User feedback driven modular housing solution dynamic generation method and system

Abstract

The application relates to the technical field of building informatization and residential industrialization design, in particular to a user feedback driven modular residential scheme dynamic generation method and system, aiming at the problems of low unstructured demand conversion accuracy, static lag of design rules and insufficient full-link collaboration capability of an existing residential design system, fusing user sketches and voice demands through multi-modal analysis, combining with an architectural field knowledge graph to resolve cross-modal semantic ambiguity; a design rule relationship graph is constructed with function areas as nodes and spatial topological relations as edges, a user negative evaluation is taken as a trigger signal, and a closed loop self-iteration of the design rule is realized through a graph convolution network; scheme generation, compliance verification, lightweight three-dimensional visualization, real-time cost evaluation, BIM parameter linkage and component tracking functions are synchronously integrated. The method can realize adaptive optimization and full-process digital collaboration of residential customized design and is suitable for a modular residential industrialization design scene.

Description

Technical Field

[0001] This application relates to the fields of building information modeling and residential industrialization design technology, specifically to a method and system for dynamically generating modular housing solutions driven by user feedback. Background Technology

[0002] With the rapid development of the prefabricated modular housing industry, the demand for customized housing products continues to increase, and users are gradually shifting from passively accepting standardized floor plans to actively participating in the design process. Current housing design systems have gradually achieved digital transformation.

[0003] For example, the Chinese patent application number CN119577885A, "A Simulation Method and System Based on User Needs of Variable Space Residential Buildings," obtains basic attribute information by parsing uploaded residential information, providing accurate foundational data for subsequent design. It divides residential information into local spaces and provides corresponding controls, enabling users to accurately express the needs of each space. This detailed method of needs collection reduces information loss during the communication of customer ideas and transforms user needs into structured design requirements data. This step effectively eliminates misunderstandings by designers. An interactive design interface is provided in a virtual environment, allowing users to directly participate in the design process, adjust and view effects in real time, thereby ensuring that the final design scheme is consistent with the customer's expectations.

[0004] However, the generation of apartment layouts relies on a static rule base preset by humans or a generation model trained offline. Rule updates require manual revision by architectural experts and retraining of the model by algorithm engineers. The iteration cycle is long and the labor cost is high. It is difficult to adapt to the dynamic changes in market preferences and user needs, resulting in low efficiency in solution iteration. Summary of the Invention

[0005] A user feedback-driven method for dynamically generating modular housing solutions includes the following steps:

[0006] S1. Multimodal requirement analysis and constraint fusion: Convolutional neural network is used to extract features from the user-input sketch image to obtain spatial feature vectors; Sequence model analysis is used to analyze the user-input voice or text requirements to obtain functional constraint vectors; After aligning the entity reference relationships of the two types of vectors through cross-modal semantic ambiguity resolution, a unified 512-dimensional design constraint vector is generated by fusion.

[0007] S2. Pre-construct a design rule relationship diagram with residential functional areas as nodes and spatial topological relationships as edges. Each node corresponds to an initial node feature matrix. The number of cluster centers is dynamically calculated based on the negative feedback rate, using the initial node feature matrix. Multiple preference rule clusters are obtained by clustering the rule nodes; then the design constraint vector is matched with the cluster center, the optimal rule cluster is selected as the generation prior, and finally the gridded house type scheme is output.

[0008] S3. Scheme Verification and Visualization Rendering: Perform physical collision detection and verification on the generated floor plan. After the verification is passed, use WebGL to perform 3D rendering and display it to the user.

[0009] S4. User Review Collection and Update Trigger: Collect user reviews of the displayed apartment layouts and calculate the negative feedback rate using the following formula. :

[0010]

[0011] in, The number of negative reviews. The total number of evaluations; when the following conditions are met. When the cumulative number of valid evaluation samples is greater than or equal to the set value, the rule self-iteration process is triggered.

[0012] S5. Graph Convolution Rule Self-Iteration: The node feature matrix of the design rule relationship graph is updated through at least two layers of graph convolutional network. Based on the updated node features, the number of cluster centers is recalculated and node clustering is completed. The updated design rule relationship graph is used for the next round of house type scheme generation, forming a closed loop iteration.

