Research On Time Series Classification Algorithm Based On Deep Learning

Time series refers to a set of data with a sequential relationship in time,in recent years,with the continuous popularization of smart devices,time series data has shown an explosive growth trend,widely existing in various fields of production and life such as transportation,medical treatment,and industry.Some changes in each thing with the advancement of time can be regarded as time series,time series contains extremely rich value information,through the analysis of the collected time series,useful information can be extracted from the past state of development of things,thus providing guidance for future practice.In the face of today’s massive time series with high dimensionality and noise characteristics,compared with the traditional time series classification algorithm,deep learning is used to solve this problem and achieve better results.Among them,the long-short-term memory network(LSTM)uses the memory cell and gate mechanism to control the transmission of sequence information,and fully extracts the correlation information of the time series;multi-scale convolution extracts the multi-scale features of the sequence;the attention module fuses feature information to obtain features importance and assign attention weights to make the network focus on important temporal features.(1)Through the analysis and study of various methods in the current deep learning,according to the characteristics of univariate time series,the method of first downsampling and then convolution is adopted to reduce the amount of parameters in the multi-scale convolutional extraction features,and then the attention mechanism is used to assign different weight values to different important features,and the other branch uses the LSTM network to extract the long-term sequence dependencies in the time series,and the feature information of the two branches is fused for classification,and the effectiveness of the proposed method is verified by experiments.(2)According to the characteristics of multivariate time series can be regarded as a combination of multiple univariate time series,this thesis first uses the attention mechanism to assign different weights to the sequences of different dimensions,then uses three different volume integral branches to extract multi-scale features,and uses the BiLSTM network to extract long-term dependencies,according to experiments and comparative analysis,finally verify the effectiveness of the proposed model in the classification of multivariate time series.