Anomaly detection in noisy and irregular time series: the "turbodiesel charging pressure" case study

In this paper we consider the problem of detecting anomalies in sample series obtained from critical train subsystems. Our study is the analysis of charging pressure in turbodiesel engines powering a fleet of passenger trains. We describe an automated methodology for (i) labelling time series samples as normal, abnormal or noisy, (ii) training supervised classifiers with labeled historical data, and (iii) combining classifiers to filter new data. We provide experimental evidence that our methodology yields error rates comparable to those of an equivalent manual process.