Research Of Multivariate Time Series Classification Based On Deep Learning

Multivariate time series is a time series with multiple variable values at each time step.Multivariate time series classification,as an important branch of multivariate time series data mining,has great research significance.The current multivariate time series classification methods mainly include feature-based approaches and data-driven approaches.Although the feature-based approach has achieved certain results in many fields,this approach consumes a lot of manpower and is easily limited by domain knowledge.The data-driven approach uses deep learning as the main technology to realize automatic mining of data features,which greatly improves the possibility of classification technology landing.Because multivariate time series data has different characteristics from image data and text data,there are certain limitations in using deep learning techniques in other fields.This thesis focuses on the problems and potential improvements in current multivariate time series classification methods based on deep learning techniques.The research content is top-down,not only the theoretical feasibility analysis is performed,but also the comparative experiments with the existing models on the public domain data set,and some results have been obtained.Aiming at the problem that the existing models still cannot well capture the time-dependent characteristics of multivariate time series,this paper proposes a multivariate time series classification method based on self-attention mechanism.The benchmark model used in this method is mainly composed of a time convolutional network.The self-attention mechanism strengthens the model’s ability to extract correlation features between arbitrary global time points of the data samples based on the benchmark model’s ability to mine variable correlation features in the local window of the original data samples.The model is short and powerful,and has strong robustness and versatility.This method can break the limitation of domain knowledge to a certain extent,and provide a simple and effective solution for non-professional entry research.Aiming at the problem that the network structure of the existing human activity recognition model based on sensor data is too single,this paper finds its potential improvement space and proposes a multivariate time series classification method that combines multiple network structures.Based on the original model using convolutional neural network and recurrent neural network,the method improved it by integrating self-attention mechanism and capsule network structure.The improvement measures make the original model focus on the characteristics of some time steps,make fuller use of the important information output by the hidden layer,and improve the final classification performance.At the same time,the improved model has a certain degree of competitiveness compared with the current best model.