A BIM-based building indoor decoration modeling method

Abstract

The application discloses a kind of based on BIM's building indoor decoration modeling method, it is related to decoration modeling technical field, by using sensor obtains point cloud data and creates building structure model M_b, ensure the reliability of basic model. The preliminary building decoration model M_ini is generated by using AI technology to analyze user reference picture, and the individual style demand is responded quickly. New building decoration model M_adj is generated by introducing natural language interaction adjustment, and the technical threshold of user participation in design modification is reduced. The physical performance simulation analysis such as light and air circulation is integrated into the design process, and the evaluation decision D_eva is obtained based on the simulation results and the comprehensive score S_fin calculation. According to the evaluation decision D_eva, optimization is carried out, so that the design is no longer limited to the aesthetic level, but also takes into account the objective performance indicators such as health and comfort, forming an efficient and intelligent indoor decoration design new paradigm, improving design quality, efficiency and user experience.

Description

Technical Field

[0001] This invention relates to the field of interior decoration modeling technology, specifically a BIM-based method for modeling architectural interior decoration. Background Technology

[0002] In today's wave of digital transformation in the architecture, engineering, and construction industries, BIM technology plays a core driving role. BIM is not merely 3D visualization; it's a digital representation and management method integrating information from the entire lifecycle of building design, construction, and operation. Through parametric components and collaborative platforms, it significantly improves the efficiency and quality of engineering projects. However, when we focus on the interior decoration stage, which is closely related to user experience within the building lifecycle, we find that the depth of BIM technology application and its level of intelligence still have room for improvement. Interior design involves complex finishing materials, furniture arrangement, equipment placement, color matching, and personalized style requirements. How to efficiently and accurately create, modify, and evaluate these decoration models within a BIM environment has become a specific and urgent need.

[0003] While BIM technology offers advantages such as 3D visualization, information integration, and parametric design for building interiors, current widely adopted BIM-based modeling methods for building interiors still reveal several shortcomings in practice. First, when dealing with existing building renovations or high-end interior design projects, the design starting point often relies on 2D drawings, requiring significant time and manpower for on-site manual measurement and model making. This is not only inefficient but also makes it difficult to guarantee the accuracy of the structural model. Second, in the design concept generation stage, traditional methods struggle to quickly respond to users' personalized, often visually inspired, style requirements. Designers typically need to rely on personal experience and extensive manual work to experiment with different material, color, and furniture combinations, transforming the "style" conveyed by a reference image into specific BIM model elements—a time-consuming and highly subjective process. Furthermore, the adjustment and optimization of design schemes usually depend on designers' proficiency in complex BIM software interfaces. For non-professional users such as owners, the interaction threshold is high, the communication feedback chain is long, and repeated modifications are not only inefficient but also prone to information transmission errors. Finally, and very importantly, existing modeling processes often focus on geometric representation and aesthetic effects, while neglecting and evaluating the physical performance of the indoor environment. There is a lack of effective means to closely integrate performance analysis with model adjustments in the early stages of design. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a BIM-based method for modeling building interior decoration, which solves the problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a BIM-based method for modeling architectural interior decoration, comprising the following steps:

[0006] S1. Obtain the original three-dimensional point cloud data of the building structure through sensors, preprocess the original three-dimensional point cloud data to obtain standardized three-dimensional point cloud data PCD, perform geometric fitting based on the standardized three-dimensional point cloud data PCD, and create the building structure model M_b.

[0007] S2. Based on the building structure model M_b and the reference images provided by the user, a style feature extraction algorithm is constructed using a convolutional neural network to extract the features of the reference images to obtain the style feature vector V_sty. A preliminary building decoration model Mini is obtained by using a 3D style transfer model of a generative adversarial network.

[0008] S3. Based on the preliminary building decoration model M_ini, the user inputs the natural language command Cmd according to the modification opinions. The NLP model is used to parse Cmd to generate the structured operation command Op_str. The new building decoration model M_adj is generated according to the operation command Op_str.

[0009] S4. Based on the new building decoration model M_adj, conduct simulation analysis of illumination and air circulation, and construct algorithms for illumination and air circulation performance indicators to calculate illumination score S_lig and air circulation score S_air;

[0010] S5. Based on the illumination score S_lig and the air circulation score S_air, construct a comprehensive scoring algorithm for the new building decoration model M_adj, calculate the comprehensive score S_fin and compare it with the preset scoring threshold T_acc to obtain the evaluation decision D_eva;

[0011] S6. Based on the evaluation decision D_eva, adjust the new building decoration model M_adj and obtain the final scheme M_fin based on the processing result.

[0012] Preferably, S1 includes S11 and S12;

[0013] S11. By using an RGB-D depth image sensor to scan multiple interior spaces of the building, multiple original 3D point cloud data are obtained. A radius filter is used to remove noise points and outliers from the multiple original 3D point cloud data. The 3D point cloud data of the living room area after removing noise points and outliers is used as the identification common point. The iterative nearest point algorithm is applied to align the multiple original 3D point cloud data into the same coordinate system to obtain standardized 3D point cloud data PCD.

[0014] Preferably, in step S12, based on standardized 3D point cloud data PCD, the geometric primitives implicit in the standardized 3D point cloud data PCD are detected by the RANSAC set recognition and fitting algorithm, including planes, cylinders and spheres. Points belonging to the same geometric primitive are segmented and clustered. Based on the geometric features and spatial relationships of the segmented and clustered points, component types are selected from the BIM template component library for instantiation. These component instances are created in the BIM environment to generate a preliminary building structure model M_b.

[0015] Preferably, S2 includes S21 and S22;

[0016] S21. Based on the reference image provided by the user, the size of the reference image is adjusted and normalized to obtain the style reference image I_ref. The processed style reference image I_ref is input into the convolutional neural network for forward calculation, and feature maps are extracted from the pre-selected layer. The extracted feature maps are vectorized and concatenated to construct a style feature extraction algorithm, and finally form the style feature vector V_sty.

[0017] The style feature extraction algorithm is expressed as follows:

[0018] V_sty=ExtFea_CNN(I_ref,Model_CNN);

[0019] In the formula, ExtFea_CNN represents the set of algorithms used in the entire process of extracting style features using CNN, and Model_CNN represents the convolutional neural network model provided by the system for feature extraction.

[0020] Preferably, in step S22, based on the spatial information of the building structure model M_b and combined with the style feature vector V_sty, the spatial information of the building structure model M_b and the style feature vector V_sty are analyzed using a 3D style transfer model of generative adversarial network. Guided by the style feature vector V_sty, a preliminary decoration scheme is obtained under the constraints of the building structure model M_b, and a preliminary building decoration model Mini is output.

[0021] The expression for the 3D style transfer model is as follows:

[0022] M_ini=Gen_Style3D(M_b, V_sty, Lib);

[0023] In the formula, Gen_Style3D represents the 3D style transfer model, and Lib represents the decoration material library provided by the system.

[0024] Preferably, S3 includes S31 and S32;

[0025] S31. Based on the preliminary building decoration model M_ini, all furniture is named and assigned a unique identifier to obtain the furniture set P = {p_1, p_2, ..., p_m}. The user inputs a natural language instruction Cmd for the furniture p_i with the unique identifier i according to their modification opinions. A natural language instruction analysis algorithm is constructed to analyze the natural language instruction Cmd input by the user, identify the user intent I, the target object set O, and the target attribute set A, and convert the parsed user intent I, the target object set O, and the target object attribute set A into a structured data format for computer execution to obtain the structured operation instruction Op_str.

