Gear pitting fault diagnosis using disentangled features from unsupervised deep learning

Effective feature extraction is critical for machinery fault diagnosis and prognosis. The use of time–frequency features for machinery fault diagnosis has prevailed in the last decade. However, more attentions have been drawn to machine learning–based features. While time–frequency domain features can be directly correlated to fault types and fault levels, data-driven features are typically abstract representations. Therefore, classical machine learning approaches require large amount of training data to classify these abstract features for fault diagnosis. This article proposed a fully unsupervised feature extraction method for “meaningful” feature mining, named disentangled tone mining. It is shown that disentangled tone mining can effectively extract the hidden “trend” associated with machinery health state, which can be used directly for online anomaly detection and prediction. Compared with wavelet transform and time domain statistics, disentangled tone mining can better extract fault-related features and reflect the fault degradation process. Shallow methods, such as principal component analysis, multidimensional scaling and single-layer sparse autoencoder, are shown to be inferior in terms of disentangled feature learning for machinery signals. Simulation analysis is also provided to demonstrate and explain the potential mechanism underlying the proposed method.