.

To support an effective health management and maintenance planning of WTs, we propose an integrated data-driven framework for Remaining Useful Life (RUL) prognostics and inspection planning of WTs. We propose a Long-short term memory (LSTM) neural network with Monte Carlo dropout to estimate the distribution of the RUL of WTs, i.e. we develop probabilistic prognostics. Different from existing studies focused on prognostics for single components, we consider the simultaneous health-monitoring of multiple components of the WTs, thus seeing the turbine as an integrated system. The obtained prognostics are further included into a stochastic planning model which determines optimal moments for inspections. For this, we pose the problem of WT inspections as a renewal reward process. We illustrate our framework for four offshore WTs which are continuously monitored by Supervisory Control and Data Acquisition (SCADA) systems. The results show that LSTMs are able to estimate well the RUL of the WTs, even in the early phase of their usage. We also show that the prognostics are informative for maintenance planning and are conducive to conservative inspections.

Testing Topological Data Analysis for Condition Monitoring of Wind Turbines

Simone Casolo — 2024-06-27

We present an investigation of how topological data analysis (TDA) can be applied to condition-based monitoring (CBM) of wind turbines for energy generation.
TDA is a branch of data analysis focusing on extracting mean- ingful information from complex datasets by analyzing their structure in state space and computing their underlying topo- logical features. By representing data in a high-dimensional state space, TDA enables the identification of patterns, anoma- lies, and trends in the data that may not be apparent through traditional signal processing methods.

For this study, wind turbine data was acquired from a wind park in Norway via standard vibration sensors at different lo- cations of the turbine’s gearbox. Both the vibration acceler- ation data and its frequency spectra were recorded at infre- quent intervals for a few seconds at high frequency and fail- ure events were labelled as either gear-tooth or ball-bearing failures. The data processing and analysis are based on a pipeline where the time series data is first split into intervals and then transformed into multi-dimensional point clouds via a time-delay embedding. The shape of the point cloud is an- alyzed with topological methods such as persistent homol- ogy to generate topology-based key health indicators based on Betti numbers, information entropy and signal persistence. Such indicators are tested for CBM and diagnosis (fault de- tection) to identify faults in wind turbines and classify them accordingly. Topological indicators are shown to be an in- teresting alternative for failure identification and diagnosis of operational failures in wind turbines.

Test-Training Leakage in Evaluation of Machine Learning Algorithms for Condition-Based Maintenance

Omri Matania — 2024-06-27

Many articles have been published utilizing machine learning algorithms for condition-based maintenance through the analysis of vibration signals. One extensively researched topic is the classification of fault types in rolling bearings. There is a fairly widespread problem in the evaluation of these learning algorithms, where the separation of examples between the test and training sets is incorrect, leading to an optimistic conclusion about the algorithm's performance even when it is not the case. In this article, we will review this issue and explain how the data should be properly divided between the test and training sets to avoid this occurrence.

Timeseries Feature Extraction for Dataset Creation in Prognostic Health Management

Thanos Kontogiannis — 2024-06-27

This study focuses on a critical aspect of implementing prognostics and health management (PHM) for assets: the creation of a descriptive dataset. In real-world applications, dealing with sparse and unlabelled big data is common, particularly in industries like production lines where complex subprocesses are monitored by multiple sensors. Moreover, selective application of quality control means that much of the data lacks information about end properties, making datasets provided by manufacturers unsuitable for PHM frameworks. This work aims to bridge the gap between raw production data and PHM frameworks, focusing on steel manufacturing management. In the context of steel manufacturing, compromised surface quality, characterized by thicker oxide layers chipping during milling, has been observed. We propose inferring compromised coils by analyzing temperature profiles directly before the coiling station to address this. Deviations from the goal temperature profile can indicate compromised surface quality, eliminating the need for tedious oxide layer thickness measurements, which are not feasible for continuous hot strip milling processes. The available dataset comprised multiple years of production, with no direct indication of the surface quality. Exploratory clustering analysis was the first step in the lack of labels. Even though indicative of the underlying pattern of the healthy/damaged coils distinction, three shortcomings were identified. Clustering was solely based on the similarity between the temperature profiles of the coils, so no domain knowledge was included regarding the goal temperature profile. Additionally, since different steel grades have different goal profiles, the model needs to be specifically trained for each grade. Also, a soft classification between healthy and damaged can provide more detailed information about the surface quality. Coils with low-confidence classifications can be identified and treated accordingly, thereby improving PHM framework performance by providing a dataset with only high-confidence samples. To tackle these issues, an expert-knowledge-based normalization technique and feature engineering, paired with synthetic labelling, contributed to the creation of a soft neural network classifier. This study presents the reality of handling real-world data for PHM applications and highlights the need for careful and informed feature extraction. This ensures the seamless integration of PHM frameworks into real-world systems, ultimately enhancing production yield by improving end-product quality.

