Enhanced Feature Extraction Network Based on Acoustic Signal Feature Learning for Bearing Fault Diagnosis

Yuanqing Luo, Wenxia Lu, Shuang Kang, Xueyong Tian, Xiaoqi Kang, Feng SunSensors 2023 · DOI

TL;DR

This paper introduces an Enhanced Feature Extraction Network (EFEN) that leverages acoustic signals for more robust and accurate rolling bearing fault diagnosis. By transforming acoustic data into images and employing novel information selection and channel attention modules, EFEN achieves a remarkable 98.52% accuracy, outperforming existing methods, especially in noisy environments.

What Problem Does This Paper Address?

Rolling bearings are critical components in mechanical equipment, and their failure can lead to significant economic losses and safety hazards. Traditional fault diagnosis methods often rely on vibration signals, which require contact measurement and can be challenging to implement in certain complex or space-constrained scenarios. Acoustic signals offer a non-contact alternative, but they are highly susceptible to background noise and interference, making accurate fault diagnosis difficult. Existing deep learning approaches for acoustic signal analysis often use overly simplistic feature extraction and struggle with noise, limiting their effectiveness. This paper aims to bridge this gap by developing a robust deep learning framework that can effectively extract fault features from noisy acoustic signals for reliable bearing fault diagnosis.

Methods

The proposed EFEN network is an end-to-end deep learning model composed of several key modules:

  • Data Preprocessing: One-dimensional acoustic signals are converted into two-dimensional grayscale images. This transformation is achieved by normalizing the signal to a 0-1 range and then reshaping segments of 1024 data points into 32x32 grayscale images. This approach aims to capture richer feature information compared to 1D signals.
  • Information Feature Selection Module (IFSM): This novel module utilizes three convolutional filters of different sizes (3x3, 5x5, and 7x7) to process the input image data. These filters extract features at various scales, and their outputs are concatenated and further processed by a 1x1 convolution. This multi-scale approach helps to attenuate noise while highlighting critical fault information.
  • Channel Attention Mechanism Module (CAMM): Recognizing that different acoustic channels (from the microphone array) contribute differently to fault diagnosis, CAMM assigns adaptive weights to each channel. It uses global average and max pooling, followed by a multi-layer perceptron (MLP), to learn channel-wise importance, thereby refining fault information.
  • Convolutional Neural Network Module (CNNM): A standard CNN architecture, comprising convolutional layers, pooling layers (specifically max pooling), and fully connected layers with a softmax activation function, is used for the final classification of bearing faults.

The entire EFENet model is trained using a backpropagation algorithm with a cross-entropy loss function.

Datasets

The study utilizes experimental data collected from a rolling bearing test bench. The dataset comprises acoustic signals from bearings in various states: normal, inner ring failure, outer ring failure, rolling element failure, and cage failure. Specifically, seven fault types (labeled C0 to C6) were generated. For each fault type, 500 samples were collected. The dataset was split with 80% for training and 20% for testing and verification. The data length for each sample is 1024.

Key Insights and Ideas

  • Acoustic Signals as a Viable Alternative: The research strongly supports the use of acoustic signals for non-contact bearing fault diagnosis, especially when vibration sensors are impractical.
  • Image Transformation for Acoustic Data: Converting 1D acoustic signals into 2D grayscale images is an effective strategy to enhance feature representation for deep learning models.
  • Multi-Scale Feature Extraction is Crucial: The IFSM's use of varying convolution kernel sizes is key to capturing both localized and global fault characteristics, leading to more comprehensive feature extraction.
  • Channel Attention for Signal Robustness: CAMM's ability to dynamically weight different acoustic sensor channels significantly improves the model's resilience to noise and its overall diagnostic accuracy.
  • EFENet's Superiority in Noise: The proposed EFENet demonstrates remarkable performance not only in clean conditions but also under noisy environments, highlighting its robustness.

Key Results

  • High Accuracy: The EFEN network achieved an impressive classification accuracy of 98.52% on the test set.
  • Noise Robustness: Even when 2 dB of white Gaussian noise was added to the acoustic signals, the EFENet maintained a high average accuracy of 96.62%.
  • Efficiency: The average iteration time for the EFENet model was a mere 0.25 seconds, indicating its computational efficiency.
  • Comparative Performance: EFENet outperformed several established deep learning models, including RCAE, CNN, SAE, DenseNet, and ResNet, by a significant margin in terms of accuracy. For instance, it was 7% higher than CNN and 4% higher than ResNet.
  • Visualization: t-SNE visualizations showed that EFENet effectively separates different fault classes, with clear boundaries and significant inter-class spacing, especially after noise injection.

Limitations and Future Work

  • Limitations: While the paper demonstrates excellent performance, the authors acknowledge the potential for computational efficiency and overfitting issues in deep learning models. The study primarily focuses on acoustic signals, and further research could explore multimodal fusion with other sensor types.
  • Future Work: The authors plan to explore lightweight pre-trained deep learning models to address computational efficiency and overfitting. Their future research will also focus on developing and refining fault life prediction methods for rolling bearings, specifically utilizing acoustic radiation signals.

Summary updated

BibTeX
@article{luo2023enhanced,
  author = {Yuanqing Luo and Wenxia Lu and Shuang Kang and Xueyong Tian and Xiaoqi Kang and Feng Sun},
  title = {{Enhanced Feature Extraction Network Based on Acoustic Signal Feature Learning for Bearing Fault Diagnosis}},
  journal = {Sensors},
  year = {2023},
  doi = {10.3390/s23218703},
}
Figure