[0026] The natural language instruction analysis algorithm flow is as follows:

[0027] First, the natural language instruction Cmd is preprocessed and its features are encoded:

[0028] Tok=segment(Cmd), POS=Tag(Tok);

[0029] DepTree=DependencyParser(Cmd);

[0030] E = [Embed(Tok)];

[0031] In the formula, segment represents the word segmentation model, Tok represents the segmented word sequence, Tag represents the part-of-speech tagging model, POS represents the part-of-speech tagging result, DependencyParser represents the dependency parsing model, DepTree represents the syntactic dependency tree, Embed represents the word vector embedding function, and E represents the word vector sequence.

[0032] Secondly, intent recognition and classification are performed based on the word vector sequence E, the syntactic dependency tree DepTree, and the part-of-speech tag POS:

[0033] H=BiGRU(E, DepTree, POS);

[0034]

[0035] I = argmax(Wc*C + Bc);

[0036] In the formula, BiGRU represents a bidirectional gated recurrent unit, H represents the context feature matrix output by the bidirectional gated recurrent unit BiGRU, Spftmax represents the activation function, Hj represents the context feature of the j-th word, αj represents the attention weight of the j-th word, C represents the weighted context feature vector, I represents the intent category, Wa represents the attention layer weight, Wc represents the classification layer weight, Ba represents the attention layer bias term, and Bc represents the classification layer bias term.

[0037] Next, the target object and attributes are extracted based on the word vector sequence E, the syntactic dependency tree (DepTree), the part-of-speech tag (POS), and the intent category I:

[0038] Lab=CRF(E, DepTree, POS), Ocan=ExtractEmtities(Lab);

[0039] oq∈Ocan, aq=Match(oq, DepTree, Tok, I);

[0040] O={oq|oq∈Lib_BIM}, A={aq|aq∈Attr};

[0041] In the formula, CRF represents Conditional Random Field, Lab represents Entity Boundary Label, ExtraactEmtities represents Entity Extraction Function, Ocan represents Candidate Target Object Set, oq represents the q-th Candidate Target Object, aq represents the Attribute Key-Value Pair of the q-th Candidate Target Object, Match represents Attribute Matching Function, Lib_BIM represents Valid Component Set, Attr represents Legal Attribute Set, O represents Target Object, and A represents Target Attribute.

[0042] Finally, the target object O and target attribute A are populated into a predefined instruction template according to the user intent I for structured output, resulting in the structured operation instruction Op_str:

[0043] Based on user intent, templates are divided into Move and Replace templates.

[0044] Move template: {"Type":"Move","Target":"$O","Params":"$A"};

[0045] Replace template: {"Type":"Replace","Target":"$O","Params":"$A"};

[0046] Output structured operation instructions Op_str based on the filled template;

[0047] Where Move represents the movement intent, Replace represents the replacement intent, Type represents the user intent field, Target represents the target object field, $O is the target object to be filled in, Params represents the target attribute field, and $A represents the target attribute to be filled in.

[0048] Preferably, in step S32, based on the structured operation instruction Op_str and the preliminary building decoration model M_ini, the structured operation instruction Op_str is analyzed, the API interface of the BIM software application is called, the analyzed structured operation instruction Op_str and the preliminary building decoration model M_ini to be modified are input into the BIM software, the actual modification operation is performed on the preliminary building decoration model M_ini to be modified, and finally a new building decoration model M_adj reflecting the user's modification intention is generated;

[0049] The generation process of the new building decoration model M_adj is as follows:

[0050] First, the structured operation instruction Op_str is parsed and its parameters are extracted:

[0051] Detecting the user intent field Type yields the action type set ACT; detecting the target object field Target yields the operation target object set AAO; detecting the target attribute field yields the operation target attribute set TA.

[0052] Secondly, target object location and attribute extraction are performed on the preliminary architectural decoration model M_ini:

[0053] Based on the preliminary architectural and decoration model M_ini, the GetGeometry function is called to locate and target the geometric state set G of the object set AAO. The geometric state set G includes position coordinates, rotation angle and size. The GetMaterial function is called to extract the attribute state set M of the object set AAO. The attribute state set M includes color, texture and material.

[0054] The expressions for target object location and attribute extraction are as follows:

[0055] G = Mini.GetGeometry(AAO);

[0056] M = Mini.GetMaterial(AAO);

[0057] Next, the initial architectural decoration model M_ini was modified using parametric modeling:

[0058] If the Type attribute of the user intent field is valued as move, then the coordinate transformation of the target object set AAO is performed based on the geometric state set G and the target attribute set TA.

[0059] If the Type attribute of the user intent field is valued as Replace, then the attribute of the target object set AAO is replaced according to the attribute state set M and the target attribute set TA.

[0060] Finally, the modified geometric state set G and attribute state set M are written into the preliminary building decoration model M_ini, and the preliminary building decoration model M_ini is updated to generate a new building decoration model M_adj that reflects the user's modification intention.

[0061] Preferably, S4 includes S41;

[0062] S41. Based on the new building decoration model M_adj, using the building physics simulation engine of BIM software, combined with the geometric information, material properties and set environmental parameters P_env of the new building decoration model M_adj, perform lighting simulation and air circulation simulation to obtain the lighting simulation result R_lig and the air circulation simulation result R_air.

[0063] R_lig=Sim_Lig(M_adj, P_env, Engine_lig, Psim_lig);

[0064] R_air=Sim_Air(M_adj, P_env, Engine_air, Psim_air);

[0065] In the formula, Sim_Lig represents the simulation process executed by the lighting simulation engine, Engine_lig represents the lighting simulation engine, Psim_lig represents the preset lighting simulation control parameters, Sim_Air represents the simulation process executed by the air circulation simulation engine, Engine_air represents the air circulation simulation engine, and Psim_air represents the preset air circulation simulation control parameters.

[0066] Based on the illumination simulation results R_lig and the air circulation simulation results R_air, the average illuminance E_avg and the illuminance uniformity value U_O are extracted from the illumination simulation results R_lig, and an illumination performance index algorithm is constructed to calculate the illumination score S_lig. The air exchange rate ACH and the regional average wind speed V_avg are extracted from the air circulation simulation results R_air, and an air circulation performance index algorithm is constructed to calculate the air circulation score S_air.

[0067] The algorithm expression for the illumination performance index is as follows:

[0068]

[0069] In the formula, w1 represents the preset weight of the average illuminance term, w2 represents the preset weight of the illuminance uniformity term, E_tar represents the target average illuminance, U_tar represents the target illuminance uniformity, and exp represents the natural exponential function.

[0070] The algorithm expression for the air circulation performance index is as follows:

[0071]

[0072] S_air = S_air1 + S_air2;

[0073] In the formula, S_air1 represents the air exchange scoring item, S_air2 represents the regional average wind speed scoring item, w3 represents the preset weight of the air exchange scoring item, w4 represents the preset weight of the regional average wind speed scoring item, Kach represents the preset air exchange rate sensitivity coefficient, Kv represents the preset regional wind speed penalty coefficient, ACH_atr represents the target air exchange rate, V_max represents the upper limit of the set comfortable wind speed range, and V_min represents the lower limit of the set comfortable wind speed range.

