A method, system, device and medium for specification detection of an IFC file

Abstract

The application discloses an IFC file specification detection method, system, device and medium, and relates to the technical field of building information model data processing. The method comprises the following steps: analyzing an IFC file, extracting component attribute information and spatial topology / functional relationship between components, constructing a graph structure with components as nodes and relationships as edges, and marking abnormal nodes to form a training sample set; a graph neural network specification detection model based on an RGCN layer is constructed, a supervised classification loss and a self-supervised reconstruction loss are fused to train the model; and the trained model is applied to an IFC file to be detected to realize abnormal detection. The method breaks through the limitation of traditional static rule detection, can accurately identify hidden errors, and improves the intelligent level of IFC file detection and the reliability of collaborative work.

Description

Technical Field

[0001] This invention relates to the field of building information modeling data processing technology, specifically to a method, system, equipment, and medium for standardizing IFC files. Background Technology

[0002] IFC (Industry Foundation Classes) files, as an open data standard widely used in Building Information Modeling (BIM), are core to achieving seamless exchange and collaborative work of building information between different disciplines and software platforms. Throughout the entire lifecycle of a construction project, IFC files carry collaborative information from multiple disciplines, including architecture, structure, and mechanical and electrical systems. The standardization of these data directly affects the usability of the BIM model, the efficiency of collaborative work, and the accuracy of subsequent engineering applications.

[0003] However, in actual engineering design and delivery processes, the generation, conversion, and transfer of IFC files often introduce various standardization issues due to software differences, version compatibility problems, and human operating habits, severely impacting the usability and collaborative efficiency of BIM data. On the one hand, differences in the compatibility of different design software and inconsistent IFC Schema versions lead to compatibility problems when parsing files across platforms. On the other hand, different human operating habits can also cause data deviations, specifically manifested as missing or incorrect file object attributes (such as missing material parameters or lost component geometric information) and chaotic data hierarchy relationships (such as missing relationships between components and spaces). These problems seriously affect the consistency and usability of BIM data, posing potential risks to engineering construction.

[0004] Currently, the quality control of IFC documents primarily relies on static rule-based methods, such as using IFC Schema for syntax validation, matching attribute formats using regular expressions, or making logical judgments based on predefined business rules. These methods can effectively identify explicit errors such as missing attribute fields, incorrect data types, and violations of explicit enumeration values, and are practical in simple scenarios.

[0005] However, the limitations of static rule-based verification methods are also quite obvious: First, the completeness of the rule base depends on prior knowledge, making it difficult to cover all potential error patterns, especially implicit errors caused by differences in modeling habits that do not conform to common sense in engineering but do not violate explicit rules, such as mismatch between components and their spatial functions, missing support relationships, and abnormal topology of electromechanical system connections; Second, static rules lack the ability to learn and evolve, and cannot mine potential and complex error patterns from massive historical data, resulting in limited intelligence in detection.

[0006] Therefore, the industry urgently needs an intelligent IFC document specification detection solution that can effectively identify hidden errors. Summary of the Invention

[0007] This invention provides a standard detection method, system, device, and medium for IFC files, to solve the technical problems of existing IFC file detection methods being unable to effectively identify hidden errors and having insufficient detection intelligence.

[0008] This invention is achieved through the following technical solution:

[0009] A first aspect of the present invention provides a method for detecting the specifications of an IFC document, comprising:

[0010] The IFC file is parsed, and a graph structure is constructed based on the component information obtained from the parsing. Anomaly annotations are performed on each node in the graph structure to form a training sample set for the graph neural network. The graph structure uses components as nodes and spatial topological or functional relationships between components as edges.

[0011] A canonical detection model based on graph neural networks is constructed. The canonical detection model includes a relational graph convolutional network (RGCN) layer, which is used to perform information aggregation and node representation learning based on the input node features, edge relationships, and edge relationship types.

