Open-Set Fault Diagnosis for Industrial Rotating Machines Based on Trustworthy Deep Learning

Detecting and diagnosing faults in rotating machines is crucial for ensuring the safety and reliability of modern industrial cyber-physical systems. Traditional data-driven fault diagnosis methods have achieved significant success when dealing with a set list of known faults and working conditions. However, they become inaccurate and overconfident when faced with new fault classes outside the training set. This paper introduces a novel Evidential Abstention Classifier based on trustworthy deep learning. It can quantify prediction uncertainty and recognize new fault classes without the need for their training data. Experiment results validated the efficacy of the proposed L1 regularization in improving uncertainty quantification. They also highlighted the proficiency of the designed auxiliary training method in generating fault-discriminative features and establishing effective decision boundaries for new fault types. EAC enables accurate open-set fault diagnosis with reduced reliance on historical data, offering improved transparency in the diagnostic process.