SCF-Net: A sparse counterfactual generation network for interpretable fault diagnosis

Interpretability of deep learning models is essential for their massification in the context of prognostics and health management (PHM), as it is useful for transparency, bias detection, and accountability. These properties help to build trust, which is necessary for deployment in industrial environments. Among different approaches, counterfactuals are minimally altered versions of the original inputs that generate a change in the outputs’ class. As such, counterfactual-based interpretations give insights on how the model calculates an output given a set of input feature values. In this paper, we present a multi-task network for fault classification and counterfactual generation as a means to increase interpretability of deep learning-based fault diagnosis. By using the proposed network, referred to as Sparse Counterfactual Generation Network (SCF-Net), the model is able to classify input values coming from different sensors into its corresponding health state and simultaneously calculate counterfactual values for all of the other possible classes, even if there are more than two. The network is tested in two case studies using real data from the Oil and Gas (O&G) industry. Results are evaluated using different performance metrics available in the literature, and compared to two other counterfactual generation frameworks.