Interpretable Convolutional Neural Network Through Layer-wise Relevance Propagation for Machine Fault Diagnosis

Abstract

As a state-of-the-art pattern recognition technique, convolutional neural networks (CNNs) have been increasingly investigated for machine fault diagnosis, due to their ability in analyzing nonlinear and nonstationary high-dimensional data that are typically associated with the performance degradation process of machines. A key issue of interest is how the inputs to CNNs that contain fault-related patterns are learned by CNNs to recognize discriminatory information for fault diagnosis. Understanding this link will help establish connection to the physical meaning of the diagnosis, contributing to the broad acceptance of CNNs as a trustworthy complement to physics-based reasoning by human experts. Using Layer-wise Relevance Propagation (LRP) as an indicator, this paper investigates the performance of a CNN trained by time-frequency spectra images of vibration signals measured on an induction motor. The LRP provides pixel-level representation of which values in the input signal contribute the most to the diagnosis results, thereby providing an improved understanding of how the CNN learns to distinguish between fault types from these inputs. Results have shown that the patterns learned by CNNs in the time-frequency spectra images are intuitive and consistent with respect to network re-training. Comparison with using raw time series and discrete Fourier transform coefficients as inputs reveals that time-frequency images allow for more consistent pattern recognition by CNNs.