Interpretable AI Enabled Intelligent Fault Diagnosis and Prognosis

Artificial intelligence (AI) is an emerging technology in mechanical fault diagnosis and prognosis. However, their limited interpretability, generalization, and reliance on extensive datasets often reduce trust and hinder broader engineering applications. Interpretable AI models have recently become a popular research topic in this field, providing a promising solution to these challenges. This chapter first reviews the main approaches for developing interpretable AI models in mechanical fault diagnosis and prognosis. Subsequently, two interpretable AI models based on a convex optimization framework and a fully interpretable neural network framework are proposed. These two models demonstrate the strategies of integrating signal processing methods and domain knowledge to design transparent network layers and loss functions for constructing interpretable AI methods. To validate their effectiveness, two experimental case studies using bearing run-to-failure datasets are presented. The results demonstrate their capability in HI construction and interpretable fault feature extraction based on the learned weights, facilitating early fault detection, diagnosis, and monotonic degradation assessment. The proposed interpretable AI models exhibit flexibility in generating HIs with diverse characteristics, outperforming existing methods in monotonicity, robustness, and trendability.