Predictive change point detection for heterogeneous data

Abstract

An unsupervised change point detection (CPD) framework assisted by a predictive machine learning model called “Predict and Compare” is introduced which is able to detect change points online under the presence of non-trivial trend patterns which must be prevented from triggering false positives. Different predictive models for the required time series forecasting (Predict) step together with different statistical tests for deciding about the proximity of predicted and actual data (Compare step) are allowed. Its performance is shown for the Predict step being carried out by either an LSTM recursive neural network or an ARIMA linear time series model together with the CUSUM rule as Compare step method. It shows to perform best in comparison to several other online CPD methods for detect times in the regime of low numbers of false positive detections. The method’s good performance is based on its ability to detect structural changes in the presence complex underlying trend patterns. The use case concerns tribological wear for which change points separating the run-in, steady-state, and divergent wear phases are detected.