Anomaly Detection of a Screw Air Compressor in a Railway Vehicle through Feature Extraction of Data

Failure of an air compressor in a railway vehicle threatens the safety of passengers; thus, anomaly detection of air compressors using deep learning techniques has gained significant attention recently. Training an LSTM-autoencoder through data partitioned using the sliding window technique results in extremely large deviations between the anomaly scores of the windows and extremely large thresholds; thus, the conventional LSTM-autoencoder method has limited anomaly detection capacity. In this study, we partitioned the data using data feature extraction to reduce the variability of the average anomaly score between datasets and detect abnormality of an air compressor on a daily basis using CNN-LSTM-autoencoder for data reconstruction performance. The proposed method was applied to air compressor data received from an Incheon Airport Line (AREX) railway vehicle, and the performance of the anomaly detection of the air compressor of the railway vehicle was confirmed.