Domain Knowledge Regularised Fault Detection

Unsupervised data-driven methods are attractive options for fault detection in rotating machinery since they do not require any failure data during training. However, in these data-driven approaches, engineering domain knowledge remains unexploited. Although engineering features are often used as inputs to machine learning models, thereby including domain knowledge, few methods exist for directly integrating domain knowledge about the expected machine fault behaviour into unsupervised fault detection methods. This paper presents a generic method for including domain knowledge into unsupervised, auto-encoder-based fault detection methods by regularising the Jacobian of the latent feature representation. This regularisation results in informative latent features that are sensitive to changes that are expected from a machine in a faulty condition. The proposed method is evaluated on a bearing fault detection task, both when using a low dimensional vector of engineering features and when using high dimensional frequency domain data. The analysis is conducted on two bearing fault data sets with different operating conditions, fault modes and signal-to-noise ratios. The proposed regularised auto-encoder yields improved ROC-AUC performance as compared to the unregularized baseline when evaluated on a latent-feature based fault indicator. The proposed method shows potential as a generic method for integrating engineering domain knowledge into fault detection problems.