[0013] Preferably, in step S1, cross-modal semantic ambiguity resolution includes:

[0014] S101, a knowledge graph library for the architectural field of pre-constructed residential design entities;

[0015] S102. Calculate the cosine similarity matrix between the spatial feature vector and the functional constraint vector. ;

[0016] S103, When the similarity matrix When the maximum value is less than the preset threshold, the knowledge graph is called to query the candidate set of synonyms, and the entity matching result with the highest probability is output to complete the entity reference alignment.

[0017] Preferably, in step S5, the node feature update process of the graph convolutional network includes:

[0018] The first layer is the neighborhood feature aggregation layer, which performs first-order neighborhood feature fusion according to the following formula:

[0019]

[0020] in, This is the initial node feature matrix; Adjacency matrix for adding self-loops; for The corresponding degree matrix; This is the first layer of trainable weight matrix; It is the ReLU activation function; This is the feature matrix of the first layer output nodes;

[0021] The second layer is the feature update layer, which outputs the final updated node features using the following formula:

[0022]

[0023] in, This is the trainable weight matrix for the second layer. This is the updated node feature matrix;

[0024] The mean squared error is used as the loss function, and the weight matrix is ​​updated iteratively through backpropagation to fit the node features with the user's negative feedback labels.

[0025] Preferably, the number of cluster centers in S2 The calculation formula is:

[0026]

[0027] in, The total number of rule nodes in the design rule relationship graph. The number of baseline nodes in a single cluster is 12. For floor operations, As a dynamic adjustment factor, based on the negative feedback rate Segmented value retrieval:

[0028]

[0029] The maximum value is 15.

[0030] Preferably, the convergence determination and fallback mechanism of rule self-iteration in S5 includes: setting the maximum number of iterations for a single update of the graph convolutional network to 20; if the loss function decreases by less than 1 / 3 consecutive iterations... If the iteration fails, it is considered converged and the iteration stops. After the update is completed, user samples from the past 7 days are used for testing. If the negative feedback rate of the test set decreases by ≥5% compared to before the update, the update is considered effective and the new rules are enabled. If the iteration fails to converge after 20 rounds or the negative feedback rate of the test set increases, it will automatically revert to the stable design rule relationship graph of the previous version.

[0031] Preferably, the physical collision detection and specification verification in S3 includes: converting the apartment building components into rigid body objects using the Cannon.js physics engine; triggering a conflict warning when the spacing between components is less than 0.1m; and automatically adjusting the spacing between components to a safe threshold using a rigid body position interpolation algorithm; and verifying the specification against the residential design requirements in real time, returning the location coordinates of the non-compliant component and the non-compliant requirements when the verification fails.

[0032] Preferably, in step S3, WebGL is used for 3D rendering, including:

[0033] S301: Employs a 6-level progressive LOD loading method, and uses a low-precision model with fewer than 500 faces when the viewing distance is greater than 10m.

[0034] S302. Accelerate model-view matrix calculation through WebAssembly;

[0035] S303. Employs octree-based spatial segmentation frustum culling to control the first frame rendering latency to less than 50ms.

[0036] S304. During the rendering process, a heatmap of cost per unit area is generated simultaneously: The residential model is divided into 1m×1m grids.

[0037] Cost ≤ Lowest Price, mapped to the blue color scheme;

[0038] The range between low price, cost, and high price is represented by the green color scheme.

[0039] Costs exceeding or equaling high prices are represented by the red color scheme.

[0040] Preferably, S3 further includes an AR real-scene overlay step, which uses the WebXR HitTestAPI to detect planar feature points in the camera image, uses the PnP algorithm to solve the spatial transformation matrix of the house model and iteratively optimizes it; when the reprojection error is greater than 2cm, SLAM sparse point cloud-assisted positioning is enabled to control the spatial anchor point positioning error to be less than 2cm, thereby realizing the real-scene overlay display of the house model on the actual site.

[0041] Preferably, it also includes BIM parameter intervention steps: opening the IFC standard BIM model data interface, receiving the component parameter change instructions input by the user, and automatically triggering three linked updates, including calling the Radiance engine to recalculate physical performance parameters, synchronously updating component cost and cost heat map, and re-verifying the compliance of the scheme after parameter changes.

[0042] A user feedback-driven, modular housing design dynamic generation system, including:

[0043] The intelligent requirement input unit is used to acquire the user's input sketch image and voice / text requirements, extract features from them, and after cross-modal semantic ambiguity resolution, fuse them to generate a design constraint vector;

[0044] The dynamic rule engine unit is used to store the design rule relationship graph with residential functional areas as nodes and spatial topological relationships as edges. It dynamically calculates the number of cluster centers based on the current negative feedback rate and completes node clustering, outputting the matching preference rule clusters.

