Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TimePred: efficient and interpretable offline change point detection for high volume data -- with application to industrial process monitoring

Simon Leszek

TL;DR

TimePred reframes high-dimensional offline change-point detection as a univariate mean-shift problem by training a time-prediction network to produce a scalar y_t from multivariate features. This yields substantial computational savings and seamless integration with existing CPD algorithms, while enabling XAI-based explanations for detected changes. Quantitative experiments show competitive detection performance with large speedups, and a case study on industrial image-based quality monitoring demonstrates practical gains and interpretable insights for root-cause analysis. Overall, TimePred offers a scalable, interpretable solution for CPD in industrial and high-volume settings, with clear avenues for online extension and uncertainty quantification.

Abstract

Change-point detection (CPD) in high-dimensional, large-volume time series is challenging for statistical consistency, scalability, and interpretability. We introduce TimePred, a self-supervised framework that reduces multivariate CPD to univariate mean-shift detection by predicting each sample's normalized time index. This enables efficient offline CPD using existing algorithms and supports the integration of XAI attribution methods for feature-level explanations. Our experiments show competitive CPD performance while reducing computational cost by up to two orders of magnitude. In an industrial manufacturing case study, we demonstrate improved detection accuracy and illustrate the practical value of interpretable change-point insights.

TimePred: efficient and interpretable offline change point detection for high volume data -- with application to industrial process monitoring

TL;DR

TimePred reframes high-dimensional offline change-point detection as a univariate mean-shift problem by training a time-prediction network to produce a scalar y_t from multivariate features. This yields substantial computational savings and seamless integration with existing CPD algorithms, while enabling XAI-based explanations for detected changes. Quantitative experiments show competitive detection performance with large speedups, and a case study on industrial image-based quality monitoring demonstrates practical gains and interpretable insights for root-cause analysis. Overall, TimePred offers a scalable, interpretable solution for CPD in industrial and high-volume settings, with clear avenues for online extension and uncertainty quantification.

Abstract

Change-point detection (CPD) in high-dimensional, large-volume time series is challenging for statistical consistency, scalability, and interpretability. We introduce TimePred, a self-supervised framework that reduces multivariate CPD to univariate mean-shift detection by predicting each sample's normalized time index. This enables efficient offline CPD using existing algorithms and supports the integration of XAI attribution methods for feature-level explanations. Our experiments show competitive CPD performance while reducing computational cost by up to two orders of magnitude. In an industrial manufacturing case study, we demonstrate improved detection accuracy and illustrate the practical value of interpretable change-point insights.

Paper Structure

This paper contains 7 sections, 3 figures.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Experiments & Results
  5. Quantitative Experiments
  6. Case Study: Image-based Quality Monitoring
  7. Discussion & Outlook

Figures (3)

  • Figure 1: Overview on our TimePred approach. The time prediction network allows to detect CPs (at $t^*$) on its univariate output and enables XAI integration.
  • Figure 2: A-D: quantitative results per CP problem (mean precision over 100 data sets each, higher is better). The mean-shift problems were detected perfectly across all methods and are therefore not displayed. E: comparison algorithm runtime in seconds (lower is better).
  • Figure 3: A: CPD results on the MVTec case study. Average precision over 15 objects (higher is better). B: examples of (anomalous) objects from after the detected change point (top) and their corresponding XAI heatmap indicating why the sample was associated with the abnormal segment (bottom). Additionally, the anomaly ground truth (GT) is shown as a grey overlay.