Trustworthy Artificial Intelligence Based on an Explicable Temporal Feature Network for Industrial Fault Diagnosis

Artificial intelligence is extensively utilized across various high-risk domains, and ensuring the safety, reliability, and trustworthiness of these systems is of paramount importance. This necessitates adherence to several imperative requirements such as fairness, explainability, accountability, reliability, and acceptability in order to establish the trustworthiness of these systems. The decision-making process of the systems relies heavily on data quality. However, existing studies in the field of industrial fault diagnosis have not fully considered the influence of noise interference on the system accuracy and the interpretability of the algorithm. Therefore, this study aims to investigate reliable and robust diagnostic techniques along with black-box model interpretation when confronted with noise interference in practical applications for industrial fault diagnosis. To solve the above problems, an explicable temporal feature network (ETFN) based on deep Shapley additive explanation (Deep SHAP) values is proposed, which increases the robustness and interpretability of the model. First, adaptive features extracted from the improved deep residual shrinkage network are combined with empirical features to increase the robustness of the model. Then, the combined features are used as input to the ETFN model for rotating device diagnosis. Deep SHAP is used to rank all the combined feature contributions and further interpret the model diagnosis by adjusting the number of features. The proposed ETFN achieves a good balance between stability, accuracy, and interpretation on three real-world datasets. Leading accuracy is achieved on all three datasets. In particular, a diagnostic accuracy of more than 97% can still be maintained when perturbed by noise, which is not achieved by alternative methods. The interoperability of the proposed method in industrial diagnostic applications is also enhanced by Deep SHAP. We implemented ETFN for extremely robust diagnosis and human-computer interaction in real noise for industrial data.