Rolling-Element Bearing Fault Diagnosis Using Improved LeNet-5 Network.

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Traditional LeNet-5 networks exhibit limitations in rolling-element bearing fault diagnosis, including low accuracy, slow convergence, and poor generalization.
• Effective fault diagnosis is crucial for machinery health and preventing catastrophic failures.

Purpose of the Study:

• To develop an improved LeNet-5 based method for enhanced rolling-element bearing fault diagnosis.
• To address the limitations of traditional LeNet-5 in accuracy, convergence speed, and generalization ability.

Main Methods:

• An improved 2D LeNet-5 network was developed with optimized convolution/pooling layers, adjusted kernel sizes, batch normalization (BN), and dropout operations.
• An end-to-end improved 1D LeNet-5 network was proposed for direct 1D convolution/pooling on raw vibration signals, eliminating preprocessing steps.

Main Results:

• Both improved 2D and 1D LeNet-5 networks demonstrated significant performance gains over the traditional LeNet-5.
• The improved 1D LeNet-5 achieved higher accuracy with reduced training time in most scenarios.
• The improved 2D LeNet-5 outperformed the improved 1D version in conditions with limited training data and high noise levels.

Conclusions:

• The proposed improved LeNet-5 networks offer substantial advancements in rolling-element bearing fault diagnosis.
• The improved 1D LeNet-5 provides an efficient and accurate solution for real-time fault detection.
• The improved 2D LeNet-5 demonstrates robustness in challenging diagnostic environments with limited samples or significant noise.