A multimodal data-driven AI house type customization scheme generation method and system

Abstract

The application relates to a multi-modal data-driven AI house type customization scheme generation method and system, comprising the following steps: acquiring multi-modal data and performing feature extraction to obtain a design target feature set; performing matching through a building specification knowledge base to generate a preliminary constraint condition set; matching a plurality of optimization strategies and generating a reinforced constraint condition set through a first intelligent agent; generating a space planning scheme set through a second intelligent agent and generating a stylized layout scheme set through a third intelligent agent; generating a parameterized building information model set through a fourth intelligent agent; in summary, the application acquires multi-modal data and performs feature extraction, matches constraints through a building specification knowledge base, generates reinforced constraints through intelligent agent optimization, and then generates a space planning scheme, a stylized scheme and a parameterized building information model, and solves the problems of lack of flexibility, controllability and engineering compliance in the prior art through multi-modal data driving and intelligent agent optimization.

Description

Technical Field

[0001] This application relates to the technical field of interior design, and in particular to a multimodal data-driven AI-based method and system for generating customized apartment layouts. Background Technology

[0002] In the fields of interior design, residential planning, and decoration, the efficient and accurate generation of customized floor plan solutions that meet the diverse and personalized needs of users has always been a core challenge. While automated intelligent floor plan generation systems have developed with the evolution of computer-aided design technology, they still have certain shortcomings in practical applications.

[0003] Currently, existing technologies mainly fall into two categories: First, there are template-based and rule-driven systems. These systems typically embed a large number of standardized floor plan templates, limiting user combinations to limited operations such as size adjustments or component replacements. Their operational logic relies on pre-defined rigid rules, such as space area ratios and basic circulation guidelines, directly restricting the output diversity limit due to the size of the solution library. When user needs deviate from the pre-defined templates or rules, these systems often lack adaptability, requiring cumbersome secondary adjustments through manual intervention, making it difficult to respond to the market's urgent demand for unique and personalized designs. Second, there are generative model-based methods. These methods directly generate visually novel interior layout images by learning from massive amounts of floor plan data. However, because the generation process is typically end-to-end image output, it is difficult to effectively embed building codes, structural safety requirements, and precise user functional constraints. While the generated results may have some innovation, they easily overlook rigid engineering requirements such as space utilization and load-bearing structures, causing the solutions to remain only at the conceptual reference stage and failing to meet the feasibility standards for actual construction and application. Summary of the Invention

[0004] To address the aforementioned shortcomings, this application provides a method and system for generating AI-driven apartment layout customization solutions based on multimodal data.

[0005] The above-mentioned objective of this application is achieved through the following technical solution:

[0006] A multimodal data-driven AI-based method for generating customized apartment layout solutions includes the following steps:

[0007] In response to the received task instruction, the corresponding multimodal data is acquired, and feature extraction is performed on the task instruction and multimodal data to obtain the design target feature set, which includes user preference vector, optimization target identifier and constraint condition identifier;

[0008] A preliminary set of constraints is generated by matching constraint identifiers with a pre-built building code knowledge base.

[0009] Based on the optimization target identifier, several optimization strategies are matched from the preset optimization strategy pool. Then, through the first intelligent agent, the matched optimization strategies and user preference vectors are combined to perform multi-strategy collaborative optimization on the initial set of constraints, generating a set of enhanced constraints.

[0010] The set of enhanced constraints is input into the second agent to generate a set of spatial planning schemes, and the third agent combines the set of spatial planning schemes with user preference vectors to generate a set of stylized layout schemes.

[0011] The stylized layout scheme set is input into the fourth intelligent agent to generate a parametric building information model set.

[0012] The second objective of this invention is achieved through the following technical solution:

[0013] A multimodal data-driven AI-based apartment layout customization solution generation system, comprising:

[0014] The feature extraction module is used to respond to the received task instructions, acquire the corresponding multimodal data, and extract features from the task instructions and multimodal data to obtain the design target feature set, which includes user preference vectors, optimization target identifiers, and constraint identifiers.

[0015] The constraint matching module is used to match constraint identifiers with a pre-built building code knowledge base to generate a preliminary set of constraints.

[0016] The constraint enhancement module is used to match several optimization strategies from a preset optimization strategy pool based on the optimization target identifier, and then, through the first intelligent agent, combine the matched optimization strategies and the user preference vector to perform multi-strategy collaborative optimization on the initial constraint set to generate a reinforced constraint set.

[0017] The scheme generation module is used to input the set of enhanced constraints into the second agent to generate a set of spatial planning schemes, and then use the third agent to combine the set of spatial planning schemes with user preference vectors to generate a set of stylized layout schemes.

[0018] The model generation module is used to input the stylized layout scheme set into the fourth intelligent agent to generate a parametric building information model set.

[0019] In summary, the multimodal data-driven AI-based method and system for generating customized apartment layouts provided in this application responds to received task instructions, acquires multimodal data and extracts features, matches constraints through a building code knowledge base, and combines intelligent agent optimization to generate enhanced constraint conditions, thereby generating spatial planning schemes, stylized layout schemes, and parametric building information models. Through multimodal data-driven and hierarchical intelligent agent optimization, it can solve the problems of lack of flexibility, controllability, and engineering compliance in existing technologies, and achieve efficient generation of customized apartment layout schemes that comply with regulations. Attached Figure Description

[0020] Figure 1 This is a flowchart of an embodiment of a multimodal data-driven AI-based apartment layout customization solution generation method according to this application;

[0021] Figure 2 This is an example diagram of the architecture and output of the second intelligent agent in an embodiment of a multimodal data-driven AI apartment layout customization solution generation method of this application;

[0022] Figure 3 This is an example diagram of the architecture and output of the third intelligent agent in an embodiment of a multimodal data-driven AI apartment layout customization solution generation method of this application. Detailed Implementation

[0023] The following is in conjunction with the appendix Figures 1-3 This application will be described in further detail.

[0024] In one embodiment, such as Figure 1 As shown, this application discloses a multimodal data-driven AI-based method for generating customized apartment layouts, which specifically includes the following steps:

[0025] S10: In response to the received task instruction, acquire the corresponding multimodal data, and perform feature extraction on the task instruction and multimodal data to obtain the design target feature set, which includes user preference vector, optimization target identifier and constraint condition identifier;

[0026] In this embodiment, the task instruction refers to the specific design request issued by the user terminal, usually presented in the form of natural language text, such as "Design a three-bedroom, two-living room apartment, requiring the living room to face south, the master bedroom to have an ensuite bathroom, and a modern minimalist style"; multimodal data refers to data from different modalities or sources, such as text descriptions, images, videos, audio, structured data, etc., where multimodal data may include user-inputted text requirements, floor plans, reference images, structural parameters, etc., describing the user's design intent and existing conditions; feature extraction refers to the process of identifying and quantifying representative information from the raw data and converting it into a machine-processable numerical form to extract feature elements related to the apartment design from complex multimodal data; the design target feature set refers to the set obtained after feature extraction, used to indicate subsequent design... The design process is a set of characteristic information. The design target feature set mainly includes the user's core needs and constraints in the design process. The user preference vector is a numerical vector that quantifies the user's personalized preferences, representing their inclinations towards style, function, materials, color, etc., such as their preference for a "modern minimalist" style. The optimization target identifier is one or more discrete labels or codes that identify the key optimization targets for this apartment design, such as "maximizing natural light," "minimizing space utilization," and "controlling overall costs." The constraint identifier is a structured set of keywords or key-value pairs that identify the various constraints that the apartment design must meet. These identifiers clarify the mandatory constraints proposed by the user or derived from their requirements, ensuring the compliance and feasibility of the generated solution.

[0027] Specifically, in response to received task instructions, the system acquires corresponding multimodal data and extracts features from the task instructions and multimodal data to obtain a design target feature set. This design target feature set includes user preference vectors, optimization target identifiers, and constraint identifiers. For example, a user can input "I want a modern three-bedroom apartment with a large living room and good lighting" and upload some pictures of their favorite floor plans and furniture styles. After receiving the user's input, the system converts it into features that the machine can understand and process. One approach is to perform text analysis on the task instructions, identify keywords, and classify and label the multimodal data to generate user preference vectors, optimization target identifiers, and constraint identifiers. Another approach is to pre-set a set of rules to map and extract features based on specific keywords input by the user or metadata of uploaded images to obtain a preliminary design target feature set.

[0028] S20: Match constraint identifiers using a pre-built building code knowledge base to generate a preliminary set of constraints;

[0029] In this embodiment, the building code knowledge base refers to a pre-built database containing a large amount of building design-related codes, standards, regulations, and experience knowledge, used to verify and generate the compliance of design schemes; the preliminary constraint set refers to a list of unoptimized constraints that is initially matched and generated based on the constraint identifiers input by the user and the building code knowledge base.

[0030] Specifically, a preliminary set of constraints is generated by matching constraint identifiers with a pre-built building code knowledge base. For example, if a constraint identifier contains information such as "the number of bedrooms is 3" or "the building area is not less than 100 square meters", the corresponding specific rules can be searched from the building code knowledge base based on these constraint identifiers. One implementation method is to build and maintain a corresponding list of building codes. When a constraint identifier is received, the preliminary constraints are searched and generated from this list. Another implementation method is to predefine several query templates. When a constraint identifier is received, keyword matching is performed in the building code knowledge base using the query templates to obtain relevant building code entries and combine them into a preliminary set of constraints.

[0031] S30: Based on the optimization target identifier, several optimization strategies are matched from the preset optimization strategy pool, and the first intelligent agent combines the matched optimization strategies and user preference vectors to perform multi-strategy collaborative optimization on the preliminary constraint set to generate a reinforced constraint set.

[0032] In this embodiment, the optimization strategy pool refers to a database storing various algorithms and methods for solving design optimization problems, such as genetic algorithms, simulated annealing, and multi-objective optimization algorithms, to select appropriate strategies based on the optimization objective. The agent refers to a software entity with perception, decision-making, and execution capabilities. In this method, the agent is designed to perform specific design tasks, such as constraint optimization, spatial planning, layout generation, and model building. The first agent receives the matched optimization strategy and, guided by user preference vectors, performs collaborative optimization, game theory, or relaxation on the initial constraint set, thereby outputting a reinforced constraint set. The reinforced constraint set refers to the constraint set obtained after multi-strategy collaborative optimization based on the initial constraint set.