[0045] The scheme generation and verification unit is used to generate gridded house type schemes by taking design constraint vectors and matching preference rule clusters as inputs, and to perform component physical collision detection and residential design code verification.

[0046] The visualization and user feedback unit is used to display the verified apartment layouts using WebGL 3D rendering, and to collect user evaluation data and calculate the negative feedback rate.

[0047] The rule self-iterative execution unit is used to update the node features of the design rule relationship graph through a two-layer graph convolutional network and recalculate the number of cluster centers to complete clustering when the negative feedback rate meets the triggering condition.

[0048] The updated design rule relationship diagram is fed back to the dynamic rule engine unit, forming a closed-loop iterative dynamic design rule system.

[0049] The beneficial effects of this invention are as follows:

[0050] 1. This invention enables closed-loop self-iterative optimization of the design rule system, continuously adapting to dynamic changes in market and user preferences. Without relying on manual, periodic rule revisions and model retraining, it automatically adjusts the weights and classification logic of design rules based on user feedback, ensuring that the apartment layout generation logic always aligns with current mainstream needs, maintaining long-term user adaptability of the solutions, while significantly reducing the manpower and time costs of rule maintenance.

[0051] 2. It establishes a seamless data pathway across the entire process, from scheme design, compliance verification, cost assessment to construction implementation. Lightweight cross-platform visualization capabilities support online collaboration among multiple roles, and the parametric building information model interface links with the supply chain and construction tracking system, avoiding data gaps and information discrepancies at each stage. This ensures the compliance and feasibility of the design scheme and effectively supports the industrial design and construction of modular housing.

[0052] 3. It can identify users' unstructured design needs, lowering the threshold for ordinary users to participate in residential customization. At the same time, the design requirements expressed by users through hand-drawn sketches and natural language can be accurately interpreted through cross-modal semantic alignment and domain knowledge disambiguation, avoiding the deviation of requirements caused by differences in professional expression. This makes the final generated plan more in line with the user's real living expectations and reduces the cost of demand communication and repeated adjustments to the plan. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the framework of the present invention;

[0054] Figure 2 This is a flowchart. Detailed Implementation

[0055] The application principle of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0056] Example 1:

[0057] like Figure 1-2 As shown, the user feedback-driven modular housing scheme dynamic generation system includes an intelligent demand input unit, a dynamic rule engine unit, a scheme generation and verification unit, a visualization and user feedback unit, and a rule self-iterative execution unit. It constructs a graph-structured design rule system with functional areas as nodes and topological relationships as edges. Simultaneously, it uses negative user feedback as a driving signal and achieves closed-loop self-iteration of the generated rules through a graph convolutional network. Furthermore, it integrates multimodal demand analysis and lightweight 3D visualization technology to form a full-link adaptive generation system for modular housing.

[0058] The intelligent requirement input unit transforms unstructured user input into standardized design constraints. It includes a sketch feature extraction module, a multimodal requirement parsing module, a semantic ambiguity resolution module, and a constraint fusion module. Specifically, it includes:

[0059] The sketch feature extraction module uses a ResNet50-FPN convolutional neural network to process user-submitted 600×600 pixel RGB format sketches, extract spatial topological relationships and functional area contours, and output a 2048-dimensional spatial feature vector.

[0060] The multimodal requirement analysis module uses a bidirectional long short-term memory network to analyze voice or text input, identify keywords such as functional area names, size constraints, and special requirements, and generate a 5-dimensional functional constraint vector that includes area requirements, lighting intensity, and streamline priority.

[0061] The semantic ambiguity resolution module incorporates a knowledge graph library in the architectural domain, storing residential design entities and their relationships. Entities are categorized into three types: functional area type, building material attributes, and regulatory requirements. The module calculates the cosine similarity matrix between sketch feature vectors and speech feature vectors using a cross-modal attention alignment unit. ;when When the maximum value is less than the preset threshold of 0.7, the ambiguity decision-maker calls the knowledge graph to query the candidate set of synonyms and outputs the entity matching result with the highest probability. For example, when a user inputs "living room lighting", the system prioritizes associating the entity of "south-facing window" rather than "lighting equipment" to eliminate referential ambiguity.

[0062] The constraint fusion module fuses the 2048-dimensional spatial feature vector and the 5-dimensional functional constraint vector through a fully connected layer to reduce the dimensionality and output a unified 512-dimensional design constraint vector, which is then transmitted to the scheme generation unit.

