Sample Augmentation Using Enhanced Auxiliary Classifier Generative Adversarial Network by Transformer for Railway...

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Railway Engineering

Background:

• Accurate diagnosis of wheelset bearing faults is crucial for high-speed train safety.
• Limited fault sample data hinders deep learning model training and diagnostic accuracy.
• Existing Auxiliary Classifier Generative Adversarial Network (ACGAN) models face challenges like complexity and computational expense.

Purpose of the Study:

• To address the limitations of current ACGAN models in generating diverse and accurate fault samples for wheelset bearing diagnostics.
• To propose a novel Transformer and Auxiliary Classifier Generative Adversarial Network (TACGAN) for enhanced fault diagnosis.

Main Methods:

• Implemented a Transformer network to replace traditional Convolutional Neural Networks (CNNs), reducing computational load.
• Integrated an independent classifier to resolve the coupling issue in ACGAN.
• Utilized Wasserstein distance in the loss function to mitigate mode collapse and vanishing gradients.

Main Results:

• The proposed TACGAN significantly increased the diversity, complexity, and entropy of generated fault samples.
• TACGAN demonstrated reduced computational expenses compared to traditional ACGAN models.
• Experimental results confirmed the high accuracy and effectiveness of TACGAN on train wheelset bearing datasets.

Conclusions:

• TACGAN offers an effective solution for generating diverse fault samples in scenarios with limited data.
• The model enhances the accuracy and efficiency of deep learning-based fault diagnosis for wheelset bearings.
• TACGAN represents a significant advancement in ensuring train safety through improved diagnostic capabilities.