[0033] Specifically, based on the optimization target identifier, several optimization strategies are matched from a preset optimization strategy pool. The first intelligent agent then combines the matched optimization strategies with the user preference vector to perform multi-strategy collaborative optimization on the initial constraint set, generating a reinforced constraint set. For example, if the optimization target identifier is "maximizing natural light" and the user preference vector shows that the user has a high preference for "open space", the constraints are adjusted according to this information. One implementation is to preset several optimization rules. For example, when the optimization target is "maximizing natural light", the constraint value of "window area ratio" is increased by a preset fixed percentage, or when the user prefers "open space", the constraint value of "number of walls" is decreased by a preset fixed number. The first intelligent agent can adjust the flexible constraints in the initial constraint set according to the above optimization rules and the user preference vector, thereby generating a reinforced constraint set.

[0034] S40: Input the set of enhanced constraints into the second agent to generate a set of spatial planning schemes, and then use the third agent to combine the set of spatial planning schemes with user preference vectors to generate a set of stylized layout schemes.

[0035] In this embodiment, the second intelligent agent treats the set of reinforced constraints as inviolable design conditions, and uses its pre-set or pre-stored algorithms to explore all compliant spatial layout possibilities, generating several spatial planning schemes that are reasonable and feasible in terms of functional zoning, area ratio, and circulation organization. The spatial planning scheme set refers to an abstract representation of the division and layout of functional areas within the apartment, usually presented in the form of topology diagrams, grid divisions, or parametric geometric models, without involving specific furniture placement and style details. The third intelligent agent treats the spatial planning scheme set as an immutable spatial architecture, and uses the user preference vector as a directional indicator to generate furniture layout, material matching, and decoration schemes that conform to the size and function of the spatial architecture and meet style requirements. It usually adopts a conditional generative AI model and includes a physical rule verification step to ensure practicality. The stylized layout scheme set refers to an interior layout scheme that can include specific furniture placement, decorative elements, and style details, generated based on the spatial planning scheme set and combined with the user preference vector.

[0036] Specifically, a set of reinforced constraints is input into a second agent to generate a set of spatial planning schemes. A third agent then combines the set of spatial planning schemes with user preference vectors to generate a set of stylized layout schemes. For example, the set of reinforced constraints might include "living room area not less than 30 square meters" or "master bedroom adjacent to bathroom." The second agent generates several possible spatial layout schemes based on the input set of reinforced constraints. One implementation is that the second agent can use a rule-based layout generator to generate several spatial planning schemes that meet the basic requirements based on constraints such as area and topological relationships in the set of reinforced constraints, through exhaustive or heuristic search. The generated spatial planning schemes only include the division and approximate location of functional areas, without involving specific styles. Furthermore, the third agent selects furniture that matches the user preference vector, such as "modern minimalist style," from a pre-set furniture model library based on the user preference vector. It then sets the furniture positions according to the area division in the set of spatial planning schemes using a pre-set placement algorithm, thereby generating a set of stylized layout schemes.

[0037] S50: Input the set of stylized layout schemes into the fourth intelligent agent to generate a set of parametric building information models.

[0038] In this embodiment, the fourth intelligent agent is used to identify design elements in the stylized layout scheme set, map and instantiate them into parametric building components with rich attributes, automatically assemble the instantiated parametric building components and establish the engineering logic relationships between them, and finally output a parametric building information model set that can be directly used for detailed design, engineering quantity calculation and construction drawing output; the parametric building information model set refers to a collection of digital representations and management of building projects based on three-dimensional models and related data, wherein the parametric building information model set includes the geometric information, attribute information and interrelationships of building components to support subsequent engineering analysis, construction and management.

[0039] Specifically, a set of stylized layout schemes is input into the fourth agent to generate a set of parametric building information models. For example, the set of stylized layout schemes may contain information such as specific wall locations, door and window dimensions, and furniture models. The fourth agent needs to convert the input information into a BIM model that can be used in building engineering. One implementation is that the fourth agent can convert the two-dimensional drawing information in the set of stylized layout schemes into three-dimensional model components one by one, and generate their attribute information, such as the material and thickness of the walls, and the type and size of the doors and windows. Then, the converted three-dimensional model components are combined into a set of parametric building information models. Another implementation is that the fourth agent can preset a component library containing various standard three-dimensional building component models. When it receives a set of stylized layout schemes, it replaces the elements in the layout schemes with three-dimensional components in the component library through matching and replacement, and adjusts the parameters of the components according to the size and position information in the set of stylized layout schemes, thereby generating a set of parametric building information models.

[0040] For example, suppose user A wants to customize an apartment layout. Their input task instruction is: "I need a three-bedroom, two-living room apartment with an area between 120 and 150 square meters. The living room should face south, the master bedroom should have an ensuite bathroom, the style should be Scandinavian minimalist, and the budget should be moderate." At the same time, user A uploads several pictures of his favorite Scandinavian-style interiors and a hand-drawn floor plan sketch.

[0041] First, in response to the received task instructions, the corresponding multimodal data is acquired, and feature extraction is performed on the task instructions and multimodal data to obtain the design target feature set. Specifically, upon receiving text instructions, Nordic-style images, and floor plan sketches from user A, this information is analyzed. Information such as "three bedrooms and two living rooms," "area 120-150 square meters," "living room facing south," and "master bedroom with ensuite bathroom" are extracted from the text as preliminary constraint identifiers. User preferences for color, material, and furniture lines are extracted from "Nordic minimalism" and the uploaded images to generate a user preference vector. "Improving natural light" is identified from implicit requirements such as "living room facing south" as an optimization target identifier. All of this information together constitutes the design target feature set.

[0042] Furthermore, by matching constraint identifiers through a pre-built building code knowledge base, a preliminary constraint set is generated. Constraint identifiers such as "three bedrooms and two living rooms", "area 120-150 square meters", "living room facing south", and "master bedroom with ensuite bathroom" are input into the building code knowledge base. This allows the knowledge base to query and generate a series of quantifiable preliminary constraints, such as "number of bedrooms = 3", "living room area ≥ 25 square meters", "master bedroom and ensuite bathroom are adjacent and have a direct passage", and "total building area ∈ [120, 150] square meters". The preliminary constraint set can ensure the basic compliance of the scheme.

[0043] Furthermore, based on the optimization target identifier, several optimization strategies are matched from a preset optimization strategy pool. The first agent then combines the matched optimization strategies with the user preference vector to perform multi-strategy collaborative optimization on the initial constraint set, generating a reinforced constraint set. Specifically, in this example, the optimization target identifier is "improving natural light," and the user preference vector reflects a preference for "Nordic minimalist style" and "open space." The first agent matches strategies from the optimization strategy pool, such as "space layout optimization algorithm," and optimizes the initial constraint set by combining the matched strategies with the user preference vector. For example, while satisfying the rigid constraint of "living room facing south," the first agent might, based on the user's preference for open space, appropriately relax the flexible constraint of "the wall between the living room and dining room," or adjust the suggested window size range, to further improve the lighting effect while ensuring structural rationality and aligning with the openness of the Nordic minimalist style. This optimization generates a more personalized and feasible reinforced constraint set.

[0044] Furthermore, the set of reinforced constraints is input into the second agent to generate a set of spatial planning schemes. A third agent then combines these spatial planning schemes with the user's preference vector to generate a set of stylized layout schemes. Specifically, the second agent receives the set of reinforced constraints, such as "living room area ≥ 30 square meters," "kitchen and dining room open-plan connection," and "master bedroom adjacent to an ensuite bathroom." Based on these constraints, it generates several abstract spatial planning schemes that meet the conditions. These schemes may be presented as two-dimensional floor plans, only including the division of functional areas and their approximate location relationships. For example, Scheme A: living room, dining room, and kitchen in an L-shaped open layout; Scheme B: living room, dining room, and kitchen in a linear open layout. The third agent then receives the spatial planning schemes and user A's Nordic minimalist style preference vector. Based on the user's preferences, it selects furniture, lighting, and decorations that conform to the Nordic minimalist style and places them appropriately in the various functional areas, generating a specific set of stylized layout schemes. For example, based on Scheme A, placing light-colored wooden furniture, a sofa with simple lines, and green plants creates a complete Nordic minimalist style layout scheme.

[0045] Furthermore, the stylized layout scheme set is input into the fourth agent to generate a parametric building information model set. Specifically, the fourth agent receives the stylized layout scheme set containing the specific dimensions, locations, and material information of walls, doors, windows, and furniture. It identifies the features of building components and interior elements within the set and extracts their attribute information. For example, it identifies walls, doors, windows, sofas, dining tables, etc., and extracts attributes such as category, geometric dimensions, spatial coordinates, and material. Based on the extracted features and attributes, the fourth agent matches the corresponding 3D parametric component families from a predefined parametric component library and instantiates several 3D parametric components. For example, it converts 2D wall segments into 3D wall models with thickness and material, and converts door and window symbols into 3D door and window models with opening methods and dimensions. Simultaneously, the fourth agent can generate component connection relationships based on the matched 3D components and their assembly logic, such as the connection relationships between walls, the embedding relationships between doors and windows and walls, and the relative positional relationships between furniture. Finally, based on all 3D parametric components and component connection relationships, the fourth agent generates a parametric building information model set that can be used for subsequent construction and management.

[0046] Through the above examples, this solution achieves a comprehensive capture of user preferences and design intentions by extracting features from task instructions and multimodal data. For example, in the case of user A, it not only identifies hard requirements such as "three bedrooms and two living rooms," but also extracts the "Nordic minimalist" style preference from the uploaded image and transforms it into a user preference vector. Through this deep fusion and feature extraction of multimodal data, this solution can handle personalized needs that far exceed the coverage of traditional template libraries, reducing the problem of single solutions or the need for a lot of manual adjustments due to template limitations.

[0047] Compared to methods that directly generate images using generative models, this approach improves the controllability, compliance, and engineering feasibility of the generated solutions by introducing a building code knowledge base, multi-agent collaborative optimization, and a phased generation strategy. In the example above, the constraint identifiers are first matched using the building code knowledge base to ensure that the initial solution complies with basic building regulations. Subsequently, the first agent combines the optimization objective and user preferences to strengthen and optimize the constraints, enabling the solution to better reflect the user's personalized needs while meeting compliance requirements. For example, under the objective of "improving lighting," the window size and wall layout are reasonably adjusted. In summary, through the aforementioned multi-agent collaboration, this approach can reduce or even avoid the problems of "aesthetically pleasing but impractical" or "non-compliant" generative models, thereby improving the feasibility of the generated solutions.

[0048] In summary, this solution, through multimodal data-driven approaches, intelligent agent collaborative optimization, and a phased generation process of parametric building information models, can effectively address the bottlenecks of existing technologies in meeting personalized needs, ensuring compliance, and ensuring engineering feasibility. It effectively improves the personalization, compliance, and feasibility of customized housing solutions.

