Research And Application Based On SDAE-DBN Hybrid Neural Network

Neural network is an intelligent system developed by imitating the information processing function of human nervous system.It can learn to obtain external knowledge and store it in the network.It can solve difficult problems that are not easily handled by computers,especially a series of essentially non-computing problems.Now,some excellent neural network algorithms have been proposed,such as Generative Adversarial Network(GAN);there are also some hybrid neural networks that perform well,such as Convolutional Recurrent Neural Network(CRNN).Although it is undeniable that these algorithms have greatly improved the accuracy in many application fields,at the same time,these algorithms have also made the neural network too large and complicated in the case of excessive pursuit of accuracy,leading to some circumstances will no longer approperiate.This paper designs a hybrid neural network model that combines two neural networks: Sparse Denoise Auto Encoder(SDAE)and Deep Belief Network(DBN).The model makes use of Sparse Denoise Auto Encoder strong abstraction,robustness and data compression capabilities,and excellent feature expression capabilities of Deep Belief Network.Combining the advantages of two neural networks,a SDAE-DBN hybrid neural network model is proposed in this paper.The core of the model is: First,train the Sparse Denoise Auto Encoder,use the Sparse Denoise Auto Encoder to extract features from the data,and compress the data;then,pass the training output data to the Deep Belief Network to train the depth Confidence network;Finally,use Back Propagation(BP)algorithm to fine-tune the overall hybrid neural network model.The model also takes advantage of the simplicity and similar framework of the Sparse Denoise Auto Encoder and Deep Belief Network models,thereby ensuring that the hybrid model is not overly complex.Finally,for the problems of stock forecasting and air quality forecasting,based on the SDAE-DBN hybrid neural network proposed in this paper,using the collected real sample data,a relevant forecasting model is established.At the same time,the SDAEDBN hybrid neural network prediction model and traditional deep neural network models such as Back Propagation Neural Network(BPNN),Auto Encoder(AE),Deep Belief Network,and Convolutional Neural Network(CNN),etc.have conducted comparative experiments,and compared and discussed their accuracy rates.The experimental application results show that the SDAE-DBN hybrid neural network prediction model has better accuracy and stability,which reflects its superiority compared with traditional deep neural networks,and proves that it can provide certain prediction projects for similar it can be used for reference,indicating that it has certain application value.