[0074] Preferably, S5 includes S51;

[0075] S51. Based on the illumination score S_lig and the air circulation score S_air, count the number of furniture and decorative elements N_adj set in the new building decoration model M_adj, construct a comprehensive scoring algorithm for the new building decoration model M_adj, and calculate the comprehensive score S_fin.

[0076] The comprehensive scoring algorithm expression is as follows:

[0077]

[0078] In the formula, A_adj represents the area of ​​the new building decoration model M_adj. D_tar represents the element density per unit area, D_tar represents the recommended element density per unit area, and tanh represents the hyperbolic tangent function.

[0079] The comprehensive score S_fin is compared with the preset score threshold T_acc to obtain the evaluation decision D_eva;

[0080] If the comprehensive score S_fin ≥ the preset score threshold T_acc, then the functionality of the current new building decoration model M_adj is determined to meet the requirements, and the evaluation decision D_eva is assigned the value Accept.

[0081] If the overall score S_fin < the preset score threshold T_acc, then the functionality of the current new building decoration model M_adj is determined to be unacceptable, and the evaluation decision D_eva is assigned the value Revise.

[0082] Preferably, S6 includes S61;

[0083] S61. Adjust the new building decoration model M_adj based on the evaluation decision D_eva;

[0084] If the evaluation decision D_eva value is Accept, then the new building decoration model M_adj is determined to be unnecessary to adjust and the current new building decoration model M_adj is marked as the final building decoration model scheme M_fin;

[0085] If the evaluation decision D_eva value is Revise, then the comprehensive scoring algorithm based on the new building decoration model M_adj is used to analyze the results of the scoring items and generate feedback information. This guides the user to communicate with the designer and returns to S3 to adjust the new building decoration model M_adj. The adjusted building decoration model with the evaluation decision D_eva value of Accept is marked as the final building decoration model scheme M_fin.

[0086] This invention provides a BIM-based method for modeling building interior decoration, which has the following beneficial effects:

[0087] (1) During system operation, point cloud data is acquired by sensors and structural models are automatically created, ensuring that the starting point of the design work is based on real physical space. This avoids subsequent design errors and rework caused by relying on drawings or manual measurement, thus improving the reliability of the basic model. Utilizing AI technology to analyze user reference images and automatically generate preliminary stylized schemes greatly shortens the conversion time from abstract inspiration to concrete visual schemes, overcoming the efficiency bottleneck of traditional design and quickly responding to personalized style demands. Introducing natural language interaction significantly lowers the technical threshold for users to participate in design modifications, making the scheme fine-tuning process as intuitive and convenient as a dialogue, improving interaction efficiency and user satisfaction. Integrating simulation analysis of key physical performance such as lighting and air circulation into the design process, and conducting automated comprehensive evaluation and decision-making based on simulation results and preset standards, ensures that the design is no longer limited to aesthetics but also takes into account objective performance indicators such as health and comfort, achieving data-driven scientific decision-making. Based on the evaluation results, scheme optimization or finalization is performed, forming a complete, efficient, and intelligent new paradigm for interior decoration design, from accurate input to intelligent generation, convenient interaction, scientific evaluation, and closed-loop optimization, comprehensively improving design quality, efficiency, and user experience.

[0088] (2) By directly capturing 3D data from physical reality and generating accurate BIM structural models, a highly reliable digital foundation is provided for all subsequent design work. This overcomes the problems of dimensional deviations and information gaps that may arise from relying on traditional 2D drawings, ensuring accurate matching between the decoration design and the building entity, and significantly reducing later construction problems caused by errors in the model source. The introduced AI style transfer technology can quickly understand the vague style preferences expressed by users through reference images and automatically convert them into preliminary BIM decoration schemes, greatly improving the speed and diversity of design concept generation, effectively stimulating design inspiration, and shortening the project start-up cycle. The natural language processing technology adopted completely changes the way design schemes are adjusted. Users do not need to master complex BIM software operations; they can modify the model simply through voice or text commands. This interactive mode greatly reduces the design threshold, promotes more direct and efficient communication and collaboration between designers and owners, and makes the expression and realization of personalized needs easier and faster than ever before, significantly accelerating the iteration and convergence process of design schemes.

[0089] (3) By embedding simulation analysis of key physical properties such as lighting and air circulation into the design process, and using preset algorithms to transform the raw simulation data into standardized performance scores, designers and users can obtain objective and quantitative feedback on the comfort and health aspects of the design scheme early in the design process. This avoids relying solely on subjective feelings or delayed post-design verification, enabling more informed design decisions based on data and allowing for the early identification and avoidance of potential performance defects, such as insufficient indoor lighting or poor natural ventilation. Subsequent actions are performed based on the evaluation decisions. If the scheme fails to meet the standards, users are guided back to the adjustment steps for targeted optimization; if the scheme meets the standards, it is confirmed. This scoring and optimization mechanism ensures that the final decoration scheme is not only aesthetically pleasing and personalized, but also validated and guaranteed in terms of key physical properties, thereby significantly improving the actual usability and comfort of the built space. Attached Figure Description

[0090] Figure 1 This is a schematic diagram illustrating the steps of a BIM-based architectural interior decoration modeling method according to the present invention.

[0091] Figure 2 This is a schematic diagram of the natural language instruction processing of the present invention;

[0092] Figure 3 This is a schematic diagram of the data processing flow of the scoring algorithm of this invention. Detailed Implementation

[0093] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0094] Example 1

[0095] This invention provides a BIM-based method for modeling architectural interior decoration. Please refer to [link / reference]. Figure 1 This includes the following steps:

[0096] S1. Obtain the original three-dimensional point cloud data of the building structure through sensors, preprocess the original three-dimensional point cloud data to obtain standardized three-dimensional point cloud data PCD, perform geometric fitting based on the standardized three-dimensional point cloud data PCD, and create the building structure model M_b.

[0097] S2. Based on the building structure model M_b and the reference images provided by the user, a style feature extraction algorithm is constructed using a convolutional neural network to extract the features of the reference images to obtain the style feature vector V_sty. A preliminary building decoration model Mini is obtained by using a 3D style transfer model of a generative adversarial network.

[0098] S3. Based on the preliminary building decoration model M_ini, the user inputs the natural language command Cmd according to the modification opinions. The NLP model is used to parse Cmd to generate the structured operation command Op_str. The new building decoration model M_adj is generated according to the operation command Op_str.

[0099] S4. Based on the new building decoration model M_adj, conduct simulation analysis of illumination and air circulation, and construct algorithms for illumination and air circulation performance indicators to calculate illumination score S_lig and air circulation score S_air;

[0100] S5. Based on the illumination score S_lig and the air circulation score S_air, construct a comprehensive scoring algorithm for the new building decoration model M_adj, calculate the comprehensive score S_fin and compare it with the preset scoring threshold T_acc to obtain the evaluation decision D_eva;

[0101] S6. Based on the evaluation decision D_eva, adjust the new building decoration model M_adj and obtain the final scheme M_fin based on the processing result.

