Mixed Model of Multivariate Timing Anomaly Detection Integrating Prediction and Reconstruction

In this study, a joint training strategy, combining LSTM, VAE reconstruction model and BiGRU-based prediction model, was used to adjust the loss function through the adaptive balance training task. The reconstruction unit combines the data reconstruction function of VAE and the LSTM's deep understanding of the sequence length relationship, which can generate the reconstruction results highly consistent with the original data. The prediction unit relies on BiGRU to accurately simulate the forward and backward time series relationship and effectively predict the future time data. By setting the data anomaly threshold, the data is compared with the results of the reconstructed model and the prediction model to achieve an accurate judgment of the anomalous fluctuations. To verify the applicability and efficiency of the model, this study was tested on two publicly available industrial data sets and cluster production data. The results show that the model improves at least 3.04^%, 3.53% and 3.28% in accuracy, recall and F1 value, which not only highlights the efficiency of the model in real industrial scenarios, but also verifies the rationality and efficiency of its component design.