Time Series Data Clustering Algorithm Based On Recurrent Neural Network And Its Parallelization

Time series data is the very important data in the real world,it is accumulated over time.As time series data has the trait of dynamic growth,it's often high-dimensional and large-scale in size.With the rapid development of science and finance,the way how to get the time series data is more and more easier.At the same time,this kind of development also drives the development of related fields,especially in the field of time-series data clustering.Because it can implicitly find the regular pattern contained in the time-series data,in order to further study the time-series data,more and more academician pay attention in this field.From the view of the current research methods in the field,it's mainly about how to measure the similarity,and how to find the key components of the sequence.Although these methods can achieve pretty good results on some datasets,these methods can not be used to model time-series properties and are difficult to understand,so their applications are very limited.Recently,deep learning has achieved remarkable results in some applications,including the successful application of recurrent neural networks in sequence learning.However,in the field of time series data mining,recurrent neural network is more frequently in the prediction of time series data,and it hasn't been applied in time-series clustering tasks.Based on the above analysis and the existing problems,this p aper proposes an method that uses the LSTM to learn the implicit expression of time-series data.Because the general expression of the recurrent neural network choose s the last hidden layer state as the learned expression,but only the last layer state can not express the original time-series data,in order to better express time series data,so this paper will use pooling technology to combine all hidden layer states.At the same time,time series data are often associated with local property,so we use a network structure similar to Siamese Network to model the nature.The original time series are randomly sampled to obtain the time segment,then GRU is used to learn the expression of the time segment,and then classification training is carried out.Finally,we input the original time series into the trained GRU network,and extract the hidden state expression.