Explainable Fault Diagnosis Model using Stacked Autoencoder and Kernel SHAP

Process monitoring plays an important role in economic and safe process operations. Various fault diagnosis models based on data with high classification performance have been presented, but they cannot provide an explanation of the prediction results. In this study, a framework for the explainable fault diagnosis model is proposed. High classification performance is achieved by using stacked autoencoder (SAE) and an explanation of the model predictions is presented by Kernel Shapley additive explanations (Kernel SHAP). The resulting explanation which includes the feature importance and the model behavior reinforces reliability of the model and enables to take manipulating actions when a fault occurs. Dataset from the Tennessee Eastman process is used to illustrate the effectiveness of the method.