Towards a Hybrid Framework for Prognostics with Limited Run-to-Failure Data

Luc Keizers — 2024-06-27

The introduction of cyber-physical systems with increased availability of sensor data creates a lot of research interest in prognostic algorithms for predictive maintenance. Although a lot of algorithms are successfully applied to benchmark case studies based on simulated data and experimental set-ups, deployment
in industry lags behind. From a comparison between three benchmark case studies with two real-world case studies based on prognostic metrics (monotonicity, prognosability and trendability), two main issues are observed: 1) the lack of run-to-failures and 2) low prognostic metrics due to a low signal-to-noise ratio of degradation trends, as a result of unexplained physical phenomena. To make prognostics feasible, a hybrid framework is proposed that focuses on improving system knowledge. The framework consists of a quantitative diagnostic assessments, guided by (modular) system models in which damage is induced. This quantitative damage assessment provides input for prognostics based on Bayesian filtering, enabling prognostics for assets in varying operational conditions. Implementation and validation of the framework requires investments, but modularity within the framework can accelerate development for new systems.

Towards a Probabilistic Fusion Approach for Robust Battery Prognostics

Jokin Alcibar — 2024-06-27

Batteries are a key enabling technology for the decarbonization of transport and energy sectors. The safe and reliable operation of batteries is crucial for battery-powered systems. In this direction, the development of accurate and robust battery state-of-health prognostics models can unlock the potential of autonomous systems for complex, remote and reliable operations. The combination of Neural Networks, Bayesian modelling concepts and ensemble learning strategies, form a valuable prognostics framework to combine uncertainty in a robust and accurate manner. Accordingly, this paper introduces a Bayesian ensemble learning approach to predict the capacity depletion of lithium-ion batteries. The approach accurately predicts the capacity fade and quantifies the uncertainty associated with battery design and degradation processes. The proposed Bayesian ensemble methodology employs a stacking technique, integrating multiple Bayesian neural networks (BNNs) as base learners, which have been trained on data diversity. The proposed method has been validated using a battery aging dataset collected by the NASA Ames Prognostics Center of Excellence. Obtained results demonstrate the improved accuracy and robustness of the proposed probabilistic fusion approach with respect to (i) a single BNN model and (ii) a classical stacking strategy based on different
BNNs.

Towards Efficient Operation and Maintenance of Wind Farms: Leveraging AI for Minimizing Human Error

Arvind Keprate — 2024-06-27

To effectively compete with other renewable energy sources, there remains a critical need to further decrease the Levelized Cost of Energy of Wind Farms (WFs). A promising way to achieve this objective is by minimizing the downtime of wind turbines (WTs) through effective Inspection and Maintenance (I&M) activities. Conventionally, I&M plans have predominantly relied on CM/SCADA data obtained from the physical components of turbines, with data analytics and machine learning (ML) techniques being employed to predict their performance and maintenance needs. However, statistics indicate that nearly 40% of WT failures can be traced back to HFs. These include aspects such as skills, knowledge, communication, and even the broader organizational culture. This paper delves into the importance of integrating HFs in the I&M of WFs to optimize turbine performance, enhance safety, and reduce downtime.

