New Fault Diagnosis Method for Rolling Bearings Based on Improved Residual Shrinkage Network Combined with Transfer Learning

Tieyang Sun, Jianxiong Gao2024 · DOI

TL;DR

This paper introduces an "Improved Adaptive Residual Shrinkage Network" (IARSN) combined with transfer learning to tackle the challenge of rolling bearing fault diagnosis with limited data. The novel approach leverages an adaptive activation function and a residual shrinkage structure with attention and soft thresholding to enhance noise resistance and feature extraction, achieving high accuracy even with small, noisy datasets.

What Problem Does This Paper Address?

Rolling bearings are critical components in many machines, and their failure can lead to significant downtime and safety hazards. Accurate and timely fault diagnosis is crucial for predictive maintenance. However, a major hurdle in developing robust diagnostic models is the scarcity of labeled fault data, especially for rare fault types or under varying operating conditions. Traditional deep learning models often struggle with such small-sample, imbalanced datasets, leading to decreased diagnostic accuracy. Furthermore, real-world industrial environments are often noisy, which can further degrade the performance of diagnostic systems. This paper aims to address these limitations by proposing a method that can effectively diagnose bearing faults even with limited data and in the presence of significant noise.

Methods

The core of the proposed solution is the Improved Adaptive Residual Shrinkage Network (IARSN). This network integrates several key components:

  • Adaptive Residual Shrinkage Unit (ARSU): This unit is designed to extract multi-scale features by considering different scales of information within the vibration signals. It aims to alleviate issues like gradient disappearance and explosion during training.
  • Adaptive Parametric Rectified Linear Unit (APReLU): This novel activation function introduces nonlinearity and automatically adjusts its threshold. This allows the network to learn more complex relationships and improves its representation ability, helping to overcome dead neurons and enhance training stability.
  • Attention Mechanism and Soft Thresholding: To combat the inherent noise in vibration signals, the model incorporates an attention mechanism (inspired by Squeeze-and-Excitation networks) to adaptively adjust feature channel weights and a soft thresholding function. The soft thresholding function is crucial for noise reduction, effectively shrinking features towards zero while preserving useful negative characteristics and preventing gradient issues.
  • Transfer Learning: To overcome the data scarcity problem, the IARSN model is trained on a large, well-established dataset (Case Western Reserve University - CWRU) and then fine-tuned on smaller, target datasets (CWRU with reduced samples and Jiangnan University - JNU). This allows the model to leverage the knowledge gained from the source domain to improve performance on the target domain. The fine-tuning process involves freezing earlier layers of the network and training the later layers.

Datasets

The study utilizes two primary datasets:

  • Case Western Reserve University (CWRU) Bearing Dataset: This is a widely used public dataset for bearing fault diagnosis. It contains vibration data under various loads and speeds, with different fault types (inner race, outer race, ball defects) and normal conditions. The authors used this dataset for initial training of the IARSN model.
  • Jiangnan University (JNU) Bearing Dataset: This dataset was collected from an industrial motor and includes data for two types of bearings (NU205 and N205) under different operating conditions and fault types (inner-sphere, rolling-element, outer-sphere, and normal). This dataset was used for migration experiments, simulating a smaller, target domain.

Key Insights and Ideas

  • Noise is a significant challenge, but manageable: The paper demonstrates that even with substantial added Gaussian white noise (down to -3 dB SNR), the proposed IARSN model maintains high diagnostic accuracy, highlighting its strong antinoise capabilities.
  • Transfer learning is essential for small-sample fault diagnosis: The experiments clearly show a dramatic improvement in accuracy when transfer learning is applied, moving from around 70-76% to over 95% on small-sample target datasets. This validates the effectiveness of transferring knowledge from a data-rich source to a data-scarce target.
  • APReLU offers a flexible and robust activation: The introduction of APReLU as an adaptive activation function appears to be a key innovation, allowing the network to better handle diverse vibration signal characteristics and improve overall model fitting.
  • IARSN's architecture is well-suited for complex signals: The combination of residual shrinkage, attention, and soft thresholding within the IARSN architecture effectively processes signals with high redundancy and noise, extracting more meaningful features.

Key Results

  • High Accuracy on Small-Sample Datasets: The IARSN model achieved an accuracy of 99.3% on the CWRU small-sample dataset after transfer learning, compared to 76.3% without it.
  • Robustness to Noise: Even with signal-to-noise ratios as low as -3 dB, the model maintained an accuracy of 85.0%, and reached 96.0% at 3 dB SNR.
  • Effective Cross-Domain Transfer: On the JNU dataset, the IARSN model achieved 95.3% accuracy with transfer learning, significantly outperforming the 70.7% accuracy without it.
  • Unidirectional Transfer Effectiveness: Migration experiments between different load conditions (A-B, B-A, etc.) showed consistently high accuracies (above 95%), indicating that the model's transferability is not limited to a single direction.
  • Comparative Performance: In comparative experiments against other state-of-the-art methods (RDPN-FCDAE, SA-ACGAN, GPLVM-mcODM, Bayesian PDL), the proposed IARSN model demonstrated the best performance.

Limitations and Future Work

  • Laboratory vs. Real-World Data: The study acknowledges that the data used was primarily from laboratory experiments. Future work should focus on collecting and testing the model with data from bearings in actual operation.
  • Limited Fault States: The current study focused on ten specific fault states. Future research could explore scenarios with a coexistence of multiple fault states and varying degrees of failure.
  • Sensor Fusion: The current method relies solely on vibration signals. Integrating other sensor data, such as temperature and current, could further enhance the robustness and diagnostic capabilities of the fault diagnosis method.

Summary updated

BibTeX
@misc{sun2024new,
  author = {Tieyang Sun and Jianxiong Gao},
  title = {{New Fault Diagnosis Method for Rolling Bearings Based on Improved Residual Shrinkage Network Combined with Transfer Learning}},
  year = {2024},
  doi = {10.3390/s24175700},
}
Figure