Research On Multi-core Fiber Bundle Positioning Detection Technology Based On Deep Learning

Large Sky Area Multi-Object Fiber Spectroscopic Telescope is a large-scale scientific research equipment the field of astronomy and astrophysics.Due to the large number of optical fibers and the open-loop design of the optical fiber positioning unit used in the LAMOST system,the precise positioning of the optical fiber has become a key factor limiting the further improvement of the spectral acquisition efficiency of the LAMOST systemThis thesis presents a technical solution for optical fiber positioning accuracy detection based on deep learning and multi-core fiber bundles to assist in the accurate positioning of optical fibers in the LAMOST system.This solution consists of two key parts.One is the fiber positioning detection method based on the multi-core fiber bundle,which is use the position guide fiber that integrates multiple cores around the central transmission fiber and can be used to provide feedback information for fiber positioning detection.The multi-core fiber replaces the transmission fiber in the LAMOST system,controls the probe positioning,so that the star spot is canned on the probe end face in a fixed manner,and uses the CCD camera to collect the probe position to guide the exit photo of the fiber pigtail.Then establish the mapping model of the image at its corresponding star spot center and probe center offset.The second is the processing method of the probe position guided fiber pigtail exit photo.This paper proposes the traditional image processing and deep learning neural network-based image processing and analysis solutionsThis thesis first uses template matching methods to process and analyze the probe position guided fiber pigtail exit photos,establish a matching matching template database of the image and its corresponding star spot center and probe center offset.And use the matching template to position detection of test data sets,whose detection error is 25?m.This verify the fiber positioning accuracy based on the multi-core fiber bundle check the feasibility.Secondly,based on the three convolutional neural networks of AlexNet,Vgg and improved ResNet50,a neural network model for mining and learning the mapping relationship between the image and its corresponding star spot center and probe center offset was established,and several experiments were conducted to select the optimal training parameters.When the learning rate is 0.0001,the number of iterations is 80 and the optimization function is RMSprop,the training result of the improved ResNet50 model is optimal and the mean square error of the predicted value and the true value of the test data reaches a minimum of 0.01.And the detection error of this network is 3.6?m.The prediction result of the model of the star spot pattern meets the 40?m error requirement of the fiber positioning of the LAMOST system.