[0012] The standardized detection model is trained using the training sample set to optimize the model parameters; wherein the objective function of the model training includes at least a supervised classification loss based on node anomaly labeling and a self-supervised reconstruction loss based on graph structure reconstruction.

[0013] Output the trained standard detection model and apply it to the standard detection of the IFC file to be detected.

[0014] Furthermore, the construction of the graph structure based on the component information obtained through parsing includes:

[0015] Extract the attribute information of each component as the node features of each node in the graph structure;

[0016] Based on the geometric positional relationships between components or the connection relationships defined in the IFC file, determine the edges between each node and assign an edge relationship type and relationship strength to each edge.

[0017] Furthermore, the node features include at least one of the following attribute information: component type, material, spatial identifier, system identifier, geometric dimensions, geometric center coordinates, bounding box parameters, component area, and component volume;

[0018] The edge relationship types include at least one of the following: contact, adjacency, shared space, support, connection, and shared system;

[0019] The relationship strength is a quantified value calculated based on the geometric metric or logical correlation degree corresponding to the relationship type.

[0020] Furthermore, before training the canonical detection model using the training sample set, the method further includes:

[0021] The categorical features in the node features are embedded and encoded, and the numerical features are standardized.

[0022] The processed node features are concatenated with the edge relationship type and relationship strength, and used as the input features of the standardized detection model.

[0023] Furthermore, the step of annotating each node in the graph structure includes:

[0024] Obtain manually confirmed abnormal components as the first annotation sample;

[0025] The system automatically modifies known normal graph structures to generate new abnormal components, which are then used as second annotation samples.

[0026] The automatic modification includes at least one of the following operations: randomly altering the spatial identifier of the component, randomly removing edges representing support or connection relationships, and applying random micro-displacements to the geometric center of the component.

[0027] Furthermore, the trained canonical detection model is applied to the canonical detection of the IFC files to be detected, including:

[0028] The IFC file to be detected is parsed and a graph structure is constructed. The constructed graph structure is input into the trained standard detection model to obtain the anomaly detection results of the components in the IFC file to be detected. The anomaly detection results include the probability value of each component being an anomaly.

[0029] Components whose probability values ​​exceed a preset threshold are identified as abnormal components;

[0030] For components identified as anomalous, a model-based interpretability method is used to extract neighboring nodes, node features, and / or edge relationships that cause the anomalousness, serving as interpretive evidence.

[0031] Furthermore, it also includes:

[0032] The IFC file to be detected is subjected to rule detection based on a predefined rule base to obtain explicit error detection results;

[0033] The explicit error detection results are fused with the anomaly detection results output by the standard detection model to generate a comprehensive error report.

[0034] A second aspect of the present invention provides a specification detection system for IFC documents, comprising:

[0035] The data preprocessing unit is used to parse the IFC file, construct a graph structure based on the component information obtained from the parsing, and anomaly labeling is performed on each node in the graph structure to form a training sample set for the graph neural network; wherein, the graph structure uses components as nodes and spatial topological or functional relationships between components as edges;

[0036] The model building and training unit is used to build a canonical detection model based on a graph neural network, and to train the canonical detection model using the training sample set to optimize the model parameters. The canonical detection model includes a relational graph convolutional network (RGCN) layer, which is used to perform information aggregation and node representation learning based on the input node features, edge relationships, and edge relationship types. The objective function of the model training includes at least a supervised classification loss based on node anomaly labeling and a self-supervised reconstruction loss based on graph structure reconstruction.

[0037] The model output unit is used to output the trained canonical detection model, which is then applied to the canonical detection of the IFC file to be detected.

[0038] A third aspect of the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the specification detection method for IFC files as described in any one of the first aspects of the present invention.

[0039] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the specification detection method for IFC files as described in any one of the first aspects of the present invention.

