AnomalyTK: A Multivariate Time Series Anomaly Detection Method

The detection of anomalies in multivariate time series is a challenging task. Existing methods often face issues such as difficulty in capturing the position information of time series, excessive smoothing, and false positives. In addition, the scarcity of anomalies in real-world datasets adds to the detection challenges. To address these issues, we propose a Transformer-based time series anomaly detection model, called AnomalyTK. In the proposed model, we use a position encoding method to capture position information of the time series, design a new normalization layer to mitigate excessive smoothing, and incorporate a smooth attention mechanism to smooth the attention weights. The model is trained in an unsupervised manner, and anomalies are simulated during data preprocessing to reduce the dependency on labeled data for supervised model training. The experimental results show that, compared with the baseline model, AnomalyTK has the F1-score improved by 10.0% and 9.0% on two public datasets, respectively.