Time Series Classification Based On Multi-scale Full Convolution Networks And Cross Clustering Algorithm

Time series data is a common form of data,are widespread in every field of daily life,work is closely linked with our life.It is of great significance to study the time series deeply by discovering some potential rules of life and applying them to business activities.The definition of time series covers a wide range,including not only the sequences that change with time,but also the sequences with certain logical relations.In Time Series Classification problem(Time Series Classification,TSC),recognized as a baseline model for the nearest neighbor algorithm based on dynamic Time neat.In the past two years,the time series classification model based on full convolutional network has been outstanding and obviously better than other methods.Therefore,this paper mainly focuses on the time series classification problem and the full convolution method to improve and optimize the model,and puts forward the multi-scale full convolution network,cross clustering data transformation,and piecewise based time series prediction method.The main contributions of this paper are as follows:1)Because the characteristics of time series classification work scale is not fixed,therefore a single full convolution network will not be able to effectively extract the characteristics of the different scales.Aiming at this problem,this paper presents a one dimensional network MFCN full convolution,makes the convolution sequence characteristics of FCN can detect a variety of network scale,raise the classification accuracy of the model.Compared to recent years better ten time series classification model,UCR data set in the 44 group,this paper puts forward the MFCN model classification performance is best.2)The crossover clustering algorithm is proposed,which is different from the traditional clustering algorithm in that it considers that an instance can belong to multiple clusters and allows for the intersection between clusters.Different from soft clustering,cross clustering does not need to calculate membership degree,but directly cluster with similar distance between instances.An important application of this clustering algorithm is data transformation.The original data is transformed by extracting the center point of the cluster to represent the cluster.Compared with other clustering algorithms,the crossover clustering algorithm can reduce the compression ratio of data.3)The data characteristics of time series after cross clustering data transformation are more obvious,which can improve the performance of time series classification model.Experiments show that,without changing the data dimension,the transformation has a significant effect on the classification of time series.First,it improves the classification accuracy of the time series classification model;second,it reduces the time complexity of training;third,it weakens the overfitting effect on the classification of neural network FCN.4)Time series segmentation feature extraction method is proposed,respectively to the FCN and LSTM-FCN network was improved,joined the segmentation mechanism,the Sub-FCN model and Sub-FCN-LSTM model.Through experiments on multiple time series data sets in UCR database,the results show that sub-fcn and sub-fcn-lstm have better fitting ability and better generalization accuracy than FCN and lstm-fcn respectively,and can learn data in samples faster.In conclusion,this paper mainly studied the time series classification task,puts forward the effective data transform method of time sequence,can improve supervision and performance of the model.In order to avoid the translation invariance of convolution structure and to ensure that the time series of local information validity,segmented full of convolution parallel LSTM model is put forward.A multi-scale full convolutional network with enhanced feature extraction capability is proposed.