MultiRC: Joint Learning for Time Series Anomaly Prediction and Detection with Multi-scale Reconstructive Contrast

TL;DR

This work proposes MultiRC to integrate reconstructive and contrastive learning for joint learning of anomaly prediction and detection, with multi-scale structure and adaptive dominant period mask to deal with the diverse reaction time.

Abstract

Many methods have been proposed for unsupervised time series anomaly detection. Despite some progress, research on predicting future anomalies is still relatively scarce. Predicting anomalies is particularly challenging due to the diverse reaction time and the lack of labeled data. To address these challenges, we propose MultiRC to integrate reconstructive and contrastive learning for joint learning of anomaly prediction and detection, with multi-scale structure and adaptive dominant period mask to deal with the diverse reaction time. MultiRC also generates negative samples to provide essential training momentum for the anomaly prediction tasks and prevent model degradation. We evaluate seven benchmark datasets from different fields. For both anomaly prediction and detection tasks, MultiRC outperforms existing state-of-the-art methods.

Figures (12)

• Figure 1: (a) Anomaly detection on historical data. (b) Yellow period indicates the precursors where anomalies have not happened yet, and the pink period indicates future anomalies. Anomaly prediction forecasts if anomalies will occur in the future given current data.
• Figure 2: The overall architecture of MultiRC.
• Figure 3: Design specific loss functions for different tasks. ${\mathrm{{\rm\bf x}_{1:t}}}$ represents the historical time series, ${\mathrm{{\rm\bf x}_{t-h:t}}}$ represents the reaction time, and ${\mathrm{{\rm\bf x}_{t+1:t+f}}}$ denotes the size of the look-forward time window. In anomaly detection, significant loss differences indicate anomalies. In anomaly prediction, large fluctuations in loss during the reaction time suggest a high likelihood of future anomalies.
• Figure 4: Performance of different datasets under different window sizes.
• Figure 5: Parameter sensitivity studies of main hyper-parameters in MultiRC.