Classification of Time Series Data based on Capsule Network

A capsule network was applied to the classification of time series data, which can efficiently develop an organization result of time series.A time series classification way on the capsule network is suggested to solve problem that deep learning cannot effectively deal with the relationship between features.Aiming at the situation that the position relationship of time series data is not obvious, GramianAngularField algorithm was used to convert the sequence data into images.The experiment was mainly based on python 3.7.1 and keras 2.2.4.Through experiments, the superparameter settings involved in this paper are as follows: the learning rate of Adam optimizer were set to 0.05, (1,2) were set to (0.9,0.99), a number of training epoch were set to 250, a batch size of training data were 128, and the quantity of routing were 3.According to the investigation of literature in this paper, there is no time series arrangement model on a capsule network.The dynamic routing algorithm continuously updated the settings in the capsule network, and then a data sets of 50 ords, Adiac, beef, coffee, ECG200, Faceall, lighting2, lighting7, Oliveoi, OSULeaf, SwedishLeaf and yoga were substituted into the model, and high classification accuracy is obtained.