Towards a General Time Series Anomaly Detector with Adaptive Bottlenecks and Dual Adversarial Decoders

TL;DR

The paper tackles the problem of generalizing time series anomaly detection across diverse domains by pre-training a single model on large multi-domain data. It introduces DADA, which combines adaptive bottlenecks (AdaBN) with a bottleneck pool and adaptive router to dynamically match information capacity to data density, and dual adversarial decoders to sharpen the boundary between normal and common anomalies during pre-training. With a mask-based reconstruction framework and complementary masking, DADA learns robust normal patterns while injecting anomaly patterns through adversarial training, enabling strong zero-shot anomaly detection on multiple downstream datasets. Empirical results show DADA achieving competitive or superior performance to dataset-specific detectors, including on NeurIPS-TS, and analyses validate the importance of AdaBN and dual decoders for cross-domain generalization and reliable anomaly scoring. The approach offers practical impact by reducing the need for target-domain training data and enabling scalable, generalizable anomaly detection across varied time series domains.

Abstract

Time series anomaly detection plays a vital role in a wide range of applications. Existing methods require training one specific model for each dataset, which exhibits limited generalization capability across different target datasets, hindering anomaly detection performance in various scenarios with scarce training data. Aiming at this problem, we propose constructing a general time series anomaly detection model, which is pre-trained on extensive multi-domain datasets and can subsequently apply to a multitude of downstream scenarios. The significant divergence of time series data across different domains presents two primary challenges in building such a general model: (1) meeting the diverse requirements of appropriate information bottlenecks tailored to different datasets in one unified model, and (2) enabling distinguishment between multiple normal and abnormal patterns, both are crucial for effective anomaly detection in various target scenarios. To tackle these two challenges, we propose a General time series anomaly Detector with Adaptive Bottlenecks and Dual Adversarial Decoders (DADA), which enables flexible selection of bottlenecks based on different data and explicitly enhances clear differentiation between normal and abnormal series. We conduct extensive experiments on nine target datasets from different domains. After pre-training on multi-domain data, DADA, serving as a zero-shot anomaly detector for these datasets, still achieves competitive or even superior results compared to those models tailored to each specific dataset. The code is made available at https://github.com/decisionintelligence/DADA.

Figures (13)

• Figure 1: General time series anomaly detection (GTSAD) and standard methods training specific detectors for each scenario.
• Figure 2: (a) The workflow during the pre-training stage. DADA mainly consists of Patch and Complementary Mask, Encoder, Adaptive Bottlenecks, and Dual Adversarial Decoders. (b) The structure of Adaptive Bottlenecks. (c) The workflow during the inference stage.
• Figure 3: The forward and backward propagation of normal data and abnormal data.
• Figure 4: Results of fine-tuning DADA under different data percentage.
• Figure 5: Model Analysis. (left) Comparison between DADA using AdaBN and using a single bottleneck. (right) Pre-training baseline model on multi-domain datasets and then evaluate them as a zero-shot detector on the target dataset.