Research On Methods Of Time Series Data For Anomaly Detection And Repair

Time series anomaly detection and repair is an important task in data analysis,and outliers can significantly affect the accuracy of predictive models.Outliers can arise for a variety of reasons,such as measurement errors,data entry errors,or actual abnormal events.The presence of outliers can adversely affect the analysis results,making it necessary to identify and repair them.In this thesis,a hybrid repair approach of HTM-Attention algorithm for time series anomaly detection and time series anomaly repair is proposed,and the main research work is as follows:(1)HTM-Attention based time series anomaly detection algorithmIn order to improve the effectiveness of anomalous data detection,this paper combines the spatial and temporal nature of temporal data and proposes the HTM-Attention algorithm.Given that the original model of HTM is not very efficient in selecting neurons and the structure of neural columns,but only performs simple random selection,and the model does not make good use of the stored information in the cells.This paper therefore proposes an improved model structure to address these problems,adding an attention mechanism to the HTM algorithm to select the activation of neurons in order to improve the accuracy and efficiency of anomalous data detection.The attention mechanism is used to analyse the data and cell storage information at the previous time and to efficiently select the activated cells at the next time.In this paper,we validate the superiority and effectiveness of the HTM-Attention algorithm by conducting experiments on real industrial time series data.(2)Time-series data anomaly repair technique based on hybrid repair method.In time series anomaly repair,this paper classifies univariate time series anomaly processing methods,one category is to find one or more sets of similar subsequences with the location of the anomaly through learning or statistical laws,and then use the most similar subsequence to replace the value of the anomaly;one category is based on machine learning,deep learning or other models to train and fit the data(or normal data)sequences,learn The more the predicted value is close to the normal value,the better the repair effect.In this paper,we combine two different types of anomaly repair processing methods and propose a time series anomaly repair technique based on a hybrid repair method,which achieves better repair results than the baseline model on the public dataset.Aiming at the detection and repair of anomalies in time series data generated during industrial production,the research results of this paper can provide a new solution for the detection and repair of anomalous data in industrial production.