Interpret what a Convolutional Neural Network learns for fault detection and diagnosis in process systems

The focus of this work is on an interpretation strategy on what a Convolutional Neural Network (CNN) has learned for fault detection and diagnosis (FDD) in process systems. Frequency spectra of process variables obtained by Continuous Wavelet Transform (CWT) are adopted as input features. Then, a CNN structure is designed to represent the mappings from input frequency features to different operation conditions. The Layer-wise Relevance Propagation (LRP) strategy is utilized to gain the relevance of each frequency feature to the classification performance. The formulations of relevance propagation for 4 types of CNN layers are presented in detail. The relevance scores are then depicted in heatmaps, where the pixels' colors denote the contribution degrees and the most significant frequency features are considered as the major bases that the CNN discriminates different operation situations. The proposed interpretation strategy is experimented on the Tennessee Eastman process benchmark. The testing results demonstrate the efficiency of the strategy in interpreting what the CNN has learned to distinguish normal or faulty conditions in the FDD task.