Anomaly Detection in Time Series Data Using Generative Models
| Title | Anomaly Detection in Time Series Data Using Generative Models |
|---|---|
| Summary | Anomaly Detection in Time Series Data Using Generative Models |
| Keywords | Anomaly detection, time series, generative modelProperty "Keywords" has a restricted application area and cannot be used as annotation property by a user. |
| TimeFrame | Fall 2025 |
| References | https://github.com/exathlonbenchmark/exathlon |
| Prerequisites | Deep learning |
| Author | |
| Supervisor | Guojun Liang |
| Level | Master |
| Status | Open |
Research Question:
How effectively can generative models detect anomalies in time series data compared to traditional statistical or supervised methods?
Type: This research is software-oriented, focusing on algorithm design, model training, and data analysis rather than hardware implementation.
Brief Description of 3–4 Related Works:
Variational Autoencoders (VAE) for Time Series: VAEs learn the normal patterns of time series data by reconstructing input sequences. Anomalies are detected when the reconstruction error exceeds a certain threshold.
Diffusion Models for Probabilistic Detection: Diffusion-based generative models capture complex temporal distributions and allow for likelihood-based anomaly detection by modeling the data generation process step by step.
Transformer-based Generative Models: Attention mechanisms in transformer architectures can capture long-term dependencies in time series, improving the ability to model and identify subtle temporal anomalies.
Hybrid Generative–Predictive Models: Combining generative models with forecasting networks (e.g., VAE + LSTM) enables learning both the underlying data distribution and predictive patterns, enhancing anomaly detection robustness.
Expected Outcome: Students will develop a working prototype of a generative-model-based anomaly detection system for time series data. They will:
Gain hands-on experience with deep learning frameworks (e.g., PyTorch or TensorFlow).
Learn how to design, train, and evaluate generative models on real or simulated time series datasets.
Analyze model performance compared to traditional methods.
Produce a short research thesis summarizing the findings, methodology, and potential applications (e.g., fault detection, health monitoring, or finance).
