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 frequency features and time-frequency features for machinery fault diagnosis has been prevalent in the last decade. However, more attentions have been drawn to machine learning based features recently. While frequency domain features can be directly correlated to fault type and fault level, statistical features are typically abstract representations. Most existing machine learning methods are based on supervised learning scheme to classify these abstract features for diagnostic purpose. Since labeled training data can be unreliable and hard to obtain, unsupervised fault diagnosis methods are often more desirable. Traditionally, only physics-based condition indicators can be used directly to indicate fault level and fault type. This paper aims to evaluate some of the unsupervised feature extraction methods by deep learning for ‘meaningful’ feature mining, i.e., disentangle and extract features that are related to the faults. We name this type of feature extraction method as disentangled tone mining (DTM), where each tone in the present work refers to certain latent structure in the spectrum. This paper has shown that fully unsupervised methods can potentially extract the ‘trend’ for machine health state, which can be used directly in on-line anomaly detection or prediction. Although not fully understood yet, the experimental results indicate that unsupervised deep sparse autoencoder can disentangle the fault related features from uncorrelated noise and self-track the potential fault evolution process.