[0049] In one embodiment, step S10 includes:

[0050] S11: Input the task instructions and multimodal data into the pre-trained feature extraction model to obtain fused semantic features;

[0051] In this embodiment, the pre-trained feature extraction model is used to learn and extract semantic information features from the original task instructions and multimodal data. Its principle is to unify data from different modalities into a shared feature space, thereby enabling a comprehensive understanding of the user's design intent and needs. The feature extraction model can be a deep learning model, such as a Transformer-based encoder, which can process sequential and non-sequential data and capture the correlation between different modalities through a self-attention mechanism. Alternatively, it can be a hybrid model consisting of multiple modality-specific encoders and a fusion layer, where the fusion layer effectively integrates the features encoded by each modality. The fused semantic features refer to the high-dimensional vector representation output by the pre-trained feature extraction model, which integrates all relevant semantic information from the task instructions and multimodal data. It aims to capture deeper meanings such as user preferences, design intent, functional requirements, and potential constraints, providing a consistent semantic foundation for subsequent analysis and decision-making. The fused semantic features can be fixed-length vectors with sufficiently high dimensions to encode complex semantic information, and in the feature space, semantically similar input data will be mapped to similar feature vectors.

[0052] S12: Parallel parsing of the fused semantic features generates a preliminary preference vector, an initial target identifier, and an initial constraint identifier;

[0053] In this embodiment, parallel parsing is used to simultaneously extract different types of design target information from the fused semantic features, namely, the preliminary preference vector, the initial target identifier, and the initial constraint identifier. Parallel processing avoids the information loss or inefficiency problems that may occur with serial parsing, ensuring that various key information types can be comprehensively and independently identified. Parallel parsing can employ a multi-head attention mechanism, where each attention head focuses on extracting a specific type of information; or, it can be designed as a neural network structure with multiple independent output branches, each branch responsible for mapping the fused semantic features to the corresponding output space. The preliminary preference vector refers to the initial representation of user preferences extracted from the fused semantic features through parallel parsing. Typically, this preliminary preference vector is a continuous vector used to quantify the user's inclination towards design style, spatial perception, material selection, etc. The generation of the preliminary preference vector can be... The fully connected layer maps the fused semantic features to a predefined preference semantic space. The initial target identifier is the initial representation of the design optimization target identified from the fused semantic features through parallel parsing. It is usually a set of discrete identifiers that indicate the aspects that need to be prioritized in the design process. The initial target identifier can be identified from the predefined set of optimization targets through a multi-label classification task. Each initial target identifier corresponds to one or more design targets. The initial constraint identifier is the initial representation of the design constraints identified from the fused semantic features through parallel parsing. It is also usually a set of discrete identifiers that represent physical constraints, functional requirements, or specification constraints. The initial constraint identifier can be obtained by identifying restrictive phrases in the text description from the fused semantic features through sequence labeling or key information extraction techniques and mapping them to standard entries in the predefined constraint library.

[0054] S13: Perform consistency verification and calibration on the initial preference vector, initial target identifier, and initial constraint identifier to obtain the design target feature set.

[0055] In this embodiment, consistency verification and calibration aim to ensure that there are no logical conflicts or inconsistencies among the initial preference vector, initial target identifier, and initial constraint identifier, and to make necessary adjustments to improve their accuracy and reliability. For example, if the user preference vector indicates "open space," but the initial constraint identifier contains "strict sound insulation requirements," a conflict may exist. Consistency verification can use a rule-based expert system to detect preset conflict patterns, or use a machine learning model to identify abnormal combinations. The calibration process can make the design target feature set more consistent by adjusting vector values, modifying identifiers, or introducing weights based on the detected conflicts or uncertainties.

[0056] Specifically, the solution in this application transforms heterogeneous input information into unified fused semantic features by inputting task instructions and multimodal data into a pre-trained feature extraction model. These fused semantic features effectively integrate semantic information from different modalities such as text, images, and structural parameters, thus providing a rich and consistent semantic foundation for subsequent analysis. Furthermore, to efficiently and comprehensively acquire the key information required for design, the fused semantic features are analyzed in parallel, simultaneously generating preliminary preference vectors, initial target identifiers, and initial constraint identifiers. The preliminary preference vectors capture the continuous expression of user preferences for design style and functionality; the initial target identifiers identify the discrete optimization objectives preset by the user or system; and the initial constraint identifiers extract specific... By addressing the physical, functional, or regulatory limitations of the object, a parallel parsing mechanism ensures that various types of information can be extracted independently and completely, reducing the potential loss or confusion of information in a single path. Finally, to improve the reliability and internal consistency of the design target feature set, consistency checks and calibrations are performed on the initial preference vector, initial target identifier, and initial constraint identifier. This identifies and resolves potential logical conflicts or inconsistencies between different features, such as contradictions between user preferences and certain constraints, or potential conflicts between different optimization objectives. Specifically, checks identify unreasonable or difficult-to-satisfy feature combinations, while calibration allows for appropriate adjustments or priority ranking of the identified feature combinations, thereby generating the design target feature set.

[0057] By introducing a pre-trained feature extraction model through the above technical solution, semantic understanding and unified representation of heterogeneous data can be achieved, thereby overcoming the limitations of multimodal information fusion processing. At the same time, the parallel parsing mechanism can ensure that user preferences, optimization goals and constraints can be fully and efficiently identified, reducing information omissions or low processing efficiency. More importantly, through the consistency verification and calibration steps, potential conflicts or inconsistencies between different features can be discovered and resolved in a timely manner, thereby generating a logically consistent and highly reliable design target feature set, which has the effect of improving the quality, feasibility and matching degree of the generated scheme with user needs.

[0058] In one embodiment, the multimodal data includes first data and second data, the feature extraction model includes a feature encoding layer, a feature fusion layer, and a feature enhancement layer, and step S11 includes:

[0059] S111: The feature encoding layer encodes the first data to generate a text feature vector, a spatial feature vector, and a parameter feature vector. The first data includes task text information, apartment layout information, and structural parameter information.

[0060] In this embodiment, multimodal data refers to data originating from different sensors or having different representations, such as text, images, and structured data. Multimodal data is divided into first data and second data to differentiate the processing of different types of data, capturing their inherent characteristics and interrelationships. The first data contains core design input information, while the second data provides auxiliary, context-sensitive, or enhancement information. The feature extraction model is a computational model used to learn and generate meaningful low-dimensional representations from the original input data. Further, the feature extraction model includes a feature encoding layer, a feature fusion layer, and a feature enhancement layer. The feature encoding layer converts the original data from different modalities into a unified feature vector representation. The fusion layer is used to integrate feature vectors from these different modalities; the feature enhancement layer further optimizes the fused features to improve their expressive power and robustness; specifically, the role of the feature encoding layer is to convert different forms of first data into machine-understandable numerical vectors. For example, the first data includes task text information, apartment layout information, and structural parameter information. For task text information, it can be encoded into text feature vectors using word embedding models or pre-trained language models; for apartment layout information, graph neural networks can be used to process its topology, or convolutional neural networks can be used to process its rasterized representation to generate spatial feature vectors; for structural parameter information, multilayer perceptrons or simple linear mappings can be used to encode it into parametric feature vectors.

[0061] S112: The feature fusion layer concatenates and fuses text feature vectors, spatial feature vectors, and parametric feature vectors to generate a fused feature vector;

[0062] In this embodiment, the feature fusion layer is used to integrate feature vectors from different modalities into a fused representation with high-level semantics. One implementation is based on concatenation and nonlinear transformation fusion, specifically concatenating text feature vectors, spatial feature vectors, and parametric feature vectors along the feature dimension to form a high-dimensional concatenated vector. This concatenated vector is then input into one or more fully connected layers, where it undergoes transformation and information interaction through a nonlinear activation function. Finally, a dimensionality reduction projection layer normalizes the feature dimension to a preset size, outputting a fused feature vector. Another implementation is based on dynamic fusion using an attention mechanism, specifically, the cross-modal attention module built into the feature fusion layer treats each modal feature vector as a feature sequence, and uses a multi-head self-attention mechanism. The system employs a cross-attention mechanism to dynamically calculate the correlation weights between feature vectors of different modalities, and even between different feature dimensions within the same feature vector of a modality. Based on these correlation weights, the original features are weighted and summed, and information is aggregated to generate a fused feature vector that more accurately reflects the semantic relationships between multimodalities. For example, when the text description mentions "open kitchen," the attention mechanism can automatically enhance the features of the corresponding kitchen area in the spatial feature vector and correlate the weights of the area allocation in the parametric feature vector. The fused feature vector output by the feature fusion layer is a deep integrated representation of text, space, and parametric information. It carries the comprehensive semantics of task instructions and multimodal data, providing a unified input for subsequent feature decoupling and parallel parsing steps.

[0063] S113: The feature enhancement layer encodes the second data to generate an enhanced feature vector, and uses the enhanced feature vector as a modulation signal to recalibrate the fused feature vector to generate fused semantic features.

[0064] In this embodiment, the feature enhancement layer is used to process the second data and optimize the fused features using its information. The second data may include user historical behavior data, environmental context information, and real-time market trends. The feature enhancement layer can be an independent neural network that encodes the second data into an enhanced feature vector. Subsequently, the enhanced feature vector is used as a modulation signal, and the weights of each channel or dimension in the fused feature vector are dynamically adjusted through, for example, a channel attention mechanism or a feature gating mechanism, thereby achieving feature recalibration and generating more discriminative and robust fused semantic features.

[0065] Specifically, the solution proposed in this application effectively addresses the complexity issues in multimodal data processing by introducing a hierarchical feature extraction model. First, the feature encoding layer encodes the task text information, apartment layout information, and structural parameter information in the first data in parallel, generating text feature vectors, spatial feature vectors, and parametric feature vectors, respectively. Parallel encoding ensures that the original information of each modality is captured and converted into a structured numerical representation. Next, the feature fusion layer receives the feature vectors generated by the feature encoding layer and performs concatenation and fusion operations on them, aiming to integrate information from different modalities to form a comprehensive fused feature vector. Based on this, the feature enhancement layer independently encodes the second data to generate an enhanced feature vector. This enhanced feature vector acts as a modulation signal to recalibrate the fused feature vector. This recalibration mechanism allows the model to dynamically adjust the weights or activations of the fused feature vector based on the auxiliary information provided by the second data, thereby improving the semantic expressiveness and robustness of the fused feature vector without introducing redundant information, ultimately generating high-quality fused semantic features.

