Research On Intelligent Auxiliary Design Of Subway Station Building Space Based On Deep Learning

With the continuous development of the architecture discipline,many problems need to be in this major and other subjects crossover study of architecture can be better solved,at the same time,along with the computer science and technology promotion and the rapid growth of the work force,appeared such as the "deep learning" and "artificial neural network" epoch-making computer cutting edge technology.Reviewing the digitization process of architectural design,from modular system,computer-aided architectural design,and then to parametric design,it has already faced the bottleneck and needs a new breakthrough.At the same time,the rail traffic in the city to ease traffic pressure and traffic transportation,the design of the subway station is becoming more and more important,and subway station has the very strong "module","bottom",potentially more obvious design logic,this feature determines the spatial layout of metro station of laws can be deep learning is learning,So will the subway station as the research object,based on the current technical level of the computer aided architectural design and parametric design,integrated the deep learning technology of artificial intelligence technology,artificial intelligence is proposed to design and architecture design of metro station intersection,in order to realize in the true sense "computer intelligent building design".In this study,Pointnet++ model,which is very advanced in deep learning,is used to train the cloud data of subway stations.The main work focuses on code repetition and parameter training of the Pointnet++ model,as well as the collection and arrangement of a large number of point cloud data sets in subway stations.Innovative interdisciplinary research method,the three-dimensional point cloud data of deep learning combined with architectural design,to break through the original 2 d image as the research object of the present situation,to avoid the "deep learning" was applied to 2 d images such as the research object and can’t accurately describe the limitations of 3 d space,provides a more intuitive and architect diversified auxiliary design,In addition,end-to-end experiments are carried out,and both the input data and the output prediction data are in point cloud format,which greatly increases the fidelity and versatility of the data.After training and validation of the forecast and the actual project,the following conclusions:(1)based on point cloud data is verified feasibility of building space depth study of the form: the whole process of training and prediction have higher can form,training and prediction effect is good,and the point cloud format under Pointnet++ model is fully compatible with excellent performance;(2)It verifies the validity of Pointnet++for semantic segmentation and prediction of cloud information of subway stations: the experimental verification results show that: in the overall condition of 20*20*20 space and block＿size of 10.0,the value of 9: The training prediction data of 1 are 60%+MIOU(Mean IOU Mean crossover ratio)and 75%+ ACC(Accuracy),which belong to relatively excellent values in the neural network operation of the limited model data of 618 subway stations.(3)The validity of semantic segmentation and prediction in the study was verified by the interchange stations of Tianjin West Railway Station Line 1 and Line 6:Taking the transfer stations of Tianjin West Railway Station Line 1 and Line 6 as an engineering example,it is transformed into a point cloud model with color labels.After the prediction,it is compared with the Ground Truth,so that the theoretical research is closer to the actual needs,and the theory is consolidated,checked and improved in the verification.The Ground Truth and Predict were compared in a visual way,and Predict showed a good prediction effect,with appropriate control over both the range and scale of functional partitions.This deep learning model can be further expanded into a "subway space design comparative evaluation system",that is,it can be used for rapid generation and comparison of schemes,as well as comparison and evaluation of original schemes.Although it is not comparable to such models as expert evaluation,it has a rare quality--objectivity.It collected a large number of existing design mode of space,spatial patterns of these existing or based on a large number of professionals of wisdom,but the deep learning model can be very objective and rational to forecast,is the result of our use of it to plan than to choose not only possess the wisdom of predecessors,also has the objective and comprehensive,It reduces the one-sidedness and contradiction brought by personal subjective evaluation.