Deformation Prediction Based On AVMD-PSO-LSTM Neural Network

With the development of measurement technology,deformation monitoring data becomes more and more complex,In order to improve the prediction accuracy of deformation data,aiming at the problem that the traditional prediction model is not accurate enough to deal with the complex nonlinear and non-stationary deformation monitoring data,a kind of deep learning model--long and short time memory neural network is applied to study the deformation prediction,a combinatorial model of particle swarm optimization(pso)based on variational modal decomposition(variational modal decomposition)is proposed and applied to the prediction of engineering deformation.The validity and reliability of the combined model experimental results are discussed by combining the simulation experiment with the actual engineering deformation prediction analysis and the comparative analysis with other models.The observed values of deformation monitoring data are easy to be interfered by external factors or human factors,which leads to measurement errors.Variational modal decomposition is a decomposition method that can provide higher time-frequency resolution than EMD.Aiming at the uncertainty of key parameter selection in variational modal decomposition,an adaptive variational modal decomposition method is proposed: The evaluation value of adjacent decomposition layers was used to guide the optimal value of decomposition layers.And a critical value determination method of noise component based on mutual information entropy.In view of the principle of wavelet threshold denoising and the adaptive variational mode decomposition-threshold denoising model,after the parameters of the model are adaptively determined,the determined noise components are eliminated,and the remaining components are recombined to obtain the denoised deformation sequence.Through the analysis of simulation experiments,under the influence of different noise levels,the evaluation indexes of adaptive variational modal decomposition are better than wavelet denoising,empirical mode decomposition and complete set empirical mode decomposition.In view of the dam deformation data to forecast model of nonlinear and non-stationary effects,introducing the idea of "decomposition forecasting-reconstruction",the original deformation sequence decomposition,then respectively forecast each subsequence after the reconstruction,in order to reduce the nonstationarity in original sequence,has the very high accuracy and strong generalization ability.The combined prediction model of adaptive variational modal decomposition and long and short time memory neural network was constructed,and the modal component of noise removal was used as the input value to predict,and then the predicted value of each modal component was reconstructed to obtain the predicted value of the combined model.For the training of long and short term memory neural network,large sample data is needed,while the sample size of deformation monitoring data is small,which limits the generalization ability of the model.The method of segmentation was introduced to amplify the samples.Considering that the value of the segmentation window affects the prediction results of the model and the selection of super parameters such as learning rate and training times in the long and short time memory neural network,the particle swarm optimization algorithm was used to optimize the super parameters of the model.The deformation prediction model of avmd-pso-lstm was constructed and compared with the single model and the combined model of ceemd-pso-lstm.The experimental results show that the avmd-pso-lstm combination model has strong adaptive ability and predictive performance,which can be used for reference and guidance in engineering applications.