Multivariable time series data anomaly detection method and system, and multivariable time series data model training method and system

The invention relates to the technical field of data detection, in particular to a multivariable time sequence data anomaly detection and model training method and system, and solves the problems of large noise influence, non-ideal detection effect and incapability of realizing real-time detection during multivariable time sequence abnormal data detection in various technologies in the prior art. According to the multivariable time series data anomaly detection model training method provided by the invention, a self-adaptive weight and filtering module for eliminating noise influence is additionally added, and a comparative learning method is adopted to learn data features, so that the generalization ability of the model is improved. According to the method, reconstruction errors are emphatically considered in the anomaly detection stage, an evaluation function used for evaluating the data anomaly degree is designed, the trained anomaly detection model has a better F1 score, and the robustness of the model is