[0066] For example, suppose a user submits a floor plan customization task. The first data includes task text information, such as "a family of three, prefers a minimalist style, needs three bedrooms and two living rooms, with a walk-in closet in the master bedroom," and floor plan structure information, including CAD data of the existing floor plan uploaded by the user or vectorized data of a hand-drawn sketch, as well as structural parameter information, such as "total area 120 square meters, floor height 2.8 meters." The second data consists of several interior design style reference images provided by the user. The feature encoding layer first processes the first and second data, specifically including: using a BERT model to encode the task text information to generate text feature vectors; and using a graph convolutional network to process the floor plan structure. The information is encoded into spatial feature vectors by room nodes and their connections. A small fully connected network is used to embed numerical parameters such as area and floor height into parametric feature vectors. Further, the feature fusion layer concatenates the above three vectors and fuses them through a multilayer perceptron to generate a preliminary fused feature vector. Finally, the feature enhancement layer uses a ResNet-50 model to encode the reference image to generate an enhanced feature vector. This enhanced feature vector is then used as a channel attention weight by the Squeeze-and-Excitation module to weight each channel of the fused feature vector, thereby achieving feature recalibration and ultimately obtaining the fused semantic features.

[0067] The above technical solution subdivides multimodal data into first data and second data, and adopts a hierarchical feature encoding, fusion, and enhancement model mechanism to effectively handle the complexity of heterogeneous multimodal data. Specifically, the feature encoding layer can perform specialized encoding for the characteristics of different modalities, ensuring the integrity and accuracy of the original information; the feature fusion layer effectively integrates the encoded features to form a preliminary comprehensive representation; the feature enhancement layer uses the second data as a modulation signal to recalibrate the fused features, enabling the feature extraction model to dynamically optimize the feature representation based on auxiliary information, enhancing the discriminative power and robustness of the fused semantic features, thereby improving the accuracy and comprehensiveness of extracting the design target feature set from task instructions and multimodal data.

[0068] In one embodiment, step S12 includes:

[0069] S121: Input the fused semantic features into a pre-built parallel parsing model, wherein the parallel parsing model includes a first parsing layer, a second parsing layer, and a third parsing layer;

[0070] In this embodiment, the pre-built parallel parsing model is a computational model used to process input data simultaneously and output multiple independent results. This parallel parsing model has been trained and optimized before application, possessing the ability to parse specific information from fused semantic features. The training and optimization process can be set and adjusted by those skilled in the art using their available knowledge. The parallel parsing model can be implemented using various architectures. For example, it can be a deep neural network containing multiple independent branches, each responsible for parsing a specific type of information; or it can be a composite model sharing a bottom-level feature extraction layer but with separate upper-level output layers. Its core purpose is to ensure that different parsing tasks do not interfere with each other, thereby improving parsing efficiency and accuracy.

[0071] S122: The first parsing layer maps the fused semantic features to a continuous preference semantic space, generating a preliminary preference vector;

[0072] In this embodiment, the first parsing layer is a component module of the parallel parsing model, used to extract user preference information from the fused semantic features. It transforms the high-dimensional fused semantic features into a low-dimensional and continuous vector representation, namely the preliminary preference vector. The preliminary preference vector can capture various user preferences in apartment design, such as preferences for style, function, and spatial layout. Its mapping process can be implemented based on autoencoders, principal component analysis, or neural network layers, aiming to transform discrete or complex semantic information into a quantifiable and continuous numerical representation for subsequent calculation and optimization. The continuous preference semantic space is a high-dimensional and continuous mathematical vector space pre-constructed based on the first parsing layer. In the continuous preference semantic space, each point coordinate vector uniquely corresponds to a comprehensive combination of design style and functional preferences.

[0073] S123: The second parsing layer performs multi-label classification on the fused semantic features and identifies at least one dominant optimization objective from a predefined set of optimization objectives, generating the corresponding initial objective identifier;

[0074] In this embodiment, the second parsing layer is a component module of the parallel parsing model, used to identify the optimization objectives of interest to the user or system from the fused semantic features. This is typically achieved through a multi-label classification task, where one fused semantic feature may correspond to multiple optimization objectives. Multi-label classification is a machine learning task that allows a single input sample to have multiple output labels simultaneously. In this step, one fused semantic feature may correspond to multiple optimization objectives such as "cost control," "space efficiency," and "lighting optimization." The predefined set of optimization objectives may include cost minimization, area utilization maximization, lighting optimization, ventilation optimization, and functional zoning rationalization. The dominant optimization objective refers to the objective identified from the predefined set of optimization objectives that has the highest relevance or priority to the current fused semantic feature. It reflects the core contradictions or value orientations that need to be considered and weighed in the current specific design task. The second parsing layer can employ a classifier based on convolutional neural networks or recurrent neural networks, or a model based on the Transformer architecture, to learn the association between the fused semantic features and the optimization objectives, and output one or more identifiers representing the dominant optimization objective.

[0075] S124: The third parsing layer performs sequence labeling and key information extraction on the fused semantic features, identifies physical and functional constraints, maps them to standard constraint entries in a predefined constraint library, and generates initial constraint identifiers.

[0076] In this embodiment, the third parsing layer is a component module of the parallel parsing model, used to extract specific constraints from the fused semantic features. This typically involves sequence labeling and key information extraction techniques in natural language processing. Sequence labeling is a natural language processing task used to label input sequences with predefined types. It identifies segments belonging to specific categories within the sequence. Specifically, it reconstructs the fused semantic features semantically into an internal representation similar to a text sequence and labels this text sequence. For example, it can label which parts represent entities, such as "load-bearing wall" or "toilet," which parts represent operations or states, such as "cannot move" or "need," and which parts represent attributes, such as "3 meters" or "facing south." Key information extraction, based on sequence labeling, identifies and combines segments with complete semantic and logical relationships from the labeled segments. Structured information units; physical constraints refer to the uncompromising hard constraints related to the building entity, spatial form, and environmental physical laws, which may include building area, floor height, structural load-bearing capacity, etc.; functional constraints refer to the subjective and clear functional requirements related to the behavior, lifestyle, and use of space of residents, which may involve the number of rooms, functional area division, circulation planning, etc.; the predefined constraint library refers to a standardized design constraint knowledge database, which maps constraints described in natural language into standard constraint entries that can be accurately understood and processed by computers. Among them, standard constraint entries include unique identifiers, standard descriptions, parameterized rules, and associated specifications; the third parsing layer can standardize and map the identified unstructured or semi-structured information to specific entries in the predefined constraint library through rule matching, semantic similarity calculation, or deep learning models, thereby generating initial constraint identifiers.

[0077] Furthermore, physical constraints include structural safety, spatial scale, and environmental physics; functional constraints include spatial function, behavioral flow, and facility configuration.

[0078] Specifically, the fused semantic features are input into a pre-built parallel parsing model, which consists of three parsing layers: a first parsing layer, a second parsing layer, and a third parsing layer. These three parsing layers process the same fused semantic features independently and in parallel, avoiding information loss or mutual interference that may occur during serial processing. Specifically, the first parsing layer focuses on user preferences, mapping the fused semantic features to a continuous preference semantic space to generate an initial preference vector. The second parsing layer identifies the optimization objectives that need to be prioritized in the design process from the fused semantic features. It identifies at least one dominant optimization objective from a predefined set of optimization objectives through multi-label classification and generates a corresponding initial objective identifier. The third parsing layer extracts specific constraints from the fused semantic features. It identifies the physical and functional constraints that must be followed in the apartment design through sequence labeling and key information extraction techniques, and standardizes them into standard constraint entries in a predefined constraint library to generate an initial constraint identifier.

[0079] For example, after receiving task instructions and multimodal data and processing them through a feature extraction model, a fused semantic feature is obtained. This fused semantic feature can be a high-dimensional vector representation containing comprehensive information about the user's preferences regarding apartment style, functional requirements, and area requirements. To extract user preferences, optimization objectives, and constraints from this fused semantic feature, it can be input into a pre-built parallel parsing model, which consists of three independent deep learning sub-networks. The first parsing layer can be a neural network composed of multiple fully connected layers, whose function is to convert the fused semantic feature into a 256-dimensional preliminary preference vector. This preliminary preference vector can quantify the user's preferences for modern minimalist style, The first layer considers preferences for open-plan kitchens, etc. The second parsing layer can be a multi-label classification network based on a Transformer encoder. It receives fused semantic features as input and outputs a vector containing several binary values, each corresponding to an optimization objective in a predefined set of optimization objectives, thus generating an initial objective identifier. The third parsing layer can be a sequence labeling model combining BERT and conditional random fields. This model identifies physical and functional constraints from the fused semantic features and maps them to corresponding standard entries in a constraint library. For example, constraints such as "bedroom area not less than 12 square meters," "living room facing south," and "kitchen must have a window" can be mapped to "bedroom area ≥ 12 square meters." 2 "Living room orientation = south", "Kitchen ventilation = window", and so on, thus generating initial constraint identifiers.

[0080] By employing the above technical solution and using a pre-built parallel parsing model, which is divided into a dedicated first parsing layer, a second parsing layer, and a third parsing layer, parallel processing of fused semantic features is achieved. This not only avoids the information loss and inefficiency that may be caused by traditional serial processing, but also improves the independence and accuracy of different types of information extraction processes, thus enhancing the intelligence level and processing efficiency of the initial information parsing stage.

[0081] In one embodiment, step S20 includes:

[0082] S21: Parse the constraint identifiers, identify the constraint type and constraint parameters, and perform reasoning based on the preset association rule set. Add implicit constraints based on the reasoning results and generate a list of constraint identifiers containing several constraint identifiers.

[0083] In this embodiment, step S21 aims to analyze the initial constraint identifiers, clarify their constituent elements, deduce and supplement implicit constraints that are not explicitly mentioned but logically necessary, thereby ensuring a multi-level understanding of user needs and specification requirements. This step can utilize natural language processing technology to semantically analyze the constraint identifiers, combining predefined grammatical rules and ontological models to identify constraint types and constraint parameters. Constraint types refer to the classification of constraints; different constraint types describe the essential attributes and scope of different constraints. Constraint types include existence constraints, numerical constraints, topological constraints, compatibility constraints, and compliance constraints. Constraint parameters refer to the constraints... Specific variables or values ​​that can be quantified and adjusted; the preset association rule set can be a rule engine built based on expert knowledge. For example, if "bedroom" is identified and "area greater than 10 square meters", the implicit constraint "must have a window" is added through reasoning; or, a reasoning mechanism based on a graph database can be used to model constraint identifiers and their related attributes as nodes and edges in a graph. The association rule set is then represented as graph traversal or pattern matching rules. When a specific pattern is detected, new constraint nodes or edges are automatically added, thereby generating a list of constraint identifiers containing explicit and implicit constraints; implicit constraints refer to supplementary constraints that are not directly expressed by the user but are derived from the identified explicit constraints based on the preset association rule set.