[0063] The dynamic rule engine unit is used to store the design rule relationship graph and to implement dynamic clustering and matching of rules, specifically including:

[0064] The design rule relationship diagram is denoted as follows: ,in, For a set of rule nodes, It is a set of topological edges.

[0065] The node set centers on 12 basic functional areas: living room, master bedroom, secondary bedroom, kitchen, bathroom, dining room, balcony, entryway, storage room, study, dressing room, and corridor. It is expanded by three dimensions: area level, lighting level, and orientation preference, forming 142 rule nodes. Each node corresponds to an 8-dimensional initial feature vector, with the dimensions being, in order: upper and lower limits of area, lighting weight, orientation priority, circulation connection weight, compliance threshold, material preference, cost range, and total area of ​​the suitable apartment type. The initial values ​​are marked by architectural design experts and calibrated through a database of 100,000 compliant commercial apartment types.

[0066] The initial eigenvectors of all nodes together constitute the initial node feature matrix. , dimension ,in The total number of regular nodes, used in the initial clustering phase. K-Means clustering is performed on the input to obtain an initial cluster of preference rules; after iterative updating by a graph convolutional network, the updated node feature matrix is ​​output. Its dimensions and Completely identical, the system uses replace Then, the number of cluster centers is recalculated and clustering is performed to obtain the updated preference rule clusters, which are used for the next round of apartment layout generation.

[0067] Edge sets represent the spatial topological relationships between nodes, categorized into four types: adjacency, containment, connectivity, and exclusion. Edge weights range from [0,1], with initial weights calculated using the co-occurrence frequency of classic apartment layouts. ,in Functional area and Number of adjacent apartment types For containing functional areas The total number of apartment types; the adjacency matrix is ​​stored using a symmetric matrix, with diagonal elements being 0.

[0068] During dynamic clustering and rule matching, based on the current system's negative feedback rate... Calculate the number of cluster centers using the following formula. :

[0069]

[0070] in, The total number of rule nodes in the design rule relationship graph. The number of baseline nodes in a single cluster is 12. For floor operations, As a dynamic adjustment factor, based on the negative feedback rate Segmented value retrieval:

[0071]

[0072] The maximum value is 15 to avoid rule fragmentation.

[0073] The K-Means algorithm is used to cluster the node feature vectors to obtain K preference rule clusters, each cluster corresponding to a user preference style. After the user inputs the design constraint vector, the cosine similarity between the constraint vector and each cluster center is calculated. The rule cluster with the highest similarity is selected as the generation prior and embedded into the conditional input layer of the generative model.

[0074] The scheme generation and verification unit includes a generative layout engine, a physics collision detection module, and a specification verification module, among which:

[0075] The generative layout engine takes a 512-dimensional design constraint vector and a matching rule cluster prior as input, and outputs a 32×32 gridded house layout scheme through the Transformer-GAN model with a grid resolution of 0.5m / grid.

[0076] The physical collision detection module converts the building components into rigid body objects using the Cannon.js engine, and sets the component density to 7850 kg / m³. When a red collision warning is triggered when the component spacing is less than 0.1 m, a rigid body position interpolation algorithm is used to automatically adjust the component spacing to a safe threshold and record the displacement trajectory.

[0077] The standard verification module has a built-in residential design standard clause library, which verifies indicators such as window-to-floor ratio, corridor width, and fire distance of the plan in real time. When the verification fails, it returns the location coordinates of the violation and the corresponding violation clause.

[0078] The visualization and user feedback unit includes a WebGL rendering module, a collaborative control module, an AR real-time overlay module, and a user evaluation collection module, among which:

[0079] The WebGL rendering module is based on WebGL 2.0 to achieve cross-platform visualization. It adopts a triple optimization strategy: 6-level progressive LOD hierarchical loading, and low-precision models with less than 500 faces are enabled when the view distance is >10m; WebAssembly is used to accelerate the calculation of the model-view matrix; and octree-based spatial segmentation frustum culling is used to reduce the rendering load by 67%; ultimately achieving a first frame rendering latency of <50ms, supporting responsive operation on desktop and mobile devices.

[0080] The rendering process synchronously generates a heatmap of the cost per unit area. The residential model is divided into 1m×1m grids. Costs ≤2000 yuan / ㎡ are mapped to blue RGB(0,0,255), 2001-5000 yuan / ㎡ are mapped to green RGB(0,255,0), and costs ≥5000 yuan / ㎡ are mapped to red RGB(255,0,0).

