Research On 3D-neural-networks-based Automatic Classification Of Bim Element

With the growing adoption of Building Information Modeling(BIM),specialized applications have been developed to perform domain-specific analyses.In order to reach the full potention of BIM,the priority is to make date excahnge smoothly between those softwares,breaking the dilemma of isolated information.The cost of communication will no more be too expensive to afford,and more valuable information are able to be digged out.The Industry Foundation Classes(IFC)provide a rich representation with which to exchange semantic entity and relationship data,which requres the IFC to include a large range of data and make sure there is no mistake in them.What ’ s more,tailored information with respect to a BIM model element ’ s attributes and relationships is needed.In particular,architectural elements need further qualification concerning their geometric and functional ‘subtypes ’ to support exact simulations and compliance checks.For example,in energy analyses,to simulate accurately,element subtypes are occurred to provide occupancy use conditions.What’ s more,automated code checks will function effectively only in the condition that BIM elements are correctly assigned to their IFC entities and the assignments of detailed subtype are significant to carry out intricate code compliance analyses.However,subtypes for individual elements are not represented by default and often require manual designation,which is not always100% positive,making the data of IFC unreliable.This paper conducted classification by importing neural networks,establishing a method to automaticlly export and identify BIM elements.What’ s following are the main conclusions of this paper.(1)A method of automatically generating point clouds and multi-view representations was built up.The 3D models in Revit were exporeted in the form of obj file by Dynamo program.Based on the obj file,the automatic productions of point clouds and multi-view representations were conducted while severe tools,PCL,Python and Matlab,were used during the process.Comapred with former way to obtin data after manual model building in Blender,the new method directly make the full use of existing BIM molel,which is confirming to the concept of BIM to save effort by applying the same model in all life circle of the construction.(2)This paper built up with a way of automatic classification about BIM element using3 D neural networks.After applying Matlab and Dynamo,along with other tools such as Visual Basic and bat,the data required by the 3D neural networks was obtianed,and pre-processing on the data was performed.Then,two original data sets were seprated into train set and test set in the rate of 7: 3 and inputed into the neural networks.(3)Several tests of specific clauses were carried out on the case of door and wall element.The results confirmed that three kinds of neural networks,Point Net,Point Net++ and MVCNN,are all able to function as a viable solution to distinguishing BIM element subtypes.When it comes to door element,the three models had excellent classification performance with ACC ’ s of 93%,94% and 92% for Point Net,Point Net++ and MVCNN,respectively.The F1 score was the same as the value of ACC.Revolving Door types had the highest AUC,and Sliding Door types had relatively lower performance,but it seems that all learning models had the most difficulty in distinguishing the Double Door types.In the terms of classification of wall subtypes,MVCNN still had the best performance during the three neural networks.Point Net recorded lower precision-recall values than Point Net,particularly for Wall with window opening(AUC= 52%)and Wall with door opening(AUC = 57%)types.According to local distinctions,certain neural network is supposed to be properly chosen,improving efficiency while the accuracy is promised.