Research map / Vision-language and multimodal / Human and embodied modeling
BearLLM: A Prior Knowledge-Enhanced Bearing Health Management Framework with Unified Vibration Signal Representation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes BearLLM, a multimodal LLM for bearing health management that unifies vibration signals (adaptive sampling, DCT-based frequency alignment, fault-free reference residuals) with text prompts, and releases the MBHM dataset of 135k vibration segments with 542k text pairs.
Problem
Bearing health management methods are specialized per working condition and task, and domain adaptation or generalization struggles with category-inconsistent or many-condition settings; a unified model across heterogeneous vibration datasets and multiple tasks is needed.
Approach
Adaptively sample fixed-duration vibration segments, convert them to a unified DCT frequency representation, and use a fault-free reference signal from the same working condition to form residual channels; a pre-trained fault classification network's features are aligned into LLM word embeddings and fed with text to Qwen2-1.5B fine-tuned with LoRA.
Key results
- BearLLM reaches 99.02% accuracy on the MBHM test set, versus 90.07% for the best baseline (BearingFM+FCN).
- BearLLM reports 100.00% on CWRU, 99.72% on DIRG, and 99.90% on HIT, exceeding individually trained specialized methods (e.g., QCNN 80.55% on CWRU).
- Zero-shot trained without JUST and IMS, BearLLM reaches 90.22% on JUST and 98.52% on IMS; zero-shot on CWRU reaches 90.26%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use BearLLM instead of the strongest listed baseline | fault diagnosis · CWRU | accuracy: 95.55 -> 100.00 | Helped |
| use BearLLM instead of the strongest listed baseline | fault diagnosis · DIRG | accuracy: 91.32 -> 99.72 | Helped |
| use BearLLM instead of the strongest listed baseline | fault diagnosis · HIT | accuracy: 99.41 -> 99.90 | Helped |
| use BearLLM instead of the strongest listed baseline | fault diagnosis · IMS | accuracy: 97.74 -> 99.39 | Helped |
| use BearLLM instead of the strongest listed baseline | fault diagnosis · JNU | accuracy: 98.96 -> 99.44 | Helped |
| use BearLLM instead of the strongest listed baseline | fault diagnosis · JUST | accuracy: 85.91 -> 98.16 | Helped |
| use BearLLM instead of the strongest listed baseline | fault diagnosis · PU | accuracy: 95.31 -> 99.41 | Helped |
| use BearLLM instead of the strongest listed baseline | fault diagnosis · XJTU | accuracy: 98.27 -> 98.79 | Helped |
| remove DCN and use fixed-length time-domain signals (ablation) | BearLLM · MBHM | accuracy: 99.02% -> 72.15% | Hurt |
| remove the fault-free reference and residual channels (ablation) | BearLLM · MBHM | accuracy: 99.02% -> 98.35% | Hurt |
| remove the residual channel (ablation) | BearLLM · JUST | accuracy: 90.22% -> 87.83% | Hurt |
| remove the fault-free reference channel (ablation) | BearLLM · JUST | accuracy: 90.22% -> 90.34% | No clear effect |
Surprising findings
- Removing the DCN alignment drops MBHM accuracy from 99.02% to 72.15%, showing time-domain inputs fail under heterogeneous conditions.
- Some specialized methods trained on individual datasets (e.g., WDCNN on CWRU at 71.60%) are outperformed by a unified model trained on the pooled MBHM data.
Limitations
- Requires a fault-free reference signal under the same working condition, so it cannot be used when no fault-free history exists or conditions vary over time.
- Relies on ChatGPT-generated text responses for the MBHM dataset rather than expert-authored labels.
- User study is small-scale with blind voting; no automatic LLM response-quality metric is reported.
Experimental setup
Datasets: CWRU DIRG HIT IMS JNU JUST MFPT PU XJTU MBHM
Benchmarks: CWRU DIRG HIT IMS JNU JUST MFPT PU XJTU MBHM
Metrics: accuracy false alarm rate missed alarm rate
Baselines: WDCNN TCNN QCNN MagNet BearingFM WDCNN+DCN QCNN+DCN MagNet+DCN BearingFM+FCN
Builds on: LoRA WDCNN MagNet BearingFM QCNN
Closest papers by meaning
- MDFG: Multi-Dimensional Fine-Grained Modeling for Fatigue Detection (AAAI 2025 · similarity 0.83)
- Towards Open-Vocabulary Industrial Defect Understanding with a Large-Scale Multimodal Dataset (CVPR 2026 · similarity 0.81)
- PICID: A Modular Evaluation Infrastructure for Reproducible PHM Across Tasks and Domains (NeurIPS 2026 · similarity 0.81)
- SensorLM: Learning the Language of Wearable Sensors (NeurIPS 2025 · similarity 0.80)
- Multi-modal Medical Diagnosis via Large-small Model Collaboration (CVPR 2025 · similarity 0.79)
- Hierarchical State Space Models for Continuous Sequence-to-Sequence Modeling (ICML 2024 · similarity 0.79)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.