Multivariate time-series anomaly detection
Qifa Wang, Qiwei Shen · 2023
Anomalies are rare items that differ significantly from the majority of the data and raise suspicion. Time series anomaly detection is of great significance in industrial applications, data mining and other fields. In this paper, we propose an autoencoder-based anomaly detection model. In the encoder part, the dependency relationship between different time series is mined through the graph attention network, the time domain data features are extracted through the GRU, and the frequency domain features are extracted through the convolutional neural network. The temporal data is reconstructed using a decoder consisting of a recurrent neural network in the encoder part. The residual between the reconstructed data and the original data is used to further judge anomalies. Data cleaning is performed in the preprocessing part to improve model performance. The effectiveness of our method is demonstrated on three publicly available datasets, and our method is found to outperform four other common anomaly detection methods.