Continual Adaptation for Unsupervised Time Series Anomaly Detection
Ting Yu, Bo Ding, Xu Wang · 2024
The application of time series anomaly detection has been widespread across various industrial scenarios. Although current algorithms for time series anomaly detection have shown promising results, they primarily focus on the characteristics of the data distribution within the current training set. This often leads to a lack of adaptability to changes in data distribution while retaining the ability to accommodate new data and preserve existing knowledge. Given the dynamic nature of real-world industrial production systems, the issue of continuous updating of time series anomaly detection algorithms warrants significant attention. However, existing time series anomaly detection algorithms predominantly rely on unsupervised methods, thereby resulting in sparse availability of relevant update strategies. Recognizing this challenge, this paper proposes an unsupervised anomaly detection algorithm update mechanism that incorporates continual learning methods. This approach aims to address the issue in time series anomaly detection algorithms stemming from the absence of label guidance, which leads to a lack of continuous updating capability in models. The method employs a strategy of retaining key information about the past data's feature distribution to ensure memory of past information, thus fulfilling the need for updating and adapting to new data features with minimal storage overhead. Compared to other potential improvement algorithms for continual updating, our approach achieves comparable performance with significantly reduced overhead. In particular, it addresses the issue in the current domain of unsupervised time series anomaly detection, where the lack of labeled data results in a deficiency of continuous updating methods.