Understand how CNN diagnoses faults with Grad-CAM

CNN-based models for fault diagnosis have achieved high prediction accuracy, but the lack of explainability makes them hardly be understood by humans. In this paper, a technique used to produce visual explanations for CNN has been introduced to a CNN-based fault diagnosis model, DCNN, to make it more transparent and understandable. Experiments on Tennessee Eastman process showed variables that DCNN pays more attention to when diagnosing faults, which makes the decision making process of DCNN more explainable and understandable.