Research On Capsule LSTM Feature Extraction Algorithm For Time Series Data

In this world,countless events are observed and recorded by humans every moment.Time series data is a collection of observations arranged in the order of observation time and widely exists in all walks of life in real life.The analysis and research of time series data not only exists in many mainstream disciplines,such as meteorology,life science,economics,and agronomy,but also is relied by other emerging fields such as Natural Language Processing,traffic management,human posture detection analysis.The extraction of eigenvalues of time series data has become one of the hotspots and difficulties in research.However,unlike static airspace data,time series data often have features such as strong temporal correlation,large data volume,and high data dimensionality.Therefore,it is of great significance to study how to extract effective features from a large number of time series data and thus obtain the knowledge and information contained in the data.In recent years,deep learning technology has performed excellently in practical applications,and significant progress has been made in research on feature extraction algorithms.As a variant of the circulatory neural network in deep learning,the long short term memory(LSTM)has long-term memory ability for time series.However,the features extracted from the traditional LSTM network structure are transmitted at various moments in scalar form,and the correlation between features is lost to a certain extent during feature extraction.To solve this problem,this paper combines the capsule network with the LSTM network and proposes a capsule LSTM network structure to extraction feature.The network transfers the feature information in the form of vectors,thus retaining the correlation between the features and applying them in the human body posture detection system,improving the accuracy of human posture detection tasks.The main work and achievements of this paper are as follows:(1)A CapsLSTM network model structure is proposed for feature extraction of time series data by combining capsule network and LSTM network.Multiple scalar features are constructed by constructing multiple LSTM blocks,then convert them into vector features by feature routing.While applying the traditional time back propagation algorithm,dynamic routing is added to update the extracted features,refining more essential feature information from the input time series data.At the same time,considering that the characteristics extracted from the data should be able to restore the idea of the data itself,this paper applied a reconstruction network to the capsule LSTM network model,and calculated the error between the reconstructed data and the input data by reducing the reconstruction network.(2)The performance of the CapsLSTM network model was verified by the language modeling experiment based on the open dataset named PTB(Penn Treebank Dataset).The feature extraction abil ity of the CapsLSTM network model is measured by the performance of the text sequence prediction.Compared with the classic deep learning algorithm,the CapsLSTM network model has advantages in the prediction of the time series data.(3)Through the human posture detection verification experiment,the performance of the CapsLSTM-based feature extraction algorithm in multivariate time series data was verified.Through self-built human pose data set,the feature extraction ability of CapsLSTM network model was explored.By comparing with the current mainstream machine learning algorithms,the advantage of CapsLSTM in the classification of time series was verified.