Diagnostic Evidence Network for Verifiable Bearing Fault Diagnosis

A new framework called Diagnostic Evidence Network (DENet) addresses the lack of physically verifiable evidence in AI-based fault diagnosis for mechanical systems. Current classifiers output only a class label with a softmax confidence score, which cannot be checked against independent physical knowledge. Additionally, generative language models risk introducing hallucinated content into maintenance reports. DENet is an encoder-agnostic multi-task framework that extends outputs to a structured evidence record, including classification and a predicted characteristic frequency comparable to theoretical values. The work uses bearing fault diagnosis as a testbed, aiming to enable trustworthy deployment in safety-critical systems.