High-precision Prediction,reconstruction And Missing Recovery Of Nonstationary Wind Signal Based On Deep Learning

With the rapid progress of science and technology as well as the development of economic prosperity,high-rise and super high-rise buildings,long-span roof structures and bridges are emerging continuously.As the main design load of high-rise flexible buildings and long-span structures,wind load has attracted increasing attention from experts.In recent years,the structural health monitoring system based on wireless sensor technology has developed rapidly.The research on wireless sensor data reconstruction and recovery is of great engineering practical value.This paper focuses on the hot topics in the current research field of civil engineering,i.e.,prediction,data reconstruction and missing data recovery for non-stationary wind signals(including wind speed and wind pressure).The main work and innovation achievements of this paper are listed as follows:(1)In the case that the forward observed data of non-stationary wind signals(including wind speed and wind pressure)are unknown,a high-precision simplifiedboost reinforced model for non-stationary wind signal forecasting is proposed.Firstly,the multi-channel conversion strategy is adopted for the wind signal time series preprocess,which can provide information to the model at different time scales.Then the adaptive residual convolutional neural network is proposed as the basic predictor,which can be adaptively simplified according to the fluctuation complexity of the wind signal.Afterwards,the reinforced forecasting is proposed to optimize the primary forecast results.Finally,in order to fully consider the effect of the historical observed signals at several time steps before the signal to be forecast at time step + 1,a simplified-boost technique is implemented to improve the universality of the reinforced forecasting strategy.The multi-step forecast results of five kinds of field mornitoring non-stationary wind signals have testified the superiority of the proposed hybrid model.(2)In view of the situation that data at an observation point in wind field(including wind speed field and wind pressure field)are completely missing,a data reconstruction algorithm based on Kriging based sequence interpolation(KSI)and statistical properties correction is proposed to reconstruct observed data missing completely with high-precision.Firstly,the primary interpolation results and probability density function of data at missing point can be obtained by KSI.Back propagation neural network is then adopted to extract the exact standard deviation of the missing data from the reconstructed probability density function.Subsequently,the exact standard deviation is used to correct the primary interpolation results.Data reconstruction of wind load considering probability distribution information can reflect load characteristics precisely,which is of great value in engineering practice.(3)When the data are completely missing at several continuous observation points in the wind field,the correction for the primary interpolation results based on the exact standard deviation cannot achieve satisfactory results,especially on the recovery of extreme value.Therefore,the quantile regression combined with deep neural network and kernel density estimation are proposed for the probability reconstruction of the extreme value of missing data.At the same time,the point estimations based on quantile regression can be adopted to the secondary correction for the reconstructed data with first correction.The structural dynamic response analysis has fully verified the effectiveness of the two correction strategies.Probability reconstruction of extreme value can provide more reliable and comprehensive information for researchers in structural design,which is benefit for decision-making.(4)A high-precision algorithm for missing data recovery based on convolutional neural network and KSI is proposed in order to solve the random missing during data transmission of non-stationary wind signals.Regarding the wind signal(wind speed)time series at observation points,the observed data at different time instants maybe missing randomly.The proposed algorithm can effectively recover the missing data.Before data transmission,the non-stationary wind signal are decomposed by wavelet transform and empirical wavelet transform to achieve a group of sub signals with uniform frequency components.The cubic spline interpolation and convolution neural network are then used to establish missing data recovery models for low frequency and high frequency signals,respectively.Finally,the KSI algorithm is used to obtain the extreme value curve of the missing data sequence,which can be referred to correct the extreme value of the primary recovered data.Numerical examples based on the field monitoring data have testified that the proposed method is of great rationality and practicability.At the same time,high economic cost and complex technology are needed to obtain a large number of field monitoring data.The samples of training set will be insufficient due to the limited field monitoring data,which may lead to over fitting of the deep learning model.In order to solve the above problem,a pre-training and finetuning algorithm based on non-stationary simulation is proposed.The accuracy and stability of deep learning model can be promoted by pre-training and fine-tuning algorithm to a large extent.Simulation stratege effectively solves the problem of lacking of field monitoring data,which testifies that the proposed algorithm is of great engineering significance and high practical value.