[0040] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0041] The method of this invention abstracts the Building Information Model (BIM) into a graph structure containing rich attributes and relationships, and utilizes Relational Graph Convolutional Networks (RGCNs), a graph neural network specifically designed to handle heterogeneous relationships, for learning. This enables the model to not only perceive the individual attributes of components but also to deeply understand the contextual relationships of components within the overall architectural space and system. By combining supervised learning for anomalous nodes with self-supervised learning based on graph reconstruction, the model can learn deep patterns of normal architectural models from labeled samples and a large number of unlabeled normal graph structures, thus becoming highly sensitive to implicit anomalies that violate these patterns. Finally, the trained model can perform efficient and intelligent specification detection on new IFC files, effectively compensating for the blind spots of traditional rule-based detection. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0043] Figure 1 This is a flowchart of a specification detection method for IFC files according to an embodiment of the present invention;

[0044] Figure 2 This is a model structure diagram of a standard detection model according to an embodiment of the present invention;

[0045] Figure 3 This is a schematic diagram of a specification detection method for IFC files according to an embodiment of the present invention;

[0046] Figure 4 This is a structural block diagram of an IFC document specification detection device according to an embodiment of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0048] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims, and accompanying drawings of this invention are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to other steps or units inherent in the device.

[0049] The terminology used in the various embodiments of the invention is for the purpose of describing particular embodiments only and is not intended to limit the various embodiments of the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of the invention pertain. The terms (such as those defined in commonly used dictionaries) are to be interpreted as having the same meaning as in the context of the relevant technical field and are not to be interpreted as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of the invention.

[0050] This invention aims to address the shortcomings of existing IFC document specification detection methods, such as weak ability to identify hidden errors and low level of intelligence. It proposes a method for detecting hidden anomaly nodes based on graph neural networks (GNNs). This method is applicable to BIM model quality inspection and data governance at all stages of building design, construction, and operation and maintenance. It helps improve the reliability of cross-platform data exchange, reduce engineering rework caused by model errors, and promote the intelligent application of BIM data.

[0051] Please see Figure 1 , Figure 1 The diagram shows a flowchart of a standard detection method for IFC files proposed in this invention, with details provided in steps S100 to S400.

[0052] S100: Parse the IFC file, construct a graph structure based on the component information obtained from the parsing, anomaly labeling is performed on each node in the graph structure, and a training sample set for the graph neural network is constructed.

[0053] The IFC file described in this invention is a STEP physical file based on the EXPRESS language, typically stored in text or binary format. It describes building components, their attributes, and relationships in object form, supporting information sharing throughout the entire BIM lifecycle, including design, construction, and operation and maintenance phases. Open-source parsing tools (such as IfcOpenShell) can be used to parse the IFC file. After parsing, all component entities (such as walls, doors, floors, windows, etc.) and their related data in the building model can be obtained.

[0054] In this step, the component information from the IFC file is parsed, and the entire Building Information Model is abstracted into a Property Heterogeneous Graph. This graph structure contains multiple types of nodes and edge relationships. Each node in the graph structure corresponds to a building component, and the edges between nodes are determined based on the spatial topological relationships (such as contact and adjacency) and functional relationships (such as related spaces and supports) between components. Each node is accompanied by a set of node features, consisting of the node's (component's) basic attributes, geometric parameters, and positional parameters.

[0055] The basic attributes of a component include, but are not limited to: component type, material name, ifcspace identifier, system identifier (such as air supply system, water supply and drainage system), floor; geometric parameters include, but are not limited to: length, width, height, area, volume; and location parameters include, but are not limited to: geometric center coordinates, bounding box parameters.

[0056] In a heterogeneous graph, edges represent relationships between related nodes (components). Each edge includes three key features: edge relationship, edge relationship type, and relationship strength. Among them, the edge relationship is an abstract representation of the existence of edges between nodes. In actual data, it is reflected as the edge index (edge_index), which records which pairs of nodes are connected. Together with the relationship type (edge_type) and relationship strength (edge_weight), it fully describes the structure and semantics of the graph.

