Research On BIM Component Collocation Recommendation Method Based On Deep Learning

Construction industry is one of the important pillar industries of China’s national economy.For a long time,the traditional construction industry has some pain points,such as extensive production management,insufficient cross-disciplinary collaboration ability,excessive resource consumption,etc.,and it is difficult to meet the needs of green and high-quality development of the construction industry in the new era.Digital architecture is the product of integrating the new generation information technology such as Building Information Modeling(BIM),big data and artificial intelligence with the elements of the whole life cycle of architecture.The development of digital architecture is an inevitable requirement to promote the transformation and upgrading of the construction industry,improve the construction level and quality,and achieve the "double carbon" strategic goal.BIM model is the carrier of building data,covering all the data from planning and design to construction operation and maintenance.It provides a reliable basis for scientific decision-making.Therefore,BIM technology is the foundation of digital architecture.At present,BIM technology has been widely used in architectural design,construction and operation and maintenance.A large number of BIM components are built and stored in order to shorten the design time.It is increasingly urgent for designers to obtain similar components quickly and accurately.Whether it is the online shared BIM component library or the designer’s own component library,they can only manually find the components that meet the requirements through the classified catalogue or search engine established by text information,and open them one by one for confirmation.This method takes a lot of time and manpower.Besides,BIM interior design is an indispensable part of architectural design.Coordinating the styles of components in the scene is the fundamental principle of BIM interior design.However,style is a high-level semantic concept,which is difficult to describe clearly by text information,so it is impossible to accurately retrieve BIM components with similar styles for designers.Obviously,the failure to obtain BIM components quickly and accurately will directly reduce the efficiency of BIM interior design.The recommendation algorithm can filter and sort the massive BIM components,and actively recommend the component resources that conform to the style consistency for interior designers.In this paper,the building information model,deep learning,knowledge map and collaborative filtering recommendation algorithm are combined,and a BIM component recommendation algorithm based on the similarity of deep style features is studied from the perspective image of components.On this basis,the problems of using this recommendation algorithm in actual scenarios are analyzed,and a knowledge-driven BIM component collocation recommendation algorithm is proposed.The main research results are as follows:(1)BIM component style feature representation based on deep learning.Most existing3 D models use geometric shapes to describe style information,but this way will lose a lot of texture and color information of the models,which will easily lead to inaccurate recommendation.BIM component images can describe style features from geometric shape,color,texture and other aspects.The method comprises the following steps: firstly,designing a BIM feature map selection model,fixing observation points with different angles to obtain a plurality of BIM component images,and setting an evaluation function to select the perspective images that can be correctly classified and contain enough information as BIM feature maps;Then,the style feature extraction model is built,and the BIM feature map is input into the convolution neural network,and the convolution features of different layers are extracted to calculate the Gram matrix,so as to obtain the style vector of BIM components.(2)A style consistency BIM component recommendation algorithm based on deep learning.Firstly,all components in BIM component library are subjected to BIM feature graph selection model and style feature extraction model to obtain the corresponding style vector,which is stored in the style vector library with the component name as the index;Secondly,calculating the similarity between the style vector of the BIM component to be recommended and all vectors in the style vector library;Finally,the Top-k BIM components are recommended according to the style similarity.In this paper,15 BIM furniture components were selected,and 20 volunteers were recruited to select a test set for each component,to verify the influence of style vectors obtained from different convolution layers on the recommendation results.Experimental results show that the combination style vector of BIM components can improve the performance of recommendation algorithm.(3)BIM component collocation knowledge map construction.In practical application,besides recommending a single BIM component,it is also necessary to consider recommending components for users according to the existing BIM scene style.Scenery style is composed of the component styles it contains.However,the proportion of each component style in the scene style depends on the matching degree between it and the recommended results.In this paper,firstly,the furniture collocation data is generated from the real scene graph by the target detection algorithm,and then the collocation rules among different types of furniture are obtained by the association mining algorithm,and the confidence of the rules is used as the style influence weight.Then,the entity and relationship of BIM component classification information and furniture collocation rules are extracted,and the knowledge map of BIM component collocation is stored in Secondary database.(4)A knowledge-driven BIM component collocation recommendation algorithm.Firstly,the BIM scene style is obtained by combining the style vector of BIM components with the influence weight of style stored in collocation knowledge map.Secondly,the similarity between BIM scene style and any vector in the style vector library is calculated;Finally,the recommendation results are sorted according to the style similarity.In this paper,10 BIM scenarios are composed,and test sets are set for each scenario to verify the influence of using single component style recommendation and overall scenario style recommendation in BIM scenarios on the results.The experimental results show that the knowledge-driven BIM scene style can more accurately recommend the most BIM components that meet the requirements.At the same time,the knowledge map also brings interpretability to the recommendation results.