Multivariate Time Series Anomaly Detection via Dual-Branch Reconstruction and Autoregressive Flow-based Residual Density Estimation

The project is a fresh research implementation (19 days old) targeting a specific failure mode in Multivariate Time Series Anomaly Detection (MTSAD): the tendency of reconstruction-based models to overfit spurious correlations and the inaccuracy of simple MSE-based anomaly scores. By introducing a dual-branch architecture (separating variable and temporal modeling) and using Autoregressive Flows for residual density estimation, it offers a more statistically rigorous way to define anomalies than traditional thresholding. However, with 0 stars and 5 forks, it currently exists as a 'paper code' repository with no ecosystem, community, or commercial tooling. It competes against established benchmarks like OmniAnomaly, AnomalyTransformer, and DeepSVDD. The primary risk comes from the rapid rise of Time Series Foundation Models (e.g., Google's TimesFM, Amazon's Chronos), which aim to solve these tasks zero-shot or via fine-tuning, potentially making specialized architectures like this one obsolete within 1-2 years if foundation models reach sufficient granularity for industrial sensor data.