Green building model compliance verification method and system

Abstract

The invention discloses a green building model compliance verification method and system, and the method comprises the steps: constructing a domain knowledge ontology and a sensor ontology, and forming a unified green building ontology; performing entity recognition and relation extraction on the unstructured data of the green building standard file to obtain an entity-relation-attribute triple structure, and constructing a visual green building knowledge graph; converting the BIM model semi-structured data and the sensor monitoring structured data into a unified RDF format, and storing the RDF format data into a knowledge base under a green building body; and defining an SHACL rule based on the green building knowledge graph, and performing green building compliance check on the RDF format data in the known library based on the SHACL rule. According to the method, the NLP technology is utilized to extract the triple to construct the green building knowledge graph, the text of the unstructured green building standard is automatically converted into the SHACL constraint, and compliance verification of the green building model is automatically realized.

Description

Technical Field

[0001] This invention relates to the field of building compliance inspection technology, and in particular to a method and system for verifying the compliance of green building models. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Green buildings, with their advantages of low energy consumption, low pollution, and high resource utilization throughout their design, construction, and operation lifecycles, have become the core carrier of sustainable development in the construction industry. Green building compliance verification is a crucial step in ensuring that a building project meets the target green building standards or specifications throughout its design, construction, and operation processes.

[0004] Traditional green building compliance verification methods rely heavily on manual review. Auditors need to manually compare massive amounts of data, such as building design drawings, material testing reports, and energy consumption monitoring data, with the requirements of the target green building standards or specifications. This process is inefficient and cannot meet the verification needs of large-scale green building projects. Furthermore, the professional level, experience, and subjective judgment of auditors can influence the process. Different auditors may have different understandings of the standard clauses, leading to different verification results for the same project in different review scenarios, making it difficult to ensure the accuracy of the verification results.

[0005] Existing technologies also disclose some automated green building assessment tools based on BIM (Building Information Modeling), such as the LEED (Leadership in Energy and Environmental Design) assessor plugin, which automates the application of LEED scoring rules. These tools achieve automatic comparison between design data and assessment requirements by embedding evaluation criteria into BIM software. However, green building standards are usually in unstructured text form, containing a large number of technical terms, complex sentences, and constraints. How to automatically extract key information from this text and transform it into structured data, and how to automatically verify building data according to standard requirements, are the core challenges in achieving automated compliance verification. Most existing solutions use Natural Language Processing (NLP) technology for semantic relationship extraction, but there are still many shortcomings; for example: The automation level of green building standard analysis is insufficient. Faced with complex green building standard clauses, it is difficult to accurately extract all constraints, especially the ability to analyze non-numerical constraints and ambiguous clauses. Semantic consistency of multi-source data integration is difficult to guarantee. There is a lack of a unified method for semantic mapping between BIM data, sensor data, and standard knowledge, which easily leads to semantic conflicts during data integration and affects the accuracy of verification results. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a method and system for verifying the compliance of green building models. It utilizes NLP technology to extract triples and construct a green building knowledge graph, automatically converting the text of unstructured green building standards into SHACL (Shape Constraint Language) constraints, thereby automatically verifying the compliance of green building models.

[0007] In some implementations, the following technical solutions are adopted: A method for verifying the compliance of a green building model includes: Based on green building standard documents, a domain knowledge ontology and a sensor ontology are constructed respectively; the domain knowledge ontology and sensor ontology are deeply integrated through semantic mapping rules to form a unified green building ontology; For the unstructured data of green building standard documents, entity recognition and relation extraction are performed to obtain entity-relation-attribute triple structure. Using the green building ontology as a semantic framework, a visualized green building knowledge graph is constructed. The semi-structured data of the BIM model and the structured data of the sensor monitoring are converted into a unified RDF format, and the RDF format data is stored in the knowledge base under the green building ontology; Based on the green building knowledge graph, SHACL rules are defined, including numerical constraints, logical constraints, and instance semantic mapping rules; based on the SHACL rules, green building compliance checks are performed on RDF format data in the knowledge base.

[0008] As a further solution, the domain knowledge ontology is constructed in a top-down manner. The top-level classes include five categories: building type, evaluation criteria, building components, building structure, and evaluation role. Each top-level class has several subclasses, forming a hierarchical structure.

[0009] As a further solution, the sensor ontology is based on the SOSA framework, defining sensor, observation object, observable attribute and measurement result class. Semantic knowledge is injected into the sensor monitoring process through semantic modeling of real-time monitoring data, and the observation behavior and measurement data are associated through the hasResult attribute.

[0010] As a further solution, deep fusion of domain knowledge ontology and sensor ontology is achieved through semantic mapping rules, specifically including: The rdfs:subClassOf rule is used to classify sensor devices into the building facilities subclass. The owl:equivalentClass rule is used to clarify the equivalence relationship between environmental risk factors and sensor monitoring objects. The core attributes of the two types of ontology are associated through the owl:equivalentProp and rdfs:subPropertyOf rules to ensure the semantic consistency between building design data and sensor monitoring data.

