A rolling bearing fault diagnosis method under insufficient samples condition based on MSLSTM transfer learning

Ping Zhang, Debo LiuJournal of Vibroengineering 2025 · DOI

TL;DR

This paper introduces a Multi-Scale Long Short-Term Memory (MSLSTM) network paired with transfer learning to solve the "small data" problem in industrial fault diagnosis. By replacing traditional pooling with multi-scale convolutions, the model captures richer vibration features and achieves over 93% accuracy even when only 5% of the target data is available for training.

What Problem Does This Paper Address?

Deep learning has revolutionized the way we monitor rotating machinery, but it has a "data hunger" problem. In real-world industrial settings, it is often difficult, expensive, or dangerous to collect large, labeled datasets for every possible bearing fault.

Existing solutions like Generative Adversarial Networks (GANs) or manual data augmentation are often time-consuming and require significant expert tuning. Furthermore, standard architectures like the Long Short-Term Memory (LSTM) network—while great for time-series data—often struggle with noisy environments because they extract only a single type of feature and lose critical information during traditional "pooling" operations. The authors aim to create a more robust, "data-efficient" model that works across different machine loads and platforms.

Methods

The authors propose a two-pronged approach: a new architecture (MSLSTM) and a transfer learning strategy.

  1. Multi-Scale Convolution: Instead of a single-sized filter, the model uses three parallel convolutional branches with kernel sizes of 1×3, 1×5, and 1×7. This allows the network to "see" vibration patterns at different resolutions simultaneously.
  2. Convolutional Downsampling: Traditional max or average pooling layers are replaced with 1×1 convolutions (stride of 2). This allows the model to learn how to reduce dimensionality rather than just throwing data away, which preserves more detailed fault information.
  3. Global Average Pooling (GAP): The final fully connected layers are replaced with a GAP layer to reduce parameters and prevent overfitting.
  4. Transfer Learning: The model is first trained on a "source domain" with abundant data (e.g., a bearing under a 1HP load). It is then "transferred" to a "target domain" (e.g., a 3HP load or a different machine) and fine-tuned using a tiny fraction (5%) of the target data.

Datasets

The study validates the method using two well-known benchmarks:

  • CWRU (Case Western Reserve University) Dataset: The industry standard for bearing faults, featuring 10 status labels (1 healthy, 9 fault types) across three different motor loads (1HP, 2HP, 3HP).
  • QPZZ-II Bearing Dataset: A vibration analysis test-bed used to verify "cross-platform" performance (e.g., training on CWRU data and testing on QPZZ-II data).

Key Insights and Ideas

  • Pooling is a Information Leak: The authors argue that traditional pooling is too "blunt" for sensitive vibration signals. By using convolutions to reduce data size, the model retains "abundant and complex feature information" that would otherwise be lost.
  • The "5% Threshold": Without transfer learning, the model's accuracy plummets to ~75% when data is reduced to 5% of the total set. However, with transfer learning, that same 5% is enough to push accuracy back above 93%.
  • Multi-Scale Robustness: Vibration signals in noisy factories are messy. Capturing features across multiple scales (1x3 to 1x7) acts as a built-in "denoising" mechanism, making the model more stable than standard LSTMs.

Key Results

  • High Baseline Accuracy: With sufficient data, the MSLSTM achieves over 99% accuracy across different workloads.
  • Superior Transfer Performance: In variable load scenarios (e.g., 1HP $\rightarrow$ 2HP), MSLSTM achieved an average accuracy of 93.7%, significantly outperforming ResNet (70.8%), GFK (85.1%), and MMD (85.8%).
  • Cross-Platform Success: In the most difficult test—transferring knowledge between entirely different machines—the model maintained high accuracy: 93.5% for CWRU $\rightarrow$ QPZZ-II and 95.2% for QPZZ-II $\rightarrow$ CWRU.
  • Ablation Proof: Removing the multi-scale component or the convolutional pooling significantly dropped performance, proving that these specific architectural choices are the "secret sauce" of the model.

Limitations and Future Work

While the model handles variable loads and different machines well, the authors acknowledge that real-world industrial conditions are even more chaotic. Specifically:

  • Simultaneous Variables: Real bearings often face simultaneous changes in both speed and load, which this paper did not test in combination.
  • Future Directions: The authors plan to focus on even more complex, non-stationary running conditions to ensure the model can survive the "noise" of a real factory floor.

Summary updated

Abstract

It usually affects the accuracy and reliability of deep learning based intelligent diagnosis methods under the condition of insufficient samples. Existing methods for handling insufficient samples often have problems such as requiring rich expert experience or consuming a lot of time. To solve the above problems, a rolling bearing fault diagnosis method under insufficient samples condition based on multi-scale long-term and short-term memory network (MSLSTM) transfer learning is proposed, which mainly consists of an improved long-term and short-term memory network named as MSLSTM and transfer learning. By introducing multi-scale convolution operation into the traditional LSTM to improve its drawback that only extracts single type of fault feature information, which leads to poor diagnostic performance in noisy environments. Besides, the pooling layer and global average pooling layer in traditional LSTM are replaced with convolution operation to avoid the problem of information loss. Subsequently, the MSLSTM is combined with transfer learning, and a rolling bearing fault diagnosis method under insufficient samples condition based on MSLSTM transfer learning is proposed, which fine tunes the model parameters using a small amount of target domain data. Feasibility of the proposed method is verified through two kinds of experiments. The proposed method has stronger feature extraction ability and training efficiency compared with other models.
BibTeX
@article{zhang2025rolling,
  author = {Ping Zhang and Debo Liu},
  title = {{A rolling bearing fault diagnosis method under insufficient samples condition based on MSLSTM transfer learning}},
  journal = {Journal of Vibroengineering},
  year = {2025},
  doi = {10.21595/jve.2024.24621},
}
Figure