Time Series Classification And Prediction Based On Image Transformation

Time series in different forms widely exist in many fields such as meteorology,medicine,industry and economy.Data mining on time series can obtain a lot of valuable information,which contains enormous economic and social benefits.Therefore,it has attracted a large number of researchers to engage in the research of time series analysis.However,in recent years,with deepening of research and continuous expansion of application fields,the limitations and shortcomings of traditional methods have gradually become prominent.Therefore,it is necessary to use new and effective analysis tools to deeply explore the important information hidden in the time series and then solve challenging problems in complex applications.To this end,this thesis adopts the deep learning techniques to solve the complex problems of EEG signal classification and air quality index prediction from the perspective of time series image transformation.The innovations achieved in this thesis are as follows:1.A classification algorithm of epileptic EEG signals based on gray recurrence plot(RP)and deep convolution neural network(CNN)is proposed.The basic idea of this method is to transform one-dimensional time series into two-dimensional gray RP,and then achieve high-accuracy classification of time series with the help of the powerful automatic feature extraction and classification capabilities of the deep CNN.The algorithm is optimized as follows: Firstly,compared with the traditional binary RP,the gray RP retains more dynamics information of time series.Secondly,the Dense Net network with excellent classification performance is adopted as the backbone network.Finally,the long time series is reasonably segmented.The test results on the public Andrzejak dataset show that the proposed method achieves the highest 100%recognition rate for different groups of EEG signals.2.A future-time air quality prediction algorithm based on meshed index map and CNN network is proposed.The method firstly divides the area into reasonable meshes and get a regional air quality map.On the basis of this,the meehod integrates long,medium and short-term evolution information to predict the future air quality.In the model construction process,by comparing different deep convolution feature-extracted network,the residual network is preferred as the backbone network of the model.Regional meshing is the graphical encoding of the geographic location of air quality monitoring stations from a spatial domain perspective.The meshed index map of regional station at different times is temporal encoding of regional air quality changes from the time domain perspective.Because the algorithm comprehensively considers the spatial-temporal correlation of regional air quality index changes,its prediction accuracy is better than the existing prediction algorithms,especially in the medium and long-term prediction.3.An algorithm for regional fine-grained air quality estimation based on road network and multivariate meteorological information is proposed.The algorithm uses road network data to reflect the impact of local pollution sources on regional air quality.Besides,the use of multivariate meteorological data mainly reflects the spread of air pollutants on different regions.The estimation algorithm in this thesis first uses the LSTM network to integrate multivariate meteorological data.Subsequently,the integrated multivariate meteorological map and the road network map are merged and fed into a densely connected convolutional network to realize regional fine-grained air quality estimation.By estimating the air quality in Beijing,the results show that the proposed algorithm is better than the latest U-Air and ADAIN methods.