Failure and Remaining Useful Life Prediction of Wind Turbine Gearboxes

The purpose of this project is to predict wind turbine gearbox incipient faults using a combination of condition monitoring data. It is expected to contribute in developing a robust framework for wind turbine gearbox component incipient failure prediction and remaining useful life estimation. It further proposes a solution on how to overcome the challenges of expert knowledge based systems using AI techniques. Wind turbine operation and maintenance decision making confidence can be therefore increased.


INTRODUCTION
A large proportion of the total cost of energy from wind in large wind farms is composed of operation and maintenance (O&M) costs, which can be reduced if incipient machinery faults are successfully detected before they become catastrophic failures.For that reason, a combination of preventive and corrective maintenance strategies is implemented in wind industry (Sinha & Steel, 2015).Preventive maintenance can be performed through SCADA (Supervisory Control and Data Acquisition) systems and Condition Monitoring Systems (CMS).In both onshore and offshore wind, the wind turbine downtime is dominated by gearbox failures (Wilkinson et al., 2010), (Carroll, McDonald, & McMillan, 2016).It is thus vital to successfully monitor this component.Furthermore, as the installed wind capacity grows, the volume of condition monitoring data increases.The variable loading conditions of wind turbines and the nonlinear relationships between the parameters render rule-based monitoring impractical.For all aforementioned reasons, manual interpretation of wind turbine data becomes challenging.Therefore, Artificial Intelligence (AI) techniques can aid the decision making process of wind turbine gearbox maintenance.
Based on the problem statement, the objective of this thesis Sofia Koukoura. et al.This is an open-access article distributed under the terms of the Creative Commons Attribution 3.0 United States License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.is to answer the following research question and a high level methodology overview is given in Figure 1.
"How can gearbox faults be predicted based on a combination of condition monitoring data before catastrophic failure occurs and how can information from similar wind turbines be used?"

ANOMALY DETECTION USING SCADA DATA
There is a large amount of SCADA data available, which can be used to give an indication of the turbine condition.These data are the cheapest source for developing condition monitoring systems for wind turbines.The data can be operational, environmental or status codes.Anomaly detection and health assessment of a wind turbine through SCADA can be performed in various ways (Tautz-Weinert & Watson, 2016).One of them is through normal behaviour modeling (Zaher, McArthur, Infield, & Patel, 2009).a measured parameter is modelled empirically based on a training phase and the residual of measured minus modelled signal acts as a clear indicator for a possible fault.An example is shown in Figure 2. SCADA data is available at different time periods progressively before failure from 199 operating wind turbines.All turbines experienced a planet bearing inner race fault.The temperature in the hollow shaft is predicted using neural networks, based on other temperatures, the power and the speed of the turbine.The mean absolute error of predicted temperature increases towards failure.System level health assessment can be easily achieved but gearbox specific component fault detection could be more challenging.Load and temperature play an important role in the the vibration signature of the gearbox components.Therefore, SCADA data information of power level, speed and temperature is incorporated in the vibration classification models.Depending on the sampling regime of vibrations and SCADA, load and environmental parameters can be used directly as input features to the pattern recognition model or some type of correlation technique can be applied.
The example presented focuses on fault detection.Since the features extracted are around specific fault frequencies and sensors, fault diagnosis can be achieved.The aim of fault diagnosis is to identify specific faults from the rest of the data (Leahy et al., 2018).Therefore, classification of different faulty components (i.e.gears or bearings) around different stages of the gearbox is to be explored in the thesis through different failure mode case studies.

REMAINING USEFUL LIFE ESTIMATION
Once an incipient fault has been detected, fault prognosis is performed.Features extracted from vibrations can be ranked and fused to create a health index (Coble, 2010).Remaining useful life is estimated using degradation models (Saidi, Ali, Bechhoefer, & Benbouzid, 2017).Either measurements or historical run-to-failure data can be utilised for that reason.Parameters are yield to prior Gaussian distributions and are updated when new observations are available.Some preliminary results from a planet bearing remaining useful life

FLEET BASED FAULT DETECTION
If enough historical data is available, turbine unit specific diagnostics and prognostics can be performed.On the other hand, fleet-based diagnostics and prognostics can provide a robust fault detection framework based on similar wind turbines.Assuming that the majority of wind turbines in the fleet are in normal operating condition, a clustering approach can identify baseline data and detect abnormalities (Lapira, 2012).The normal operation identified from the fleet data is taken as baseline, and is used to train a global fault detection model using data-driven approaches.

CONCLUSION
The thesis provides a framework for fault prediction of wind turbine gearbox components, using inputs from various sensors on the wind turbine drivetrain.The features extracted from those sensors are used to train models that determine the health state and the remaining useful life of the components.Fleet based condition monitoring is also explored, since it can significantly improve the decision making process.

Figure 2 .
Figure 2. Normal behaviour model mean absolute error when predicting a gearbox hollow shaft temperature.

Sofia
Koukoura received her degree in Mechanical Engineering at the National Technical University of Athens in 2015.She then joined the University of Strathclyde Wind and Marine Energy Systems Centre for Doctoral Training.Her PhD focuses on diagnostics and prognostics of wind turbine gearboxes using signal processing and machine learning techniques.

Table 1 .
Table 1, months before before the alarm was activated .Approximately 700 samples are used for training this specific model.Binary classification results