[0102] In this embodiment, a laser scanner is used to capture the actual spatial form of the living room, generating raw 3D point cloud data. After preprocessing such as automatic denoising, registration, and downsampling, standardized 3D point cloud data PCD is obtained. Based on this PCD, the system automatically identifies the main structural surfaces such as walls, floors, and ceilings through geometric fitting, and creates an architectural model M_b in the BIM environment that is highly consistent with physical reality, effectively overcoming the error risks caused by relying on inaccurate drawings or manual measurements in the past. Subsequently, the user provides a reference image I_ref that matches their modern minimalist style preference. The system uses a convolutional neural network to analyze the reference image I_ref, extract its deep features, and encode them into a style feature vector V_sty, which quantifies the visual style desired by the user. Next, a 3D style transfer model based on a generative adversarial network is launched. It uses the accurate architectural model M_b as spatial constraints, the style feature vector V_sty as aesthetic guidance, and combines the built-in BIM component library to quickly generate a preliminary architectural decoration model M_ini containing preliminary decoration elements, greatly shortening the conversion time from abstract style to specific scheme. After reviewing the M_ini, if users need to make adjustments, they can interact through the natural language command Cmd. The integrated NLP model will parse this Cmd and generate a structured operation command Op_str, which will then drive the BIM software API to accurately modify the model M_ini and obtain the new building decoration model M_adj after the user's adjustments. This interaction method significantly reduces the participation threshold for non-professional users.

[0103] After the scheme is adjusted, the system automatically performs a performance evaluation on the new building decoration model M_adj, calling the simulation engine to calculate the lighting score S_lig and the air circulation score S_air. These scores are then substituted into a preset comprehensive scoring formula to calculate the comprehensive score S_fin. This comprehensive score S_fin is compared with a preset acceptance threshold T_ac, and the system arrives at an evaluation decision D_eval. Finally, based on the evaluation decision D_eval, the system decides whether to optimize the current new building decoration model M_adj and determines the final scheme M_fin, completing a full design iteration from accurate modeling, AI-assisted creative generation, convenient interactive adjustments to automated performance and style comprehensive evaluation, ensuring that the final result meets both personalized needs and objective performance standards.

[0104] Example 2

[0105] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically: S1 includes S11 and S12;

[0106] S11. By using an RGB-D depth image sensor to scan multiple interior spaces of the building, multiple original 3D point cloud data are obtained. A radius filter is used to remove noise points and outliers from the multiple original 3D point cloud data. The 3D point cloud data of the living room area after removing noise points and outliers is used as the identification common point. The iterative nearest point algorithm is applied to align the multiple original 3D point cloud data into the same coordinate system to obtain standardized 3D point cloud data PCD.

[0107] S12. Based on standardized 3D point cloud data PCD, the implicit geometric primitives in the standardized 3D point cloud data PCD are detected by the RANSAC set recognition and fitting algorithm, including planes, cylinders and spheres. Points belonging to the same geometric primitive are segmented and clustered. According to the geometric characteristics and spatial relationships of the segmented and clustered points, component types are selected from the BIM template component library for instantiation. These component instances are created in the BIM environment to generate a preliminary building structure model M_b.

[0108] In this embodiment, RGB-D depth image sensors are used to scan multiple spaces inside the building to acquire raw 3D point cloud data. A radius filter is then used to remove noise and outliers. The iterative nearest point algorithm is applied to align multiple point cloud data sets to a unified coordinate system, thereby generating standardized 3D point cloud data (PCD). This ensures high-precision, comprehensive digital capture of the physical reality of the existing building structure, resulting in reliable standardized PCD. Based on this, the RANSAC ensemble recognition and fitting algorithm is further used to detect implicit geometric primitives, such as planes, cylinders, and spheres, from the standardized PCD. Points belonging to the same primitive are segmented and clustered. Based on the geometric features and spatial relationships of these points, corresponding component types are selected from the BIM template component library for instantiation. Finally, a preliminary building structure model M_b containing precise geometric dimensions and location information is created in the BIM environment. This method automates the conversion from unstructured, standardized 3D point cloud data (PCD) to a structured, parametric preliminary building structure model (M_b). In particular, through integration with a BIM template component library, it directly generates BIM component instances with semantic information and correct geometric attributes. This workflow, from precise scanning to intelligent recognition and then to automated BIM component instantiation, significantly improves the accuracy and automation level of the initial building structure model creation stage.

[0109] Example 3

[0110] This embodiment is an explanation based on Embodiment 2. Please refer to it. Figure 1 and Figure 2 Specifically: S2 includes S21 and S22;

[0111] S21. Based on the reference image provided by the user, the size of the reference image is adjusted and normalized to obtain a style reference image I_ref that meets the input requirements of the convolutional neural network. The processed style reference image I_ref is input into the convolutional neural network for forward computation, and feature maps are extracted from the pre-selected layer. The extracted feature maps are vectorized and concatenated to construct a style feature extraction algorithm, and finally form a style feature vector V_sty.

[0112] The style feature extraction algorithm is expressed as follows:

[0113] A Q =GAM(Model_CNN, I_ref, Q);

[0114]

[0115] In the formula, Model_CNN represents an existing mature convolutional neural network for feature extraction, Q_sty represents a pre-defined set of CNN layers for style feature extraction, Q represents a specified layer identifier of the CNN layer set, GAM represents the process of inputting the style reference image I_ref into the mature convolutional neural network Model_CNN for feature extraction and obtaining the activation map output of the Qth layer, A Q Indicates dimension C Q *H Q *W Q The activation graph of the Q-th layer, C Q H represents the number of channels. Q W represents height. Q F represents the width. Q Indicates that by A Q The decomposition size is H Q *W Q A two-dimensional matrix, (F Q ) T F represents Q The transpose matrix, G Q Let represent the Gram matrix of the Q-th layer, vec denotes matrix vectorization, and Concat denotes vector concatenation.

[0116] The purpose of this formula is to quantify the abstract style expressed by the user-provided reference image and transform it into a computable representation. Directly using the pixel information of the original image cannot effectively separate style from content. Therefore, deep features are extracted through a convolutional neural network, and the correlation of these features is quantified using the Gram matrix G. Style information from different layers is vectorized and concatenated through the Gram matrix GQ representation, which can comprehensively capture style features at different scales and ultimately form a comprehensive style feature vector V_sty.

[0117] S22. Based on the spatial information of the building structure model M_b and combined with the style feature vector V_sty, the 3D style transfer model of generative adversarial network is used to analyze the spatial information of the building structure model M_b and the style feature vector V_sty. According to the style feature vector V_sty as guidance, a preliminary decoration scheme is obtained under the constraints of the building structure model M_b and a preliminary building decoration model Mini is output.

[0118] The expression for the 3D style transfer model is as follows:

[0119] M_ini=Gen_Style3D(M_b, V_sty, Lib);

[0120] In the formula, Gen_Style3D represents an existing mature 3D style transfer model, and Lib represents the decoration material library provided by the system;

[0121] S3 includes S31 and S32;

[0122] S31. Based on the preliminary building decoration model M_ini, all furniture is named and assigned a unique identifier to obtain the furniture set P = {p_1, p_2, ..., p_m}. The user inputs a natural language instruction Cmd for the furniture p_i with the unique identifier i according to their modification opinions. A natural language instruction analysis algorithm is constructed to analyze the natural language instruction Cmd input by the user, identify the user intent I, the target object set O, and the target attribute set A, and convert the parsed user intent I, the target object set O, and the target object attribute set A into a structured data format for computer execution to obtain the structured operation instruction Op_str.

