Explainable AI methodology for understanding fault detection results during Multi-Mode operations

Multi-mode operations are prevalent in the chemical industry. Various methods have been proposed for monitoring multi-mode operations. Of these, AI-based approaches such as Deep Neural Networks (DNN) are becoming popular due to their higher accuracy. However, the lack of transparency of DNNs hinders their widespread adoption in safety-critical applications such as process monitoring. This work addresses this limitation by proposing an Explainable AI (XAI) methodology for multi-mode operations. The proposed methodology encompasses a supervisory system that identifies the current operational mode. This information is used by an Integrated Gradient (IG) based XAI method to configure mode-specific baselines and thus generate DNN explanations corresponding to each operational mode. The ability of this methodology to generate reliable explanations to aid plant operators is illustrated through a simulated CSTR process, the Tennessee-Eastman process, and the pilot-scale Multiphase Flow Facility case study.