基于长期监测系统和无监督学习的道岔钢轨健康监测OA

A turnout is an important infrastructure in the railway system,and it is also a major challenge in track system maintenance.The track structure in the turnout area is complex and the rails are relatively weak,making them more prone to damage than rails in other sections.In addition,the inability of rail flaw detection vehicles to fully assess the condition of the rails in the turnout area makes these areas a blind spot for automatic detection.Therefore,monitoring the condition of turnout rails and detecting damage in a timely manner is an urgent need for China's railway operation and development.Based on a health monitoring system for turnout switch rails,multi-sensor data dominated by acoustic emission were collected,and the models were established based on two unsupervised learning algorithms:one-class support vector machine(SVM)and isolation forest(IF).These models were used to classify and identify the single-sensor acoustic emission data and multi-sensor data of turnout rails.The classification accuracy,recall rate,and F1 score of the multi-sensor algorithm all exceeded 80%,and the AUC exceeded 70%,outperforming the single-sensor acoustic emission-based method by more than 10%.This method made full use of the advantages of various sensors to address the complex and variable noise characteristics of acoustic emission signals in turnout areas.The proposed method was compared with the traditional supervised learning K-nearest neighbor algorithm,highlighting the advantages of isolation forest and one-class SVM models in terms of limited measured damage data or even no damage data.Long-term monitoring of the service performance of the turnout rail has been achieved,addressing the challenges of rail detection in the turnout areas.