Firstly, we briefly discussed various Human Reliability Analysis (HRA) methods with special emphasis on Performance Shape Factors (PSFs). We then identify key human factors (HFs) that are vital for performing O&M tasks. For this, we have prepared a questionnaire to get qualitative input from technicians and also done a thorough literature review. E.g., some of the HFs that stand out include the ergonomics of tools and workspace designs tailored to technicians' needs, the cognitive load placed on operators during system monitoring and diagnostics, continuous training to handle evolving challenges, effective communication channels, and safety protocols designed with human behavior in mind. We then propose a novel framework for developing a computer vision-based recommendation system that can guide the technicians to perform the maintenance effectively thus minimizing the HE.

Towards Physics-Informed PHM for Multi-component degradation (MCD) in complex systems

Atuahene Barimah — 2024-06-27

This study seeks to address the challenge of limited degradation data in developing Fault Detection and Isolation (FDI) models for multi-component degradation (MCD) scenarios. Utilizing a small fraction (0.05%) of a previously utilized water distribution testbed dataset in a previous publication, a weighted ensemble hybrid approach is proposed and evaluated against more established modelling approaches used in the previous publication. The proposed approach combines heuristic approximation and Physics-Informed Neural Network (PINN) methods with a recurrent neural network (RNN) model to enhance diagnostic performance for predicting MCD scenarios. The hybrid model generally outperformed other algorithms when tested on an MCD dataset, demonstrating improved diagnostic accuracy in such scenarios. Future research aims to optimize ensemble weights based on model uncertainty, further enhancing diagnostic capabilities.

Transfer Learning-based Adaptive Diagnosis for Power Plants under Varying Operating Conditions

Jiwoon Han — 2024-06-27

Transfer learning is a method that transfers knowledge learned from a source domain to a similar target domain to improve learning. In power plants, obtaining sufficient anomaly data is difficult due to the characteristics of the systems. Transfer learning enables learning with only a small amount of data from the target domain by using a model trained in a similar domain. By applying transfer learning, models developed for one power plant can be expanded and used in other power plants where available data are limited.

Using actual data from an operating combined-cycle power plant, an anomaly diagnosis model was developed and tested. Its applicability to different operating conditions and anomaly cases was evaluated through transfer learning. The fine-tuned pre-trained model was effectively adapted with limited target domain data. Transfer learning was applied despite the limitations of data and distribution differences. The expandability of anomaly diagnosis models to different power plant systems was demonstrated by applying transfer learning.

Ultrafast laser damaging of ball bearings for the condition monitoring of a fleet of linear motors

Abdul Jabbar — 2024-06-27

Machine learning-based condition monitoring of mechanical systems, such as bearings, employs two primary approaches: unsupervised and supervised methods. Unsupervised approaches aim to characterize the healthy state of the machine and monitor deviations from this state. The advantage lies in requiring only the health condition of the component without the need for historical data until breakdown. However, the disadvantage is the lack of information regarding the root cause of any potential malfunction.
On the other hand, supervised methods consider both healthy and faulty cases, aiming to maximize the difference between them through post-processing, as well as among different fault types. The advantage is the ability to analyze the specific signature of a particular fault type. Nonetheless, the disadvantage is that available data usually do not cover all possible faults that may occur.
Typically, obtaining a faulty bearing involves either a time-consuming run-to-failure test or the artificial induction of faults using drills, electro-discharge pens, etc. While artificial faults offer a quicker procedure, they often fail to replicate real faults faithfully. This paper suggests using picosecond laser technology to engrave the surface of the bearing and create artificial faults. Modern laser technology allows for precise control over the dimensions of injected faults, enhancing the understanding of fault progression at various stages in the life of bearings. These measurements are crucial parameters for evaluating the robustness of diagnostic algorithms. This paper focuses on artificially damaging a ball bearing used in an independent cart systems application, which comprises a fleet of linear motors moving on the same rail. These systems have recently been proposed by different manufacturers and adopted in the field of packaging machines for their flexibility. For such systems, no prior instances of faulted bearings are available, and the size of a real fault is also unknown. Hand-made faults with drills did not produce discernible faults appreciable in post-processing of the data. Therefore, a picosecond laser with a pulse duration of 10 ps and a maximum energy per pulse of approximately 100 µJ is utilized to create a set of test bearings with increasing fault sizes on the outer race. Post-processing of the data enables the qualification of the minimum fault severity detectable in this specific application.

