Bearing Fault Diagnosis via Improved One-Dimensional Multi-Scale Dilated CNN

Jiajun He, Ping Wu, Yizhi Tong, Xujie Zhang, Meizhen Lei, Jinfeng Gao2021 · DOI

TL;DR

This paper introduces an improved Convolutional Neural Network (CNN) architecture, the Improved Multi-Scale Dilated CNN (IMSCNN), for diagnosing bearing faults in rotating machinery. By incorporating dilated convolutional kernels and a preliminary 1D convolutional layer, IMSCNN effectively extracts multi-scale features while mitigating noise, leading to superior fault diagnosis accuracy compared to existing methods.

What Problem Does This Paper Address?

Bearings are critical components in rotating machinery, and their failure can lead to significant downtime and costs. Early and accurate fault diagnosis is crucial for ensuring operational safety and reducing maintenance expenses. While vibration analysis is a common method for detecting bearing faults, traditional signal analysis techniques often require extensive expert knowledge. Data-driven methods, particularly those leveraging deep learning, have shown promise by automatically extracting features. However, existing CNN-based approaches often use fixed-size convolutional kernels, which may not fully capture the diverse informative features present in vibration signals, especially when these signals are contaminated with noise. This paper aims to address this by developing a CNN architecture that can effectively extract multi-scale features from noisy vibration data.

Methods

The proposed method, IMSCNN, builds upon the strengths of CNNs by introducing two key enhancements:

  • Improved Multi-Scale Feature Extraction: The core of the IMSCNN is a "dilated multi-scale convolutional (DMSConv)" layer. This layer integrates four dilated convolutional kernels with varying dilation rates. Dilated convolutions allow for a larger receptive field without increasing computational cost or kernel size, enabling the capture of features at different scales. Specifically, two of these kernels have a dilation rate of 1, effectively acting as standard convolutional kernels, while others have higher dilation rates to capture broader patterns. This multi-scale approach is inspired by the inception structure.
  • Noise Reduction with 1D Convolution: To combat the common issue of noise in vibration signals, a 1D convolutional layer is employed before the DMSConv layer. This initial layer acts as a filter, helping to mitigate the impact of noise and extract cleaner features for subsequent processing.

The network also includes standard components like pooling layers (specifically max pooling) and fully connected layers for classification.

Datasets

The proposed IMSCNN method was evaluated on two widely used benchmark datasets for bearing fault diagnosis:

  • CWRU Dataset: Provided by the Case Western Reserve University bearing data center, this dataset contains vibration data collected under various faulty conditions (inner race, outer race, ball faults) and normal conditions. The study utilized data with a 12 kHz acquisition frequency.
  • PU Dataset: This dataset comes from the Paderborn University Bearing Data Center and includes vibration data from 14 different faulty conditions. The experiments used data collected under specific operating conditions (1500 rpm, 0.7 Nm load torque, 1000 N radial force).

Key Insights and Ideas

  • Dilated Convolutions for Multi-Scale Feature Capture: The paper highlights how dilated convolutions, by introducing spacing in the kernel, can effectively expand the receptive field and capture features at different scales without increasing computational complexity. This is crucial for complex signals like vibrations where faults manifest at various frequency bands.
  • The Importance of Pre-processing for Noise: The inclusion of an initial 1D convolutional layer to filter noise before feature extraction is a significant insight. It demonstrates that addressing signal quality upfront can substantially improve the performance of subsequent deep learning models.
  • Synergy of Multi-Scale and Noise Reduction: The combination of multi-scale feature extraction (via dilated convolutions) and noise reduction (via the initial 1D CNN) proves to be more effective than either technique alone, leading to a robust diagnostic system.
  • Visualizing Feature Separability: The use of t-SNE to visualize extracted features clearly shows that the proposed IMSCNN method results in more distinct and separable clusters for different fault types compared to simpler MLP or standard CNN models. This provides a qualitative understanding of why IMSCNN achieves higher accuracy.

Key Results

The IMSCNN method demonstrated superior performance on both datasets compared to several baseline methods, including MLP, CNN, MSCNN, and SimpleIMSCNN.

  • CWRU Dataset:
    • MLP achieved 94.63% accuracy.
  • CNN achieved 99.77% accuracy.
  • MSCNN, SimpleIMSCNN, and the proposed IMSCNN all achieved 100% accuracy.
  • PU Dataset:
    • MLP achieved 69.69% accuracy.
  • CNN achieved 85.64% accuracy.
  • MSCNN achieved 95.53% accuracy.
  • SimpleIMSCNN achieved 92.10% accuracy.
  • The proposed IMSCNN achieved the highest accuracy at 96.55%.

The paper also notes that while MSCNN and SimpleIMSCNN also performed well, IMSCNN's initial 1D convolutional layer was particularly effective in filtering noise, leading to its improved performance on the PU dataset.

Limitations and Future Work

  • Limitations: The paper primarily focuses on 1D vibration signals. While the proposed method is effective, it doesn't explicitly explore the fusion of multi-sensor data or other signal types. The computational cost of training deep learning models, even with optimizations, can still be a consideration.
  • Future Work: The authors suggest that the proposed IMSCNN method can be further improved by exploring more advanced feature extraction techniques and potentially incorporating multi-sensor data fusion for even more comprehensive fault diagnosis. The effectiveness of the noise filtering layer could also be investigated further with different types and levels of noise.

Summary updated

BibTeX
@misc{he2021bearing,
  author = {Jiajun He and Ping Wu and Yizhi Tong and Xujie Zhang and Meizhen Lei and Jinfeng Gao},
  title = {{Bearing Fault Diagnosis via Improved One-Dimensional Multi-Scale Dilated CNN}},
  year = {2021},
  doi = {10.3390/s21217319},
}
Figure