The Research Of Features Extraction Based On Improved Deep Neural Network

Features extraction is an important part of many algorithms and becomes a hot spot of many scholars' studies.The shallow features extracted by traditional method cannot meet the requirement in many cases as the problem is complicated.In contrast,the deep features can reflect the internal characteristics of the data better,so the effect of control or classification is better.As the method of deep features extraction,deep learning(DL)algorithm has been successfully applied to many fields.The algorithm uses a multilayer neural network structure to extract deep features step by step.This can also reduce data dimension in order to avoid "dimension disaster" problem.In this thesis,a new method based on improved deep network is proposed for high dimension input data classification problem and the controller design problem.The paper research content is as follows:1.High dimension input data classification problem is researched.The traditional classification method is complex because it needs to combine with some other methods to achieve classification.An improved CNN is proposed to achieve features extraction and classification at the same time.Convolution layers and sub-sampling layers are used to extract features and finally features are put into the multilayer perceptron(MLP)to achieve classification.l2 regularization is added to the network in order to prevent the over-fitting problem.The accelerated proximal gradient(APG)is used to train the CNN,which can accelerate the training rate.Finally,EEG is employed to the experimental verification,EEG is often affected by external and internal factors and has the characteristics of randomicity,irregular,no cycle,nonlinear.The signal dimension is high in the time domain.The experimental results show that the classification method effect is better than the traditional classification method.2.In addition,the learning control problem based on data driven is also studied.In the field of industrial control,design of data-driven learning control which depends on the feature extracted from raw data is usually used when the system is complex or the model is unknown.The quality of the extracted features is the important factor that affects the result of control.Improved auto-encoder neural network is used to extract the internal features of the system data.l1 regularization and l2 regularization are added to the network and APG is used to train the network.According to the features extracted by the improved auto-encoder neural network,reinforcement learning(RL)algorithm can learn the useful control strategy.The paper realizes the design of the controller on the inverted pendulum system.The experiment shows that this method has better control effect and lower computational complexity.