Research On The Multi-granularity Analysis Method Of Time Sequence Signal Using Convolutional Long Short Term Memory Network

The timing signal is a very important signal,which is the multi component signal whose frequency and amplitude change with time.Typical timing signals include speech signals,bioelectricity signals,radar and sonar signals,mechanical vibration and seismic signals[1],and so on.Timing signals are nonlinear and non-stationary characteristics,most studies are based on the current signal is the short-time stationarity hypothesis,feature extraction in frequency domain characteristics,level and particle size analysis is relatively single.Moreover,most of the most important information in the signal is ignored,which greatly affects the ability to extract the time-varying signals and limits their performance in practical applications.This paper aims at extracting and modeling of temporal information in r the timing signal,automatic optimization and integration capability of multi period,multi granularity and multi level features of human cognitive process,is proposed to extract the feature fusion method of multi granularity and framework,we will extract feature from the signal in three levels,frame,segment and the global.This preserves the global features of the existing methods commonly used,and increase the frame size and size of the two section contains timing information signal in dynamic characteristics,effectively to timing signals from multiple angles were extracted,the ability to express information in the signal is more abundant.In the segmentation of segment size,we refer to the rules of the human brain in cognitive activity to classify the window length.After that,we fuse the three granularity features at the frame level and fuse them according to the order of time.Then we use the LSTM neural network model which has strong temporal information modeling ability to classify.In the concrete implementation of multi granularity features,we employ two methods.One is the use of traditional time-frequency analysis methods to extract features of the frame timing signal,using the Gauss function group in the window of the frame segment granularity characteristics were calculated by convolution feature,global feature is obtained by the statistic of all computing frame feature.Another method is a combination of deep learning technology which has a breakthrough in various fields at present,with the help of convolutional neural network can extract the ability of end to end information in the original data,and the characteristics of multiple levels of feature extraction,feature extraction to complete the multi granularity of timing signal.We will be the original signal timing analysis of direct input to the convolutional network,through convolution preset sliding convolution in the signal,obtaining the frame size characteristics in the shallow layer of CNN,and continue to use higher CNN for further processing of frame size characteristics,respectively output size and granularity in the global feature the middle and senior CNN.Finally,three granularities of feature information are integrated at the frame level according to time series,and multi granularity fusion features are obtained.This is our proposed network structure of C-LSTM,and achieves the extraction and analysis of multi granularity features of end-to-end timing signals.Finally,the proposed method,framework and network structure model are used to classify speech emotion recognition problems on speech signals,and to classify and recognize moving imaginary signals on EEG signals.In the speech emotion classification problem,we adopted the Chinese Academy of Sciences Institute of automation announced the more than 2016 modal emotion recognition in the competition data set included anxiety,anger,disgust,joy,sadness,surprise,fear and neutral of the eight emotional categories,compared with the base line system of data set,the recognition rate is improved by more than 4% and,more than the first prize of the contest by a method.In the classification of brain motor imagery,we use the BCI2008 dataset,which is the two category of left and right hand motor imagery.We according to the characteristics of multi channel EEG,with spatial distribution characteristics,on the basis of C-LSTM was improved,the spatial information of the electrodes by methods of data integration and wavelet transform fusion of brain networks in which the establishment of the 3D-C-LSTM model,and compared with other methods in the recognition rate increased by nearly 10%,to 92%,that in the EEG in addition to temporal information,spatial information is also very important.Provide effective solutions of this research work for some key technical problems existing in the analysis field in timing signal,as proved by the experiments of speech signal and EEG signal,the C-LSTM network structure is universal for processing timing signal,has a certain promotion value.It also provides new ideas and directions for the application and development of deep learning methods such as convolutional neural networks in time series signal processing.