Segmentation Improves Prediction Without Increasing Statistical Separability: Evidence from Industrial Time Series

Abstract Change Point Detection (CPD) techniques are widely applied in predictive maintenance systems to improve performance in detecting anomalies within multivariate industrial time series. Previous studies on a steam turbine monitoring system demonstrated that segmentation based on the ChangeFinder algorithm, combined with heterogeneous ensemble models, substantially enhances classification accuracy. Despite these results, the underlying cause of this improvement remains unclear. A common interpretation is that segmentation increases statistical separability between healthy and faulty operational conditions, thereby facilitating the learning process of classification models.In this study, we critically examine this hypothesis by employing a classifier-independent approach that relies on the Sliced Wasserstein Distance (SWD). We analyze and compare data representations derived from both raw process variables and segmentation-related features. Statistical robustness are ensured through stratified bootstrap resampling and repetitive distance estimations.Our findings reveal that segmented representations consistently show reduced statistical separability compared to the original feature spaces. This trend persists even with the addition of extra clustering-derived descriptors. The results suggest that previously observed improvements in predictive performance are not due to an increase in statistical distance between classes. Rather, segmentation appears to assist learning by diminishing temporal ambiguity and coherently structuring the temporal evolution of process dynamics.