Fault Diagnosis Method for Shearer Arm Gear Based on Improved S-Transform and Depthwise Separable Convolution.

Area of Science:

• Mechanical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Standard Convolutional Neural Networks (CNNs) face challenges in shearer arm gear fault diagnosis due to parameter redundancy and low training efficiency.
• Effective time-frequency feature representation is crucial for accurately identifying gear faults.

Purpose of the Study:

• To propose an improved diagnostic method for shearer arm gear faults using an enhanced S-transform and a Depthwise Separable Convolutional Neural Network (DSCNN).
• To improve the efficiency and accuracy of gear fault diagnosis by optimizing feature representation and network architecture.

Main Methods:

• Utilized an improved S-transform for time-frequency analysis, converting 1D vibration signals into 2D time-frequency images.
• Developed a DSCNN model combining standard and depthwise separable convolutions for fault identification.
• Compared performance against Wavelet Transform, Fourier Transform, S-transform, and Gramian Angular Field (GAF) using various CNN architectures.

Main Results:

• Frequency-domain representations significantly outperformed raw time-domain signals in fault diagnosis.
• Grad-CAM visualization confirmed the model's accurate focus on critical fault features.
• The proposed DSCNN method achieved high classification accuracy with reduced training time and fewer parameters compared to other CNNs.

Conclusions:

• The improved S-transform and DSCNN offer a robust and efficient solution for shearer arm gear fault diagnosis.
• Time-frequency image representations are superior to time-domain signals for this application.
• The method effectively balances diagnostic accuracy with computational efficiency.