[0081] The collaborative control module synchronizes multi-terminal operations based on an operation conversion algorithm, generates a colored collaborative cursor with a user ID for each user, and resolves conflicting operations based on millisecond-level timestamp priority.

[0082] The AR real-time overlay module registers the solution model space anchor points in the mobile camera image through the WebXR API; it uses the HitTest API to detect planar feature points, solves the model space transformation matrix through the PnP algorithm and iteratively optimizes it; when the reprojection error is >2cm, it enables SLAM sparse point cloud assisted positioning to control the overall positioning error to <2cm.

[0083] The user feedback collection module records users' 5-point ratings and sub-item evaluations of the solution, and calculates the negative feedback rate using the following formula:

[0084]

[0085] in, The number of valid negative reviews within the statistical period is defined as follows: a score of ≤2 out of 5 is considered a negative review. The total number of valid evaluations within the same statistical period, i.e., the total number of all valid samples participating in the calculation of the negative feedback rate. The system generates a list of projects to be evaluated based on the residential design plan.

[0086] A valid evaluation must meet the following criteria: the rating is submitted after a complete review of the 3D design; only one valid evaluation is counted for the same user on the same design; and invalid data, such as those identified by the system as abnormal score manipulation or duplicate submissions, is excluded.

[0087] The above indicators are calculated using a rolling statistical window, with the statistical period being all cumulative valid evaluations from the last rule update to the current time. And the cumulative effective evaluation sample size, i.e. When there are ≥300 rules, an update instruction is sent to the rule self-iteration execution unit.

[0088] The rule-based self-iterative execution unit is used to achieve automated rule updates, including negative feedback attribution and supervision signal generation, graph convolution update, and convergence determination and fallback mechanism, wherein:

[0089] Negative feedback attribution and monitoring signal generation are used to map user negative reviews to node-level monitoring tags: for itemized negative reviews, they are directly mapped to the corresponding functional area node and its associated edges, with a node monitoring weight set to 1.5; for global negative reviews, negative weights are allocated according to the functional area area ratio, and adjusted based on the frequency of mention of demand keywords. Finally, monitoring tags are generated for each node. This represents the actual negative feedback rate of the solution corresponding to that node.

[0090] Graph convolutional update uses a two-layer graph convolutional network to update node features, and the calculation formula is as follows:

[0091] First layer: Neighborhood feature aggregation layer

[0092]

[0093] in, This is the initial node feature matrix; Adjacency matrix for adding self-loops; for The corresponding degree matrix; This is the first layer of trainable weight matrix; It is the ReLU activation function; This is the feature matrix of the output nodes of the first layer.

[0094] Second feature update layer:

[0095]

[0096] in, This is the trainable weight matrix for the second layer. This is the updated node feature matrix.

[0097] Mean squared error is used as the loss function:

[0098]

[0099] in, The negative feedback rate of nodes predicted by the model. The actual monitoring labels are updated iteratively through backpropagation. , Minimize the loss function to reduce the weight of topological relationships with high negative feedback and increase the weight of topological relationships with high satisfaction.

[0100] The convergence determination and fallback mechanism are implemented by setting the maximum number of iterations per update to 20; if the loss decreases by less than 20 for three consecutive iterations... If convergence is detected, iteration is stopped. After the update is completed, the system tests with samples from the past 7 days. If the negative feedback rate of the test set decreases by ≥5% compared to before the update, the update is deemed effective and the new rules are enabled. If convergence is not achieved after 20 iterations or the negative feedback rate of the test set increases, the system automatically reverts to the previous stable rule graph and records the anomaly log.

[0101] The supporting functional units include a real-time cost assessment unit, a precast component inventory prediction unit, a BIM parameter intervention interface, and a component tracking and simulation unit, among which:

[0102] The real-time cost assessment unit connects with publicly available quotations from building materials websites, guidance prices from local housing and construction departments, and real-time quotation systems from suppliers. Based on BIM models, it accurately measures component volumes and, combined with daily updated price data, calculates actual costs using a loss coefficient calibration formula. ,in This is the loss factor, with a default value of 0.03, and the cost calculation error is controlled within 5%.

