BearLLM: A Prior Knowledge-Enhanced Bearing Health Management Framework with Unified Vibration Signal Representation

Haotian Peng, Jiawei Liu, Jinsong Du, Jie Gao, Wei Wang

Abstract

We propose a bearing health management framework leveraging large language models (BearLLM), a novel multimodal model that unifies multiple bearing-related tasks by processing user prompts and vibration signals. Specifically, we introduce a prior knowledge-enhanced unified vibration signal representation to handle various working conditions across multiple datasets. This involves adaptively sampling the vibration signals based on the sampling rate of the sensor, incorporating the frequency domain to unify input dimensions, and using a fault-free reference signal as an auxiliary input. To extract features from vibration signals, we first train a fault classification network, then convert and align the extracted features into word embedding, and finally concatenate these with text embedding as input to an LLM. To evaluate the performance of the proposed method, we constructed the first large-scale multimodal bearing health management (MBHM) dataset, including paired vibration signals and textual descriptions. With our unified vibration signal representation, BearLLM using one set of pre-trained weights achieves state-of-the-art performance on nine publicly available fault diagnosis benchmarks, outperforming specific methods designed for individual datasets.

Introduction

In industrial manufacturing, bearings are critical components whose failure can lead to significant downtime and safety risks. While AI has advanced predictive maintenance, current models suffer from fragmentation—they are often siloed for specific machines or tasks—and a lack of interpretability, typically outputting cryptic fault codes rather than actionable advice.

Comparison between existing fragmented bearing health management frameworks and the proposed unified BearLLM framework.

BearLLM addresses these challenges by creating a foundation model for bearing health that generalizes across diverse industrial environments and communicates through natural language, providing context and maintenance recommendations directly to technicians.

Methodology

The core of BearLLM is its ability to bridge the gap between raw sensor data and natural language processing through a unified architecture.

Unified Vibration Signal Representation

To handle varying sensor sampling rates and operating speeds, BearLLM utilizes Discrete Cosine Normalization (DCN). This method adaptively samples signals based on duration and converts them into the frequency domain, creating a "universal language" for vibration data. Furthermore, the model incorporates a fault-free reference signal from the same machine as prior knowledge, allowing it to focus on the residual signature of damage rather than normal mechanical noise.

Model Architecture

As shown in the architecture diagram, the system consists of three primary stages:

  1. Feature Encoder (FCN): A specialized convolutional network that extracts high-level features from the unified vibration representation.
  2. Alignment Layer: A multi-layer perceptron (MLP) that translates vibration features into word embeddings compatible with the LLM.
  3. LLM Backbone: A fine-tuned Qwen2-1.5B model that processes the combined vibration and text embeddings to generate multi-task responses.
The BearLLM architecture, detailing the flow from vibration signals and user instructions to multi-task natural language responses.

The MBHM Dataset

To train and evaluate BearLLM, we introduced the Multimodal Bearing Health Management (MBHM) dataset. It is the first large-scale dataset of its kind, aggregating data from nine major public benchmarks (including CWRU, IMS, and XJTU). It covers 262 different working conditions and contains over 135,000 vibration-text pairs, supporting tasks such as anomaly detection, fault diagnosis, and maintenance planning.

Key Results and Insights

BearLLM demonstrates significant improvements over traditional specialized models:

  • State-of-the-Art Accuracy: The model achieved 99.02% accuracy on the MBHM dataset, outperforming models like WDCNN.
  • Generalization: In zero-shot testing on unseen datasets, BearLLM maintained high performance (e.g., 90.26% on CWRU), whereas traditional models often fail when moved to new environments.
  • Interpretability: User studies show a strong preference (55%) for BearLLM’s detailed natural language responses over simple classification outputs, as the model can reason about fault severity and provide specific maintenance steps.

Conclusion

BearLLM represents a shift toward more capable, interpretable industrial multimodal models. By unifying vibration signals and leveraging the reasoning power of LLMs, it provides a robust framework for cross-dataset bearing health management. Future work will focus on removing the requirement for a fault-free reference signal and expanding the framework to handle time-varying operating conditions and additional sensor modalities like infrared imaging.

Summary updated

Abstract

We propose a bearing health management framework leveraging large language models (BearLLM), a novel multimodal model that unifies multiple bearing-related tasks by processing user prompts and vibration signals. Specifically, we introduce a prior knowledge-enhanced unified vibration signal representation to handle various working conditions across multiple datasets. This involves adaptively sampling the vibration signals based on the sampling rate of the sensor, incorporating the frequency domain to unify input dimensions, and using a fault-free reference signal as an auxiliary input. To extract features from vibration signals, we first train a fault classification network, then convert and align the extracted features into word embedding, and finally concatenate these with text embedding as input to an LLM. To evaluate the performance of the proposed method, we constructed the first large-scale multimodal bearing health management (MBHM) dataset, including paired vibration signals and textual descriptions. With our unified vibration signal representation, BearLLM using one set of pre-trained weights achieves state-of-the-art performance on nine publicly available fault diagnosis benchmarks, outperforming specific methods designed for individual datasets. We provide a dataset, our model, and code to inspire future research on building more capable industrial multimodal models https://github.com/SIA-IDE/BearLLM.
BibTeX
@inproceedings{peng2025bearllm,
  author = {Haotian Peng and Jiawei Liu and Jinsong Du and Jie Gao and Wei Wang},
  title = {{BearLLM: A Prior Knowledge-Enhanced Bearing Health Management Framework with Unified Vibration Signal Representation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i19.34188},
}
Figure