A Gear Health Indicator Based on f-AnoGAN

Abstract

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.