The Linearization And Application Of Two-Dimensional PSD Based On BP Neural Network

The position sensitive(PSD) in this paper is an photo electronic sensor which can detect the continuous position of a light spot traveling over its surface, and convert the position of light spot to simple electric current signal. Based on PSD, many types of precision and contactless motion detection instruments could be constructed. The most important problem to use the PSD is how to overcome the influence of non-linear action os the PSD. Therefore to improve the precision and reliability of the instrument. Based on artificial neural network,a non-linear compensation method of PSD is presented in this paper. In order to non-linear compensation over a full range, the neural network is trained to represent the non-linear mapping between sensor reading and their represent output accurately properly. So the correction leads to prominent improvement of linearity of B-area,and the usable area of PSD is thus extended. The reliability of data is also improved without increasing the complexity of hardware by the method.In this thesis, the working principle , structure and performance analysis of PSD were introduced; Many measures of eliminating background light of PSD are presented. Based on neural network, mathematical model of BP algorithm is analyzed particularly; Two-dimensional adjustment platform has been designed; Optical system of autocollimator has been established and the method of evaluating straightness has been introduced.In the course of training neural network, two hidden layers' structure is applied. Conclusion, 33 and 39 are respectively two hidden layers' neurons; tansig, tansig and purelin are respectively activation function of first, second and output layer in network; times of maximum training is 500,training function is Trainlm; With untrained data network is tested ,and the error of the network output is almost within 0.001mm, 0.003mm of maximum error.And in this project, the development complexity and cycle were decreased by taking advantage of mix_programe of MATLAB and Visual Basic, and it has improved portability of neural network.