[0123] The natural language instruction analysis algorithm flow is as follows:

[0124] First, the natural language instruction Cmd is preprocessed and its features are encoded:

[0125] Tok=segment(Cmd), POS=Tag(Tok);

[0126] DepTree=DependencyParser(Cmd);

[0127] E = [Embed(Tok)];

[0128] In the formula, segment represents the word segmentation model, Tok represents the segmented word sequence, Tag represents the part-of-speech tagging model, POS represents the part-of-speech tagging result, DependencyParser represents the dependency parsing model, DepTree represents the syntactic dependency tree, Embed represents the word vector embedding function, and E represents the word vector sequence.

[0129] The purpose of this formula is to decompose the natural language instruction Cmd into basic semantic units by using the segmentation model segmen and the part-of-speech tagging model Tag, providing structured input for subsequent analysis; to reveal the grammatical structure of the instruction and the relationship between auxiliary attributes and the target by using the dependency parsing model DependencyParser; and to convert discrete words into continuous vectors by using the word embedding function Embed, thus preserving semantic information.

[0130] Secondly, intent recognition and classification are performed based on the word vector sequence E, the syntactic dependency tree DepTree, and the part-of-speech tag POS:

[0131] H=BiGRU(E, DepTree, POS);

[0132]

[0133] I = argmax(Wc*C + Bc);

[0134] In the formula, BiGRU represents the bidirectional gated recurrent unit used to capture the contextual dependencies of instructions, H represents the context feature matrix output by the bidirectional gated recurrent unit BiGRU, Spftmax represents the activation function, Hj represents the contextual feature of the j-th word, αj represents the attention weight of the j-th word used to reflect the importance of the j-th word to the intent classification, C represents the weighted context feature vector used for classification, I represents the intent category used to clarify the user operation type and provide direction for subsequent steps, Wa represents the attention layer weight, Wc represents the classification layer weight, Ba represents the attention layer bias term, and Bc represents the classification layer bias term;

[0135] Next, the target object and attributes are extracted based on the word vector sequence E, the syntactic dependency tree (DepTree), the part-of-speech tag (POS), and the intent category I:

[0136] Lab=CRF(E, DepTree, POS), Ocan=ExtractEmtities(Lab);

[0137] oq∈Ocan, aq=Match(oq, DepTree, Tok, I);

[0138] O={oq|oq∈Lib_BIM}, A={aq|aq∈Attr};

[0139] In the formula, CRF represents Conditional Random Field, Lab represents Entity Boundary Label, ExtraactEmtities represents Entity Extraction Function, Ocan represents Candidate Target Object Set, oq represents the q-th Candidate Target Object, aq represents the Attribute Key-Value Pair of the q-th Candidate Target Object, Match represents Attribute Matching Function, Lib_BIM represents Valid Component Set, Attr represents Legal Attribute Set, O represents Target Object, and A represents Target Attribute.

[0140] Finally, the target object O and target attribute A are populated into a predefined instruction template according to the user intent I for structured output, resulting in the structured operation instruction Op_str:

[0141] Based on user intent, templates are divided into Move and Replace templates.

[0142] Move template: {"Type":"Move","Target":"$O","Params":"$A"};

[0143] Replace template: {"Type":"Replace","Target":"$O","Params":"$A"};

[0144] Output structured operation instructions Op_str based on the filled template;

[0145] Where Move represents the movement intent, Replace represents the replacement intent, Type represents the user intent field, Target represents the target object field, $O is the target object, Params represents the target attribute field, and $A represents the target attribute.

[0146] S32. Based on the structured operation instructions Op_str and the preliminary building decoration model M_ini, the structured operation instructions Op_str are analyzed, the API interface of the BIM software application is called, the analyzed structured operation instructions Op_str and the preliminary building decoration model M_ini to be modified are input into the BIM software, the actual modification operation is performed on the preliminary building decoration model M_ini to be modified, and finally a new building decoration model M_adj reflecting the user's modification intention is generated;

[0147] The generation process of the new building decoration model M_adj is as follows:

[0148] First, the structured operation instruction Op_str is parsed and its parameters are extracted:

[0149] Detecting the user intent field Type yields the action type set ACT; detecting the target object field Target yields the operation target object set AAO; detecting the target attribute field yields the operation target attribute set TA.

[0150] Secondly, target object location and attribute extraction are performed on the preliminary architectural decoration model M_ini:

[0151] Based on the preliminary architectural and decoration model M_ini, the GetGeometry function is called to locate and target the geometric state set G of the object set AAO. The geometric state set G includes position coordinates, rotation angle and size. The GetMaterial function is called to extract the attribute state set M of the object set AAO. The attribute state set M includes color, texture and material.

[0152] The expressions for target object location and attribute extraction are as follows:

[0153] G = Mini.GetGeometry(AAO);

[0154] M = Mini.GetMaterial(AAO);

[0155] Next, the initial architectural decoration model M_ini was modified using parametric modeling:

[0156] If the Type attribute of the user intent field is valued as move, then the coordinate transformation of the target object set AAO is performed based on the geometric state set G and the target attribute set TA.

[0157] If the Type attribute of the user intent field is valued as Replace, then the attribute of the target object set AAO is replaced according to the attribute state set M and the target attribute set TA.

[0158] Finally, the modified geometric state set G and attribute state set M are written into the preliminary building decoration model M_ini, and the preliminary building decoration model M_ini is updated to generate a new building decoration model M_adj that reflects the user's modification intention.

[0159] In this embodiment, a convolutional neural network is used to perform deep analysis on the style reference image I_ref provided by the user. Based on the convolutional neural network and the Gram matrix GQ, a quantified style feature vector V_sty is extracted and constructed. Subsequently, using a 3D style transfer model based on a generative adversarial network, under the spatial constraints of the building structure model M_b and the support of the decoration material library Lib provided by the system, a preliminary building decoration model M_ini that conforms to the user's style preferences is automatically generated. On this basis, to facilitate user modifications, the system names and assigns unique identifiers to the decoration furniture elements in the preliminary building decoration model M_ini, forming a decoration furniture set P. The user can input the natural language command Cmd to express their modification intentions for the furniture p_i with the unique identifier i. The system then uses an NLP model to parse the natural language command Cmd, converting it into a structured operation command Op_str that the computer can understand. Finally, by calling the BIM software API, the corresponding modification operation is performed on the preliminary building decoration model M_ini, thereby generating a new building decoration model M_adj that reflects the user's adjustment intentions. The unique advantage of this method lies in its use of convolutional neural networks and generative adversarial networks to automate and intelligently generate specific 3D decoration schemes from abstract visual style references, greatly improving the efficiency and creative transformation capabilities of the design concept stage. At the same time, the introduction of NLP models for processing natural language as the core of human-computer interaction enables non-professional users to intuitively and conveniently make personalized adjustments to the design schemes, significantly lowering the technical threshold and promoting an efficient communication and feedback loop. Its core advantage lies in significantly improving the intelligence level of design creative expression and the convenience of user interaction adjustments.

