Data Driven Methods For Signal Reconstruction Of Compressive Sampling In Structural Health Monitoring

The number of large-scale public infrastructure projects in China has been increasing since the establishment of our country,especially since the reform and opening up.However,in the process of long-term service,structures often need to bear long-term effects such as environmental erosion,and even encounter extreme effects such as typhoons,tsunamis and earthquakes.Therefore,it is necessary to study the structural health monitoring technology to evaluate the service state of the structure to ensure the structural safety.However,in the structural health monitoring system,there is a prominent contradiction between the massive data collected by a large number of sensors and the limited transmission and storage capacity,so it is necessary to develop efficient data compression technology.The emergence of compressive sampling technology has brought revolutionary changes to the field of signal processing.This technique takes advantage of the sparseness of signals in certain domains to collect compressed data directly at a rate far lower than Nyquist sampling frequency,and then reconstructs the original signal from the projection by optimization method,which has broad development prospect in the field of structural health monitoring.However,in the process of application,it is necessary to find a suitable sparse base to make the signal sparse enough,and to embed the sparse regular term explicit expression for sparse optimization,which limits its application effect.Therefore,it is necessary to develop reconstruction method of compressive sampling signal to break through the limitation of sparse constraint,which can be realized by the powerful feature extraction capability based on deep learning.In this paper,the research on data-driven signal compression and reconstruction method based on the deep learning is carried out.In the second chapter,the one-dimensional signal compression and reconstruction method for structural health monitoring based on discriminative deep neural network model is established.Firstly,the compressive sampling measurement matrix is established to compress the one-dimensional structural health monitoring signals.Then,the discriminative deep neural network models,such as Multi-Layer Perceptron,Waveform Transposed Convolution Neural Network and Vector Quantised-Variational Auto Encoder(modified version)are constructed,and the feature learning based on signal big data is done,which is to establish the mapping relationship between low-dimensional compressive sampling signals and high-dimensional original signals.Based on this,the one-dimensional signal compression and reconstruction of structural health monitoring based on discriminative deep neural network model is realized.In the third chapter,a compression and reconstruction method of structural health monitoring signal based on generative deep neural network model is established.Based on the compressive sampling data of one-dimensional structural health monitoring signals,this paper constructs generative deep neural network models,such as Waveform Generative Adversarial Network and Vector Quantised-Variational Auto Encoder.The feature learning is carried out based on the compressive sampling data of one-dimensional structural health monitoring signals,so as to realize the generation of quasi-genuine structural health monitoring signals from random vectors.The data are input into the generative network model to realize the compression and reconstruction of onedimensional structural health monitoring signals based on the generative deep neural network model.In the fourth chapter,the discriminative and generative deep neural network models are verified based on real health monitoring signals,and compared with the common compression sampling methods based on sparse constraint.The results show that the discriminant and generative network model methods do not need to select appropriate sparse basis for the signal,and the compressive sampling method based on sparse constraint has better performance in signal reconstruction of cosine basis and wavelet basis.Among them,the method based on Vector Quantised-Variational Auto Encoder(modified version)has the best signal reconstruction effect among all methods,and when the compression ratio is 0.25,the acceptance rate of accurate reconstruction is more than 88%.