[0084] S22: Traverse the list of constraint identifiers, query the parameterization rules, value ranges and boundary conditions corresponding to all constraint identifiers in the building code knowledge base, and instantiate each constraint identifier into at least one atomic constraint condition based on the query results.

[0085] In this embodiment, step S22 aims to transform abstract constraint identifiers into concrete and operable atomic constraints, enabling them to be understood and processed by the computer system, thereby providing standardized input for subsequent automated design and optimization. This step can be performed using a structured query language or NoSQL query interface to search for corresponding records in a building code knowledge base stored in a relational database or document database for each identifier in the constraint identifier list. These records contain parameterization rules, value ranges, and boundary conditions, and this information is combined into specific atomic constraint objects. Alternatively, knowledge graph technology can be used to integrate building code knowledge... The library is constructed as a semantic network. When traversing the list of constraint identifiers, it retrieves the attributes, relationships, and values ​​associated with each identifier through knowledge graph queries. For example, querying the identifier "bedroom area" retrieves its "minimum area" attribute value and, combined with other contextual information, instantiates it into the atomic constraint condition "bedroom area ≥ 9 square meters". Parametric rules refer to mathematical expressions or program logic stored in the building code knowledge base that are bound to a constraint identifier and can be computed. They define how to apply the constraint to specific design variables. Atomic constraints refer to the basic constraint units obtained after instantiation, which are indivisible and can be directly used to drive optimization algorithms or perform satisfaction verification.

[0086] S23: Perform logical conflict detection on all atomic constraints. If a conflict is detected, adjust and optimize the conflicting atomic constraints according to the constraint priority rules and conflict resolution strategies preset in the building code knowledge base to generate a preliminary set of constraints.

[0087] In this embodiment, step S23 aims to ensure that all constraints are logically consistent and feasible, avoiding design process obstruction or the generation of unreasonable solutions due to contradictory constraints. This step can be implemented by constructing a constraint satisfaction problem solver. Specifically, all atomic constraints are treated as variables and constraints in the constraint satisfaction problem, and an attempt is made to find a solution that satisfies all conditions. If the solver cannot find the required solution, it indicates a conflict. Logical conflict refers to the state where two or more atomic constraints cannot be satisfied simultaneously within the context of the same design problem. Logical conflicts include numerical conflicts, existence conflicts, and resource competition conflicts. The constraint priority rule defines the ordering rules for which constraints need to be or can be prioritized when conflicts occur between constraints. The priority is usually based on the rigidity of the constraints. The conflict resolution strategy refers to the strategy of prioritizing constraints... Guided by hierarchical rules, specific strategies are adopted to resolve specific conflicts. Among them, conflict resolution strategies can be preset as a series of rules. For example, when "bedroom area ≥ 10 square meters" and "total area ≤ 80 square meters" conflict under a specific apartment type, one of them is prioritized based on "user preference" or "standard priority", or a compromise adjustment is made between the two. Alternatively, a priority-based conflict detection and resolution mechanism can be adopted, which specifically includes pre-setting a priority for each atomic constraint in the building code knowledge base. When a logical contradiction is detected between two or more atomic constraints, for example, through Boolean logic expressions or interval overlap detection, the priority rules of atomic constraints will be used to judge, prioritizing the retention of higher priority constraints and adjusting or removing lower priority constraints until all constraint sets achieve logical consistency.

[0088] Specifically, upon receiving the initial constraint identifiers, they are parsed to identify their inherent constraint types and parameters. Based on this, reasoning is performed using a pre-defined set of association rules to automatically discover and add implicit constraints not explicitly mentioned in the initial identifiers but logically present, generating a constraint identifier list containing several constraint identifiers. Further, the constraint identifier list is traversed, and for each identifier, a query is performed in a pre-built building code knowledge base. This query not only obtains the parameterized rules corresponding to the identifier but also its specific value range and boundary conditions, thus transforming the abstract constraint identifiers into a series of concrete atomic constraints. After the atomic constraints are instantiated, logical conflict detection is performed to avoid introducing contradictory requirements in the early design stages, which could lead to difficulties in subsequent scheme generation or unreasonable results. Once a conflict is detected, the conflicting atomic constraints are intelligently adjusted and optimized according to the constraint priority rules and conflict resolution strategies pre-defined in the building code knowledge base, until the entire initial constraint set achieves logical consistency and feasibility.

[0089] By parsing and reasoning through the above technical solutions, the initial constraint identifiers can not only identify the types and parameters of explicit constraints, but also automatically discover and supplement potential implicit constraints, thereby improving the comprehensiveness and completeness of constraint information. On this basis, abstract constraint identifiers are transformed into specific atomic constraints. Furthermore, by introducing a logical conflict detection and intelligent resolution mechanism, potential contradictions between different constraints can be effectively identified and resolved. This avoids the inability to generate design schemes or the generation of unreasonable schemes due to constraint conflicts, thus improving the quality and reliability of the initial constraint set. Consequently, it enhances the efficiency, accuracy, and user satisfaction of generating customized apartment layout schemes, while reducing manual intervention and repeated adjustments to requirements.

[0090] In this regard, this application further proposes that, in one embodiment, step S30 includes:

[0091] S31: Divide the initial set of constraints to obtain a set of rigid constraints and a set of flexible constraints;

[0092] In this embodiment, step S31 aims to classify and manage constraints of different natures. Rigid constraints refer to constraints that must be strictly adhered to and cannot be compromised, such as mandatory regulations regarding building structural safety, fire lane width, and minimum daylighting area. Flexible constraints refer to constraints with a certain adjustable range or priority, such as users' desired bedroom area, living room orientation preferences, and storage space size. This classification of constraints can be based on a preset rule base. For example, all constraints derived from building codes can be automatically classified as rigid constraints, while constraints derived from user preferences or industry recommended standards can be classified as flexible constraints. Alternatively, machine learning models, combined with historical design data and expert annotations, can automatically identify and distinguish the rigid and flexible attributes of constraints.

[0093] S32: Input the matched optimization strategy, user preference vector, and preliminary constraint set into the first agent, causing the first agent to execute a hierarchical optimization process, wherein the hierarchical optimization process includes:

[0094] Keeping the rigid constraint set unchanged, a multi-objective optimization problem is constructed with the flexible constraint set as the optimization variable. The multi-objective optimization problem is solved iteratively, and the Pareto optimal relaxation interval of the flexible constraint set is calculated under the premise of satisfying the rigid constraint set. The optimization objective of the multi-objective optimization problem is defined by the matching optimization strategy.

[0095] Based on the user preference vector, the corresponding re-constraint value is selected for each flexible constraint in the flexible constraint set within the Pareto optimal relaxation interval, and a constraint adjustment scheme is generated.

[0096] Update the parameters of the flexible constraint set according to the constraint adjustment scheme, and generate the set of enhanced constraint conditions.

[0097] In this embodiment, the first intelligent agent is a computational entity with complex decision-making and optimization capabilities. Its core lies in its ability to systematically handle different types of constraints and adjust according to user preferences. The first intelligent agent can be built based on a reinforcement learning framework, learning the optimal optimization path through interaction with a simulated environment; alternatively, it can employ an optimization engine based on heuristic or metaheuristic algorithms to efficiently search for solutions that meet the conditions in a multi-dimensional design space. In the hierarchical optimization process, the rigid constraint set remains unchanged, and a multi-objective optimization problem is constructed with the flexible constraint set as the optimization variable. This means that throughout the optimization process, rigid constraints are treated as insurmountable boundaries, ensuring… The compliance and security of the solution; the multi-objective optimization problem aims to simultaneously optimize multiple conflicting or mutually influential flexible objectives, such as satisfying users' demand for a large living room while also considering the comfort of bedrooms or the economy of total area. The optimization objective of the multi-objective optimization problem is defined by a matching optimization strategy. For example, if the matching optimization strategy is "maximizing space utilization" and "minimizing construction costs," then the adjustment of flexible constraints will revolve around these two objectives. The multi-objective optimization problem is solved iteratively, calculating the Pareto optimal relaxation interval of the flexible constraint set under the premise of satisfying the rigid constraint set. Various multi-objective optimization algorithms can be used in the iterative solution process, such as non-dominated sorting. Genetic Algorithm II or multi-objective particle swarm optimization can effectively explore the solution space and generate Pareto optimal solutions. The Pareto optimal relaxation interval represents the set of possible values ​​for flexible constraints without sacrificing any optimization objective. For example, under the premise of satisfying all rigid specifications, the living room area can vary within a certain interval, and any value within that interval will not worsen other flexible objectives. In the stage of generating constraint adjustment schemes, the user preference vector serves as a decision weight or preference function, guiding the first agent to select the specific constraint value that best meets the user's personalized needs from the Pareto optimal solution set. For example, if the user preference vector indicates that the user prefers... Regarding "spaciousness," the first agent might choose a solution with a larger living room and a moderately sized bedroom on the Pareto front. This selection process can be completed by a decision support module, which transforms the user preference vector into a specific selection strategy, thereby converting abstract user preferences into executable constraint parameters. Based on the constraint adjustment scheme, the parameters of the flexible constraint set are updated to generate a set of reinforced constraints. This means replacing the corresponding flexible constraint parameters in the initial constraint set with the flexible constraint values ​​after user preference guidance and multi-objective optimization. Then, the updated flexible constraint set is integrated with the unchanged rigid constraint set to form the set of reinforced constraints.

[0098] Specifically, the initial constraint set is divided into a rigid constraint set and a flexible constraint set, clarifying which constraints are insurmountable bottom lines and which are adjustable flexible spaces. Furthermore, during the hierarchical optimization process, the first intelligent agent maintains the rigid constraint set unchanged, treating it as a hard boundary condition to ensure the compliance and security of the generated solution. Based on this, the first intelligent agent uses the flexible constraint set as optimization variables to construct a multi-objective optimization problem. This multi-objective optimization problem is defined by a matching optimization strategy; for example, it may simultaneously pursue maximizing space utilization and optimizing lighting effects. The solution is then iteratively solved... The solution involves calculating the Pareto optimal relaxation interval of the flexible constraint set while satisfying the rigid constraints. This provides a solution set that balances multiple optimization objectives, avoiding local optima or unintended consequences that may result from single-objective optimization. Furthermore, the user preference vector is introduced into the Pareto optimal relaxation interval as a basis for selecting the optimal solution. From a series of optimal trade-offs, the re-constraint value that best meets the user's personalized needs is selected to generate a constraint adjustment scheme. Finally, the parameters of the flexible constraint set are updated according to the constraint adjustment scheme, and the updated flexible constraint set is integrated with the rigid constraint set to form a set of enhanced constraints.

