Anomaly Detection With Transformer for a Railway Vehicle Air Compressor

Abstract

Recently, research on condition-based maintenance (CBM) using artificial intelligence has attracted attention to reduce the maintenance costs of railway vehicles. The air compressor of a railway vehicle that adopts an air brake system is a target for CBM, and in order to reduce maintenance costs and ensure driving stability, technology to detect anomalies in the air compressor is necessary. The long short-term memory (LSTM) autoencoder is used to process sequence data. However, LSTM has limitations in that it cannot perform parallel processing due to its recurrent neural network characteristics, and it is difficult to learn the dependency between data as the sequence data lengthens. In this paper, we propose a novel transformer architecture as an anomaly detection model to learn dependency between air compressor data of a railway vehicle and we demonstrate superior data reconstruction and generalization ability of proposed transformer. We conduct simulation tests based on actual railway vehicle air compressor data and confirm improved anomaly detection performance. We propose an anomaly score definition method using mean squared error to perform reconstruction error-based anomaly detection. With the successful anomaly detection results in the air compressor of a railway vehicle, we demonstrate the effectiveness of a proposed anomaly detection algorithm that applies the moving average of anomaly scores and defines anomaly criteria using three-sigma. We test the proposed anomaly detection method using sensor data received from actual urban railway air compressors and prove its usefulness.