[0057] Edge relationship types represent the spatial topological or functional relationships between components. Spatial topological relationships can be obtained based on geometric calculations or IFC semantic relationships, and include the following edge relationship types:

[0058] 1. Contact: There is geometric contact between the surfaces of the two components (determined by collision detection);

[0059] 2. Adjacent: Spatial distance less than a threshold (e.g., 0.5 meters) but not in contact;

[0060] 3. Same Space: Two components are associated with the same IfcSpace entity;

[0061] 3. Support: One component is located directly below another component and their vertical projections overlap;

[0062] 4. Connection: Determined based on the IfcRelConnects entity in the IFC;

[0063] 5. Same System: Two components belong to the same functional system;

[0064] 6. Belonging to the same floor: The two components belong to the same floor.

[0065] Edge weight is used to quantify the tightness of the relationship between components and is a value between [0,1]. For example, for a "contact" type relationship, the edge weight can be the normalized contact area, and for a "adjacent" type relationship, the edge weight can be inversely proportional to the normalized component distance.

[0066] The parsed component information and relationship data are standardized into a unified file format, including node information (nodes.csv), edge relationship information (edges.csv), and metadata (meta.json). For each component, attribute information is extracted and output as columns, including the following fields: component type (id, type), material, room_id, system, floor, geometric center (center_x, center_y, center_z), bounding box parameters (bbox_min_x, bbox_min_y, bbox_min_z, bbox_max_x, bbox_max_y, bbox_max_z), geometric dimensions (length, width, height, area, volume), text description, and label. Preliminary topological relationships for the components are generated and output as columns, including the following fields: source node ID (src), target node ID (dst), edge relationship type (rel_type), relationship strength (weight), and metadata (meta).

[0067] To construct a supervised learning task, it is necessary to annotate the nodes (components) in the graph. Obtain manually confirmed anomalous components (such as components with missing attributes or incorrect topological relationships) and directly label them as anomalous nodes, using them as the first annotation sample. Assign different labels to anomalous and normal components, such as 0 for normal and 1 for anomalous. Merge the labels into the label field of the node information file nodes.csv.

[0068] Furthermore, to expand the training set, especially to address the problem of scarce hidden anomalous samples, the known normal graph structure is automatically modified to generate new anomalous components, which serve as second labeled samples.

[0069] Among them, the modification strategy simulates common hidden errors, including the following handling:

[0070] (1) Randomly tamper with the room ID of a component to make it inconsistent with the room ID of most surrounding components, for example, by swapping the room IDs of two components;

[0071] (2) Randomly remove edges representing support or connection relationships to simulate design oversights;

[0072] (3) Apply a small geometric displacement to randomly selected components to disrupt their original alignment or contact relationship.

[0073] By automatically modifying and expanding the implicit anomaly samples, the model's ability to identify various implicit errors is improved. The first and second labeled samples are used as negative samples, and normal nodes are used as positive samples. Finally, each node obtains a label, and the multiple graph structures and their node labels obtained from processing multiple IFC files together constitute the training sample set.

[0074] S200, construct a canonical detection model based on graph neural networks. This canonical detection model uses the relational graph convolutional network RGCN layer to perform information aggregation and node representation learning on the input features (node ​​features, edge relations and edge relation types).

[0075] like Figure 2 As shown, the canonical detection model constructed in this invention uses a relational graph convolutional network (RGCN) as its core framework. RGCN can explicitly handle different types of edges and learn independent weight matrices for each relation type to propagate information, making it very suitable for the heterogeneous graph structure constructed in step S100.

[0076] The model structure includes an input layer, a feature encoding layer, an RGCN layer, and an output layer, with the data flow as follows.

[0077] Input layer: Receives graph structure data obtained from step S100, including: node features, edge relationship edge_index, and edge relationship type edge_type.

[0078] The input features include a feature matrix of all nodes in a graph structure. The node features X are extracted from the attribute information of each component obtained by parsing (including the basic attributes, geometric parameters, and positional parameters mentioned above). These features can be a combination or all parameters among component type, material name, spatial identifier, system identifier, geometric dimensions, geometric center coordinates, bounding box parameters, area, and volume.