[0160] Example 4

[0161] This embodiment is an explanation based on Embodiment 3. Please refer to it. Figure 1 and Figure 3 Specifically: S41 and S42;

[0162] S41. Based on the new building decoration model M_adj, using the building physics simulation engine of BIM software, combined with the geometric information, material properties and set environmental parameters P_env of the new building decoration model M_adj, perform lighting simulation and air circulation simulation to obtain the lighting simulation result R_lig and the air circulation simulation result R_air.

[0163] R_lig=Sim_Lig(M_adj, P_env, Engine_lig, Psim_lig);

[0164] R_air=Sim_Air(M_adj, P_env, Engine_air, Psim_air);

[0165] In the formula, Sim_Lig represents the simulation process executed by the lighting simulation engine, Engine_lig represents the lighting simulation engine, Psim_lig represents the preset lighting simulation control parameters, Sim_Air represents the simulation process executed by the air circulation simulation engine, Engine_air represents the air circulation simulation engine, and Psim_air represents the preset air circulation simulation control parameters.

[0166] S42. Based on the illumination simulation results R_lig and the air circulation simulation results R_air, extract the average illuminance E_avg and the illuminance uniformity value U_O from the illumination simulation results R_lig and construct an illumination performance index algorithm to calculate the illumination score S_lig. Extract the air exchange rate ACH and the regional average wind speed V_avg from the air circulation simulation results R_air and construct an air circulation performance index algorithm to calculate the air circulation score S_air.

[0167] The algorithm expression for the illumination performance index is as follows:

[0168]

[0169] In the formula, w1 represents the preset weight of the average illuminance term, w2 represents the preset weight of the illuminance uniformity term, E_tar represents the target average illuminance, U_tar represents the target illuminance uniformity, and exp represents the natural exponential function.

[0170] The purpose of this formula is to use a natural exponential function for the average illuminance term. By amplifying the relative deviation between actual and target values ​​through exponential characteristics, the importance of key indicators is highlighted, while the degree of deviation from the target can be smoothly assessed. The illuminance uniformity value uses a linear scaling factor. It directly reflects the uniformity standard. The weighted sum of the two values ​​reasonably balances the brightness level and the distribution uniformity. The higher the S_lig value of the illumination score, the closer the illumination conditions are to the ideal target and the more uniform the distribution.

[0171] The algorithm expression for the air circulation performance index is as follows:

[0172]

[0173] S_air = S_air1 + S_air2;

[0174] In the formula, S_air1 represents the air exchange scoring item, S_air2 represents the regional average wind speed scoring item, w3 represents the preset weight of the air exchange scoring item, w4 represents the preset weight of the regional average wind speed scoring item, Kach represents the preset air exchange rate sensitivity coefficient, Kv represents the preset regional wind speed penalty coefficient, ACH_atr represents the target air exchange rate, V_max represents the upper limit of the set comfortable wind speed range, and V_min represents the lower limit of the set comfortable wind speed range.

[0175] The purpose of this formula is to achieve the air exchange scoring item S_air1 passing... Will The values ​​are mapped to scores of 0 to 1, and Kach adjusts the sensitivity to avoid a sudden drop in scores due to small deviations. The regional wind speed score item S_air2 uses a Gaussian function. The system penalizes situations where the wind speed exceeds the comfortable range, ensuring that the wind speed remains stable within a reasonable range. The higher the air circulation score S_air, the fresher the indoor air and the more comfortable the perceived wind speed.

[0176] Specific examples for calculating the illumination score S_lig and the air circulation score S_air:

[0177] The parameters are set as follows:

[0178] Analysis of the lighting simulation results R_lig and the airflow simulation results R_air yields the following results:

[0179] Average illuminance E_avg: 420, illuminance uniformity U_O: 0.68;

[0180] Air exchange rate (ACH): 5.2; Average wind speed (V_avg): 0.15.

[0181] Scoring parameters configured by the system:

[0182] Target average illuminance E_tar: 500, target uniformity U_tar: 0.7;

[0183] The weight of the average illuminance term w1 is 0.6, and the weight of the uniform illuminance term w2 is 0.4.

[0184] Target air exchange rate ACH_tar: 6.0, lower limit of comfortable airflow V_min: 0.1, upper limit of comfortable airflow V_max: 0.25;

[0185] The weight of the air exchange scoring item w3 is 0.5, and the weight of the regional average wind speed scoring item w4 is 0.5.

[0186] Air exchange rate sensitivity coefficient KachH: 1.5, wind speed score penalty coefficient Kv: 0.6;

[0187] Calculate the illumination score S_lig:

[0188]

[0189] Calculate the airflow score S_air:

[0190]

[0191] S_air=0.3638+0.2744=0.6382.

[0192] In this embodiment, by deeply embedding illumination simulation and airflow simulation into the design process and constructing a standardized performance scoring algorithm based on the simulation results, early and quantitative evaluation of the physical performance of the decoration scheme is achieved. Specifically, based on the geometric information and material properties of the new building decoration model M_adj, the system calls the illumination simulation engine and airflow simulation engine of the BIM software to calculate the illumination simulation result R_lig and the airflow simulation result R_air, respectively, and extracts key physical parameters, including average illuminance E_avg, illuminance uniformity U_O, air exchange rate ACH, and regional average wind speed V_avg. By constructing illumination performance index algorithms and airflow performance index algorithms, the system transforms the simulation data into quantifiable illumination scores S_lig and airflow scores S_air. The special advantage of this method is that it flexibly balances the priority of different performance indicators through the introduction of dynamic weights and sensitivity coefficients. For example, by adjusting the air exchange rate sensitivity coefficient Kach, the sensitivity to deviations of the air exchange rate from the target value can be strengthened, while the design of the natural exponential function can non-linearly map the deviation between the actual value and the target value, avoiding the mechanical nature of linear scoring. Meanwhile, the scoring algorithm is directly linked to the target parameters set by the user, ensuring that the evaluation criteria are consistent with the specific project requirements. Through a data-driven scientific mechanism, designers can identify and correct potential physical performance defects in the early stages of design, rather than relying on later verification. This significantly reduces rework costs, improves the comfort and health of the design solutions, and forms an intelligent design paradigm that excels in both aesthetics and performance.

[0193] Example 5

[0194] A BIM-based method for modeling architectural interior decoration is provided below. Figure 1 and Figure 3 Specifically: S5 includes S51;

[0195] S51. Based on the illumination score S_lig and the air circulation score S_air, count the number of furniture and decorative elements N_adj set in the new building decoration model M_adj, construct a comprehensive scoring algorithm for the new building decoration model M_adj, and calculate the comprehensive score S_fin.

[0196] The comprehensive scoring algorithm expression is as follows:

[0197]

[0198] In the formula, A_adj represents the area of ​​the new building decoration model M_adj. D_tar represents the element density per unit area, D_tar represents the recommended element density per unit area, and tanh represents the hyperbolic tangent function.

[0199] The purpose of this formula is to reflect the performance of the design scheme in terms of physical comfort and health by incorporating the light simulation score S_lig and the air quality simulation score S_air. By introducing scoring items related to adjacency, it quantifies the rationality of the indoor space layout, reflects whether the functional zoning is compact, and avoids the one-sidedness of evaluation by a single indicator.

[0200] The comprehensive score S_fin is compared with the preset score threshold T_acc to obtain the evaluation decision D_eva;

[0201] If the comprehensive score S_fin ≥ the preset score threshold T_acc, then the functionality of the current new building decoration model M_adj is determined to meet the requirements, and the evaluation decision D_eva is assigned the value Accept.

