Reliability-aware dynamic graph fusion and LLM-based diagnostic assistant for bearing faults

Fusing and interpreting heterogeneous evidence remains a major challenge in multi-source bearing fault diagnosis, especially under noise, sensor outages, and operating-condition shifts. An interpretable diagnostic framework is presented that integrates graph-based agent collaboration, hierarchical evidence fusion, and a large language model (LLM)-based diagnostic assistant. First, a sample-adaptive dynamic diagnostic graph is constructed, where sensor-specific base models are formulated as interactive answerer-reviewer agents. A variational autoencoder (VAE) extracts a global operating-state representation to condition graph attention, enabling sample-wise agent weighting and topology adaptation. Second, a two-layer Dempster-Shafer (DS) fusion scheme with offline reliability calibration is introduced: inner-layer fusion consolidates reviewer evidence to suppress highly conflicting opinions, followed by an outer-layer fusion that aggregates committee-level decisions to progressively absorb uncertainty. Finally, an LLM-based assistant converts structured diagnostic trajectories into maintenance-oriented, verifiable explanations, supported by retrieval from a domain knowledge base. Experiments indicate that the framework consistently improves diagnostic accuracy and reduces output entropy, while enhancing process traceability and supporting human-AI collaborative decision-making.