Research On Time Series Feature Representation Method Based On Deep Neural Network

In the past few decades,the research on automatic production system of medical application,biotechnology and manufacturing technology has made great progress.In order to produce high-quality automatic production system,it is necessary to identify and diagnose the behavior of the system.One feasible method is time series method.Time series data is a group of real variables obtained in time sequence,which can help us to obtain meaningful and unknown knowledge in the system.The research on time series tasks is of great significance in the fields of telecommunications,medicine,web,motion capture and sensor networks.However,due to the high dimension,complexity and multiple outliers of time series,it is necessary to represent the features of time series.The feature representation of time series is a method to reduce the dimension of time series data while maintaining the spatial similarity of original time series data.Traditional time series feature representation methods mainly use imaginary mathematical model to stampede and sample time series data,which requires a lot of computing resources.Although this kind of mathematical model has strong feature representation ability,it will inevitably lead to the loss of time series information.Therefore,in the context of massive data mining,using the traditional feature representation method will be difficult to take into account the efficiency and accuracy.Aiming at the defects of time series such as large dimension,high computational complexity and traditional feature representation,this paper proposes a method of feature representation using deep learning technology to improve the efficiency of time series mining task.First of all,in order to get a data set suitable for deep learning network training,this paper preprocesses the time series data.Secondly,considering that the single variable time series is an unlabeled data set,this paper uses clustering algorithm to deal with the false labels of the data sets to guide the unsupervised learning of the time series.Thirdly,the model of deep learning is improved by training time series data with feature enhancement and pseudo label.Finally,the accuracy of the final deep learning feature representation model is obtained by comparing the motif obtained from the original time series analysis and the motif obtained from the deep learning feature representation.The experimental results show that deep learning has superior performance for feature representation of time series,and can obtain stable and accurate feature representation model by adjusting parameters to improve the efficiency of time series mining task.