[0099] The above technical solution divides the initial constraint set into rigid and flexible constraint sets, ensuring strict adherence to mandatory specifications and avoiding compliance risks caused by optimization. Simultaneously, the hierarchical optimization process executed by the first intelligent agent, while maintaining rigid constraints, constructs and iteratively solves a multi-objective optimization problem with flexible constraint sets as optimization variables, calculating the Pareto optimal relaxation interval to find the best balance among multiple mutually influential flexible objectives. Based on user preference vectors, re-constraint values ​​are selected within the Pareto optimal relaxation interval, ensuring that the final strengthened constraint set is not only technically feasible but also highly aligned with users' personalized needs, thus improving user satisfaction. This constraint strengthening mechanism provides more precise input for the subsequent generation of spatial planning scheme sets and stylized layout scheme sets, thereby improving the overall scheme's rationality, diversity, and user acceptance.

[0100] In one embodiment, such as Figure 2 As shown, the second intelligent agent includes a constraint parsing layer, a space optimization layer, and a scheme decoding layer. Step S40 includes:

[0101] S41A: The constraint parsing layer parses the set of enhanced constraint conditions, extracts spatial functional constraint features, area constraint features, and topological relationship constraint features, and generates a spatial parameterized representation.

[0102] In this embodiment, the second intelligent agent is an intelligent decision-making or processing unit whose main function is to transform the reinforced set of constraints into a specific set of spatial planning schemes. The second intelligent agent can be a generative model based on deep learning, such as a generative adversarial network or variational autoencoder, which generates new spatial layouts by learning from a large amount of apartment type data. Alternatively, the second intelligent agent can be a hybrid intelligent system combining rule-based reasoning, optimization algorithms, and machine learning, utilizing expert knowledge and data-driven methods to jointly complete the spatial planning task. The constraint parsing layer is a functional module layer of the second intelligent agent, used to parse the reinforced set of constraints, thereby extracting spatial functional constraint features, area constraint features, and topological relationship constraint features, and transforming them into a spatial parameterized representation that can be processed by a computer. For example, the constraint parsing layer can use natural language processing technology to process the constraints in text form. Semantic analysis and entity recognition are performed, or rule engines are used to predefine rules to identify and classify different types of constraints. Spatial function constraints, area constraints, and topological relationship constraints are all specific spatial design requirements that guide the subsequent spatial optimization process. For example, spatial function constraints can specify that the living room needs natural lighting and the bedroom needs to be quiet; area constraints can stipulate that the living room area is not less than 20 square meters; topological relationship constraints can require that the bathroom cannot directly face the living room door. Spatial parametric representation refers to an abstract data structure of spatial layout that can be processed and optimized by computers. It serves as the object of iterative optimization in the spatial optimization layer. It can be represented by a grid, that is, the space is divided into discrete grids, and each grid cell represents its attributes; it can also be represented by a graph structure, that is, the rooms are regarded as nodes and the relationships are regarded as edges; or it can be represented by parametric geometry, that is, the layout is described by geometric parameters.

[0103] S42A: The spatial optimization layer executes a preset iterative optimization process, which includes:

[0104] (a) Compute the matching loss between the spatial parameterized representation and the set of reinforced constraints;

[0105] (b) Based on the matching loss, the spatial parameterized representation is updated through a pre-defined backpropagation algorithm;

[0106] (c) Using steps (a) and (b) as one iteration process, repeat the iteration process until any termination condition is reached, and record the intermediate optimization states generated during the iteration process. The termination conditions include the matching loss converging to a preset loss threshold and the number of iterations reaching the maximum number of iterations.

[0107] In this embodiment, the spatial optimization layer is a functional module layer of the second intelligent agent, used to execute a preset iterative optimization process. This process continuously adjusts the spatial parameterized representation to satisfy a set of reinforced constraints. For example, the spatial optimization layer can employ a gradient descent-based optimizer, an evolutionary algorithm, or a simulated annealing algorithm. The iterative optimization process is a computational process that gradually approaches the optimal solution by repeatedly executing specific steps, aiming to progressively adjust the spatial parameterized representation to better satisfy the set of reinforced constraints. The matching loss between the spatial parameterized representation and the set of reinforced constraints is calculated; it is a quantitative indicator measuring the degree of non-compliance between the current spatial parameterized representation and the set of reinforced constraints, used to guide the optimization direction. For example, a loss function can be defined, generating a penalty term when the actual area is less than the required area. Based on the matching loss, the spatial parameterized representation is updated through a preset backpropagation algorithm. Its function is to adjust the magnitude and direction of the matching loss. Adjusting the spatial parameterized representation to more closely approximate the constrained state can be achieved through backpropagation of a neural network or parameter updates based on gradient descent. Backpropagation is an algorithm that uses the chain rule to efficiently calculate the gradient of the loss function with respect to all adjustable parameters. In this embodiment, the adjustable parameters refer to all parameters in the spatial parameterized representation. Taking steps (a) and (b) as an iteration process, the iteration process is repeated until any termination condition is reached, and intermediate optimization states generated during the iteration process are recorded. Its purpose is to ensure that the optimization process converges to a satisfactory result within a reasonable time and to provide diverse solutions. Intermediate optimization states refer to snapshots of the spatial parameterized representation recorded and saved during the iterative optimization process, other than the final result, at a specific number of iterations or when specific conditions are met. Termination conditions may include the matching loss converging to a preset loss threshold or the number of iterations reaching the maximum number of iterations.

[0108] S43A: The scheme decoding layer generates a spatial planning scheme set containing at least one main scheme and several variant schemes based on the spatial parameterized representation obtained after the execution of the iterative optimization process and the intermediate optimization states recorded during the iteration.

[0109] In this embodiment, the scheme decoding layer is a functional module layer of the second intelligent agent, used to transform the optimized spatial parameterized representation and the intermediate optimization states recorded during the iteration process into a spatial planning scheme set. For example, the scheme decoding layer can render the parameterized representation into a two-dimensional plan or a three-dimensional model based on a geometric rendering engine. The spatial planning scheme set is a collection of multiple apartment layout schemes that meet the constraints, providing users with diverse choices. The main scheme is usually the scheme that converges best during the optimization process, while the variant schemes are the embodiment of different stages or different local optima during the optimization process, aiming to provide diversity and meet the subtle differences in user preferences.

[0110] Specifically, the constraint parsing layer parses the set of reinforced constraints, extracting spatial functional constraint features, area constraint features, and topological relationship constraint features, and transforming them into a spatial parameterized representation that can be processed by a computer. Further, the spatial optimization layer receives the spatial parameterized representation and initiates an iterative optimization process. In this process, the matching loss between the current spatial parameterized representation and the set of reinforced constraints is calculated to quantify the degree of conformity between the layout and the constraints. Based on the matching loss, the spatial parameterized representation is updated using a pre-defined backpropagation algorithm, thereby gradually adjusting the layout to better fit the various constraints. This iterative process continues until a pre-defined termination condition is reached, and intermediate optimization states are recorded during the optimization process. Finally, the scheme decoding layer uses the optimized spatial parameterized representation and the recorded intermediate optimization states to generate a spatial planning scheme set containing at least one main scheme and several variant schemes.

[0111] For example, as a specific implementation, the second intelligent agent can be an intelligent system built on a deep learning framework. The constraint parsing layer can employ a graph neural network-based model to parse the input set of reinforced constraints into spatial functional constraint features, area constraint features, and topological relationship constraint features, generating an initial spatial parameterized representation. This spatial parameterized representation can be a tensor containing the center coordinates, length, width, rotation angle, and functional label of each room. The spatial optimization layer can be implemented as an iterative solver based on the Adam optimizer. In each iteration, the matching loss can be defined as a weighted sum of the degree of constraint violation. For example, if the area of ​​a room exceeds its specified range, an area penalty term will be generated; if two rooms that should not be adjacent are adjacent, a topological penalty term will be generated. Backpropagation... The broadcast algorithm fine-tunes the room coordinates, dimensions, and angles in the tensor based on the gradient information of the matching loss on the spatial parameterized representation. Furthermore, the termination condition of the iterative optimization process can be set as the matching loss decreasing by less than 0.001 for 100 consecutive iterations, or the total number of iterations reaching 5000. During the iterative optimization process, the current spatial parameterized representation can be saved as an intermediate optimization state every 500 iterations. The scheme decoding layer can be a module based on the geometry generation algorithm, which converts the final optimized spatial parameterized representation into a two-dimensional planar layout diagram and uses the saved intermediate optimization state to generate several variant schemes by fine-tuning key parameters. For example, under the premise of satisfying all core constraints, it can provide scheme A with a slightly different kitchen location, scheme B with a slightly adjusted living room layout, etc., for users to choose from.

[0112] Through the above technical solution, this application can effectively transform a complex set of enhanced constraints into a specific and feasible spatial planning scheme. By coordinating the multi-layer structure of the second intelligent agent, it ensures the accurate understanding of constraints, iterative optimization of spatial layout, and generation of diverse schemes, thereby achieving the processing of multi-dimensional constraints and providing a spatial planning set that includes the main scheme and variant schemes, thus improving the intelligence level and user satisfaction of the customized housing scheme.

[0113] In one embodiment, such as Figure 3 As shown, the third intelligent agent includes a style encoding layer, a layout condition generation layer, a physical verification layer, and an adjustment processing layer. Step S40 includes:

[0114] S41B: The style coding layer maps user preference vectors to style latent vectors;

[0115] In this embodiment, the third intelligent agent can be implemented through a multi-modal deep learning system or a hybrid intelligent system combining a rule engine and optimization algorithms. The style encoding layer is a functional module layer of the third intelligent agent, used to transform the abstract design intent and style tendency represented by the user preference vector into a machine-understandable and operable numerical representation, i.e., a style latent vector. This can be implemented through a pre-trained neural network model, such as an autoencoder or variational autoencoder, whose input is the user preference vector and output is a vector in a low-dimensional and continuous style latent space. Another implementation method can be a rule engine based on semantic analysis and feature extraction, mapping keywords and attributes in user preferences to predefined style feature dimensions.

[0116] S42B: The layout condition generation layer encodes the spatial planning scheme set into the corresponding spatial structure condition vector, and merges the style potential vector with the spatial structure condition vector to generate a preliminary layout scheme.

[0117] In this embodiment, the layout condition generation layer is a functional module layer of the third agent, used to encode the geometric and topological information of the spatial planning scheme set into structured spatial structure condition vectors, and fuse them with style latent vectors to generate a preliminary layout scheme. The layout condition generation layer can use graph neural networks to process the unstructured data in the spatial planning scheme and convert it into vector representation. Furthermore, the vector fusion process can be carried out through vector concatenation or through complex neural network structures such as attention mechanisms and gating mechanisms to ensure that style information can effectively guide the generation of layout elements.

