Multivariate Time Series Anomaly Detection: a Hybrid Method Based on GRU-SAE and GAIN

Anomaly detection in multivariate time series data has increasingly grown in importance as a result of the Industrial Internet's ongoing development. The abnormal operation condition of equipment, in particular, may result in component damage and pipeline obstruction, leakage, and other defects in the context of pump unit fault detection. Based on this, this research suggests a hybrid anomaly detection method for multidimensional time series data based on GAIN and GRU-SAE. The initial step of the method is filling up the missing data as a result of issues like sensor or communication failure using GAIN. Additionally, in contrast to conventional anomaly detection techniques, the method directly treats the reconstruction error produced by SAE as the anomaly score, removing the subjectivity of manually calculating the threshold value and simplifying the model. The proposed method is then used to the pump dataset for the identification of pump faults, and the experiments reveal that it outperforms other methods in terms of Precision, Recall, F1-score, AUC, and other performance evaluation metrics.