Uncertainty in Aircraft Turbofan Engine Prognostics on the C-MAPSS Dataset

Mariana Salinas-Camus — 2024-06-27

Prognostics and Health Management (PHM) plays a crucial role in maximizing operational efficiency, minimizing maintenance costs, and enhancing system reliability. Predicting Remaining Useful Life (RUL) is a key aspect of PHM, inherently incorporating uncertainty. This paper focuses on uncertainty quantification (UQ) within Data-Driven Models (DDMs), particularly Machine Learning (ML), such as Long Short-Term Memory (LSTMs), and stochastic models namely Hidden Markov Models (HMMs). While ML models emphasize accuracy, stochastic models offer a different paradigm for prognostics, directly addressing uncertainty. Traditional categorizations of uncertainty as aleatory and epistemic face challenges in practical implementation. This paper explores how, in prognostics, HMMs primarily tackle aleatory uncertainty, whereas LSTMs predominantly address epistemic uncertainty. It also discusses the complexities of uncertainty
management in prognostics and analyzes further an already proposed alternative approach to categorize uncertainties. Despite theoretical advancements, practical implementation remains challenging, especially for DL models due to their limited interpretability. This study sheds light on UQ challenges and offers insights for future research directions in prognostics.

Unsupervised Learning for Bearing Fault Identification with Vibration Data

Gianluca Nicchiotti — 2024-06-27

Machine learning methods are increasingly used for rotating machinery monitoring. Usually at system set up, only data of the machinery in healthy conditions, the so-called nominal data, are available for the machine learning phase. This type of training data enables fault detection capabilities and several methods such as Gaussian Mixture Model, One Class Support Vector Machines and Auto Associative Neural Networks (Autoencoders) have been already proved successful for this task.

However, in some predictive maintenance applications, information on the type of defect may represent a key element for producing actionable information, e.g. to reduce diagnostic burden and optimize spare procurement. This requires to define classification strategies based on machine learning even in absence of data representing the behaviour of the system with defects.

In this study we present an approach that uses only nominal data to train an autoencoder which will enable at same time fault identification and fault classification tasks.

An autoencoder is a network which learns to duplicate the input at the output. Even if this replication task may seem trivial, the presence of a “bottleneck” in the hidden layer forces the network to learn the significant features of the input data.

As faulty data are expected to possess information content which is structured differently from the healthy ones their reconstruction at output will result inaccurate. In conventional anomaly detection approaches, the module of the reconstruction error, defined as the difference between output and input, is uses to determine an unusual input such as faults.

The proposed approach represents a step forward as here a single autoencoder is used both for detection and classification.

The underlying idea is that the components of the reconstruction error vector whose module is used to trigger fault identification in classical autoencoder approaches contain the information of the fault type. This way the analysis of the different components of the reconstruction error allows to differentiate the different types of fault.

Two methods to analyse the components of the reconstruction error vector will be presented.

In the first strategy autoencoder input features more sensitive to each kind of fault are identified on the basis of a-priori knowledge and an expected ranking pattern associated to each fault type is defined. Once a new input signal is presented to the autoencoder, in case the module of the reconstruction error detects an anomaly, the ranking of the magnitudes of reconstruction error components is established and the Rank Biased Overlap (RBO) algorithm is used to measure the similarity between the computed data and the expected fault templates. This method can be easily generalised to other cases, since it is not the amplitudes of the errors that serve as a reference but only the ranking.

The latter method plots the reconstruction error components on a polar diagram and uses their position to compute the position of the barycentre of resulting star-shaped figure. The angular position of the barycentre is then used to classify the fault type.

A machine fault simulator has been used to generate 3 different types of bearing defects with different load, speed and noise conditions and a dataset of about 10000 signals has been employed to test the classification algorithms.

The results obtained using the autoencoder method do not achieve the same performances as the conventional supervised learning algorithms. However, they proved to be 88% accurate in classification when SNR is above 0dB with the ranking based method overperforming the barycentre one.