[0103] The precast component inventory forecasting unit analyzes supplier inventory data using an LSTM time series model and calculates a safety stock threshold based on the average daily consumption over the past 30 days. When the required component inventory is lower than the safety threshold, it automatically recommends alternative materials based on the material equivalence parameter table, such as the ALC board having a strength equivalence ratio of 1:1.3 with reinforced concrete.

[0104] The BIM parameter intervention interface adopts the open IFC standard data interface, which supports professional users to modify component parameters. After the parameter changes, three linked updates are automatically triggered: calling the Radiance engine to recalculate physical properties, synchronously updating cost and heat map, and re-verifying compliance.

[0105] The component tracking and simulation unit assigns a unique RFID code to each prefabricated component, linking it to information throughout the entire production, transportation, and installation process; when the transportation delay exceeds 24 hours, a red highlight warning is displayed in the 3D view; at the same time, it predicts the construction progress based on discrete event simulation, and triggers a scheme adjustment suggestion when the deviation exceeds 10%.

[0106] Example 2:

[0107] Based on Example 1, this example provides a complete process for customized residential design and rule iteration, including the following steps:

[0108] Step 1: User Requirements Input and Analysis: Users access the system through a cross-platform browser, submit a 600×600 pixel RGB format floor plan sketch in the sketch drawing area on the left side of the operation interface, and simultaneously input "Add a steel structure balcony, and ensure good lighting in the living room" via the voice button.

[0109] System execution: The sketch feature extraction module outputs a 2048-dimensional spatial feature vector through the ResNet50-FPN network; the speech parsing module generates a 5-dimensional functional constraint vector through BiLSTM; the semantic ambiguity resolution module calculates the cosine similarity between the two types of vectors to be 0.62 < 0.7, calls the architectural knowledge graph to associate "lighting" with the entity "south-facing window" to complete the ambiguity resolution; at the same time, the constraint fusion module reduces the dimensionality of the two types of vectors and fuses them into a 512-dimensional design constraint vector.

[0110] Step Two: Dynamic Rule-Driven Solution Generation: The dynamic rule engine loads a design rule relationship graph containing 142 nodes, and the current system's cumulative negative feedback rate. ,belong Interval, take dynamic adjustment factor Substitute the values ​​into the formula to calculate the number of cluster centers:

[0111]

[0112] Thirteen preference rule clusters were obtained through K-Means clustering. The system calculated the similarity between the 512-dimensional constraint vector and the center of each cluster, matched the 7th transparent rule cluster, and embedded its rule prior into the generation conditions.

[0113] Subsequently, the generative layout engine inputs constraint vectors and rule priors, and outputs a 32×32 gridded apartment layout scheme through Transformer-GAN; the physical collision detection module detects that the spacing between balcony railings is 0.08m < 0.1m safety threshold, triggers a red warning, automatically adjusts the spacing to 0.12m and records the displacement.

[0114] Step 3: Real-time Visualization and Collaborative Verification: The solution model is loaded via the WebGL rendering module, using a 6-level LOD hierarchical loading approach. A low-precision model is enabled when the viewing distance is greater than 10m. Matrix operations are accelerated using WebAssembly, resulting in a first-frame rendering latency of 42ms. The specification verification module verifies that the master bedroom window area is 2.1㎡, and the window-to-floor ratio is 0.175, which is greater than 1 / 7, conforming to residential design specifications.

[0115] In addition, the collaborative control module synchronizes the operations of the designer's end and the client end based on the OT algorithm, generating blue and red ID cursors respectively, and resolving conflicting operations with millisecond-level timestamps. Users can overlay the design onto the actual site using the mobile AR function, with a planar positioning error of 1.7cm, which meets the accuracy requirements.

[0116] Step Four: Cost Assessment and Feedback Collection: The real-time cost assessment unit connects to the building material price API to obtain the unit price of 304 stainless steel at 15.2 yuan / kg. Based on the BIM model, the volume of the balcony components is calculated to be 3.2m³. Substitute these values ​​into the loss factor formula:

[0117] Actual cost = (3.2 × 7850 × 15.2) × (1 + 0.03) = 386.5 yuan / ㎡

[0118] The system then generates a 1m×1m cost heat map. The cost of the balcony area is >5000 yuan / ㎡, mapped to the red RGB (255,50,50); the cost of the bedroom area is 1800 yuan / ㎡, mapped to the blue RGB (50,50,255).

[0119] The end user gave a rating of 4.2 stars, which was entered into the database; combined with the current total of 1247 valid reviews, the overall negative feedback rate is 17%, which has not reached the 30% update threshold.

