Research On Automated Filling System Of Vehicle Liquefied Natural Gas Based On Deep Reinforcement Learning

Traditionally,liquefied natural gas(LNG)filling operations for vehicles are completed by manpower,which requires certain professional skills for operators and has potential safety hazards in the filling operations.With the rapid development of industrial automation,it is becoming a trend that intelligent robots replace manpower operations in industry.In this context,according to the characteristics of vehicle LNG filling,an improved deep reinforcement learning control algorithm is adopted to enable robots to make behavior decisions through perceived high-dimensional raw input data,and to achieve self-realization from original input to output.Operation control of dynamic filling.The specific work of the paper includes the following parts:1.Firstly,according to the characteristics of robots that can automatically inject liquefied natural gas into vehicles,the trajectory planning control problem of robots filling guns in the process of filling is solved,and the search algorithm of guidance strategy is studied and compiled.By establishing cost function,fitting local environment dynamic linear Gauss model and using the optimal control method to generate and optimize the controller on the basis of the model,the robot can be realized.Self-regulated learning ability;2.According to the uncertainties of parking position and posture of vehicles to be filled,the automatic filling robot is required to have a good adaptability to the vehicle position and posture within a certain range of detection.A general depth convolution neural network controller is designed to represent the Gaussian trajectory distribution within the detection range,so as to improve the generalization ability of the robot of the automatic filling system to the target vehicle position.In view of the fact that the convolutional neural network fitted under the controller of the guided strategy search algorithm can not effectively process the implicit information of the robot's front and back states,a long-term and short-term memory network module is added to process the time series information to improve the positioning accuracy of the robot.3.Aiming at the problem of long training time of deep convolution neural network strategy,the weights of target feature extraction layer in convolution neural network image are pre-trained.The training data set is an image set containing detection targets,and the weights of extraction layer are initialized by MSAR to accelerate model training.4.Designing simulation experiment and building real scene platform and establishing control system,verifying the end-to-end control effect of the deep convolution neural network control strategy proposed in this paper,and the generalization ability of the robot to the target position of the filling vehicle.