Long and short term fault prediction using the VToMe-BiGRU algorithm for electric drive systems

Area of Science:

• Automotive Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Electric drive systems are vital for new energy vehicles, requiring reliable operation.
• Ensuring the stability and efficiency of these systems is critical to prevent accidents.
• Existing fault prediction methods may lack the speed or accuracy needed for real-time applications.

Purpose of the Study:

• To develop an innovative fault prediction architecture for electric drive systems.
• To combine the strengths of Token Merging (ToMe) within Vision Transformer (ViT) and Bidirectional Gated Recurrent Unit (BiGRU) for enhanced performance.
• To improve the accuracy, reliability, and efficiency of electric vehicle fault prediction.

Main Methods:

• Embedding the Token Merging (ToMe) algorithm into Vision Transformer (ViT) to create the VToMe algorithm.
• Combining the VToMe algorithm with a Bidirectional Gated Recurrent Unit (BiGRU) network, forming the VToMe-BiGRU architecture.
• Applying the VToMe-BiGRU architecture to real-world electric vehicle maintenance datasets for experimental validation.

Main Results:

• The VToMe-BiGRU architecture achieved 93.49% average accuracy in multi-class fault classification, outperforming ViT++ by 0.12%.
• Inference speed was enhanced by 28% (32 FPS vs. 25 FPS), balancing precision and real-time efficiency.
• Short-term prediction showed a root-mean-square error (RMSE) of 6.33 and 74.7% accuracy for complex faults, surpassing traditional models by over 20%.

Conclusions:

• The VToMe-BiGRU architecture provides a robust framework for real-time diagnosis of EV drive systems.
• The method effectively detects critical faults with minimized false positives, enhancing system reliability and reducing maintenance costs.
• This approach supports proactive safety measures and improves the overall performance of electric vehicles.