Train wheel-rail force collaborative calibration based on GNN-LSTM

Abstract

Accurate wheel-rail force data serves as the cornerstone for analyzing the wheel-rail relationship. However, achieving continuous and precise measurement of this force remains a significant challenge in the field. This article introduces a calibration algorithm for the wheel-rail force that leverages graph neural networks and long short-term memory networks. Initially, a comprehensive wheel-rail force detection system for trains was constructed, encompassing two key components: an instrumented wheelset and a ground wheel-rail force measuring system. Subsequently, utilizing this system, two distinct datasets were acquired from the track inspection vehicle: instrumented wheelset data and ground wheel-rail force data, a feedforward neural network was employed to calibrate the instrumented wheelset data, referencing the ground wheel-rail force data. Furthermore, ground wheel-rail force data for the locomotive was obtained for the corresponding road section. This data was then integrated with the calibrated instrumented wheelset data from the track inspection vehicle. Leveraging the GNN-LSTM network, the article establishes a mapping relationship model between the wheel-rail force of the track inspection vehicle and the locomotive wheel-rail force. This model facilitates continuous measurement of locomotive wheel-rail forces across three typical scenarios: straight sections, long and steep downhill sections, and small curve radius sections.