Explainability: Relevance based dynamic deep learning algorithm for fault detection and diagnosis in chemical processes

The focus of this work is on Statistical Process Control (SPC) a manufacturing process based available measurements. Two important applications SPC in industrial settings are fault detection and diagnosis (FDD). In work, deep learning (DL) methodology proposed for FDD. We investigate the application an explainability concept (explainable artificial intelligence (XAI)) to enhance FDD accuracy neural network model trained with dataset relatively small number samples. quantified by novel relevance measure input variables that calculated from Layerwise Relevance Propagation (LRP) algorithm. It shown relevances can be used discard redundant feature vectors/ iteratively thus resulting reduced over-fitting noisy data, increasing distinguishability between output classes superior test accuracy. efficacy method demonstrated benchmark Tennessee Eastman Process.