[0120] Step 5: Professional Intervention with BIM Parameters: The designer modifies the thermal conductivity of the exterior wall through the BIM parameter interface, adjusting it from 0.38 W / (m²・K) to 0.42 W / (m²・K). Subsequently, the system automatically triggers a linked update, calling the Radiance engine to recalculate the daylight flux; the cost module refreshes the material list and heat map; and the specification verification module re-verifies that the K value of 0.42 < 0.45, meeting the energy-saving design standards.

[0121] Step Six: Construction Progress Monitoring and Early Warning: The component tracking unit assigns RFID code SP20240715-0001 to the prefabricated balcony component. A transportation delay of 26 hours is detected and highlighted in yellow (RGB(255,255,0)) in the 3D view as a warning. The discrete event simulation model predicts a construction progress deviation of 16.7% > 10%, and the system pushes an optimization suggestion: enable the local inventory of steel bars as a substitute.

[0122] Step 7: Rule Adaptive Update: If the system's negative feedback rate is >20% for three consecutive days and the cumulative sample size meets the trigger condition, rule self-iteration is initiated. A dynamic adjustment factor is then used. The number of cluster centers was calculated. The loss decreased by less than 140% after updating the feature matrix of 142 nodes using a two-layer GCN and iterating for 14 rounds. The convergence is determined.

[0123] Subsequent testing on the test set verified that the negative feedback rate decreased from 25% to 19% after the update, a reduction of 24%, indicating the update was effective. Specifically, the adjacent topology weight from balcony to living room increased from 0.75 to 0.82, and the adjacent weight from kitchen to secondary bedroom decreased from 0.31 to 0.24. The updated rule diagram was officially launched and put into use. The final system output includes a complete solution package containing a BIM model, cost list, and construction progress forecast, supporting export in IFC format for integration with production lines. In this embodiment, the total solution generation time was 3.1 seconds, and the violation correction time was 1.8 minutes, resulting in a 62% reduction in mobile memory usage compared to the traditional Unity engine.

Claims

1. A user feedback-driven method for dynamically generating modular housing solutions, characterized by: Includes the following steps: S1. Spatial feature vectors are extracted from user-input sketches using a convolutional neural network, and functional constraint vectors are extracted from speech or text requirements using a sequence model. After cross-modal semantic ambiguity resolution and entity referencing alignment, the vectors are fused to generate a unified design constraint vector. S2. Pre-construct a design rule relationship diagram with residential functional areas as nodes and spatial topological relationships as edges. Each node corresponds to an initial node feature matrix. The number of cluster centers is dynamically calculated based on the negative feedback rate, using the initial node feature matrix. Clustering the rule nodes yields multiple preference rule clusters; Then, the design constraint vector is matched with the cluster center, the optimal rule cluster is selected as the generation prior, and finally the gridded house type scheme is output. S3. Verify the generated floor plan, and then perform 3D rendering and display after it passes the verification. S4. Collect user evaluation data on the displayed apartment layouts and calculate the negative feedback rate. When the preset trigger conditions are met, start the rule self-iteration. S5. The node feature matrix of the design rule relationship graph is processed through a two-layer graph convolutional network. The first layer performs first-order neighborhood feature fusion according to the following formula: in, This is the initial node feature matrix; Adjacency matrix for adding self-loops; for The corresponding degree matrix; This is the first layer of trainable weight matrix; It is the ReLU activation function; This is the feature matrix of the first layer output nodes; The second layer updates node features by outputting the following formula: in, This is the trainable weight matrix for the second layer. The updated node feature matrix is ​​used; mean squared error is used as the loss function, and the number of cluster centers is recalculated and the node clustering is completed based on the updated node features. The updated design rule relationship graph is used for the next round of house type scheme generation, forming a closed loop iteration.

2. The user feedback-driven modular housing scheme dynamic generation method according to claim 1, characterized in that: In S1, cross-modal semantic ambiguity resolution includes: S101, a knowledge graph library for the architectural field of pre-constructed residential design entities; S102. Calculate the cosine similarity matrix between the spatial feature vector and the functional constraint vector. ; S103, When the similarity matrix When the maximum value is less than the preset threshold, the knowledge graph is called to query the candidate set of synonyms, and the entity matching result with the highest probability is output to complete the entity reference alignment.

