FAE-LLM: Active physical-evidence reasoning and record-aware trustworthy fusion for bearing fault diagnosis

Accurate bearing diagnosis requires informative vibration representations and reliable confidence estimates. Fixed evidence sets and independent fragment processing can overlook signal-dependent features and record-level relationships. This paper proposes a fragment-aware evidence large language model (FAE-LLM) that combines adaptive physical-evidence acquisition with trustworthy record-level fusion. A learned policy uses time-domain descriptors and waveform features to select spectral, envelope-spectrum, or autocorrelation evidence. Continuous projections map the selected evidence into a partially fine-tuned GPT-2 backbone. An order-aware relational transformer captures dependencies among adjacent fragments, and Dirichlet fusion integrates their evidence according to uncertainty and conflict. A validation-based rejection rule defers uncertain predictions. Across five runs, FAE-LLM achieves classification accuracies of 100 . 00 % on XJTU-SY and 96 . 35 % on GDUPT, with all GDUPT comparison methods receiving 4096-point records. Ablation studies show the benefit of longer record context, while fusion controls demonstrate improved probability calibration. The framework connects physical signal analysis with record-aware classification and selective decision support for bearing fault diagnosis.