PRISM: New Method Maps Multivariate Time Series to Images for Anomaly Detection
A new meta-workflow named PRISM has been developed by researchers for the creation and assessment of image-based representations of multivariate time series aimed at detecting anomalies. This approach is outlined in a paper available on arXiv (2608.03926) and tackles the issue of converting high-dimensional series into multi-channel images, while evaluating the performance of vision backbones against time-domain benchmarks. In more than 7,000 tests, PRISM configurations demonstrated competitiveness with 24 time-domain baselines, securing the highest VUS-PR on 10 out of 14 datasets, and exhibiting an average enhancement of 41% compared to the leading competing method on those datasets. Additionally, the paper highlights channelization as a crucial element. This research holds significance for fields such as predictive maintenance, finance, and cloud computing.
Key facts
- PRISM is a plug-and-play meta-workflow for multivariate time series anomaly detection.
- It maps time series to images and evaluates vision backbones.
- Over 7,000 experiments were conducted.
- PRISM was competitive with 24 time-domain baselines.
- It achieved the best VUS-PR on 10 of 14 datasets.
- Average improvement of 41% over the best competing method on those datasets.
- Channelization is identified as a key factor.
- Applications include predictive maintenance, finance, and cloud computing.
Entities
Institutions
- arXiv