An Order-Invariant and Interpretable Dilated Convolution Neural Network for Chemical Process Fault Detection and Diagnosis

Although convolution neural network (CNN) has achieved certain success in fault detection and diagnosis (FDD) tasks the chemical engineering industry, performance credibility of CNN-based FDD methods are greatly limited by two factors. First, CNN relies upon strong temporal/spatial correlation data, which is very difficult to obtain generic tabular data. Second, most have poor interpretability due encapsulation mechanism feature extraction, thus there great difficulty identifying root-cause features from results obtained using these methods. To address difficulties, we propose an order-invariant interpretable dilated (OIDLCNN), composed clustering, convolution, a deep Shapley additive explanations (SHAP) explainer. Specifically, clustering technique adopted reorder those with correlations placed be adjacent each other. The large receptive field can capture long-range so then it further recover improve classification performance. Last but not least, proposed method provides including SHAP values quantify contribution identify as one highest contribution. Computational experiments conducted on Tennessee Eastman process benchmark dataset. Compared other state-of-the-art methods, OIDLCNN-SHAP achieves better capturing correlations, detecting faults, features. Note Practitioners —Fault significant for reducing maintenance costs improving safety processes. In this paper, investigate faults multivariate This work was motivated fact that existing algorithms designed data fail vital information about correlations. addition, commonly used bayesian network-based root cause analysis expensive since they require much prior knowledge expert rules. We present novel framework unique without any knowledge. automatic low-cost way tasks. It integrated into condition monitoring system real processes analyze potential corresponding causes time.