Bearing fault diagnosis based on efficient cross space multiscale CNN transformer parallelism

Area of Science:

• Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Wind turbine bearing fault diagnosis is critical for operational safety and reliability.
• Traditional serial deep learning models struggle with spatio-temporal feature extraction in noisy environments, causing information loss.

Purpose of the Study:

• To propose an Efficient Cross Space Multiscale CNN Transformer Parallelism (ECMCTP) model for enhanced wind turbine bearing fault diagnosis.
• To improve the extraction of spatio-temporal features from vibration signals, especially under noisy conditions.

Main Methods:

• 1D vibration signals are converted to 2D time-frequency images using Continuous Wavelet Transform (CWT).
• Parallel CNN and Transformer branches extract features: CNN uses multiscale modules, Reversed Residual Structure (RRS), and Efficient Multiscale Attention (EMA); Transformer uses Bidirectional Gated Recurrent Units (BiGRU) and Transformer.
• Features are concatenated and classified via a softmax classifier.

Main Results:

• The ECMCTP model achieved 100% accuracy on fault diagnosis under noise-free conditions.
• The model demonstrated superior robustness and generalization capabilities under low signal-to-noise ratio (SNR) conditions.

Conclusions:

• The proposed ECMCTP model effectively addresses the limitations of traditional methods for wind turbine bearing fault diagnosis.
• The parallel CNN-Transformer architecture offers excellent robustness and generalization, particularly in challenging noisy environments.