A New Perspective on Time Series Anomaly Detection: Faster Patch-based Broad Learning System

TL;DR

This work tackles time-series anomaly detection by challenging the notion that deep learning is strictly necessary for strong performance. It introduces CPatchBLS, a patch-based Broad Learning System that learns via pseudo-inverse and integrates a dual-branch contrastive framework (PatchBLS vs SKP-PatchBLS) with multi-scale patch ensembles to capture temporal semantics efficiently. Across five real-world datasets, CPatchBLS achieves superior anomaly detection accuracy with favorable time costs compared to twelve baselines, highlighting a compelling balance between robustness, speed, and scalability. The approach offers a practical option for industry where fast, reliable TSAD is critical, without sacrificing performance relative to complex deep learning models.

Abstract

Time series anomaly detection (TSAD) has been a research hotspot in both academia and industry in recent years. Deep learning methods have become the mainstream research direction due to their excellent performance. However, new viewpoints have emerged in recent TSAD research. Deep learning is not required for TSAD due to limitations such as slow deep learning speed. The Broad Learning System (BLS) is a shallow network framework that benefits from its ease of optimization and speed. It has been shown to outperform machine learning approaches while remaining competitive with deep learning. Based on the current situation of TSAD, we propose the Contrastive Patch-based Broad Learning System (CPatchBLS). This is a new exploration of patching technique and BLS, providing a new perspective for TSAD. We construct Dual-PatchBLS as a base through patching and Simple Kernel Perturbation (SKP) and utilize contrastive learning to capture the differences between normal and abnormal data under different representations. To compensate for the temporal semantic loss caused by various patching, we propose CPatchBLS with model level integration, which takes advantage of BLS's fast feature to build model-level integration and improve model detection. Using five real-world series anomaly detection datasets, we confirmed the method's efficacy, outperforming previous deep learning and machine learning methods while retaining a high level of computing efficiency.

Figures (8)

• Figure 1: By analyzing key metrics like ROC-AUC, PR-AUC, and PA-F1 for twelve state-of-the-art models and our methods on five datasets, the figure shows an association between training time and effectiveness. Models closer to the top-left in the heatmap (indicated by darker red shades) achieve a superior trade-off between time efficiency and performance, whereas those near the bottom-right (darker blue shades) indicate suboptimal performance or higher time consumption. Our methods, represented by red pentagonal stars (Ours-PE, parallel execution) and blue pentagonal stars (Ours-SE, serial execution), are positioned closest to the top-left red corner among the fourteen approaches. This demonstrates clear advantages of ours in both time efficiency and performance, highlighting the competitiveness against advanced baselines.
• Figure 2: Overview of CPatchBLS. First, the Patching Module segments multivariate data into univariate time series. Second, SKP-PatchBLS alongside Basic-PatchBLS are developed for dual-branch analysis. This method compares different representations of the same time series data. Third, CPatchBLS with varying patch sizes effectively utilizes temporal information and achieves ensemble learning at the model level.
• Figure 3: The overview of PatchBLS.
• Figure 4: Comparison of average time consumption of machine learning methods (ML-AVG.) and deep learning methods (DL-AVG.) with our proposed method reveals significant performance advantages.
• Figure 5: Average Training and Testing Time Comparison Results on MSL, SMAP, SWAT, WADI, and PSM Datasets.