TLDR: CAPMix is a novel framework for time series anomaly detection that tackles the issues of ‘patchy generation’ and ‘anomaly shift’ in synthetic anomaly injection. It introduces a CutAddPaste mechanism to create diverse anomalies, a DTW-based label revision to refine anomaly labels, and a dual-space mixup within a Temporal Convolutional Network for robust decision boundaries. Extensive experiments on five benchmark datasets demonstrate that CAPMix significantly outperforms state-of-the-art methods, showing enhanced accuracy and robustness even with contaminated training data.
Time series anomaly detection (TSAD) is a critical task across many industries, from monitoring industrial sensors to tracking patient vital signs and network events. Identifying unusual patterns or points that deviate from the norm can prevent failures, ensure safety, and optimize performance. However, this task is particularly challenging because real-world anomalies are often rare and difficult to label, and the temporal dependencies within time series data are complex.
Recent approaches have tried to overcome the scarcity of labeled anomalies by injecting synthetic anomalous samples into the training data, a concept known as anomaly assumption (AA). While promising, these methods often face two significant limitations: ‘patchy generation’ and ‘anomaly shift’. Patchy generation occurs when the synthetic anomalies are too simplistic or incoherent, failing to represent the true diversity of real anomalies. Anomaly shift, on the other hand, happens when synthetic anomalies either resemble normal data too closely or diverge unrealistically from actual anomalies, leading to distorted classification boundaries and unreliable detection.
Introducing CAPMix: A Robust Framework for Time Series Anomaly Detection
A new research paper titled CAPMIX: Robust Time Series Anomaly Detection Based on Abnormal Assumptions with Dual-Space Mixup, authored by Xudong Mou, Rui Wang, Tiejun Wang, Renyu Yang, Shiru Chen, Jie Sun, Tianyu Wo, and Xudong Liu, introduces CAPMix, a controllable anomaly augmentation framework designed to address these fundamental issues. CAPMix aims to inject diverse and complex anomalies in a targeted manner, adaptively refine anomaly labels, and enforce smoother, more robust decision boundaries.
How CAPMix Works: Three Core Mechanisms
CAPMix integrates three key mechanisms to achieve its robust anomaly detection capabilities:
1. Anomaly Injection via CutAddPaste: This mechanism extends existing image-based augmentation techniques to time series data. It generates five types of complex anomalies—shape, correlation, seasonality, trend, and point-wise anomalies—by cutting a patch from one time series, adding incremental or decremental trend values, and pasting it into a random position of another time series. This targeted approach ensures that the synthetic anomalies are diverse and complex, avoiding the problem of patchy generation.
2. Label Revision via DTW: To combat anomaly shift, CAPMix introduces a label revision strategy. It calculates the Dynamic Time Warping (DTW) distance of each generated synthetic anomaly to the ‘normality center’ of the training data. Based on this distance, it assigns a ‘soft label’ (a value between 0 and 1) instead of a rigid ‘1’ for anomalous. This prevents samples that are too similar to normal data from being incorrectly labeled as confidently anomalous, thereby reducing mislabeling and improving the model’s ability to learn meaningful decision boundaries.
3. Dual-Space Mixup Integrated TCN: CAPMix employs a Temporal Convolutional Network (TCN) combined with a dual-space mixup strategy. This mechanism regularizes the model by applying mixup operations in both the input data space and the latent feature space. This helps prevent generated anomalies from becoming unrealistically abnormal, ensuring they remain plausible within the anomaly distribution. By interpolating and transforming features across these spaces, the model learns smoother and more robust decision boundaries, further mitigating anomaly shift.
Experimental Validation and Robustness
The researchers conducted extensive experiments on five benchmark datasets: AIOps, UCR, SWaT, WADI, and ESA. These datasets cover both univariate (single variable) and multivariate (multiple variables) time series, representing various anomaly types and complexities. CAPMix consistently demonstrated significant improvements over state-of-the-art baselines, achieving enhanced robustness, particularly against contaminated training data where real anomalies are present in the training set.
The results highlight CAPMix’s ability to balance the detection of both point-wise (sudden spikes) and pattern-wise (changes in trends or seasonality) anomalies. Its performance on multivariate datasets, which are notoriously challenging due to complex inter-variable correlations, was particularly strong, surpassing other methods by a considerable margin. Visualization of UMAP embeddings further confirmed that CAPMix generates pseudo-anomalies that align much better with real anomalies, avoiding overlap with normal data and preventing unrealistic deviations.
Also Read:
- A New Self-Supervised Approach for Network Intrusion Detection Redefines Anomaly Learning
- Deepfake Audio Detection Under Siege: A Look at Adversarial Attacks
Conclusion
CAPMix represents a significant advancement in time series anomaly detection. By systematically addressing the challenges of patchy generation and anomaly shift through its innovative CutAddPaste, label revision, and dual-space mixup mechanisms, it offers a robust and unified framework for more reliable anomaly detection. This work provides valuable insights into the limitations of current anomaly injection methods and paves the way for more flexible and distribution-aware augmentation strategies in the future.
