Research On C-ADS Injector ? Beam Calibration Technology Based On Neural Network

It is an important issue in nuclear industrial applications how to safely and permanently treat nuclear waste from nuclear energy.At present,the most advanced method is to use fast neutrons to convert long-lived,high-radiation nuclides into nuclides with short life,good stability and low radioactivity.Accelerator Driven Sub-critical System(ADS)is one of the most advanced methods of transmuting nuclides in the world.China has been developing ADS systems since the 1990 s,and has now developed particle accelerators that can produce high-energy proton beams,it is called C-ADS Injector II.The particle accelerator is very complex and contains thousands of components.Because of engineering errors,noise interference,measurement errors,control errors and other factors,the beam has deviated from the ideal orbit during the transmission process,which seriously affects the beam quality and the operation safety.Although scientists have improved the beam drift and improved the accuracy of beam control through theoretical analysis,empirical judgment,and equipment improvement,the problem of beam offset is still serious.How to solve the problem of beam offset and improve the quality of beam transmission is still a very important scientific issue in the field of accelerators and control at home and abroad.For the problem of beam offset,the thesis starts from different levels to explore the problem of particle accelerator offset calibration based on Neural Network.The specific content includes the following aspects:(1)Aiming at the problem that the traditional beam offset correction method in C-ADS Injector II is not effective,taking the Medium Energy Beam Transmission Line(MEBT)as an example,a beam offset correction modeling method based on neural network is proposed.By abstracting the actual control process of MEBT,combining the characteristics and structure of neural network,a MEBT beam offset correction model based on neural networks was established,and a one-dimensional beam offset correction model of MEBT was experimented with an improved deep BP neural network.After the parameter design and training,the optimal parameters were determined.The results show that it is feasible to perform offline beam offset correction with neural networks,and the results can be used as a reference for researchers to manually control the accelerator.(2)Due to the influence of various noises,the x-axis and y-axis of the quadrupole magnet in the particle accelerator do not independently control the beam trajectory,and the one-dimensional beam offset correction model cannot be used to perform twodimensional beam offset correction.For this purpose,a two-dimensional offline beam offset correction technique for MEBT was studied based on the Monte Carlo neural network(MCNN).The traditional MCNN algorithm has good generalization ability,but the training speed is slow.In order to speed up the training speed of the traditional MCNN,the partial derivative is introduced to determine the direction of the parameter variation,and a Multi-dimensional Derivative-based Monte Carlo(MDMC)optimization algorithm is proposed.In order to further improve the stability and prediction accuracy of MCNN,it is proposed to combine multiple homogeneous or heterogeneous MCNNs and use a decision maker to decide the output of the network——Joint-Monto Carlo Neural Network(J-MCNN)optimization method.Analyze and establish the control process of MEBT two-dimensional offline beam offset correction,a two-dimensional beam offset correction model based on the optimized MCNN was carried out.Finally,experiments were performed on the model to determine the optimal parameters.The experimental results show that the two optimization algorithms can effectively improve the network training and prediction of MCNN.Accuracy and stability,a two-dimensional offline beam offset correction model based on MCNN can help researchers to manually adjust beam control parameters.(3)The beam transmission speed is extremely fast,the beam calibration control time is required to be completed in the order of milliseconds,and the beam offset calibration cannot meet the timeliness requirements.In response to this requirement,the Monte Carlo neural network model used for two-dimensional offline beam offset correction was implemented in Field Programmable Gate Array(FPGA)to achieve twodimensional online beam offset correction.In the research,a combined optimization method composed of a three-part neural network model reduction method,a vectormatrix multiplication parallel optimization method,a sigmoid function and softmax layer implementation methods are proposed.Based on the above optimization method,a MCNN accelerator based on FPGA is designed,optimization and performance testing are performed in the online beam offset calibration.The results show that the implementation of MCNN in FPGA shortens the prediction time,improves timeliness,the MCNN accelerator can be applied to C-ADS Injector II online beam offset calibration.