Virtual Sensor for Real-Time Bearing Load Prediction Using Heterogeneous Temporal Graph Neural Networks

Mengjie Zhao — 2024-06-27

Accurate bearing load monitoring is essential for their Prognostics and Health Management (PHM), enabling damage assessment, wear prediction, and proactive maintenance. While bearing sensors are typically placed on the bearing housing, direct load monitoring requires sensors inside the bearing itself.
Recently introduced sensor rollers enable direct bearing load monitoring but are constrained by their battery life.
Data-driven virtual sensors can learn from sensor roller data collected during a battery's lifetime to map operating conditions to bearing loads.
Although spatially distributed bearing sensors offer insights into load distribution (e.g., correlating temperature with load), traditional machine learning algorithms struggle to fully exploit these spatial-temporal dependencies. To address this gap, we introduce a graph-based virtual sensor that leverages Graph Neural Networks (GNNs) to analyze spatial-temporal dependencies among sensor signals, mapping existing measurements (temperature, vibration) to bearing loads. Since temperature and vibration signals exhibit vastly different dynamics, we propose Heterogeneous Temporal Graph Neural Networks (HTGNN), which explicitly models these signal types and their interactions for effective load prediction.
Our results demonstrate that HTGNN outperforms Convolutional Neural Networks (CNNs), which struggle to capture both spatial and heterogeneous signal characteristics. These findings highlight the importance of capturing the complex spatial interactions between temperature, vibration, and load.

A novel prognostics solution for accurate identification of degradation patterns in turbo machines with variable observation window

Carmine Allegorico — 2024-06-27

The degradation of a system is a time bound phenomenon, which leads to the deterioration of turbomachinery, in terms of performance and reliability. If undetected and not acted upon in time, this could also lead to sudden system failure, resulting in unplanned unit downtime and maintenance. Unplanned downtime of a turbomachine leads to severe production loss for the end customer and consequent economic damages. Early detection of a degradation pattern would provide the customer with the opportunity to timely carry out corrective actions, preventing an unscheduled down time. The paper evaluates degradation identification methodology currently known from literature and finds them not accurate enough for general purpose application required by the solution. The paper discusses a novel methodology which can accurately detect degradation patterns of timeseries data. Critical features of this methodology are novel time-based correlation enabled regression model with variable observation window, autonomous training, and automatic adjusting capability to incorporate operating behavior change or physical system replacement. This leads to high accuracy, high generalization, and domain agnostic application capability. Moreover, particular focus is given to achieving high probability of detection and a low probability of false alarm. The paper demonstrates the performance achieved by the methodology when applied to the field of prognostics and diagnostics of IoT connected turbomachines through 50+ real application cases.

Case Study of Product Development through Generative Design according to Anemometer Replacement Cycles

Joongyu CHOI — 2024-06-27

Product Lifecycle Management (PLM) systems are commonly used to manage various product data generated throughout the product lifecycle. This paper explains the results obtained by multiple participants using commercial software within the PLM environment to perform structural and vibration analyses of an Anemometer. Generative design techniques were employed for 3D CAD modeling of the Anemometer, and the commercial analysis software NASTRAN was used for simulation analyses. The open-source PLM system ARAS Innovator's project and workflow management modules were utilized to manage the generated design data, allocate tasks among participants, and control schedules. Through this approach, we propose a method to predict and manage the replacement cycle of Anemometer.

Feature Selection Method for Gear Health Indicator Using MIC Ranking

Hongliang Song — 2024-06-27

In the construction of health indicator for electromechanical equipment, selecting features that exhibit monotonicity, trend characteristics, and a strong correlation with equipment health is paramount to accurately reflect these indices. With the advent of numerous libraries and models for time-series data feature extraction, the range of potential features has expanded significantly. Despite this proliferation, there is a lack of extensive research on effective feature selection. This paper investigates the efficacy of the Maximum Information Coefficient (MIC) method in extracting features that align with the monotonicity and trend-related requirements of electromechanical equipment health indicator. Our experiments indicate that the MIC method adeptly identifies pertinent features for the construction of these indices, underlining its utility in the field of health monitoring for electromechanical systems.

