MTAD RF: Multivariate Time-series Anomaly Detection based on Reconstruction and Forecast

Anomaly detection in multivariate time series is an important research direction, which helps to improve the security of industrial systems by detecting abnormally unreliable devices. Multivariate (MTS) anomalies not only need pay attention correlation between different but also consider abnormal changes relationship variables. Once influence two variables that each other ignored, it will likely lead false positives or positives. At same time, degree features inconsistent, just like what happened recently have radically influences on present. Furthermore, most existing models are weak no abnormality. To tackle these issues, this paper, we propose a new model anomaly based reconstruction and forecast, named MTAD RF. First, capture temporal feature correlations MTS through parallel GAT layers, at distinguish coefficients. Second, leverage generative power VAE single-step forecast MLP jointly detect known unknown reconstructed predicted models. Major practical implications proposed approach missing. Finally, detected explained scores. Experiments demonstrate our outperforms current state-of-the-art methods 4 real-world datasets, with average F1 score about 95% excellent diagnostic ability.