Concepts de base
A novel predictive machine learning framework called "Predict and Compare" can effectively detect gradual changes in heterogeneous time series data by differentiating true change points from regular trend patterns.
Résumé
The paper introduces a novel online change point detection framework called "Predict and Compare" (P&C) that can effectively handle heterogeneous time series data containing non-trivial trend patterns.
The key idea of P&C is to use a predictive machine learning model to forecast the trend in the immediate future based on the recent past data. It then compares the actual observed data in the prediction window to the forecasted trend. Significant deviations between the prediction and the actual data are flagged as change points.
This approach allows P&C to differentiate true change points from regular trend patterns that the predictive model has learned to recognize as "normal". This helps reduce the number of false positive detections, which is a common challenge for change point detection in the presence of complex, non-stationary trends.
The authors evaluate P&C using both LSTM neural networks and ARIMA models for the prediction step, combined with the CUSUM statistical test for the comparison step. They compare the performance of P&C to several state-of-the-art online change point detection methods, including Bayesian approaches, CUSUM, BFAST, and OCD.
The experiments are conducted on real-world heterogeneous time series data from tribological wear experiments, which exhibit complex run-in, steady-state, and divergent wear patterns. The results show that P&C outperforms the benchmark methods in terms of minimizing false positive detections while maintaining good detection times, especially in the presence of non-trivial trend patterns.
Stats
The wear of a critical machine part is continuously monitored via an in-situ technique based on radioactive isotopes.
The fast detection of the change points when divergent wear starts is crucial for wear testing and machine maintenance.
Citations
"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."
"The key idea and contribution of Predict and Compare (P&C) is the ability to differentiate between true CPs and patterns belonging to previously learned regular ('in-control') trend patterns."