[0118] S43B: The physical verification layer, based on a predefined physical rule base and furniture knowledge base, performs collision detection and compliance verification on the preliminary layout plan and generates a verification report;

[0119] In this embodiment, the physical verification layer is a functional module layer of the third intelligent agent, used to ensure that the generated preliminary layout scheme conforms to real-world physical constraints and design specifications. Specifically, the physical verification layer can be implemented using a rule-based expert system. This involves using a predefined physical rule base and furniture knowledge base to perform geometric collision detection and logical compliance checks on the layout scheme. For example, computational geometry algorithms can be used to detect overlaps between furniture, or topology analysis algorithms can be used to check the accessibility of functional areas. The physical rule base refers to a structured database of engineering physics common sense and design principles, which stipulates the rules that objects, spaces, and people must follow in the real physical world and human activities. The objective laws and safety and comfort baselines; the furniture knowledge base refers to a structured database of furniture and interior furnishing products, which includes the functional attributes, usage methods, common configurations, and spatial relationships of each type of furniture; collision detection refers to the calculation process of determining whether the geometric models of two or more objects illegally overlap in three-dimensional space; compliance verification refers to the process of evaluating whether the design scheme meets the preset standards in multiple dimensions such as functionality, safety, comfort, and standardization based on the physical rule base and the furniture knowledge base; the verification report refers to the structured diagnostic document output by the physical verification layer, which includes a list of problems and their location, violation types and their severity levels, and correction suggestions.

[0120] S44B: The adjustment processing layer adjusts the initial layout scheme based on the verification report and generates a set of stylized layout schemes.

[0121] In this embodiment, the adjustment processing layer is a functional module layer of the third intelligent agent, used to iteratively optimize and correct the preliminary layout scheme based on the verification report generated by the physical verification layer, so as to resolve the detected conflicts and non-compliance. The adjustment processing layer can use optimization algorithms, such as genetic algorithms, simulated annealing algorithms, or reinforcement learning-based strategies, to automatically adjust the size, position, or orientation of the layout elements according to the error type and location in the verification report until all constraints are met.

[0122] Specifically, this application utilizes a multi-layered collaborative architecture with a third intelligent agent to generate stylized layout schemes from spatial planning. Specifically, the style encoding layer transforms user preference vectors into quantifiable style latent vectors, providing style guidance for subsequent layout generation. The layout condition generation layer encodes the spatial planning scheme set into spatial structure condition vectors and merges them with the style latent vectors to generate preliminary layout schemes, ensuring initial coordination in structure and style. Furthermore, the physical verification layer performs physical compliance checks and collision detection on the preliminary layout schemes, identifying potential problems and generating a verification report. Finally, the adjustment processing layer adjusts the schemes based on the verification report, ensuring that the final generated stylized layout scheme set not only conforms to user preferences and spatial planning but is also physically feasible and conflict-free.

[0123] As a specific implementation, upon receiving the user preference vector and the spatial planning scheme set generated by the second agent, the third agent initiates its workflow. Specifically, the style encoding layer can be an encoder based on the Transformer architecture, which encodes information such as text descriptions and image features in the user preference vector into a 128-dimensional style latent vector, capturing the user's preferences for color, material, furniture style, etc. The layout condition generation layer can use a graph convolutional network to process the spatial planning scheme set, specifically treating each room as a node in a graph and the connections between rooms as edges, thereby generating a spatial structure condition vector. Then, a multilayer perceptron fuses the style latent vector with the spatial structure condition vector to generate a preliminary layout scheme, which may include the preliminary placement and orientation of furniture. The physical verification layer calls a Unity3D or Unreal Engine-based system. The physics engine's simulation module performs collision detection on all furniture and walls in the initial layout plan. Based on pre-stored building design codes and furniture industry standards, it checks compliance such as minimum spacing between furniture, obstruction of door and window opening, and unobstructed fire exits. For example, if issues such as the sofa and coffee table being too close, or the refrigerator door not being fully open, are detected, these issues are recorded in the verification report. Finally, after receiving the verification report, the adjustment processing layer uses a genetic algorithm-based optimizer. Using the conflicts and non-compliance items pointed out in the verification report as penalty functions and user preference matching as fitness functions, it iteratively adjusts the furniture positions and rotation angles in the initial layout plan. For example, it might fine-tune the X and Y coordinates of the sofa or rotate the coffee table until all collision and compliance issues are resolved. This generates multiple stylized layout plans that meet the conditions and have different subtle differences for the user to choose from.

[0124] Through the above technical solutions, this application can effectively solve the problems that may occur when generating stylized layout schemes, such as insufficient extraction of style features, inconsistency between the layout scheme and actual physical rules, and lack of effective adjustment mechanisms. By introducing style coding, physical verification and adjustment processing, it ensures that the generated stylized layout scheme can not only accurately reflect user preferences, but is also physically reasonable and feasible, thereby improving the quality of the scheme and user satisfaction.

[0125] In one embodiment, step S50 includes:

[0126] S51: The fourth intelligent agent identifies the architectural component features and interior element features in the stylized layout scheme set, and extracts attribute information, including category labels, geometric parameters, spatial location and associated attributes;

[0127] In this embodiment, the fourth intelligent agent can be a deep learning-based model, such as a convolutional neural network or a graph neural network, or a rule-based expert system, or a combination of both. Its core function is to understand and parse the input design data and execute the transformation logic. Building component features refer to the basic components that make up a building entity, such as walls, doors, windows, columns, beams, and floor slabs. Identifying building component features means resolving the type, shape, size, and material information of these building components from the layout scheme. Interior element features refer to the non-structural elements inside the apartment, such as furniture, lighting fixtures, decorations, and appliances. Identifying interior element features means resolving the type, style, and size of these elements from the layout scheme. Information such as placement location; attribute information is data that provides a detailed description of the identified building components and interior elements. Among them, category labels indicate the specific type of component or element, which helps to match the correct model from the component library later; geometric parameters describe the size, shape, scale, etc. of the component or element, which is the basis for instantiating the 3D model; spatial location describes the precise coordinates and orientation of the component or element in the apartment space, such as the start and end points of the wall, the center coordinates and rotation angle of the door, which ensures the correct placement of the component in 3D space; association attributes describe the logical or functional relationship between the component or element and other components or elements, such as the relationship between the opening of the door and the wall, the embedding relationship of the window and the wall, and the attachment relationship of the socket and the wall.

[0128] S52: The fourth intelligent agent matches the corresponding parametric component family from the predefined parametric component library based on the features of building components and interior elements, and instantiates and generates several three-dimensional parametric components according to the extracted attribute information.

[0129] In this embodiment, the parametric component library refers to a pre-established database that stores a large number of reusable and configurable 3D model templates of building components and interior elements. These 3D model templates are typically parametric, meaning their geometry and attributes can be adjusted by input parameters. The parametric component library can contain industry-standard components, manufacturer-specific components, or custom components. A parametric component family refers to a collection of components in the library that share the same basic type and function but can be modified by adjusting parameters to generate different sizes, styles, or material variations. For example, a door is a component family that can generate different specific door models based on parameters such as width, height, and material. For example, instantiating and generating 3D parametric components refers to selecting a suitable parametric component family from a parametric component library based on the architectural component features, interior element features, and attribute information extracted from a set of stylized layout schemes, and using the extracted geometric parameters, spatial position, and other attribute information to generate an actual 3D model instance with specific dimensions, position, and attributes. Assembly logic refers to the rules and principles of how components connect, combine, and interact with each other. For example, how walls are connected at corners, how doors and windows are embedded in walls, and how furniture is placed. Assembly logic can be a preset set of rules or a pattern learned from a large number of design cases through machine learning.

[0130] S53: The fourth intelligent agent generates component connection relationships based on three-dimensional parametric components and their assembly logic. The component connection relationships include spatial topological relationships, hierarchical assembly relationships, and engineering constraint relationships.

[0131] In this embodiment, the component connection relationship describes the interaction and dependency between different three-dimensional parametric components. The component connection relationship includes spatial topological relationship, hierarchical assembly relationship and engineering constraint relationship. The spatial topological relationship describes the relative position and geometric connection of components in three-dimensional space, such as adjacent, intersecting, containing, and dependent. The hierarchical assembly relationship describes the parent-child or master-slave relationship between components, reflecting the composition hierarchy of the building. The engineering constraint relationship describes the engineering or design rules that components must meet, such as structural load transfer, pipe connection, electrical wiring, fire compartmentation, etc.

[0132] S54: The fourth intelligent agent generates a set of parametric building information models based on three-dimensional parametric components and the connection relationships between components.

[0133] In this embodiment, the parametric building information model set is a digital model that contains all detailed building components, interior elements and their interrelationships. It is not only a three-dimensional geometric model, but also contains rich attribute information, connection relationships and engineering data. Parametric means that the components and relationships in the model can be adjusted and updated by modifying parameters, thereby realizing rapid design iteration and change management.

[0134] Specifically, the fourth agent parses the received set of stylized layout schemes, identifies the features of building components and interior elements in the schemes, and extracts detailed information such as category labels, geometric parameters, spatial locations, and associated attributes corresponding to these features. Further, using the identified building component and interior element features, it matches the parametric component family that best meets the design requirements from a predefined parametric component library. Based on the previously extracted geometric parameters, spatial locations, and associated attributes, it instantiates the matched parametric component family, generating a series of three-dimensional parametric components with precise dimensions, locations, and specific attributes. After generating all independent three-dimensional parametric components, it further establishes connections between components based on the generated components and their inherent assembly logic, enabling the originally isolated three-dimensional components to form a logically complete and structurally sound whole. Finally, it integrates the instantiated three-dimensional parametric components with their connection relationships to generate a parametric building information model set.

[0135] Through the above technical solutions, this application can transform a set of stylized layout schemes into a set of parametric building information models with rich information and engineering logic. Specifically, by identifying the characteristics of building components and interior elements and extracting detailed attribute information, the refinement and accuracy of model construction are ensured. Matching and instantiation based on a parametric component library improve the efficiency and standardization of model generation, avoiding the tediousness and errors of manual modeling. More importantly, by generating component connection relationships such as spatial topology relationships, hierarchical assembly relationships, and engineering constraint relationships, the final building information model set is not just a simple stacking of three-dimensional geometry, but also has intelligent structural logic and editability, which can truly reflect the design intent and engineering requirements. This improves the intelligence level and application breadth of AI-based apartment customization solutions, effectively solving the problems of incomplete model information, difficulty in maintenance, and failure to meet actual engineering needs that may exist in traditional methods.

