Abstract:Accurately diagnosing bearing faults is crucial for maintaining the efficient operation of rotating machinery. However, traditional diagnosis methods face challenges due to the diversification of application environments, including cross-condition adaptability, small-sample learning difficulties, and cross-dataset generalization. These challenges have hindered the effectiveness and limited the application of existing approaches. Large language models (LLMs) offer new possibilities for improving the generalization of diagnosis models. However, the integration of LLMs with traditional diagnosis techniques for optimal generalization remains underexplored. This paper proposed an LLM-based bearing fault diagnosis framework to tackle these challenges. First, a signal feature quantification method was put forward to address the issue of extracting semantic information from vibration data, which integrated time and frequency domain feature extraction based on a statistical analysis framework. This method textualized time-series data, aiming to efficiently learn cross-condition and small-sample common features through concise feature selection. Fine-tuning methods based on LoRA and QLoRA were employed to enhance the generalization capability of LLMs in analyzing vibration data features. In addition, the two innovations (textualizing vibration features and fine-tuning pre-trained models) were validated by single-dataset cross-condition and cross-dataset transfer experiment with complete and limited data. The results demonstrated the ability of the proposed framework to perform three types of generalization tasks simultaneously. Trained cross-dataset models got approximately a 10% improvement in accuracy, proving the adaptability of LLMs to input patterns. Ultimately, the results effectively enhance the generalization capability and fill the research gap in using LLMs for bearing fault diagnosis.
What problem does this paper attempt to address?
### Problems the Paper Attempts to Solve
This paper aims to address several key challenges in rolling bearing fault diagnosis, specifically including:
1. **Cross-Condition Adaptability**:
- **Problem Background**: In actual operating environments, the operating conditions of rotating machinery are complex and variable, leading to significant differences in data distribution under different conditions. This makes it difficult for fault diagnosis models trained under specific conditions to maintain high accuracy under other conditions.
- **Specific Manifestations**: For example, in wind turbines, different wind speeds, temperatures, loads, and operating times can affect bearing performance, making it difficult to uniformly analyze fault data under different conditions.
2. **Few-Shot Learning**:
- **Problem Background**: Industrial systems have high reliability requirements, and faults occur infrequently, resulting in a much smaller number of fault samples compared to normal samples.
- **Specific Manifestations**: For example, commercial aircraft engines may record only a few faults over thousands of flight hours, leading to insufficient fault samples and making the model prone to misclassifying fault samples as normal samples.
3. **Cross-Dataset Generalization**:
- **Problem Background**: In practical applications, the types of bearings used in different equipment are diverse. Even for the same type of bearing, models produced by different manufacturers may vary due to different installation positions and uses.
- **Specific Manifestations**: For example, fault diagnosis methods suitable for submarine rotating machinery may not be applicable to aviation machinery, requiring re-adaptation of feature extraction and classification methods for different datasets, leading to operational complexity and weak generalization ability.
### Solutions
To address the above challenges, the paper proposes a bearing fault diagnosis framework based on large language models (LLM), specifically including the following aspects:
1. **Signal Feature Quantification Method**:
- A time-domain and frequency-domain feature extraction method is proposed, integrating features through a statistical analysis framework to textualize time-series data for efficient learning of common features across conditions and few-shot samples.
2. **Model Fine-Tuning Method**:
- Using LoRA and QLoRA fine-tuning methods to enhance the generalization ability of LLM in analyzing vibration data features.
3. **Experimental Validation**:
- Through single dataset cross-condition experiments and cross-dataset transfer experiments (complete data and limited data), the framework's capability in three generalization tasks is validated. Experimental results show that the accuracy of cross-dataset models improved by about 10%, demonstrating the adaptability of LLM to input patterns.
### Main Contributions
1. **Proposed a bearing fault diagnosis feature system that vectorizes vibration data into text**, extending traditional fault diagnosis methods and enabling integration with LLM.
2. **Introduced an efficient vibration data model fine-tuning framework**, validated through case studies of cross-condition, few-shot, and cross-dataset scenarios, showcasing the generalization diagnostic capability of the study under various bearing fault modes.
Through these improvements, the study significantly enhances the efficiency and accuracy of bearing fault diagnosis systems, stabilizing the overall operation of rotating machinery.