Research On Reinforcement Avoidance Of Prefabricated Beam-column Joint Based On Deep Reinforcement Learning

As a critical force material in civil engineering,reinforcement is widely used in all elements of the structure,for which the installation accuracy has become a key factor affecting the quality of construction.With the mainstream of new construction methods,prefabricated buildings are now being assembled on site,which requires higher precision for the production of prefabricated components in factories.Especially at the joints of beams and columns of frame structures,the lack of precision in the deepening design of the members and the relative positions between the outreach bars of the members are not fully considered in the production and processing,which will have an impact on the solid connection of the members in the assembly and make it difficult to guarantee the construction quality.The large amount of manual decision making,even though the current application of BIM technology is largely able to solve the collision problem in engineering,makes the designers fall into mechanical repetitive work.To address the problem of insufficient reinforcement deepening design at beamcolumn joints of prefabricated buildings,this paper focuses on the following.(1)Develop a parametric reinforcement program based on Revit platform.Firstly,a graphical interface is designed,in which the flat annotation of 2D drawings is submitted to the backend of the program as form information.By cross-operating with the geometric positioning coordinates of the components,the starting and ending positions of the reinforcement in the prefabricated components and the constraints are determined,while the interface is reserved for the algorithm program.(2)With the introduction of deep deterministic policy gradient algorithms,we build deep reinforcement learning models capable of autonomous learning and autonomous decision making with the help of the decision making ability of reinforcement learning and the fitting ability of deep learning.On the basis of the algorithmic principles and applicability of the model,the standard Gym training environment is established.(3)With the recognition of the joints environment of the current member,the intelligent model is trained to transform the modeling problem of rebar into a path planning problem from the starting point to the end point for different joints forms and reinforcement information.After continuous learning and summarizing in the training environment,the curve coordinates for rebar modeling positioning are finally generated.For this paper,a deep reinforcement learning model is proposed for the reinforcement automatic avoidance problem by improving on the traditional DDPG algorithm called PA-DDPG algorithm.After training with a large number of collision adjustments,the algorithm is embedded into a parametric reinforcement allocation plug-in as a background program for data processing of reinforcement placement and avoidance problems.This plug-in,unlike traditional reinforcement collision adjustment,does not require collision detection or human-specified offset directions,but rather,will automatically bend the reinforcement during curve generation after multiple rounds of training.Instead of falling into over-avoidance,the determination of model convergence is based on maximizing the reward value of the reinforcement agent in the action path.By the actual application in a high school project,the model’s avoidance accuracy and operational efficiency are verified,which also shows that AI technology can provide a more efficient solution to structural design.