[0079] Feature encoding layer: Categorical features (such as component type, material, etc.) in the node features are embedded and encoded, transforming them into dense vectors. Numerical features (such as geometric dimensions, coordinates) are standardized (e.g., Z-score or min-max standardization for individual buildings). Geometric center coordinates (center_x, center_y, center_z) can be processed by relative room centering, combined with sin / cos frequency encoding or learnable position embedding (learnablepos-emb) to improve the model's spatial perception ability. Then, all input feature vectors are concatenated to form the initial feature representation of each node.

[0080] Specifically, the standardized node information (nodes.csv) and edge information (edges.csv) are converted into a DGL graph data structure for training. The edge relationship type (rel_type) is encoded as an integer, for example, contact = 0, adjacent = 1, same_room = 2, support = 3, connection = 4, same_system = 5, etc.

[0081] RGCN Layers: A multi-layered RGCN stacked structure is adopted. Each layer of RGCN aggregates information about each node's various relational neighbors based on the edge's connection index (edge_index) and edge type (edge_type). For example, the representation of a wall will obtain information from its neighbors with different relationships, such as the floor slabs it contacts, furniture in the same space, and supporting beams, and update its own representation accordingly. After propagation through multiple layers of RGCN, each node obtains a final representation containing rich local and global contextual information. .

[0082] Optionally, the input data also includes the relation strength edge_weight, which is calculated based on the geometric metric or logical correlation degree corresponding to the edge relation type. For example, the edge_weight of a contact relation is calculated by the contact area, and the weight of an adjacent relation is calculated by the reciprocal of the distance. Finally, the edge_weight is normalized to the interval [0,1].

[0083] During model training, the relation strength edge_weight is mapped to a vector through the MLP layer and used as the attention weight or edge feature during aggregation in the RGCN layer.

[0084] The output layer includes two branches:

[0085] (1) Anomaly detection head: Represents nodes The probability that a node is an anomaly is output through a fully connected layer and a sigmoid function. .

[0086] The corresponding supervised classification loss for this process By employing binary cross-entropy loss, Focal Loss or class weights can be used to address the problem of imbalanced samples.

[0087] (2) Self-supervised reconstruction head: In order to enhance the model’s learning of normal graph structure patterns, a graph autoencoder (GAE) structure is introduced, using RGCN as the encoder and another decoder (such as inner product based on node representation) to attempt to reconstruct the original edges, especially important relationships such as contact and connection relationships.

[0088] Reconstruction error As a self-supervised reconstruction loss term, the total loss during model training can be expressed as: ,in, This represents the ranking loss based on anomaly scores, used to improve the model's ability to distinguish anomaly nodes. To balance hyperparameters, This indicates the loss from the supervision and classification process.

[0089] While supervising classification, self-supervised constraints such as neighborhood reconstruction and contrastive learning are added to uncover hidden anomalies such as spatial topological inconsistencies or semantic biases.

[0090] S300 inputs the training sample set into the standardized detection model for training to optimize the model parameters.

[0091] The training sample set is divided into a training set (train), a test set (test), and a validation set (dev) according to a preset ratio, generating independent `graph_train.py` files and recording the sampling seeds to ensure the reproducibility of the experiments. The model is trained using the divided training set, and the objective function of the model training must include at least a supervised classification loss based on node anomaly annotations. And self-supervised reconstruction loss based on graph structure reconstruction .

[0092] A stochastic gradient descent optimizer is used to minimize the total loss. With the goal of iteratively updating the trainable model parameters such as the RGCN layer, feature encoding layer, and output layer.

[0093] During training, monitor the evaluation values ​​on an independent validation set, adjust the learning rate, and use an early stopping strategy to prevent overfitting. Save the model parameters that perform best on the validation set.

[0094] S400 outputs the trained canonical detection model, which is then used for canonical detection of the IFC file to be detected.

