Multivariate Time Series Analysis Technology And System Implementation

With the improvement of information collection technology,various time series data have emerged,and their information is stored in chronological order,containing valuable information related to time.Multivariate time series analysis is widely used in nature,medicine,transportation,industry and other fields in life,and has very important practical significance.However,multivariate time series data often have complex patterns,such as spatiotemporal correlation,trend change,noise,and anomaly,which results in the unsatisfactory performance of existing algorithms.It is very important to design efficient algorithms for time series data analysis.Therefore,in view of the challenges faced in multivariate time series,this thesis has tried a lot of algorithms and experiments to verify the effectiveness of the model in traffic flow prediction scenarios.After verifying the effectiveness of the model in real traffic datasets,it was migrated to the epilepsy prediction task in medical scenarios and achieved significant performance improvement on this task.The main contributions of this thesis are as follows:1.Aiming at the problem of implicit non-Euclidean spatial structure modeling in multivariate time series of traffic flow prediction scenarios,a spatiotemporal selfattention mechanism based on the self-attention mechanism is proposed,which can not only capture both temporal and spatial context information but also automatically model spatial hidden information.In addition,for special high-frequency timing information,the model has designed additional Patch Embedding modules and window context diverting attention,which not only solves the problem of longterm timing modeling but also improves the timing fragmentation problem caused by sliding windows,After the validity of the model is verified by the real traffic data validation,effective improvement is made in the epilepsy prediction task in the medical scene.2.In view of the nonstationarity and low signal-to-noise ratio of time series,a comparative learning framework based on variational information bottleneck was designed.First,we introduce the variational information bottleneck theory and minimize the mutual information of encoder information input and output to extract more robust features and enhance the generalization ability of the model; Then,by introducing frequency domain information and comparing and learning the tem- poral pattern drift problem in the entangled time domain,we align the consistency of time-frequency information,improving the model effect while also improving the overfitting and generalization issues,On real patient data,our model significantly improves the accuracy of diagnosis and treatment,significantly improving the accuracy of diagnosis and treatment.3.To address the current lack of multiple time series analysis tools based on deep learning,we have packaged functions such as data preprocessing,data alignment,model training,and effect analysis into a Python based toolkit.In addition to being able to call the preprocessing module with one click,the toolkit can also perform various types of transformations on multivariate time series,such as timefrequency conversion,time-graph conversion,and so on.In addition to built-in baseline models for multiple multivariate time series,the toolkit can also define new models through interfaces.