Explainable convolutional neural network for gearbox fault diagnosis

The rapid advancement of data analytics has opened up new opportunities for improving the life cycle of engineered products and enhancing sustainability by intelligent monitoring and fault diagnosis of the related manufacturing processes and systems. Recently, Deep Neural Networks (DNNs) have demonstrated improved accuracy and robustness in classifying machine fault types and severities, when compared with conventional machine learning techniques. A major constraint of DNNs is that they operate as 'black boxes', which do not provide insight into how fault classification decisions are made. This not only raises questions on the trustworthiness of the decisions themselves, but also limits further improvement of DNNs for adaptation to a broader range of applications. This paper presents an explainable Deep Convolutional Neural Network (DCNN), which has been developed on the basis of Layer-wise Relevance Propagation (LRP), for fault diagnosis of gearboxes. Vibration signals as time series data are first converted to time-frequency spectra images through wavelet transform, which are then classified by a DCNN. To explain the rationale for classification decision, LRP decomposes contributions from local regions in the spectra images to the classification results, and determines which time-frequency points in the spectra image contribute the most to fault type and severity identification. Results of the analysis are then cross-checked with the time-frequency analysis. The effectiveness of the developed explainable DCNN is evaluated by experiments on a gearbox testbed, where gears with different types and degrees of faults are evaluated. LRP results have revealed that a trained DCNN is selective to different frequency bands in the time-frequency spectra for classification of gearbox fault type and severity.