Anomaly Detection of an Air Compressor from Time-series Measurement Data

Abstract

In railway vehicles, air compressor operates and stops aperiodically according to the pressure usage. LSTM-AE and anomaly score can be used to detect abnormality of an air compressor using time-series measurement data. However, it is hard to detect abnormal symptom in a short-term, without considering tendency of time-series data, because slight abnormal symptom occurs repetitively in the system if there is little failure. In this paper, a smoothing method using anomaly-score-moving average (ASMA) is proposed for anomaly detection from time-series measurement data. LSTM-AE model is trained from actual air compressor data at a railway vehicle and ASMA-based anomaly detection is performed using abnormal data.