DEEP LEARNING BASED PREDICTIVE CONTROL FOR RFT-30 CYCLOTRON

Successful construction of the control system is an important problem in the accelerator. The control behavior still relies on the human operators and the operators should directly manipulate the devices to control the accelerator system. To operate the accelerator well, the human operators should carefully manipulate the control parameters. If the control does not function properly, it becomes difficult to handle the accelerator and cannot perform the accurate operations for the control. In this work, we propose a deep learning based model predictive control approach for solving the nonlinear control problem of the accelerator. The proposed approach constructs the predictive model of the accelerator using the deep neural network (DNN). In the control design stage, the model predictive control (MPC) finds the optimal control inputs by solving the optimization problem. To analyze the performance of the proposed approach, we applied the proposed approach into the RFT-30 cyclotron. INTRODUCTION Recently, particle accelerators become more promising devices for industrial, environmental, and medical applications. Easy operation and minimum maintenance are indispensable parts for the convenient use of the accelerator. In particular, a control problem is an important and critical issue for easy and effective operation of the accelerator. If the control does not function properly, the accelerator may become difficult to handle and cannot perform the accurate operations. In this work, we propose a deep neural network (DNN) based model predictive control (MPC) approach for the RFT-30 cyclotron. To control highly non-linear and timevarying system, we applied the deep learning model with deep neural network (DNN) into the control approach. The proposed approach constructs the beamline model based on the deep belief network (DBN) [1]. Based on the DBN-DNN model, the predictive controller finds the optimal control parameters of the beamline for the desired output. We analyzed the performance of the proposed approach for the RFT-30 cyclotron beamline system. The proposed DNN MPC approach can minimize the beam tuning time and enables effective beam control. Moreover, combined with other control techniques, the proposed approach enables beam auto-tuning and control automation. DEEP BELIEF NETWORK DEEP NEURAL NETWORK (DBN-DNN) ARCHITECTURE Figure 1: The overview of DBN-DNN structure. Hinton et al proposed the deep belief networks (DBPs) for deep learning [1]. Figure 1 shows the basic overview of the DBN-DNN structure. The DBN learning is composed of pre-training procedure and fine tuning procedure. The pre-training is to decide the weight between the layers to train the model accurately before the main training. After pre-training procedure, the proposed approach performs the fine tuning with the error back propagation algorithm. Figure 2: The RBM structure and training procedure. Figure 2 shows the RBM structure and the learning procedure. The pre-training procedure trains the model using the restricted boltzman machine (RBM). The pretraining procedure is an unsupervised learning to find the optimum weights (W) between the input layer (visible layer) and the hidden layer. The learning is to find the weights W which maximize the log likelihood based on the energy function and it can be written as ___________________________________________ † hur09@kaeri.re.kr 9th International Particle Accelerator Conference IPAC2018, Vancouver, BC, Canada JACoW Publishing ISBN: 978-3-95450-184-7 doi:10.18429/JACoW-IPAC2018-WEPAL030