# Research on multi-stage on-board detection algorithm of track defects of high-speed railway based on the influence mechanism of track defects

## TL;DR

This paper proposes a multi-stage on-board detection algorithm for track defects in high-speed railways to improve train safety by analyzing vibration responses and using machine learning techniques.

## Contribution

A novel multi-stage detection algorithm combining machine learning, deep learning, and transfer learning for track defect recognition is introduced.

## Key findings

- A multi-scale track defect model and segmental numerical solution method improve solving efficiency.
- A track defect characteristic decoupling algorithm using signal decomposition and time-frequency analysis is developed.
- A transfer learning network enables track defect detection algorithm transferability across different vehicles.

## Abstract

The ballastless track defects, such as rail corrugation, fastener looseness, and CA mortar layer disengagement, will affect the safety of high-speed train operation. Hence, a multi-stage on-board detection algorithm of track defects based on the influence mechanism of track defects is proposed for track condition monitoring. Firstly, based on refined and discrete track models, a multi-scale track defect model and a segmental numerical solution method are developed, which greatly improve the solving efficiency. Then, based on the multi-scale vehicle-track coupling model with track defects, the effect of track defects on wheelset vibration response is analysed from multiple dimensions. Simultaneously, a track defect characteristic decoupling algorithm based on signal decomposition and time-frequency analysis is suggested, and a series of characteristic indexes representing track defects are constructed. Finally, a multi-stage intelligent recognition algorithm of track defects based on the nested combination of machine learning, deep learning, and transfer learning is proposed‌. In the first stage, the identification of track defect types is achieved using wheelset vibration feature vectors and the nonlinear classifier model. In the second stage, a multi-channel track defect degree recognition model is designed based on a deep residual network and multiple wheelset data, achieving accurate estimation of rail corrugation depth, fastener looseness number, and CA mortar disengagement length. In the third stage, a transfer learning network for the track defect degree recognition model based on multiple moment matching is put forward, which enables the transfer of the track defect detection algorithm for different vehicles.