Research And Application Of Multi-dimensional Signal Prediction Model Based On Sparse Auto Encoder And Deep Neural Network

With the advent of the era of big data and the rapid development of artificial intelligence,the study of predictive models has become an important research field of artificial intelligence,and has been widely used in many fields such as smart industry,smart medical treatment,and smart finance.In recent years,intelligent predictions based on machine learning algorithms have begun to be applied to natural disaster monitoring,industrial data early warning,medical auxiliary diagnosis and other fields,and have achieved good results.Traditional machine learning models have problems such as insufficient feature extraction and generalization ability,which lead to low prediction performance of the models.Deep neural networks represented by deep learning algorithms have outstanding performance in solving complex prediction problems by virtue of their efficient deep feature extraction advantages and excellent generalization capabilities.From the perspective of theoretical research and practical application,deep neural networks such as sparse auto-encoding and long short-term memory are used as the theoretical basis,and one-dimensional landslide data and threedimensional brain tumor medical data are used as the research objects.Three prediction models with high accuracy and good generalization performance are proposed,and applied to predict landslide susceptibility and predict the prognostic survival time of patients with brain tumors.The main research contents of this thesis are as follows.(1)A prediction model based on sparse auto-encoding network and multiclassifiers(Sparse Feature Extraction,SFE+)is proposed.Firstly,a sparse feature extraction network for one-dimensional data is constructed,and some sample features in the input layer of the network are randomly discarded.The network does not rely too much on certain features,alleviating overfitting,and improving the generalization performance of the model.Secondly,the lifetime sparsity is introduced in the hidden layer to solve the problem of nonlinear coupling of data to a certain extent,and sparse predictive features are extracted.Finally,the extracted sparse features are passed through support vector machine(SVM),logistic regression(LR)and stochastic gradient descent(SGD)classifiers,namely: SFE-SVM,SFE-LR and SFE-SGD prediction models.Those models are collectively referred to as SFE+ models.The SFE+ models are applied to the collected geological data of Shicheng County to predict the susceptibility of landslides.Experimental results show that the SFE+ network proposed in this thesis can effectively improve the prediction performance.The prediction accuracy of landslide susceptibility and the area under curve(AUC)of the prediction rate curve are SFE-SVM(74.52%,0.809),SFE-LR(72.98%,0.819),SFESGD(72.68%,0.808),respectively.Therefore,the proposed prediction models based on sparse auto-encoding network and multi-classifiers can better solve the problem of data coupling,effectively extract the sparse features of one-dimensional data and improve the accuracy of prediction.(2)A prediction model based on one-dimensional convolution and long short-term memory network(1DCNN-LSTM)is proposed.Firstly,construct a network consisting of two convolutional layers,one LSTM module and three fully connected layers.The network can fully correlate the data and extract one-dimensional prediction features.Secondly,the cross entropy and Adam optimizer are used to optimize the network and speed up the convergence of the model.Finally,the binary classification task is completed through the softmax function.The 1DCNN-LSTM model is applied to predict the landslide susceptibility in Shicheng County.Experimental results show that the prediction accuracy of 1DCNN-LSTM is 74.65%,and the AUC value is 0.875,which are significantly higher than the SVM,LR and SGD,and slightly higher than the SFE-SVM proposed in the previous chapter.It can be seen that the 1DCNN-LSTM prediction model proposed in this thesis has better feature extraction capabilities,faster convergence,and can significantly improve the prediction performance.(3)A prediction model of one-dimensional convolution and long short-time memory network based on principal component analysis dimension reduction(PCA-1DCNN-LSTM)is proposed.Compared with the one-dimensional features studied in the previous two chapters,images have higher-dimensional features and require more complex networks to extract predictive features.The proposed PCA-1DCNN-LSTM prediction model for images is based on 1DCNN-LSTM combined with a principal component analysis feature selection module,which can effectively filter out feature subsets,thereby avoiding the influence of redundant information.In addition,L2 regularization and dropout are added to the network,which improves the generalization ability of the model.The PCA-1DCNN-LSTM model is applied to the multimodal brain tumor segmentation competition dataset Bra TS2020(Multimodal Brain Tumor Segmentation Challenge 2020,Bra TS2020).Firstly,those regions of interest in the multimodal medical image data are segmented.Secondly,278 features are extracted from the segmented area using pyradiomics software.Finally,the extracted features are passed through PCA-1DCNN-LSTM to predict the prognostic survival time of brain tumor patients.Experimental results show that the prediction accuracy of the PCA-1DCNN-LSTM model is 52.78%,which is higher than that of several traditional machine learning models(KNN: 44.44%,SVM: 47.22%,LR: 30.56%).Therefore,the proposed PCA-1DCNN-LSTM prediction model can reduce the dimensionality of high-dimensional features,fully extract prediction information,improve prediction accuracy,and expand the application range of the prediction model.In summary,in this thesis three prediction models with high performance are proposed based on deep learning according to different signal characteristics,and applied to predict landslides and the prognosis of patients with brain tumors.The proposed SFE+ prediction model can better solve the problem of nonlinear data coupling and effectively improve the prediction performance;the proposed 1DCNNLSTM prediction model has better feature extraction capabilities leading to improvement of the performance of one-dimensional landslide prediction;the proposed PCA-1DCNN-LSTM prediction model is able to take images as input and extract prediction features from high-dimensional predictive features.The three proposed prediction models are applied to the actual collected landslide geological data for landslide susceptibility prediction,and also to multimodal images of brain tumor for prognostic survival time prediction.Therefore,the prediction models proposed in this thesis are theoretically innovative and have application value in disaster prevention and mitigation,intelligence medical care and other fields.