Fault Diagnosis of the Electric Multiple Unit Door System by Machine Learning Using Sensor Signal of the Simulator

Fault diagnosis and prediction are important to prevent traffic congestion during rush hour due to door failures of urban railway vehicles. This paper is a study on improving failure classification performance through machine learning using the data set collected by installing a displacement sensor on a door simulator. First, the durability test of the sensor and the developed simulator was verified through 147,000 no-failure tests. For machine learning, 11,225 sets of normal and abnormal data of the door were collected and supervised learning was performed. In order to overcome the difficulty of fault diagnosis of the existing pressure sensor or acoustic sensor, pre-processing was performed that converted to speed-based data. In addition, feature extraction was compared with the single zone method by testing the 2-zone segmentation method. Feature selection was made using the principal component analysis algorithm developed for feature dimensionality reduction. As a result, the classification performance of the method using the single zone method with open and close data was better than the 2-zone segmentation method by acceleration and deceleration. Among the machine learning models, the LGBM model showed the highest prediction accuracy of 99.55%, which is expected to be applied to actual vehicles.