A Novel Transfer Learning Method Utilizing Acoustic and Vibration Signals for Rotating Machinery Fault Diagnosis

Zhongliang Chen, Zhuofei Huang, Wenxiong KangarXiv 2023 · DOI · arXiv

Abstract

Identifying faults in rotating machinery is critical for industrial safety and operational efficiency. However, existing diagnostic models often struggle with domain shift between laboratory and real-world environments and the limitations of single-modal data. This paper introduces a multi-modal transfer learning approach that fuses acoustic and vibration signals into a novel feature representation called MAVgram. By integrating spectral and temporal information and utilizing a pre-trained MobileFaceNet backbone with ArcFace loss, the proposed method achieves high diagnostic accuracy even with minimal labeled data in target environments.

Introduction

In modern industrial settings, the health monitoring of rotating components like bearings and gears is vital. Two primary challenges hinder the deployment of deep learning models in this field: distribution discrepancy and single-modal limitations. Models trained in controlled settings often fail when faced with different operating conditions (domain shift). Furthermore, while most methods rely solely on vibration signals, these can be susceptible to external interference. Acoustic signals, though underutilized, offer complementary high-frequency information. This research addresses these gaps by leveraging multi-modal data and transfer learning.

Methodology

MAVgram Feature Fusion

The core of the proposed framework is the MAVgram, a fused feature representation that concatenates three distinct components:

  • Mgram: A Log-Mel spectrogram of the acoustic signal, capturing frequency-domain characteristics.
  • Agram: Temporal features of the acoustic signal extracted via TgramNet, a specialized CNN block designed to capture high-frequency information often lost in standard spectrograms.
  • Vgram: Temporal features of the vibration signal, also extracted using a TgramNet block.
The proposed MAVgram feature fusion and classification framework.

Model Architecture and Training

The system utilizes MobileFaceNet (MFN) as the primary classifier due to its efficiency. To enhance class discriminability, the researchers employed ArcFace loss, which maximizes the angular distance between different fault categories in the feature space. The transfer learning strategy involves pre-training the backbone on a large source dataset (iFlytek) and then fine-tuning the TgramNet and final classification layers on a smaller target dataset (University of Ottawa), while keeping other layers frozen.

Experiments and Results

Datasets

The model was evaluated using two synchronized acoustic-vibration datasets:

  1. iFlytek (Source): Contains five categories across 16 operating conditions, used for initial training.
  2. University of Ottawa (UO) (Target): Features different sampling rates and load conditions, representing the real-world deployment task.

Performance and Ablation Studies

The MAVgram-MFN method achieved a peak accuracy of 94.98% on the target dataset. A key finding is the model's efficiency; with only 15% of the target data available for fine-tuning, it reached 93.47% accuracy.

Ablation studies highlight the necessity of the multi-modal approach. As shown in the table below, removing specific components of the MAVgram (such as spectral information in AVgram-MFN) leads to a significant drop in performance, particularly when training data is scarce.

Ablation study comparing MAVgram (baseline) against single-modal and partial-feature variants across different data percentages.

Conclusion

The MAVgram-MFN framework demonstrates that fusing acoustic and vibration signals significantly improves fault diagnosis robustness. The use of TgramNet to compensate for temporal information and ArcFace for feature separation allows the model to perform exceptionally well even in data-scarce scenarios. Future work will focus on optimizing computational trade-offs for real-time applications and exploring more advanced fusion mechanisms, such as attention-based modules, to further improve performance in high-noise environments.

Summary updated

Abstract

Fault diagnosis of rotating machinery plays a important role for the safety and stability of modern industrial systems. However, there is a distribution discrepancy between training data and data of real-world operation scenarios, which causing the decrease of performance of existing systems. This paper proposed a transfer learning based method utilizing acoustic and vibration signal to address this distribution discrepancy. We designed the acoustic and vibration feature fusion MAVgram to offer richer and more reliable information of faults, coordinating with a DNN-based classifier to obtain more effective diagnosis representation. The backbone was pre-trained and then fine-tuned to obtained excellent performance of the target task. Experimental results demonstrate the effectiveness of the proposed method, and achieved improved performance compared to STgram-MFN.
BibTeX
@misc{chen2023novel,
  author = {Zhongliang Chen and Zhuofei Huang and Wenxiong Kang},
  title = {{A Novel Transfer Learning Method Utilizing Acoustic and Vibration Signals for Rotating Machinery Fault Diagnosis}},
  year = {2023},
  eprint = {2310.14796},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2310.14796},
}
Figure