Multiple calibration single classification method for unsupervised time sequence anomaly detection

The invention relates to a multiple calibration single classification method for unsupervised time sequence anomaly detection. Firstly, a self-adaptive reconstruction strategy is provided, a reconstruction target is calibrated by punishing a sample with a high outlier feature, and the learning quality of a normal feature is improved; secondly, in order to solve the problem of abnormal pollution in training samples, uncertainty modeling and time series data enhancement are combined, and prediction with high uncertainty is adaptively punished through modeling paired samples (an original time series and an enhanced time series), so that the negative influence of potential abnormal pollution is eliminated, and learning of normal features is calibrated; and finally, artificial abnormal samples are introduced to calibrate inaccurate and prejudice normal boundaries, and a memory module is further introduced to avoid error generalization of abnormal information. Experimental results on seven real data sets show that