Several Applications Of Neural Network Method On Displacement Measurement

The continuous development of scientific research and engineering has put forward higher requirements for the high precision,large range and real-time performance of optical mechanics methods.Among them,the two-dimensional digital image correlation method is widely used in in-plane displacement measurement.At present,the measurement accuracy of 0.01 pixel can meet most of the displacement measurement requirements,but the real-time performance of digital image correlation method(DIC)can't meet industrial online monitoring requirement and other application scenarios.Interferometry technique,including laser interferometry and electronic speckle interferometry(ESPI),has been widely used in off-plane displacement measurement.However,due to the constraints of phase extraction speed and precision,speckle decorrelation and other factors,it always can't meet the requirements of high-precision,large-range and real-time measurement at the same time.Fringe projection method,as an active measurement method,is often used in the field of three-dimensional shape measurement.The core step is the analysis of the fringe pattern and then the extracted phase can be transformed into the physical parameter to be measured.Therefore,the measurement efficiency depends on the accurate and efficient phase extraction method.In addition,as an important part of machine learning,neural network method has been widely used in image,speech and other fields.This method has also been introduced into optical measurement,providing a new solution for many optical measurement problems such as vibration signal analysis and fringe pattern analysis.The aim of this dissertation is to realize high precision,large range,real-time measurement at the same time,The following improvements have been made to the existing measurement methods:(1)The Newton method(NR)and inverse compositional Gauss-Newton method(IC-GN)as representatives of traditional DIC method are introduced,and we point out that the iterative calculation of sub-pixel displacement is the key reason of the DIC time-consuming problem.Based on this,a temporal sequence method is proposed to obtain sub-pixel displacement by using weighted moving least square method to fit the integer pixel displacement along the time axis.Furthermore,GPU parallel computing is introduced.The computation amount of the temporal sequence method and the IC-GN algorithm is compared and analyzed.Also,the computational efficiency of the proposed algorithm is verified by a simulated fatigue loading experiment and a tensile experiment,and the computation speed achieves 230000 POI/S.(2)Based on the Michelson interferometry displacement measurement system,an out-of-plane displacement tracking measurement algorithm is proposed.In this measurement algorithm,the state of the interference fringe pattern is monitored in real time.The object displacement magnitude and direction are obtained by the back-propagation neural network(BP)and the convolutional neural network(CNN),respectively,and then piezoelectric ceramic nanometer translation platform is actuated in compensation path,finishing the displacement tracking and cumulative measurement.A mirror interference measurement system and an ESPI measurement system are built,respectively.They both realize accurate measurement within 210 microns,and its positioning accuracy within 200ms is 10 nm.(3)Aiming at the key problem of phase extraction in fringe pattern analysis,a single pattern extraction algorithm based on U-Net neural network is proposed,which converts phase extraction into image mapping problem,taking fully advantages of neural network algorithm in image processing field.The simulated fringe patterns,experimental fringe patterns coming from laser interferometry and fringe projection are used to test the trained U-Net network,respectively.The results show that compared with the wavelet transform method,this algorithm owns the advantages of speed,low fringe quality and stronger phase processing ability of complex shape object.(4)To solve the problem that the effective information in the early stage of rolling bearing fault is concealed by the noise,a method of monitoring the bearing's working state is proposed by combining the envelope spectrum autocorrelation curve with BP network.