[0136] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0137] In one embodiment, a multimodal data-driven AI-based apartment layout customization solution generation system is provided. This system corresponds one-to-one with the multimodal data-driven AI-based apartment layout customization solution generation method described in the previous embodiment. The multimodal data-driven AI-based apartment layout customization solution generation system includes:

[0138] The feature extraction module is used to respond to the received task instructions, acquire the corresponding multimodal data, and extract features from the task instructions and multimodal data to obtain the design target feature set, which includes user preference vectors, optimization target identifiers, and constraint identifiers.

[0139] The constraint matching module is used to match constraint identifiers with a pre-built building code knowledge base to generate a preliminary set of constraints.

[0140] The constraint enhancement module is used to match several optimization strategies from a preset optimization strategy pool based on the optimization target identifier, and then, through the first intelligent agent, combine the matched optimization strategies and the user preference vector to perform multi-strategy collaborative optimization on the initial constraint set to generate a reinforced constraint set.

[0141] The scheme generation module is used to input the set of enhanced constraints into the second agent to generate a set of spatial planning schemes, and then use the third agent to combine the set of spatial planning schemes with user preference vectors to generate a set of stylized layout schemes.

[0142] The model generation module is used to input the stylized layout scheme set into the fourth intelligent agent to generate a parametric building information model set.

[0143] For specific limitations regarding the multimodal data-driven AI-based apartment layout customization system, please refer to the limitations of the multimodal data-driven AI-based apartment layout customization method described above, which will not be repeated here. Each module in the aforementioned multimodal data-driven AI-based apartment layout customization system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0144] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for generating AI-driven apartment layout customization solutions based on multimodal data, characterized in that, Including the following steps: In response to the received task instruction, the corresponding multimodal data is acquired, and feature extraction is performed on the task instruction and multimodal data to obtain the design target feature set, which includes user preference vector, optimization target identifier and constraint condition identifier; A preliminary set of constraints is generated by matching constraint identifiers with a pre-built building code knowledge base. Based on the optimization target identifier, several optimization strategies are matched from the preset optimization strategy pool. Then, through the first intelligent agent, the matched optimization strategies and user preference vectors are combined to perform multi-strategy collaborative optimization on the initial set of constraints, generating a set of enhanced constraints. The set of enhanced constraints is input into the second agent to generate a set of spatial planning schemes, and the third agent combines the set of spatial planning schemes with user preference vectors to generate a set of stylized layout schemes. The stylized layout scheme set is input into the fourth intelligent agent to generate a parametric building information model set. Specifically, the fourth intelligent agent identifies the building component features and interior element features in the stylized layout scheme set and extracts attribute information, including category labels, geometric parameters, spatial location and related attributes. The fourth intelligent agent matches the corresponding parametric component family from the predefined parametric component library based on the features of building components and interior elements, and instantiates and generates several three-dimensional parametric components according to the extracted attribute information. The fourth intelligent agent generates component connection relationships based on three-dimensional parameterized components and their assembly logic. The component connection relationships include spatial topological relationships, hierarchical assembly relationships, and engineering constraint relationships. The fourth intelligent agent generates a set of parametric building information models based on three-dimensional parametric components and the connection relationships between the components; The second intelligent agent includes a constraint parsing layer, a spatial optimization layer, and a scheme decoding layer. The step of inputting a set of enhanced constraints into the second intelligent agent to generate a set of spatial planning schemes includes: The constraint parsing layer parses the set of enhanced constraints, extracts spatial functional constraint features, area constraint features, and topological relationship constraint features, and generates a spatial parameterized representation. The spatial optimization layer executes a pre-defined iterative optimization process; Based on the spatial parameterized representation obtained after the iterative optimization process is completed and the intermediate optimization states recorded during the iteration, the scheme decoding layer generates a spatial planning scheme set containing at least one main scheme and several variant schemes. The third intelligent agent includes a style encoding layer, a layout condition generation layer, a physical verification layer, and an adjustment processing layer. The generation of a stylized layout scheme set by combining the spatial planning scheme set and user preference vectors through the third intelligent agent includes: The style encoding layer maps user preference vectors to style latent vectors; The layout condition generation layer encodes the spatial planning scheme set into the corresponding spatial structure condition vector, and merges the style potential vector with the spatial structure condition vector to generate a preliminary layout scheme. The physical verification layer, based on a predefined physical rule base and furniture knowledge base, performs collision detection and compliance verification on the preliminary layout plan and generates a verification report. The adjustment processing layer adjusts the initial layout scheme based on the verification report and generates a set of stylized layout schemes.

2. The method for generating AI-driven apartment layout customization solutions based on multimodal data according to claim 1, characterized in that: The steps of responding to the received task instruction, acquiring the corresponding multimodal data, and extracting features from the task instruction and multimodal data to obtain a design target feature set, wherein the design target feature set includes user preference vectors, optimization target identifiers, and constraint identifiers, include the following steps: The task instructions and multimodal data are input into a pre-trained feature extraction model to obtain fused semantic features; Parallel parsing of the fused semantic features generates a preliminary preference vector, an initial target identifier, and an initial constraint identifier. The initial preference vector, initial target identifier, and initial constraint identifier are subjected to consistency verification and calibration to obtain the design target feature set.

3. The method for generating AI-driven apartment layout customization solutions based on multimodal data according to claim 2, characterized in that: The multimodal data includes first data and second data. The feature extraction model includes a feature encoding layer, a feature fusion layer, and a feature enhancement layer. The step of inputting the task instructions and multimodal data into the pre-trained feature extraction model to obtain fused semantic features includes the following steps: The feature encoding layer encodes the first data to generate text feature vectors, spatial feature vectors, and parameter feature vectors. The first data includes task text information, apartment layout information, and structural parameter information. The feature fusion layer concatenates and fuses text feature vectors, spatial feature vectors, and parametric feature vectors to generate a fused feature vector. The feature enhancement layer encodes the second data to generate an enhanced feature vector, and uses the enhanced feature vector as a modulation signal to recalibrate the fused feature vector to generate fused semantic features.

4. The method for generating AI-driven apartment layout customization solutions based on multimodal data according to claim 2, characterized in that: The step of performing parallel parsing of the fused semantic features to generate a preliminary preference vector, an initial target identifier, and an initial constraint identifier includes the following steps: The fused semantic features are input into a pre-built parallel parsing model, wherein the parallel parsing model includes a first parsing layer, a second parsing layer, and a third parsing layer; The first parsing layer maps the fused semantic features to a continuous preference semantic space, generating a preliminary preference vector; The second parsing layer performs multi-label classification on the fused semantic features and identifies at least one dominant optimization objective from a predefined set of optimization objectives, generating the corresponding initial objective identifier; The third parsing layer performs sequence labeling and key information extraction on the fused semantic features, identifies physical and functional constraints, maps them to standard constraint entries in a predefined constraint library, and generates initial constraint identifiers.

5. The method for generating AI-driven apartment layout customization solutions based on multimodal data according to claim 1, characterized in that: The step of matching constraint identifiers through a pre-built building code knowledge base to generate a preliminary constraint set includes the following steps: The constraint identifiers are parsed to identify the constraint type and constraint parameters. Reasoning is performed based on a preset set of association rules. Implicit constraints are added based on the reasoning results, and a list of constraint identifiers containing several constraint identifiers is generated. Iterate through the list of constraint identifiers, query the building code knowledge base for the parameterization rules, value ranges and boundary conditions corresponding to all constraint identifiers, and instantiate each constraint identifier into at least one atomic constraint condition based on the query results. Logical conflict detection is performed on all atomic constraints. If a conflict is detected, the conflicting atomic constraints are adjusted and optimized according to the constraint priority rules and conflict resolution strategies preset in the building code knowledge base, and a preliminary set of constraints is generated.

6. The method for generating AI-driven apartment layout customization solutions based on multimodal data according to claim 1, characterized in that: The step of matching several optimization strategies from a preset optimization strategy pool based on the optimization target identifier, and then using a first intelligent agent to perform multi-strategy collaborative optimization on the initial constraint set to generate a reinforced constraint set, includes the following steps: The initial set of constraints is divided into a rigid constraint set and a flexible constraint set. The matched optimization strategy, user preference vector, and preliminary constraint set are input into the first agent, causing the first agent to execute a hierarchical optimization process, wherein the hierarchical optimization process includes: Keeping the rigid constraint set unchanged, a multi-objective optimization problem is constructed with the flexible constraint set as the optimization variable. The multi-objective optimization problem is solved iteratively, and the Pareto optimal relaxation interval of the flexible constraint set is calculated under the premise of satisfying the rigid constraint set. The optimization objective of the multi-objective optimization problem is defined by the matching optimization strategy. Based on the user preference vector, the corresponding re-constraint value is selected for each flexible constraint in the flexible constraint set within the Pareto optimal relaxation interval, and a constraint adjustment scheme is generated. Update the parameters of the flexible constraint set according to the constraint adjustment scheme, and generate the set of enhanced constraint conditions.

7. The method for generating AI-driven apartment layout customization solutions based on multimodal data according to claim 1, characterized in that: The iterative optimization process includes: (a) Compute the matching loss between the spatial parameterized representation and the set of reinforced constraints; (b) Based on the matching loss, the spatial parameterized representation is updated through a pre-defined backpropagation algorithm; (c) Using steps (a) and (b) as one iteration process, repeat the iteration process until any termination condition is reached, and record the intermediate optimization states generated during the iteration process. The termination conditions include the matching loss converging to a preset loss threshold and the number of iterations reaching the maximum number of iterations.

8. A multimodal data-driven AI-based apartment layout customization solution generation system, used to execute the steps of the multimodal data-driven AI-based apartment layout customization solution generation method as described in any one of claims 1-7, characterized in that, include: The feature extraction module is used to respond to the received task instructions, acquire the corresponding multimodal data, and extract features from the task instructions and multimodal data to obtain the design target feature set, which includes user preference vectors, optimization target identifiers, and constraint identifiers. The constraint matching module is used to match constraint identifiers with a pre-built building code knowledge base to generate a preliminary set of constraints. The constraint enhancement module is used to match several optimization strategies from a preset optimization strategy pool based on the optimization target identifier, and then, through the first intelligent agent, combine the matched optimization strategies and the user preference vector to perform multi-strategy collaborative optimization on the initial constraint set to generate a reinforced constraint set. The scheme generation module is used to input the set of enhanced constraints into the second agent to generate a set of spatial planning schemes, and then use the third agent to combine the set of spatial planning schemes with user preference vectors to generate a set of stylized layout schemes. The model generation module is used to input the stylized layout scheme set into the fourth intelligent agent to generate a parametric building information model set.

Images