Filter-based feature selection for prognostics incorporating cross correlations and failure thresholds

Alexander Loewen — 2024-06-27

Historical condition monitoring data from technical systems can be utilized to develop data-driven models for predicting the remaining useful life (RUL) of similar systems, whereas the Health Index (HI) often is a crucial component. The development of robust and accurate models requires meaningful features that reflect the system’s degradation process, enabling an accurate prediction of the system's HI. Traditionally, the identification of those is supported by one of various feature ranking methods. In literature, feature interdependencies and their transferability across various similar systems are not sufficiently considered in feature selection, exacerbating the challenge of HI prediction posed by the scarcity of data and system diversity in real-world applications. This work addresses this gaps by demonstrating how filter-based feature selection, incorporating failure thresholds and cross correlations, enhances feature selection leading to improved HI prediction. The proposed methodology is applied to a novel dataset* obtained from run-to-failure experiments on geared motors conducted as part of this study, which presents the aforementioned challenges. It is revealed that classical feature selection, consisting of feature ranking only, leaves potential untapped, which is utilized by the proposed selection methodology. It is shown that the proposed feature selection methodology leads to the best result with a RMSE of 0.14 in predicting the HI of a constructive different gearbox, while the features, determined by classical feature selection, lead to a RMSE of 0.19 at best.

Integrated design of negative stiffness honeycomb structures considering performance and operational degradation

Hyeong-Do Kim — 2024-06-27

This study introduces an integrated framework for conceptualizing the design of negative stiffness honeycomb (NSH) structures, specifically considering the durability and performance of their unit cells. Unlike conventional energy-absorbing structures that rely on plastic deformation, NSH offers a promising alternative for reusable energy absorption (EA) and high initial stiffness, making it suitable for a wide range of engineering applications. The research considers the variability in characteristics of NSH based on the shape of the configured negative stiffness beam (NSB), selecting a single curved-beam unit cell as the focal point. Extensive testing, including quasi-static and cyclic compression tests, is conducted on NSH unit cell fabricated using polylactic acid/polyhydroxy alkenoate (PLA/PHA) filament, to analyze performance under stress and to assess degradation over time. Central to the study is the use of multi-objective optimization (MOO) to explore the trade-off between performance and operational durability, thereby emphasizing the significance of degradation in the design process. The results demonstrate the potential for NSH structures, particularly in terms of their reusability and efficiency, highlighting the viability of incorporating durability considerations in the early stages of design, especially for structures intended for additive manufacturing processes.

Mastering Training Data Generation for AI - Integrating High- Fidelity Component Models with Standard Flight Simulator Software

Andreas Löhr — 2024-06-27

The German state-funded aviation research project “Real- time Analytics and Prognostic Health Management” (RTAPHM) envisioned fully automated urban air services executed by autonomous drones and infrastructure controlled by a digital system. Research was focused on utilizing onboard real-time diagnostics to enable AI-driven UAV capability predictions. These predictions increased the reliability of upfront service commitments. The use case selected to demonstrate these elements was organ transport. The project delivered an end-to-end demonstrator incorporating a virtual fleet of drones with onboard diagnostics to provide data for the platform decision logic.

The project followed a „digital-twin-first” approach to overcome a common bootstrapping problem faced by data- driven applications. That is, the lack of in-service data for exploration, prototyping and training of diagnostic and prognostic approaches during the concept and early development phases. Due to the upfront development of physical high-fidelity simulation models for the monitored components, a digital twin – of the portion of the twin that resembles the physical behavior – was used to generate data and facilitate preliminary exploration, prototyping and training. Digital twins were further employed to allow evaluation of what-if scenarios and identify the optimal future operation parameters of a drone.