[0095] After training, a set of optimized model parameters (including RGCN layer weights, embedding layer parameters, loss function correlation coefficients, feature preprocessing normalization parameters, etc.) are obtained. These parameters are then packaged into a standardized model service, and the preprocessing scripts required for model inference are configured to obtain a standardized detection model that can be directly used in engineering practice.

[0096] Obtain the IFC file to be detected, and perform file parsing, graph structure construction, and input feature extraction using the same preprocessing methods as in step S100. Input the input features into the deployed standardized detection model to obtain the detection results output by the model.

[0097] One detection method is described in steps S401 to S403.

[0098] S401, Graph Construction of the File Under Test: For the IFC file under test, perform the same parsing and graph structure construction steps as in S100 to obtain its corresponding graph structure G_new.

[0099] S402, Model Inference: Input the node features, edge relationships, and edge relationship types of the graph structure G_new into the trained model. After forward propagation, the model outputs the anomaly probability of each node in the graph. .

[0100] S403, Anomaly Detection and Interpretation: Set a probability threshold (e.g., 0.85) and... Nodes exceeding this threshold are considered abnormal components.

[0101] Based on the threshold that maximizes the F1 score (a comprehensive evaluation metric for model classification performance) on the validation set or the recall threshold set by business requirements (e.g., recall ≥90%), components with probability values ​​exceeding the threshold are identified as anomalous components, triggering two levels of alarms: general anomalous (warn) and severe anomalous (alert). For example, a warn alarm is triggered when the anomalous score is in the range [threshold, 0.8); an alert alarm is triggered when the anomalous score is ≥0.8.

[0102] Furthermore, for each anomalous component, an interpretability method (such as GNNExplainer) is used to analyze and extract the neighboring nodes, node features, and / or edge relationships that caused its anomalousness as interpretive evidence.

[0103] The GNNExplainer tool extracts explanatory evidence for anomalous components, including the subgraph (key neighbor nodes and edges) that contribute most to the anomaly prediction of the node, as well as the node feature dimensions, and generates explanatory evidence in natural language or visualization. For example: "The 'toilet' was judged as anomalous mainly because more than 90% of its spatial neighbors belong to 'public corridors' rather than 'toilets'."

[0104] Finally, a detailed detection report is generated by summarizing the IDs, probabilities, inferred anomaly types (based on interpretive analysis and classification) of all anomalous components.

[0105] In a further embodiment, such as Figure 3 As shown, the specification check of IFC documents also includes rule engine checks, specifically:

[0106] The IFC file to be detected is subjected to rule detection based on a predefined rule base to obtain explicit error detection results;

[0107] The explicit error detection results are fused with the anomaly detection results output by the standard detection model to generate a comprehensive error report.

[0108] A detection rule base is built based on the IFC standard schema definition. The rule base covers syntax validation rules (such as data type compliance and field format correctness), semantic validation rules (such as component type and attribute matching), and attribute constraint rules (such as required fields not being empty and numerical range compliance). It performs rapid detection on IFC files and identifies explicit errors. The two detection results are merged, deduplicated, and integrated into a final comprehensive error report output.

[0109] Furthermore, a confidence-weighted fusion algorithm is used to fuse the detection results from the rule engine and the graph neural network. Since rule detection targets explicit errors, a higher confidence level, such as 0.9, can be set. The fusion formula is as follows:

[0110]

[0111] In the formula, Indicates the fusion confidence level. These are the weighting coefficients for the rule-based detection results. To validate the confidence level of the detection results according to the rules, This represents the confidence level of the graph neural network detection results (i.e., the anomaly probability value output by the model).

[0112] The fusion mechanism achieves comprehensive coverage of explicit errors and implicit anomalies, and error handling priorities are determined by confidence ranking.

[0113] In a further embodiment, an adaptive repair step is also included.

