AI Diagnosis for Bearings Offers Verifiable Evidence and Reduces LLM Hallucinations.

AI-driven fault diagnosis in critical mechanical systems, such as bearing fault detection, often struggles with trustworthiness. Current methods provide only internal confidence scores, which cannot be independently verified against physical reality. Furthermore, the increasing use of large language models (LLMs) for maintenance reporting introduces the risk of generating hallucinated content, leading to unreliable decisions. To address these issues, a new framework called the Diagnostic Evidence Network (DENet) has been developed. This system extends the standard AI output to include a structured evidence record. This record comprises the fault classification, a predicted characteristic frequency that can be compared to theoretical values based on bearing geometry, and a temporal localization of transient impulses visible in raw data. The DENet framework, tested across various encoders and datasets, maintains diagnostic accuracy while significantly improving verifiability. It also incorporates a QLoRA-adapted language model specifically designed to translate diagnostic content without generating new, potentially false information, thereby drastically reducing unsupported claims in reports. This approach provides a crucial, label-free validation signal that effectively detects misclassifications, even in scenarios where traditional confidence scores fail.