Research on temperature prediction method for rail transit train inverters based on spatial and timing improving...

Area of Science:

• Rail transit engineering
• Artificial intelligence
• Predictive maintenance

Background:

• Inverter overheating is a critical failure point in rail transit systems.
• Challenges include sparse low-voltage data and high-dimensional features, hindering accurate prediction.
• Existing methods struggle with data scarcity and complexity.

Purpose of the Study:

• To develop a hybrid prediction framework for accurate inverter temperature forecasting in rail transit.
• To address data sparsity and high dimensionality using advanced AI techniques.
• To enhance the safety and efficiency of rail operations through intelligent fault prevention.

Main Methods:

• Utilized Random Masked Dual DCGAN (RTDG) for low-voltage data augmentation.
• Employed Gaussian Markov Random Field (GMRF) for dimensionality reduction and key variable identification.
• Developed an enhanced Transformer architecture (STTr) integrating state space modeling and temporal normalization for spatio-temporal dependency capture.
• Fused components using a weighted stacking strategy and validated on real-world rail transit data.

Main Results:

• The proposed hybrid model significantly outperformed conventional approaches in inverter temperature prediction.
• Achieved a 4.93% improvement over single models and a 9.73% gain compared to non-augmented training.
• Demonstrated effectiveness in handling sparse data and complex features.

Conclusions:

• The developed hybrid framework offers a robust solution for inverter overheating prediction in rail transit.
• Supports intelligent fault prevention, contributing to enhanced safety and operational efficiency.
• Provides a valuable tool for maintaining the integrity of modern rail systems.