TShape: Rescuing Machine Learning Models from Complex Shapelet Anomalies

TL;DR

Time series anomaly detection often misses complex shapelet-level anomalies that humans can recognize. TShape addresses this by combining patch-based processing with multi-scale convolution to capture local shape variations and a patch-wise dual-attention mechanism to integrate intra-patch and inter-patch dependencies, gated for robust fusion. Across five benchmarks, TShape achieves a mean F1 of about 0.933 and an Event F1 around 0.817, outperforming 16 strong baselines by roughly 10% and showing robustness in noisy, non-stationary environments. The approach provides a practical, shapelet-aware TSAD solution for industrial IT monitoring, with ablations and visualizations confirming each component's value and code publicly available.

Abstract

Time series anomaly detection (TSAD) is critical for maintaining the reliability of modern IT infrastructures, where complex anomalies frequently arise in highly dynamic environments. In this paper, we present TShape, a novel framework designed to address the challenges in industrial time series anomaly detection. Existing methods often struggle to detect shapelet anomalies that manifest as complex shape deviations, which appear obvious to human experts but prove challenging for machine learning algorithms. TShape introduces a patch-wise dual attention mechanism with multi-scale convolution to model intricate sub-sequence variations by balancing local, fine-grained shape features with global contextual dependencies. Our extensive evaluation on five diverse benchmarks demonstrates that TShape outperforms existing state-of-the-art models, achieving an average 10\% F1 score improvement in anomaly detection. Additionally, ablation studies and attention visualizations confirm the essential contributions of each component, highlighting the robustness and adaptability of TShape to complex shapelet shapes in time series data.

Figures (6)

• Figure 1: Performance evaluation of TShape . (a) Top section: Detection effectiveness of time series anomaly detection methods across datasets (see Table \ref{['table:results']} for quantitative metrics). (b) Bottom section: Illustration of anomaly scoring where the black curve represents a sample time series, the pink background denotes ground-truth anomaly intervals, and red curves indicate time-point anomaly scores generated by each method.
• Figure 2: Overview of TShape .
• Figure 3: Introduction of metrics.
• Figure 4: Effectiveness Evaluation of Multi-scale Convolution
• Figure 5: Effectiveness Evaluation of Patch-wise Dual‑Attention