Research On Damage Diagnosis Method Of Final Optics Assembly(FOA) Based On Deep Learning

This topic is derived from the "Shenguang III" project.The goal of the project is to focus energy on nuclear fusion.The project system is designed to operate at energy density levels near the optical material damage threshold for maximum efficiency.Therefore,optical material components under such working environment conditions are highly susceptible to damage.At the same time,it has been found from laser-related damage to optical components that this laser damage has an exponential effect,that is,the growth trend of damage is exponential.Therefore,in the system work,it is necessary to perform damage inspection on large-aperture optical components to minimize the risk of personnel and reduce the equipment risk caused by the implosion of vacuum-loaded optical devices,thereby reducing the loss of additional components caused by diffraction of damaged parts.The risk of reducing the performance of the system and providing the ability to remove the optics(and subsequently refurbish)before the optical components are damaged and cannot be repaired.In this paper,the large-scale optical components in the final optical assembly(FOA)are studied.Based on the deep learning theory,the damage level diagnosis and the residual life(RUL)prediction method are studied,and the FOA is designed for the damage diagnosis work.Damage diagnosis system.Firstly,for the input of the deep learning model training,the collected raw data images are processed,and the corresponding solutions are given for the noise and fuzzy problems through experimental comparison and verification.Next,a data set for model training is established.Here,the different classification methods of the data set are discussed respectively for the damage level diagnosis problem and the RUL prediction problem.After completing the above preparatory phase,the method was explored.For the damage level diagnosis method,three model structures are constructed,namely: simple convolutional neural network(CNN),CNN+ support vector machine(SVM),and migration learning based on pre-training network.Through experimental verification,the optimal model was selected.For the RUL prediction work,the countermeasures for the over-fitting problem of small sample size are firstly proposed.Next,the simple CNN structure and the migration learning method based on VGG-16 model are compared.It is verified by experiments that the method of fine-tuning is more suitable for the problem to be solved in this study,and has higher accuracy.Finally,based on Qt Creator 4.6.1 development platform,using PYTHON,C++,SERVER SQL and other programming languages,design and implement FOA damage diagnosis system,realize the damage level diagnosis of FOA large-sized optical components and the visual operation of RUL prediction.At the same time,the data storage function is realized for the result,which is convenient for providing data support for future research work.