Variational transformer-based anomaly detection approach for multivariate time series

•Take the lead in combining transformer and VAE for anomaly detection in telemetry data.•Use the improved positional encoding to help transformer capture long-range dependencies.•Apply the proposed multi-scale feature fusion to obtain a more robust feature expression.•A residual structure is proposed, which can alleviate the KL divergence vanishing problem. Due to the strategic importance of satellites, the safety and reliability of satellites have become more important. Sensors that monitor satellites generate lots of multivariate time series, and the abnormal patterns in the multivariate time series may imply malfunctions. The existing anomaly detection methods for multivariate time series have poor effects when processing the data with few dimensions or sparse relationships between sequences. This paper proposes an unsupervised anomaly detection model based on the variational Transformer to solve the above problems. The model uses the Transformer's self-attention mechanism to capture the potential correlations between sequences and capture the multi-scale temporal information through the improved positional encoding and up-sampling algorithm. Then, the model comprehensively considers the extracted features through the residual variational autoencoder to perform effective anomaly detection. Experimental results on a real dataset and two public datasets show that the proposed method is superior to the mainstream and state-of-the-art methods.