[0114] For explicit errors detected by the rules engine (such as missing attributes or version incompatibility), standard repair strategies from the repair strategy library are invoked to automatically fix them. For example, if a component lacks material attributes, the default material is automatically filled in; if semantic incompatibility occurs when importing an IFC2x3 file into the IFC4 platform, version mapping rules are automatically invoked to convert the relevant fields to the IFC4 standard format.

[0115] For implicit anomalies detected by the graph neural network (such as unreasonable topological relationships or semantic deviations), the model generates targeted repair suggestions (such as adjusting the position of components or supplementing missing support edges). After the user selects and confirms the suggestions through an interactive verification mechanism, the repair operation is executed.

[0116] In addition, user-confirmed "error-fix" samples are accumulated into the fix strategy library, and the fix rules are continuously optimized through model learning to achieve dynamic evolution of the fix strategy library.

[0117] Based on the same inventive concept, a specification detection system for IFC documents is also proposed, such as... Figure 4 As shown, it includes:

[0118] The data preprocessing unit is used to parse the IFC file, construct a graph structure based on the component information obtained from the parsing, and anomaly labeling is performed on each node in the graph structure to form a training sample set for the graph neural network; wherein, the graph structure uses components as nodes and spatial topological or functional relationships between components as edges.

[0119] The model building and training unit is used to build a canonical detection model based on a graph neural network. The canonical detection model is trained using a training sample set to optimize the model parameters. The canonical detection model includes a relational graph convolutional network (RGCN) layer, which is used to aggregate information and learn node representations based on the input node features, edge relationships, and edge relationship types. The objective function of the model training includes at least a supervised classification loss based on node anomaly labeling and a self-supervised reconstruction loss based on graph structure reconstruction.

[0120] The model output unit is used to output the trained canonical detection model, which is then applied to the canonical detection of the IFC file to be detected.

[0121] Furthermore, it also includes a rule engine unit, which is used to perform rule detection on IFC files based on a predefined rule base, identify explicit errors, and output the explicit error detection results and corresponding confidence levels.

[0122] Furthermore, it also includes a hybrid detection module, which is used to fuse the explicit error detection results output by the rule engine unit with the anomaly detection results output by the model building and training unit through a confidence-weighted fusion algorithm to generate a comprehensive error report and trigger an alarm of the corresponding level.

[0123] Furthermore, it also includes an adaptive repair unit, which is used to perform automatic repair on deterministic errors based on the detection results of the hybrid detection unit, generate repair suggestions for uncertain errors and support interactive repair, and at the same time realize dynamic iterative optimization of the repair strategy library.

[0124] Embodiments of the present invention also propose an electronic device comprising a processor and a memory, wherein the number of processors may be one or more. The memory, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules. The processor executes various functional applications and data processing of the electronic device by running the software programs, instructions, and modules stored in the memory, thereby implementing the specification detection method for IFC files according to any of the above embodiments of the present invention.

[0125] The memory may primarily comprise a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a given function; the data storage area may store data created based on terminal usage. Furthermore, the memory may include high-speed random access memory (RAM) and non-volatile memory, such as at least one disk storage device, flash memory, or other non-volatile solid-state storage device. In some instances, the memory may further include memory remotely located relative to the processor, which can be connected to the electronic device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks (LANs), mobile communication networks, and combinations thereof.

[0126] Embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the specification detection method for IFC files according to any embodiment of the present invention.

[0127] The computer storage medium of this invention can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0128] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0129] Embodiments of the present invention also provide a computer program product that, when run on a computer, causes the computer to execute the specification detection method for IFC files according to any of the above embodiments of the present invention.

[0130] The above embodiments are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the appended claims.