[0202] If the overall score S_fin < the preset score threshold T_acc, then the functionality of the current new building decoration model M_adj is determined to be unacceptable, and the evaluation decision D_eva is assigned the value Revise.

[0203] S6 includes S61;

[0204] S61. Adjust the new building decoration model M_adj based on the evaluation decision D_eva;

[0205] If the evaluation decision D_eva value is Accept, then the new building decoration model M_adj is determined to be unnecessary to adjust and the current new building decoration model M_adj is marked as the final building decoration model scheme M_fin;

[0206] If the evaluation decision D_eva value is Revise, then the comprehensive scoring algorithm based on the new building decoration model M_adj is used to analyze the results of the scoring items and generate feedback information. This guides the user to communicate with the designer and returns to S3 to adjust the new building decoration model M_adj. The adjusted building decoration model with the evaluation decision D_eva value of Accept is marked as the final building decoration model scheme M_fin.

[0207] In this embodiment, a dynamic balance between functionality, aesthetic density, and user needs is achieved in the decoration scheme by constructing a comprehensive scoring algorithm and a closed-loop optimization mechanism. Specifically, the system calculates a comprehensive score S_fin based on the lighting score S_lig and the air circulation score S_air, combined with the decorative element density score. After comparing the comprehensive score S_fin with a preset scoring threshold T_acc, an evaluation decision D_eva is generated. If the evaluation decision D_evaRevise, the system automatically analyzes the shortcomings of the scoring items, such as S_lig or S_air not meeting the standard, or element density exceeding the limit, and generates feedback information to guide the user back to the adjustment steps. The special advantage of this method is that it quantifies physical performance indicators and decorative aesthetic density in a unified manner through a comprehensive scoring algorithm, and introduces the hyperbolic tangent function tanh to flexibly penalize density deviations, which avoids the rigidity of linear threshold judgment and prevents the spatial oppression caused by excessive decoration. Furthermore, it achieves synergistic optimization of multi-dimensional indicators. When the illumination score S_lig meets the standard but the element density score is lowered by excessive decoration, the system will clearly prompt that the number of decorations should be reduced, rather than simply optimizing physical performance, thereby ensuring that the final solution M_fin achieves global optimization in terms of functionality, comfort and visual aesthetics.

[0208] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A BIM-based method for modeling architectural interior decoration, characterized in that: Includes the following steps: S1. Obtain the original three-dimensional point cloud data of the building structure through sensors, preprocess the original three-dimensional point cloud data to obtain standardized three-dimensional point cloud data PCD, perform geometric fitting based on the standardized three-dimensional point cloud data PCD, and create the building structure model M_b. S2. Based on the building structure model M_b and the reference images provided by the user, a style feature extraction algorithm is constructed using a convolutional neural network to extract the features of the reference images to obtain the style feature vector V_sty. A preliminary building decoration model Mini is obtained by using a 3D style transfer model of a generative adversarial network. S3. Based on the preliminary building decoration model M_ini, the user inputs the natural language command Cmd according to the modification opinions. The NLP model is used to parse Cmd to generate the structured operation command Op_str. The new building decoration model M_adj is generated according to the operation command Op_str. S4. Based on the new building decoration model M_adj, conduct simulation analysis of illumination and air circulation, and construct algorithms for illumination and air circulation performance indicators to calculate illumination score S_lig and air circulation score S_air; S5. Based on the illumination score S_lig and the air circulation score S_air, construct a comprehensive scoring algorithm for the new building decoration model M_adj, calculate the comprehensive score S_fin and compare it with the preset scoring threshold T_acc to obtain the evaluation decision D_eva; S5 includes S51; S51. Based on the illumination score S_lig and the air circulation score S_air, count the number of furniture and decorative elements N_adj set in the new building decoration model M_adj, construct a comprehensive scoring algorithm for the new building decoration model M_adj, and calculate the comprehensive score S_fin. The comprehensive scoring algorithm expression is as follows: ； In the formula, A_adj represents the area of ​​the new building decoration model M_adj. D_tar represents the element density per unit area, D_tar represents the recommended element density per unit area, and tanh represents the hyperbolic tangent function. The comprehensive score S_fin is compared with the preset score threshold T_acc to obtain the evaluation decision D_eva; If the comprehensive score S_fin ≥ the preset score threshold T_acc, then the functionality of the current new building decoration model M_adj is determined to meet the requirements, and the evaluation decision D_eva is assigned the value Accept. If the overall score S_fin < the preset score threshold T_acc, then the functionality of the current new building decoration model M_adj is determined to be unacceptable, and the evaluation decision D_eva is assigned the value Revise. S6. Based on the evaluation decision D_eva, adjust the new building decoration model M_adj and obtain the final scheme M_fin based on the processing result.

2. The BIM-based architectural interior decoration modeling method according to claim 1, characterized in that: S1 includes S11 and S12; S11. Use an RGB-D depth image sensor to scan multiple interior spaces of the building to obtain multiple raw 3D point cloud data. Use a radius filter to remove noise points and outliers from the multiple raw 3D point cloud data. Use the 3D point cloud data of the living room area after removing noise points and outliers as the identification common point. Apply the iterative nearest point algorithm to align the multiple raw 3D point cloud data into the same coordinate system to obtain standardized 3D point cloud data PCD.

3. The BIM-based architectural interior decoration modeling method according to claim 2, characterized in that: S12. Based on standardized 3D point cloud data PCD, the implicit geometric primitives in the standardized 3D point cloud data PCD are detected by the RANSAC fitting algorithm, including planes, cylinders and spheres. Points belonging to the same geometric primitive are segmented and clustered. According to the geometric features and spatial relationships of the segmented and clustered points, component types are selected from the BIM template component library for instantiation. These component instances are created in the BIM environment to generate a preliminary building structure model M_b.

4. The BIM-based architectural interior decoration modeling method according to claim 1, characterized in that: S2 includes S21 and S22; S21. Based on the reference image provided by the user, the size of the reference image is adjusted and normalized to obtain the style reference image I_ref. The processed style reference image I_ref is input into the convolutional neural network for forward computation and feature map is extracted. The extracted feature map is vectorized and concatenated to construct the style feature extraction algorithm, and finally the style feature vector V_sty is formed. The style feature extraction algorithm is expressed as follows: ； In the formula, ExtFea_CNN represents the set of algorithms used in the entire process of extracting style features using CNN, and Model_CNN represents the convolutional neural network model provided by the system for feature extraction.

5. The BIM-based architectural interior decoration modeling method according to claim 4, characterized in that: S22. Based on the spatial information of the building structure model M_b and combined with the style feature vector V_sty, the 3D style transfer model of generative adversarial network is used to analyze the spatial information of the building structure model M_b and the style feature vector V_sty. Using the style feature vector V_sty as a guide, a preliminary decoration scheme is obtained under the constraints of the building structure model M_b, and a preliminary building decoration model Mini is output. The expression for the 3D style transfer model is as follows: ； In the formula, Gen_Style3D represents the 3D style transfer model, and Lib represents the decoration material library provided by the system.

