Abstract
PCRTA is a novel self-supervised framework designed to enhance multivariate time series anomaly detection. By training a classifier to recognize specific data perturbations and using the resulting pseudo-labels to guide contrastive learning, the model generates highly discriminative representations. PCRTA addresses common issues in contrastive learning, such as class collapse and the lack of nuance in binary anomaly labels. Evaluation across six real-world datasets shows that PCRTA consistently outperforms state-of-the-art methods, yielding an average improvement of 5.14% in F1-score and 8.24% in AUC-PR.
Introduction
Detecting anomalies in multivariate time series is critical for maintaining the integrity of complex systems, from industrial plants to healthcare monitoring. However, the scarcity of labeled data makes supervised learning impractical. While self-supervised contrastive learning has emerged as a powerful alternative, existing approaches often simplify the task into a binary classification of "normal" vs. "anomaly."
This binary approach suffers from two primary limitations. First, it ignores the specific characteristics of different perturbation types, such as sudden spikes or gradual drifts. Second, it often leads to "class collapse," where samples that should be grouped together are pushed apart in the feature space due to a lack of class-level information. PCRTA addresses these challenges by leveraging the difficulty of identifying specific perturbations as a signal for anomaly detection.
Methodology
The PCRTA framework operates in three distinct stages to build robust representations for anomaly detection.
1. Data Perturbation and Classification
The model defines six perturbation functions—covering point, contextual, and collective anomalies—to create seven "pseudo-classes" (including the original data). A Perturbation Classification Network (PCN) is then trained to identify which transformation was applied to a given segment.
2. Perturbation Guiding Contrastive Learning
Instead of relying on ground-truth labels, PCRTA uses the pseudo-labels generated by the PCN to guide the contrastive learning process. This ensures that features belonging to the same perturbation class are clustered together, while different classes are separated, effectively mitigating class collapse. A key insight here is the "Misclassification Hypothesis": data points that are difficult for the PCN to classify are more likely to be true anomalies, as they already contain inherent irregularities that confuse the classifier.
3. Representation and Scoring
A Temporal Convolutional Network (TCN) serves as the backbone to capture long-term temporal dependencies. To combat the curse of dimensionality, a projection network maps high-dimensional features into a lower-dimensional space for the contrastive task. Finally, anomaly scores are calculated using the Local Outlier Factor (LOF) algorithm on the learned embeddings.
Key Results
PCRTA was evaluated against 13 state-of-the-art competitors on six diverse datasets, including WaQ, DSADS, and SWaT.
As shown in the results table, PCRTA achieved the highest performance on five out of six datasets. Notably, on the Epilepsy dataset, the model achieved a 13% increase in F1-score and a 17% rise in AUC-PR. The use of pseudo-labels proved more effective than ground-truth labels, as they prevent overfitting to synthetic anomalies and focus the model on samples that are genuinely difficult to distinguish.
Conclusion and Future Work
PCRTA demonstrates that guiding contrastive learning with perturbation-based pseudo-labels significantly improves the discriminative power of time series representations. While highly effective, future work could explore more specialized perturbation strategies for niche domains and investigate alternative backbone architectures beyond TCNs to further optimize performance and computational efficiency.
Summary updated