Deep learning-based identification of track structure health using multivariate data

Authors:

Abstract
The safe operation of rail transit systems requires accurate identification of track structure health conditions. Traditional track inspection methods, which mainly involve manual inspection and single-sensor identification, often suffer from high labor costs, low efficiency, and missed detections. To address these challenges, this paper proposes a track structure health identification method called CATDNet (Convolutional Attention for Track structure health identification Network), which integrates one-dimensional convolutional neural networks (1D CNN) with signal analysis techniques. The method designs a convolutional attention module that combines CNN with attention mechanisms to evaluate track structure health status through automatic extraction and classification of multiple signal features including rail acceleration, track bed vibration acceleration, tunnel wall Acceleration and tunnel wall sound pressure. To verify the effectiveness of the method, this study analyzes simulated vibration and sound pressure data collected in laboratory conditions. Additionally, to address noise interference issues, signals with different signal-to-noise ratios (SNR) of white noise are validated. Furthermore, the method is tested on public voiceprint datasets to verify its robustness across different data types. Experimental results demonstrate that CATDNet achieves good performance in time—frequency feature extraction and multi—source signal fusion, maintaining reliable identification performance even under challenging noise conditions. The proposed method provides a new perspective for intelligent health identification in rail transit systems, aiming to reduce manual inspection costs and improve maintenance efficiency through automated early defect detection with rapid response capability.
Funding Information
  • Research Project of Beijing Infrastructure Investment Co., Ltd (2023-GY-08)

This publication has 18 references indexed in Scilit: