SmartPP: Speeding Up Diffusion-Based Time Series Anomaly Detection via Smart Patch Parallelization

Time series anomaly detection is of great significance in various application scenarios ranging from industrial control to financial risk management. Recently, the diffusion-based deep generative model has become the state-of-the-art for time series anomaly detection due to its powerful generation capability. However, the diffusionbased deep generative model can not be easily deployed on multiple computing nodes for accelerating its inference speed limited by its original model structure. To tackle this challenge, an approach to speeding up the inference stage of the diffusion-based deep generative model is proposed by designing a Smart Patch Parallelization (denoted as SmartPP for short) among distributed diffusion components. Specifically, SmartPP enhances the diffusion-based time series anomaly detection efficiency by jointly designing the data block splitting and periodic state synchronization. To evaluate the proposed SmartPP, comprehensive experiments are conducted on 4 benchmark datasets. According to the experimental results, distributed diffusion-based deep generative model with SmartPP outperforms other baselines by achieving 1.4-1.9 times and 2.6-3.4 times speedup ratio on 2 and 4 computing nodes while maintaining 99 % or even higher time series anomaly detection precision.