Speaker
Description
Predictive Method for Track Quality Index Based on Multi-Scale Causal Convolutional Neural Networks
Xiaolin Li 1, Yuanjie Tang 2*,Yong Zhuang 1, liFen Yuan 2
1School of Traffic and Transportation, Beijing Jiaotong University, Beijing, China
2Key Laboratory of Transport Industry of Big Data Application Technologies for Comprehensive Transport, Beijing Jiaotong University, Beijing, China
Correspondence Yuanjie Tang, Key Laboratory of Transport Industry of Big Data Application Technologies for Comprehensive Transport, Beijing Jiaotong University, Beijing 100044, China. Email: tangyj@bjtu.edu.cn
Abstract: The efficient operation and safety maintenance of railway systems depend on precise predictions of the Track Quality Index (TQI). Existing forecasting methodologies often fail to adequately capture multi-scale temporal dynamics and essential causal relationships. To address these deficiencies, we introduce a novel predictive framework, MSCANet (Multi-Scale Causal Attention Network), which integrates causal inference with deep learning techniques. For the first time, causal convolutional networks are employed to uncover the deep causal mechanisms underlying variations in TQI. Through causal discovery and causality testing, a comprehensive causal relationship graph of the system’s TQI is constructed. The model is equipped with causally-consistent convolution layers, a multi-scale analysis framework, and attention mechanisms, enabling effective capture of complex temporal dependencies. Experimental results show that MSCANet significantly surpasses traditional time series models in both short-term and long-term TQI predictions. This study clearly identifies the primary drivers of TQI variations, offering more interpretable and accurate predictive support for railway track maintenance.
Keywords: Track Quality Index, Causal Inference, Convolutional Neural Networks, Multi-Scale Analysis, Time Series Forecasting