6. The BIM-based architectural interior decoration modeling method according to claim 5, characterized in that: S3 includes S31 and S32; S31. Based on the preliminary building decoration model M_ini, name and assign unique identifiers to obtain the decoration furniture set P={p_1, p_2, ..., p_m}. The user inputs natural language instruction Cmd for the furniture p_i with unique identifier i according to their modification opinions. Construct a natural language instruction analysis algorithm to analyze the natural language instruction Cmd input by the user, identify the user intent I, the target object set O of the operation, and the target attribute set A. Convert the parsed user intent I, the target object set O of the operation, and the target object attribute set A into a structured data format for computer execution to obtain the structured operation instruction Op_str. The natural language instruction analysis algorithm flow is as follows: First, the natural language instruction Cmd is preprocessed and its features are encoded: ； ； ； In the formula, segment represents the word segmentation model, Tok represents the segmented word sequence, Tag represents the part-of-speech tagging model, POS represents the part-of-speech tagging result, DependencyParser represents the dependency parsing model, DepTree represents the syntactic dependency tree, Embed represents the word vector embedding function, and E represents the word vector sequence. Secondly, intent recognition and classification are performed based on the word vector sequence E, the syntactic dependency tree DepTree, and the part-of-speech tag POS: ； ； ； In the formula, BiGRU represents a bidirectional gated recurrent unit, H represents the context feature matrix output by the bidirectional gated recurrent unit BiGRU, Spftmax represents the activation function, Hj represents the context feature of the j-th word, αj represents the attention weight of the j-th word, C represents the weighted context feature vector, I represents the intent category, Wa represents the attention layer weight, Wc represents the classification layer weight, Ba represents the attention layer bias term, and Bc represents the classification layer bias term. Next, the target object and attributes are extracted based on the word vector sequence E, the syntactic dependency tree (DepTree), the part-of-speech tag (POS), and the intent category I: ； ； ； In the formula, CRF represents Conditional Random Field, Lab represents Entity Boundary Label, ExtraactEmtities represents Entity Extraction Function, Ocan represents Candidate Target Object Set, oq represents the q-th Candidate Target Object, aq represents the Attribute Key-Value Pair of the q-th Candidate Target Object, Match represents Attribute Matching Function, Lib_BIM represents Valid Component Set, Attr represents Legal Attribute Set, O represents Target Object, and A represents Target Attribute. Finally, the target object O and target attribute A are populated into a predefined instruction template according to the user intent I for structured output, resulting in the structured operation instruction Op_str: Based on user intent, templates are divided into Move and Replace templates. Move template: { "Type": "Move", "Target": "$O", "Params": "$A"}; Replace template: { "Type": "Replace", "Target": "$O", "Params": "$A"}; Output structured operation instructions Op_str based on the filled template; Where Move represents the movement intent, Replace represents the replacement intent, Type represents the user intent field, Target represents the target object field, $O represents the target object to be filled in, Params represents the target attribute field, and $A represents the target attribute to be filled in.

7. The BIM-based architectural interior decoration modeling method according to claim 6, characterized in that: S32. Based on the structured operation instructions Op_str and the preliminary building decoration model M_ini, the structured operation instructions Op_str are analyzed, the API interface of the BIM software application is called, the analyzed structured operation instructions Op_str and the preliminary building decoration model M_ini to be modified are input into the BIM software, the actual modification operation is performed on the preliminary building decoration model M_ini to be modified, and finally a new building decoration model M_adj reflecting the user's modification intention is generated; The generation process of the new building decoration model M_adj is as follows: First, the structured operation instruction Op_str is parsed and its parameters are extracted: Detecting the user intent field Type yields the action type set ACT; detecting the target object field Target yields the operation target object set AAO; detecting the target attribute field yields the operation target attribute set TA. Secondly, target object location and attribute extraction are performed on the preliminary architectural decoration model M_ini: Based on the preliminary architectural and decoration model M_ini, the GetGeometry function is called to locate and extract the geometric state set G of the target object set AAO. The geometric state set G includes position coordinates, rotation angle and size. The GetMaterial function is called to extract the attribute state set M of the target object set AAO. The attribute state set M includes color, texture and material. The expressions for target object location and attribute extraction are as follows: ； ； Next, the initial architectural decoration model M_ini was modified using parametric modeling: If the Type attribute of the user intent field is valued as move, then the coordinate transformation of the target object set AAO is performed based on the geometric state set G and the target attribute set TA. If the Type attribute of the user intent field is valued as Replace, then the attribute of the target object set AAO is replaced according to the attribute state set M and the target attribute set TA. Finally, the modified geometric state set G and attribute state set M are written into the preliminary building decoration model M_ini, and the preliminary building decoration model M_ini is updated to generate a new building decoration model M_adj that reflects the user's modification intention.

8. The BIM-based architectural interior decoration modeling method according to claim 7, characterized in that: S4 includes S41; S41. Based on the new building decoration model M_adj, using the building physics simulation engine of BIM software, combined with the geometric information, material properties and set environmental parameters P_env of the new building decoration model M_adj, perform lighting simulation and air circulation simulation to obtain the lighting simulation result R_lig and the air circulation simulation result R_air. ； ； In the formula, Sim_Lig represents the simulation process executed by the lighting simulation engine, Engine_lig represents the lighting simulation engine, Psim_lig represents the preset lighting simulation control parameters, Sim_Air represents the simulation process executed by the air circulation simulation engine, Engine_air represents the air circulation simulation engine, and Psim_air represents the preset air circulation simulation control parameters. Based on the illumination simulation results R_lig and the air circulation simulation results R_air, the average illuminance E_avg and the illuminance uniformity value U_O are extracted from the illumination simulation results R_lig, and an illumination performance index algorithm is constructed to calculate the illumination score S_lig. The air exchange rate ACH and the regional average wind speed V_avg are extracted from the air circulation simulation results R_air, and an air circulation performance index algorithm is constructed to calculate the air circulation score S_air. The algorithm expression for the illumination performance index is as follows: ； In the formula, w1 represents the preset weight of the average illuminance term, w2 represents the preset weight of the illuminance uniformity term, E_tar represents the target average illuminance, U_tar represents the target illuminance uniformity, and exp represents the natural exponential function. The algorithm expression for the air circulation performance index is as follows: ； ； ； In the formula, S_air1 represents the air exchange scoring item, S_air2 represents the regional average wind speed scoring item, w3 represents the preset weight of the air exchange scoring item, w4 represents the preset weight of the regional average wind speed scoring item, Kach represents the preset air exchange rate sensitivity coefficient, Kv represents the preset regional wind speed penalty coefficient, ACH_atr represents the target air exchange rate, V_max represents the upper limit of the set comfortable wind speed range, and V_min represents the lower limit of the set comfortable wind speed range.

9. A BIM-based architectural interior decoration modeling method according to claim 8, characterized in that: S6 includes S61; S61. Adjust the new building decoration model M_adj based on the evaluation decision D_eva; If the evaluation decision D_eva value is Accept, then the new building decoration model M_adj is determined to be unnecessary to adjust and the current new building decoration model M_adj is marked as the final building decoration model scheme M_fin; If the evaluation decision D_eva value is Revise, then the comprehensive scoring algorithm based on the new building decoration model M_adj is used to analyze the results of the scoring items and generate feedback information. This guides the user to communicate with the designer and returns to S3 to adjust the new building decoration model M_adj. The adjusted building decoration model with the evaluation decision D_eva value of Accept is marked as the final building decoration model scheme M_fin.

Images