SPIRAL: A Novel Time Series to Image (TS2I) Transformation Method for Vision-Based Anomaly Detection
TL;DR
SPIRAL is introduced, a novel TS2I transformation designed for TSAD that maps time series windows onto a two-arm Archimedean spiral, producing asymmetric representations that preserve temporal locality while eliminating redundancy with linear time complexity with respect to window length.
Abstract
Time series anomaly detection (TSAD) underpins critical applications in manufacturing, healthcare, finance, and cloud computing. Recently, time series to image (TS2I) transformations have emerged as a promising approach that enables leveraging pretrained vision models for time series analysis. However, the majority of existing TS2I methods were designed for forecasting and classification, and suffer from parameter sensitivity, symmetric redundancy, and quadratic computational complexity. In this study, we introduce SPIRAL, a novel TS2I transformation designed for TSAD. SPIRAL maps time series windows onto a two-arm Archimedean spiral, producing asymmetric representations that preserve temporal locality while eliminating redundancy with linear time complexity with respect to window length. We additionally propose a standardized workflow with ACF-based window selection and point-wise scoring. Extensive evaluation (24,430 experiments) on 23 datasets from the TSB-AD benchmark, comparing SPIRAL against 9 TS2I transformations, 3 vision backbones, and 32 time-domain baselines across 9 metrics, shows that SPIRAL achieves the best average rank among 102 configurations in challenging metrics such as VUS-PR while exhibiting 40% lower training instability than its closest TS2I competitor. Our code is publicly available at: https://github.com/Smendowski/SPIRAL.