A Lightweight Transformer Edge Intelligence Model for RUL Prediction Classification.

Area of Science:

• Engineering
• Computer Science
• Artificial Intelligence

Background:

• Remaining Useful Life (RUL) prediction is vital for predictive maintenance.
• Current models struggle with accuracy and efficiency on resource-constrained devices.
• Complex degradation feature extraction poses a challenge for existing methods.

Purpose of the Study:

• To propose TBiGNet, a lightweight Transformer-based network for RUL prediction.
• To overcome the limitations of existing models in balancing accuracy and computational cost.
• To enable effective RUL prediction on edge devices.

Main Methods:

• Developed an encoder-decoder Transformer architecture (TBiGNet).
• Optimized the attention mechanism for reduced memory access (60% reduction).
• Integrated an adaptive feature pruning module and BiGRU in the decoder for enhanced feature capture.

Main Results:

• TBiGNet achieved over 15% higher accuracy than traditional Transformers.
• Reduced computational load, memory access, and parameter size by over 98%.
• Demonstrated significant improvements in accuracy (6% from pruning, 7% from decoder).

Conclusions:

• TBiGNet offers superior performance in accuracy, model size, and memory access.
• The model shows significant technical advantages and potential for edge device deployment.
• TBiGNet advances RUL prediction for predictive maintenance applications.