Claims

1. A method for detecting the specifications of an IFC document, characterized in that, include: The IFC file is parsed, and a graph structure is constructed based on the component information obtained from the parsing. Anomaly annotations are performed on each node in the graph structure to form a training sample set for the graph neural network. The graph structure uses components as nodes and spatial topological or functional relationships between components as edges. A canonical detection model based on graph neural networks is constructed. The canonical detection model includes a relational graph convolutional network (RGCN) layer, which is used to perform information aggregation and node representation learning based on the input node features, edge relationships, and edge relationship types. The standardized detection model is trained using the training sample set to optimize the model parameters; wherein the objective function of the model training includes at least a supervised classification loss based on node anomaly labeling and a self-supervised reconstruction loss based on graph structure reconstruction. Output the trained standard detection model and apply it to the standard detection of the IFC file to be detected.

2. The method for detecting the standardization of IFC documents according to claim 1, characterized in that, The graph structure constructed based on the component information obtained through parsing includes: Extract the attribute information of each component as the node features of each node in the graph structure; Based on the geometric positional relationships between components or the connection relationships defined in the IFC file, determine the edges between each node and assign an edge relationship type and relationship strength to each edge.

3. The method for detecting the standardization of IFC documents according to claim 2, characterized in that, The node features include at least one of the following attribute information: component type, material name, space identifier, system identifier, floor, geometric dimensions, geometric center coordinates, bounding box parameters, area, and volume; The edge relationship types include at least one of the following: contact, adjacency, shared space, support, connection, and shared system; The relationship strength is a quantified value calculated based on the geometric metric or logical correlation degree corresponding to the relationship type.

4. The method for detecting the standardization of IFC documents according to claim 2, characterized in that, Before training the canonical detection model using the training sample set, the method further includes: The categorical features in the node features are embedded and encoded, and the numerical features are standardized. The processed node features are concatenated with the edge relationship type and relationship strength, and used as the input features of the standardized detection model.

5. The method for detecting the standardization of IFC documents according to claim 1, characterized in that, The anomaly annotation of each node in the graph structure includes: Obtain manually confirmed abnormal components as the first annotation sample; The system automatically modifies known normal graph structures to generate new abnormal components, which are then used as second annotation samples. The automatic modification includes at least one of the following operations: randomly altering the spatial identifier of the component, randomly removing edges representing support or connection relationships, and applying random micro-displacements to the geometric center of the component.

6. The method for detecting the standardization of IFC documents according to claim 1, characterized in that, The trained canonical detection model is applied to the canonical detection of the IFC files to be detected, including: The IFC file to be detected is parsed and a graph structure is constructed. The constructed graph structure is input into the trained standard detection model to obtain the anomaly detection results of the components in the IFC file to be detected. The anomaly detection results include the probability value of each component being an anomaly. Components whose probability values ​​exceed a preset threshold are identified as abnormal components; For components identified as anomalous, a model-based interpretability method is used to extract neighboring nodes, node features, and / or edge relationships that cause the anomalousness, serving as interpretive evidence.

7. The method for detecting the standardization of IFC documents according to claim 6, characterized in that, Also includes: The IFC file to be detected is subjected to rule detection based on a predefined rule base to obtain explicit error detection results; The explicit error detection results are fused with the anomaly detection results output by the standard detection model to generate a comprehensive error report.

8. A specification detection system for IFC documents, characterized in that, include: The data preprocessing unit is used to parse the IFC file, construct a graph structure based on the component information obtained from the parsing, and anomaly labeling is performed on each node in the graph structure to form a training sample set for the graph neural network; wherein, the graph structure uses components as nodes and spatial topological or functional relationships between components as edges; The model building and training unit is used to build a canonical detection model based on a graph neural network, and to train the canonical detection model using the training sample set to optimize the model parameters. The canonical detection model includes a relational graph convolutional network (RGCN) layer, which is used to perform information aggregation and node representation learning based on the input node features, edge relationships, and edge relationship types. The objective function of the model training includes at least a supervised classification loss based on node anomaly labeling and a self-supervised reconstruction loss based on graph structure reconstruction. The model output unit is used to output the trained canonical detection model, which is then applied to the canonical detection of the IFC file to be detected.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the specification detection method for IFC files as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the specification detection method for IFC files as described in any one of claims 1-7.

Images