Research On Intelligent Inspection Method Of Weak Feature Damage In Large Aperture Final Optics

The final optics assembly(FOA)is located at the output end of the inertial confinement fusion(ICF)experimental facility,and each FOA contains several expensive large-aperture final optics.Laser-induced damage(LID)is easily generated under high-power laser irradiation,which seriously shortens the service life of the final optics.This will not only cause significant economic losses,but also seriously reduce the operational efficiency of ICF facility.It has been found that the final optics with initial small size LID can be effectively repaired,and the performance of the repaired final optics will not be reduced.As the LID size increases,the optics will be increasingly difficult to repair,and even cannot be repaired.Therefore,it is necessary to detect and track the LID on the final optics timely to take corresponding repair actions according to the inspection results,which is of great significance for the healthy operation of ICF facility.The main difficulties in the damage online inspection for final optics are as follows:(1)The size of initial damage is small,the signal is weak,and it is difficult to detect;(2)Due to the influence of stray light and other factors,lots of false damage are generated,which interfere with the detection and evaluation of damage;(3)The input and exit surface damage need to be distinguished due to their different growth laws,but the weak feature damage are difficult to distinguish at low resolution;(4)When imaging a large aperture optics,the damage image is small,even at sub-pixel level,and accurate size measurement is very difficult.For the above problems,this thesis conducts research on high-precision and fast damage online inspection technology for final optics based on machine learning.The main contents are as follows:(1)The zoom imaging system of damage online inspection are established,with the aim of solving the problem of high signal-to-noise ratio imaging technology for tiny defects in the large field of view.The system can perform rapid online collection of damage images for FOA at various orientations in the target chamber center.The cavity-type total internal reflection side light homogeneous illumination in dark field(CTS)is proposed for inhomogeneous illumination problem.Illumination system parameters are optimized and illumination uniformity is improved by Monte Carlo ray tracing.For damage classification problem,the imaging model is used to theoretically demonstrate that the true and false damage and the input and exit surface damage are separable.(2)An automatic classification method based on kernel-based extreme learning machine(K-ELM)is proposed to distinguish true and false damage sites,which is under low resolution,weak feature imaging conditions.Monte Carlo ray tracing is used to analyze the causes of false damage sites.The feature vector constructed by multiple weak features is used to represent the candidate sites to be classified.KELM is used for automatic recognition of the true sites and false sites.Compared with the existing algorithms,only one type of false site can be identified at a time,the proposed method classifies all false sites into one class,and classify the true sites into another class,which is more convenient to operate.(3)An autoencoder-based extreme learning machine(A-ELM)is proposed to solve the problem of determining the location of surface damage sites,which is under low resolution,weak feature imaging conditions.A-ELM designs a coding layer to fully exploit the information in the training samples.It avoids the problem of excessive increase in the number of neurons in the single hidden layer.A-ELM is used to identify the surface on which a damage site resides in the case of small sample data.The speed and accuracy are improved compared to existing classification methods.(4)A hierarchical kernel extreme learning machine(MFHK-ELM)is proposed to solve the problem that the damage size cannot be accurately measured under the condition of inhomogeneous illumination.Under the inhomogeneous illumination condition,it is difficult to establish an analytical physical model for solving the damage size through optical theoretical analysis.At the same time,the existing radiometric method has limitations and cannot accurately measure the damage size.MFHK-ELM uses multiple weak features as its input data and establishes a multiinput to single-output regression model through a deep neural network.The regression model is used to approximate an unknown target function from the LID image to its actual size.Since FODI system resolution is only 125?m ~ 140?m,the ultra-resolution measurement of the LID above 50 ?m is achieved to some extent.