Development of the RTAPHM digital twin involved a multi- disciplinary team of members distributed across different organizations and locations. Successful realization of the digital twin depended on early integration testing, performed in high frequencies, which generated continuous feedback regarding technical and conceptual issues. Within the research project we developed MOLE, an engineering tool for automating the integration of distinct simulation components, into a single system simulation driven by commercially available flight simulator software. Here, we showcase the internal mechanisms of the tool and demonstrate its abilities to generate a Docker-based executable for efficient data generation in the cloud. We also show our approach to online visualization, fault insertion, batch integration testing and debugging the digital twin executable. We also report on the utilization of MOLE in assembling the final RTAPHM demonstrator

Model-Based Loads Observer Approach for Landing Gear Remaining Useful Life Prediction

Jonathan Jobmann — 2024-06-27

Implementing health monitoring methods for aircraft landing gears holds the potential to prevent premature component replacements and optimize maintenance scheduling. Therefore, this paper introduces a fundamental framework for fatigue monitoring and subsequent steps for predicting the remaining useful life of landing gears. A key component of this framework is the model-based load observer, which lays the groundwork for subsequent remaining useful life prediction steps. This load observer will be analysed in detail in this paper. The model-based approach is specifically designed for observing the loads on civil aircraft landing gears during touchdown, utilizing signals from in-service sensors. To evaluate the load observation method, a flexible multibody simulation model is introduced to generate synthetic data sets of aircraft in-service data and the corresponding landing gear loads, given the unavailability of real in-service and recorded landing gear load data. The load observation method is applied to synthetic in-service data across various virtually performed landing scenarios, offering a proof of concept along with extensive analysis of parameter uncertainties and additional factors influencing observation quality. Through this analysis, certain challenges to the observation method are identified that require further investigation in subsequent research efforts.

Process Quality Monitoring Through a LSTM Network Derived from a Rule-Based Approach

Andreas Bernroitner — 2024-06-27

The railway infrastructure condition is a crucial factor for the
safe and efficient operation of trains. Regular maintenance
is inevitable as the track geometry degrades over time due
to traffic and environmental effects. To restore the ideal position
and provide sufficient durability of ballasted track so
called tamping machines are used. These machines lift the
track, correct the longitudinal level and the alignment of the
track panel and tamp the ballast. During the tamping process
the tamping tines penetrate the ballast bed, fill voids and compact
the ballast underneath the sleepers by a squeezing movement
with superimposed vibration. A detailed description of
the tamping cycle can be found on section 2. Monitoring and
evaluating this tamping process is essential for maintaining
process quality. This can be achieved through a variety of
sensors, such as incremental encoders, angle encoders, temperature,
pressure, and acceleration sensors, coupled with a
measurement unit (DAQ and edge device) to collect, locally
store and transmit the data to a cloud. This paper explores the
development of a rule-based algorithm for assessing the quality
of the tamping process execution in reference to its nominal
chronological sequence. The focus is on identifying tamping
occurrences and classifying them into acceptable (OK)
or non-acceptable (NOK) categories. This involves selecting
relevant measurement parameters and processing them,
considering the inherent imprecision in real-world processes.
Empirical thresholds are established to differentiate between
good and bad outcomes. The classification approach has to
be sufficiently generic in order to cover a high variety of customized
tamping machine types. As each machine is individually
designed, the process of generalization is challenging
and complex. The paper demonstrates the accuracy and universal
applicability of the developed rule set across different
tamping machines. The model’s effectiveness is validated using
the Hold-Out-Test-Set method. Furthermore, the rule-set-
achieved outcomes are compared with results gained from an
LSTM network. Both the rule-based approach and the neural
network demonstrate precision, but the latter requires significantly
more effort.

Threshold Selection for Classification Models in Prognostics

Rohit Deo — 2024-06-27

In this study, we evaluate the performance of a prognostic classification model for NOX sensors in diesel engines over one month by comparing its predictions against actual outcomes. We then construct a validation dataset to assess the model's performance. By analyzing instances where the model's predictions were incorrect, we determine new threshold values that could potentially reduce errors for each false positive (FP) and false negative (FN). Subsequently, we create a dataset where the threshold varies for each observation and train a regression model with the modified threshold as the target variable. Our findings indicate that incorporating this approach, where the model's performance is iteratively refined using the validation dataset, leads to a reduction in both false positives and false negatives.

PHME 2024 Management Team and Publisher Information

DO — 2024-06-27