Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Segmentation over Complexity: Evaluating Ensemble and Hybrid Approaches for Anomaly Detection in Industrial Time Series

Emilio Mastriani, Alessandro Costa, Federico Incardona, Kevin Munari, Sebastiano Spinello

TL;DR

This work tackles anomaly detection in a high-dimensional industrial time series from a steam turbine, where temporal labeling is uncertain and anomalies are imbalanced. It benchmarks a simple segmentation-based Random Forest + XGBoost ensemble against more complex feature engineering (change point statistics) and hybrid architectures (PCA-based and SVM-based combos). The key finding is that the simple RF+XGBoost ensemble on segmented data achieves the best overall performance (AUC-ROC ~ 0.976, F1 ~ 0.41), while advanced features and hybrids often degrade performance due to noise, overfitting, or overlapping biases. The results advocate for model simplicity paired with domain-informed segmentation, offering robustness, interpretability, and practical applicability in industrial anomaly detection, and suggest focusing on segmentation quality rather than architectural complexity in similar settings.

Abstract

In this study, we investigate the effectiveness of advanced feature engineering and hybrid model architectures for anomaly detection in a multivariate industrial time series, focusing on a steam turbine system. We evaluate the impact of change point-derived statistical features, clustering-based substructure representations, and hybrid learning strategies on detection performance. Despite their theoretical appeal, these complex approaches consistently underperformed compared to a simple Random Forest + XGBoost ensemble trained on segmented data. The ensemble achieved an AUC-ROC of 0.976, F1-score of 0.41, and 100% early detection within the defined time window. Our findings highlight that, in scenarios with highly imbalanced and temporally uncertain data, model simplicity combined with optimized segmentation can outperform more sophisticated architectures, offering greater robustness, interpretability, and operational utility.

Segmentation over Complexity: Evaluating Ensemble and Hybrid Approaches for Anomaly Detection in Industrial Time Series

TL;DR

This work tackles anomaly detection in a high-dimensional industrial time series from a steam turbine, where temporal labeling is uncertain and anomalies are imbalanced. It benchmarks a simple segmentation-based Random Forest + XGBoost ensemble against more complex feature engineering (change point statistics) and hybrid architectures (PCA-based and SVM-based combos). The key finding is that the simple RF+XGBoost ensemble on segmented data achieves the best overall performance (AUC-ROC ~ 0.976, F1 ~ 0.41), while advanced features and hybrids often degrade performance due to noise, overfitting, or overlapping biases. The results advocate for model simplicity paired with domain-informed segmentation, offering robustness, interpretability, and practical applicability in industrial anomaly detection, and suggest focusing on segmentation quality rather than architectural complexity in similar settings.

Abstract

In this study, we investigate the effectiveness of advanced feature engineering and hybrid model architectures for anomaly detection in a multivariate industrial time series, focusing on a steam turbine system. We evaluate the impact of change point-derived statistical features, clustering-based substructure representations, and hybrid learning strategies on detection performance. Despite their theoretical appeal, these complex approaches consistently underperformed compared to a simple Random Forest + XGBoost ensemble trained on segmented data. The ensemble achieved an AUC-ROC of 0.976, F1-score of 0.41, and 100% early detection within the defined time window. Our findings highlight that, in scenarios with highly imbalanced and temporally uncertain data, model simplicity combined with optimized segmentation can outperform more sophisticated architectures, offering greater robustness, interpretability, and operational utility.

Paper Structure

This paper contains 16 sections, 1 equation, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. The Allure and Peril of Complexity in ML Research
  3. Our Established Baseline and Research Challenge
  4. Study Context and Dataset
  5. Methods
  6. Phase 1: Change Point Statistics Feature Engineering
  7. Phase 2: Advanced Clustering
  8. Comparative Analysis of Density-Based Segmentation: OPTICS vs. HDBSCAN
  9. Phase 3: Feature Relevance Analysis: Random Forest and Segment-Level Permutation Importance
  10. Phase 4: Hybrid Model Architectures
  11. Experimental Results
  12. The Unbeatable Baseline
  13. The Complexity Penalty
  14. The Trade-off Analysis
  15. Conclusion
  16. ...and 1 more sections

Figures (4)

  • Figure 1: Violin plots of feature groups (dist_last_cp, mean_score_pre_cp, std_score_pre_cp, max_score_pre_cp, cp_freq) comparing Normal and anomalous samples, showing distribution differences and class-separating patterns.
  • Figure 2: Radar plot of normalized clustering metrics (Silhouette, Calinski–Harabasz, and Davies–Bouldin) for all evaluated algorithms (KMeans, BIRCH, GMM, OPTICS, MeanShift, and HDBSCAN). The plot provides a visual comparison of each algorithm’s overall performance, with larger enclosed areas indicating superior clustering quality across the combined criteria.
  • Figure 3: Top 10 segment-wise feature importance values from permutation analysis. The feature pv_dist_last_cp in segment COVA.ABB.V470PT001.pv emerges as the most informative, confirming the relevance of proximity-based and pre-change point metrics in model discrimination.
  • Figure 4: Normalized feature importance by category. Comparison of global Random Forest importance and segment-level permutation importance highlights key contributions of segmented variables, raw process variables, derived indicators, and system efficiency.