Robust Group Anomaly Detection for Quasi-Periodic Network Time Series

Kai Yang, Shaoyu Dou, Pan Luo, Xin Wang, H. Vincent PoorIEEE Trans. Network Science & Engineering 2022 · DOI · OpenReview

Abstract

Many real-world multivariate time series are collected from a network of physical objects embedded with software, electronics, and sensors. The quasi-periodic signals generated by these objects often follow a similar repetitive and periodic pattern, but have variations in the period, and come in different lengths caused by timing (synchronization) errors. Given a multitude of such quasi-periodic time series, can we build machine learning models to identify those time series that behave differently from the majority of the observations? In addition, can the models help human experts to understand how the decision was made? We propose a sequence to Gaussian Mixture Model (seq2GMM) framework. The overarching goal of this framework is to identify unusual and interesting time series within a network time series database. We further develop a surrogate-based optimization algorithm that can efficiently train the seq2GMM model. Seq2GMM exhibits strong empirical performance on a plurality of public benchmark datasets, outperforming state-of-the-art anomaly detection techniques by a significant margin. We also theoretically analyze the convergence property of the proposed training algorithm and provide numerical results to substantiate our theoretical claims.

Introduction: The Challenge of Quasi-Periodic Signals

In domains ranging from healthcare (ECG monitoring) to industrial IoT, sensors frequently capture signals that are "quasi-periodic." These signals repeat a general pattern but lack the rigid regularity of a perfect sine wave. A major hurdle in analyzing these signals is the presence of timing and synchronization errors.

Illustration of Type-1 (jitter) and Type-2 (deletion) timing errors in quasi-periodic signals.

As shown in the figure above, these errors manifest in two primary ways:

  1. Type-1 (Clock Jitter): Instantaneous time shifts where the signal starts slightly earlier or later than expected.
  2. Type-2 (Signal Deletion): Missing data points due to unreliable sampling or hardware glitches.

Traditional anomaly detection methods often struggle with these variations, frequently misidentifying normal temporal shifts as anomalies. Furthermore, many existing deep learning models act as "black boxes," providing a global anomaly score without explaining which specific part of a signal is problematic.

The seq2GMM Framework

To address these challenges, the authors propose seq2GMM, a deep learning framework that integrates temporal segmentation with a sequence-to-GMM architecture. The framework is designed to be robust against timing errors while providing local interpretability.

The seq2GMM architecture: TS Segmentation, Temporal Compression (Encoder-Decoder), and GMM Estimation.

The architecture consists of three core modules:

  1. Temporal Segmentation: The model uses piecewise linear regression to divide the time series into semantically meaningful segments. For example, in an ECG signal, it might isolate the P, QRS, and T waves. This allows the model to handle varying lengths and focus on the dynamics of specific subsections.
  2. Temporal Compression Network: An encoder-decoder architecture (utilizing Gated Recurrent Units and an Attention mechanism) maps these segments into a low-dimensional latent space. This process compresses the "essence" of the signal while filtering out noise.
  3. GMM Estimation Network: In the latent space, the model fits a Gaussian Mixture Model (GMM). Normal segments cluster into high-density regions, while anomalies fall into low-density "outlier" regions.

To ensure stability during training, the authors developed a surrogate-based optimization algorithm that alternates between clustering in the latent space and refining the neural network parameters.

Key Insights: Interpretability and Robustness

A standout feature of seq2GMM is its ability to identify "anomaly shapelets." By assigning an anomaly score to each segment rather than just the whole sequence, the model provides local interpretability. For instance, a clinician could see exactly which part of a heartbeat triggered an alarm.

Furthermore, the framework is intentionally designed for robustness. During training, the authors employ data augmentation—randomly deleting 5-10% of points—to force the model to learn features invariant to Type-2 timing errors. Unlike standard autoencoders that rely solely on reconstruction error, seq2GMM learns the underlying probability distribution of normal behavior in the latent space, making it far more resilient to complex, shifting signals.

Experimental Results

The model was evaluated against ten baselines (including LOF, DAGMM, and BeatGAN) across diverse datasets such as the UCR Time Series Archive and the CMU Motion Capture Database.

  • State-of-the-Art Performance: Seq2GMM achieved the highest average rank across all tested datasets.
  • ECG Accuracy: On the TwoLeadECG dataset, it provided a 6% performance gain in AUC over the next best algorithm.
  • Resilience: While competitors' performance degraded significantly as signal lengths varied, seq2GMM maintained high accuracy even with 10% data deletion.

Conclusion

Seq2GMM represents a significant step forward in robust and interpretable anomaly detection for network time series. By combining segmentation with latent space density estimation, it effectively navigates the noise and timing errors inherent in real-world quasi-periodic data. Future work may extend this framework to handle spatially correlated sensors in even larger multivariate networks.

Summary updated

Abstract

Many real-world multivariate time series are collected from a network of physical objects embedded with software, electronics, and sensors. The quasi-periodic signals generated by these objects often follow a similar repetitive and periodic pattern, but have variations in the period, and come in different lengths caused by timing (synchronization) errors. Given a multitude of such quasi-periodic time series, can we build machine learning models to identify those time series that behave differently from the majority of the observations? In addition, can the models help human experts to understand how the decision was made? We propose a sequence to Gaussian Mixture Model (seq2GMM) framework. The overarching goal of this framework is to identify unusual and interesting time series within a network time series database. We further develop a surrogate-based optimization algorithm that can efficiently train the seq2GMM model. Seq2GMM exhibits strong empirical performance on a plurality of public benchmark datasets, outperforming state-of-the-art anomaly detection techniques by a significant margin. We also theoretically analyze the convergence property of the proposed training algorithm and provide numerical results to substantiate our theoretical claims.
BibTeX
@misc{yang2022robust,
  author = {Kai Yang and Shaoyu Dou and Pan Luo and Xin Wang and H. Vincent Poor},
  title = {{Robust Group Anomaly Detection for Quasi-Periodic Network Time Series}},
  year = {2022},
  eprint = {2506.16815},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2506.16815},
}
Figure