철도 차량 제동 실린더 압력 예측을 위한 Branch-Attention DeepONet

초록

Brake cylinder pressure (BCP) is a key variable for railway safety, yet electronic control unit (ECU) data are high-dimensional, redundant, and noisy, making accurate prediction challenging. This study proposes an exploratory variable analysis and a lightweight Branch-Attention DeepONet to achieve high predictive accuracy and diagnostic applicability. Considering correlation, multicollinearity, and physical characteristics, 17 representative variables were selected from 66. The DeepONet architecture employs a multi-head attention-based branch network to learn temporalvariable dependencies and a trunk network to capture operational context. Power-Sigmoid and ReLU activations ensure non-negativity and physical upper bounds of pressure. Experiments show the proposed model accurately predicts BCP with interpretability, demonstrating potential as an AI-based digital twin diagnostic system for real-time brake monitoring and early anomaly detection.

키워드