3. The user feedback-driven modular housing scheme dynamic generation method according to claim 1, characterized in that: The number of cluster centers in S2 The calculation formula is: in, The total number of rule nodes in the design rule relationship graph. The number of baseline nodes in a single cluster is 12. For floor operations, As a dynamic adjustment factor, based on the negative feedback rate Segmented value retrieval: The maximum value is 15.

4. The user feedback-driven modular housing scheme dynamic generation method according to claim 1, characterized in that: In S4, the negative feedback rate The calculation formula is as follows: in, The number of negative reviews. This represents the total number of evaluations. The triggering condition is: satisfied And the cumulative number of valid evaluation samples is greater than or equal to the set value.

5. The user feedback-driven modular housing scheme dynamic generation method according to claim 3, characterized in that: The convergence determination and fallback mechanism of rule-based self-iteration in S5 includes: setting the maximum number of iterations for a single update of the graph convolutional network to 20; if the loss function decreases by less than 1 / 3 consecutive iterations... If the iteration fails, it is considered converged and the iteration stops. After the update is completed, user samples from the past 7 days are used for testing. If the negative feedback rate of the test set decreases by ≥5% compared to before the update, the update is considered effective and the new rules are enabled. If the iteration fails to converge after 20 rounds or the negative feedback rate of the test set increases, it will automatically revert to the stable design rule relationship graph of the previous version.

6. The user feedback-driven modular housing scheme dynamic generation method according to claim 1, characterized in that: In S3, the verification includes: converting the apartment building components into rigid body objects using the Cannon.js physics engine; triggering a conflict warning when the spacing between components is less than 0.1m; and automatically adjusting the spacing between components to a safe threshold using a rigid body position interpolation algorithm; and verifying the specifications against the residential design requirements in real time, returning the location coordinates of the non-compliant components and the non-compliant requirements when the verification fails.

7. The user feedback-driven modular housing scheme dynamic generation method according to claim 1, characterized in that: The S3 uses WebGL for 3D rendering, including: S301: Employs a 6-level progressive LOD loading method, and uses a low-precision model with fewer than 500 faces when the viewing distance is greater than 10m. S302. Accelerate model-view matrix calculation through WebAssembly; S303, Employ octree-based spatial segmentation frustum culling; S304. During the rendering process, a heatmap of cost per unit area is generated simultaneously: The residential model is divided into 1m×1m grids. Cost ≤ Lowest Price, mapped to the blue color scheme; The range between low price, cost, and high price is represented by the green color scheme. Costs exceeding or equaling high prices are represented by the red color scheme.

8. The user feedback-driven modular housing scheme dynamic generation method according to claim 1, characterized in that: The S3 also includes an AR real-scene overlay step, which uses the WebXR HitTestAPI to detect planar feature points in the camera image, uses the PnP algorithm to solve the spatial transformation matrix of the house model and iteratively optimizes it; when the reprojection error is greater than 2cm, SLAM sparse point cloud-assisted positioning is enabled.

9. The user feedback-driven modular housing scheme dynamic generation method according to claim 1, characterized in that: It also includes BIM parameter intervention steps: opening the IFC standard BIM model data interface, receiving the component parameter change instructions input by the user, automatically triggering three linked updates, including calling the Radiance engine to recalculate physical performance parameters, synchronously updating component cost and cost heat map, and re-verifying the compliance of the scheme after parameter changes.

10. A user feedback-driven modular housing scheme dynamic generation system, characterized in that: include: The intelligent requirement input unit is used to acquire the user's input sketch image and voice / text requirements, extract features from them, and after cross-modal semantic ambiguity resolution, fuse them to generate a design constraint vector; The dynamic rule engine unit is used to store the design rule relationship graph with residential functional areas as nodes and spatial topological relationships as edges. It dynamically calculates the number of cluster centers based on the current negative feedback rate and completes node clustering, outputting the matching preference rule clusters. The scheme generation and verification unit is used to generate gridded house type schemes by taking design constraint vectors and matching preference rule clusters as inputs, and to perform component physical collision detection and residential design code verification. The visualization and user feedback unit is used to display the verified apartment layouts using WebGL 3D rendering, and to collect user evaluation data and calculate the negative feedback rate. The rule self-iterative execution unit is used to update the node features of the design rule relationship graph through a two-layer graph convolutional network and recalculate the number of cluster centers to complete clustering when the negative feedback rate meets the triggering condition. The updated design rule relationship diagram is fed back to the dynamic rule engine unit, forming a closed-loop iterative dynamic design rule system.

Images