Lightweight Gearbox Fault Diagnosis Under High Noise Based on Improved Multi-Scale Depthwise Separable Convolution...

Area of Science:

• Mechanical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Gearbox fault diagnosis is hindered by extracting weak fault features from noisy vibration signals and inefficient multi-scale modeling.
• Existing methods struggle with diagnostic stability across varying noise levels and limited computational resources.

Purpose of the Study:

• To propose a lightweight and efficient fault diagnosis model (DSMC-ECA) for gearboxes operating under strong-noise conditions.
• To enhance the extraction of fault-related features and improve diagnostic stability despite significant noise interference.

Main Methods:

• Developed a dual-branch architecture integrating improved multi-scale depthwise separable convolution (SMC and SMDC branches) for feature extraction.
• Incorporated Efficient Channel Attention (ECA) for channel-wise recalibration, emphasizing fault features and suppressing noise.
• The proposed DSMC-ECA model is computationally efficient, with 0.204 M parameters and 10.037 M FLOPs.

Main Results:

• DSMC-ECA demonstrated superior performance compared to baseline methods across a wide range of signal-to-noise ratios (-6 dB to noise-free).
• Achieved highest average diagnostic accuracies of 95.11% on the XJTU dataset and 86.84% on the SEU dataset at -6 dB SNR.
• The model offers a favorable trade-off between diagnostic performance and computational efficiency.

Conclusions:

• The proposed DSMC-ECA model effectively addresses the challenges of gearbox fault diagnosis in strong-noise environments.
• Its lightweight design and robust feature extraction capabilities make it suitable for practical industrial applications with limited resources.