Abstract
Industrial bearing health management (BHM) has traditionally been hindered by fragmented datasets and specialized models that lack cross-domain generalization. This paper introduces BearLLM, a novel multimodal framework that leverages the reasoning capabilities of Large Language Models (LLMs) for industrial diagnostics. By implementing a unified vibration signal representation and introducing the Multimodal Bearing Health Management (MBHM) dataset—the first large-scale dataset of its kind—BearLLM achieves state-of-the-art performance across 262 working conditions. The framework not only automates fault diagnosis but also provides human-interpretable maintenance recommendations, bridging the gap between raw sensor data and actionable industrial intelligence.
Introduction
Bearings are fundamental components in industrial machinery, yet their failure remains a primary cause of downtime. Current BHM strategies typically rely on "siloed" models trained on specific sensors, sampling rates, and operating conditions. This approach suffers from two critical limitations: a significant generalization gap when models encounter unseen mechanical environments, and task fragmentation, where separate models are required for detection, diagnosis, and maintenance planning.
To address these challenges, we propose BearLLM, a foundation model approach for bearing health. By translating diverse mechanical vibrations into a unified format compatible with LLMs, BearLLM serves as a single, versatile system capable of multi-task reasoning in natural language.
The BearLLM Framework
The core innovation of BearLLM lies in its ability to bridge the gap between high-frequency mechanical signals and textual reasoning. The architecture consists of three primary stages:
- Unified Vibration Signal Representation: To handle varying sampling rates, the framework employs adaptive sampling to extract fixed-duration segments. These are transformed into the frequency domain using Discrete Cosine Normalization (DCN), resulting in a standardized dimension ($n_f = 24,000$). A "fault-free reference signal" is utilized to calculate residuals, allowing the model to isolate fault-related anomalies from background mechanical noise.
- Feature Extraction and Alignment: A Fault Classification Network (FCN) serves as the feature encoder. An MLP-based alignment layer then maps these mechanical features into the embedding space of the LLM, effectively treating vibration data as a specialized token for the language model.
- Multimodal LLM Backbone: BearLLM utilizes a pre-trained Qwen2-1.5B model as its reasoning engine. By fusing signal embeddings with user instructions (e.g., "Provide maintenance recommendations based on the fault location"), the system generates comprehensive natural language responses.
The MBHM Dataset
A significant contribution of this work is the creation of the Multimodal Bearing Health Management (MBHM) dataset. Aggregating data from nine prominent public benchmarks (including CWRU, IMS, and XJTU), MBHM covers 262 distinct working conditions. It contains over 135,000 vibration signal pairs and 542,000 text cue/response pairs. Each entry includes four task types: anomaly detection, fault diagnosis, maintenance recommendations, and risk analysis, providing a rich training ground for multimodal industrial AI.
Key Results and Insights
BearLLM demonstrates exceptional performance across diverse benchmarks, achieving 100% accuracy on the CWRU dataset and 99.72% on DIRG. Its generalization capabilities are particularly noteworthy; in zero-shot evaluations on the unseen JUST dataset, BearLLM maintained 90.22% accuracy, significantly outperforming traditional models that lack unified signal representation.
The transition to the frequency domain via DCN acts as a "great equalizer," mitigating the issues caused by varying sampling rates. Furthermore, the inclusion of a "clean" reference signal allows the model to ignore machine-specific noise, focusing purely on fault signatures. Beyond technical metrics, a blind user study revealed that 58% of participants preferred BearLLM’s descriptive natural language outputs over traditional raw fault codes, highlighting its practical utility for non-expert operators.
Conclusion and Future Directions
BearLLM represents a paradigm shift in industrial health management, moving from narrow, task-specific classifiers to versatile, reasoning-capable foundation models. While the current 1.5B parameter model shows immense promise, future work will focus on deepening the model's industrial "intuition" and integrating more complex tasks like Remaining Useful Life (RUL) prediction. As industrial hardware evolves, optimizing these models for edge deployment will be crucial for widespread adoption in real-time monitoring systems.
Summary updated