| The demand for energy is growing larger and larger in China with the rapid development of economic level and the promotion of people’s living standard.As a kind of clean energy,nuclear energy becomes more and more favored by people all over the world.The nuclear fuel pellet is the key component for the generation of nuclear power and the damage on the surface will lead to accidents related to the leakage of nuclear fuel,which brings huge losses in terms of the lives and assets of human beings.The automatic detection system for the nuclear fuel pellet which is based on machine vision is the key linkage for the generation of nuclear power.In this regard,if the posture of the nuclear fuel pellet is incorrect in the detection system,it will not only bring the ineffective detection of nuclear fuel pellet,but also lead to mechanical failure when the push rod positions the pellet.Since the nuclear fuel pellet is radioactive and toxic,we are supposed to put it inside the assembly line by grabbing with a mechanical arm while the correct recognition of posture of the nuclear fuel pellet is the premise for the completion of the grabbing task.Therefore,the exploration for the posture recognition of the nuclear fuel pellet is of great research value.Currently,as a rapid-developed research field,deep learning has made great achievements in speech recognition,image classification,target detection and other fields.The application of deep learning to posture recognition and improvement of grasp efficiency of industrial production are of great research value.Therefore,this subject proposes the nuclear fuel pellet posture recognition algorithm based on convolutional neural network and achieves the deep information of the nuclear fuel pellet by adopting 3D depth camera and recognizes the positions and postures of nuclear fuel pellet in sight by taking advantage of the algorithm combining otsu algorithm with convolutional neural network model.The main research contents of the article are as follows:(1)A posture image collection system of the nuclear fuel pellet is established,and the performance of 3D camera adopted in this installation is explored.The 16-bit pseudo-color model of this camera is used to acquire the deep posture images of nuclear fuel pellet.(2)The principles of image segmentation by otsu algorithm are explored,and nuclear fuel pellet in view through otsu algorithm and morphological processing are positioned and segmented to obtain the pseudo-color images of nuclear fuel pellet and provide the data sets for the training of posture identification model framework.(3)Compared with traditional posture recognition methods,convolutional neural network can independently extract image posture characteristics.Therefore,this paper replaces the late three full connections with an average pooling layer and fully-connected layer by referring to VGG16 network model structure and establishes a posture recognition model based on convolutional neural network.By taking advantage of coding operation,posture images labeled are put into binary files;the training of network models is realized by adopting the five-folded cross validation and using this data set;the performance of the posture recognition model is measured by the evaluation indexes of loss function,accuracy rate and recall rate;eventually,the network model capable of recognizing nuclear fuel pellet postures is obtained. |