[0011] As a further solution, entity identification is performed on the unstructured data of green building standard documents, specifically as follows: The LERT-CNN-BiLSTM-MHA-CRF model is used to extract target entities from green building standard documents. The green building standard documents are input into the LERT pre-trained model to generate deep semantic word vectors, which capture the contextual information of professional terms. The deep semantic word vectors are input into the CNN model to extract local phrase features and enhance the recognition of domain terms. The context dependency is modeled bidirectionally by BiLSTM and combined with a multi-head attention mechanism to capture long-distance semantic associations. Finally, the entity label sequence is optimized by CRF.

[0012] As a further step, relation extraction is performed on the unstructured data of green building standard documents, specifically as follows: LERT semantic encoding is used to generate text feature vectors, and TF-IDF and PMI are used to calculate edge weights to construct a text graph. A graph convolutional network (GCN) is used to aggregate the features of neighbor nodes, strengthen the semantic association between entities and perform relationship prediction. Softmax classification and subject-object probability calculation are used to predict potential subject-object pairs and preliminary relationship types. A global correspondence matrix is ​​introduced to filter out correct subject-object relationship pairs and invalid relationship pairs to ensure accurate alignment of entities and relationships.

[0013] As a further solution, the instance semantic mapping rule realizes the association between entities, attributes and BIM model elements in the knowledge graph by constructing a matching dictionary between Chinese terms and ifcOWL ontology; Load the defined SHACL rules, traverse each building instance in the BIM model, match the corresponding constraints, determine whether the rules are met, and output a list of compliance and non-compliance. The list clearly indicates the inspection object, the corresponding standard rule, whether the attribute value is compliant, and the reason for non-compliance.

[0014] In other embodiments, the following technical solutions are adopted: A green building model compliance verification system, comprising: The ontology construction module is configured to construct a domain knowledge ontology and a sensor ontology based on green building standard documents; and to achieve deep integration of the domain knowledge ontology and the sensor ontology through semantic mapping rules to form a unified green building ontology. The knowledge graph construction module is configured to perform entity recognition and relation extraction on the unstructured data of green building standard documents to obtain entity-relation-attribute triple structure, and construct a visualized green building knowledge graph using the green building ontology as a semantic framework. The data conversion module is configured to convert semi-structured data from BIM models and structured data from sensor monitoring into a unified RDF format, and store the RDF format data in the knowledge base under the green building entity. The compliance check module is configured to define SHACL rules based on the green building knowledge graph, including numerical constraints, logical constraints, and instance semantic mapping rules; and to perform green building compliance checks on RDF format data in the knowledge base based on the SHACL rules.

[0015] In other embodiments, the following technical solutions are adopted: A terminal device includes a processor and a memory, the processor being used to implement instructions; the memory being used to store multiple instructions adapted to be loaded and executed by the processor to perform the green building model compliance verification method described above.

[0016] In other embodiments, the following technical solutions are adopted: A computer-readable storage medium storing a plurality of instructions adapted to be loaded by a processor of a terminal device and executed by the above-described green building model compliance verification method.

[0017] Compared with the prior art, the beneficial effects of the present invention are: (1) The conceptual semantic granularity of design specifications and sensor outputs is different. Concepts in design specifications, such as walls, waterproofing ratings, and energy consumption limits, are completely different from those in sensor outputs, such as NH3 concentration, temperature, and observation timestamps, in terms of abstraction level, temporal and spatial granularity, and unit system. If they are forcibly placed in the same ontology, class explosion and attribute contamination will occur. This invention constructs a domain knowledge ontology and a sensor ontology separately. Each ontology only focuses on terms of the same granularity, resulting in a clear model and simple maintenance.

[0018] (2) This invention directly treats the unstructured clauses in the green building standard documents as free text to be processed, and uses natural language processing (NLP) technology to extract the entity relationships in the text so that the machine can read and execute them. The original standard clauses that required domain experts to read, disassemble and enter one by one are transformed into one-click generation of SHACL rules with zero human intervention. The degree of automation is improved from the traditional manual or semi-manual mode to end-to-end machine parsing, solving the core problems of low standard parsing efficiency, easy error and difficulty in expansion.

[0019] (3) This invention uses the LERT-CNN-BiLSTM-MHA-CRF model to extract entity information, integrating five capabilities at once: domain vocabulary prior, local phrase template, long-range semantic dependency, cross-entity attention weight, and global legal path. This allows for the accurate extraction of complex entities from green building standard documents with dense terminology, stringent numerical values, and lengthy sentence structures using very few labeled samples. Among them, LERT continues to use Chinese parts of speech and dependency relations for mask training on the basis of general BERT, which is equivalent to having the model memorize professional terms such as waterproof layer and fire resistance limit in advance, significantly reducing the need for domain labeling. CNN accurately captures fixed collocations of qualifiers such as numbers and units, avoiding the Transformer from only looking at the dot product and separating numerical values ​​and quantifiers. Bidirectional LSTM is responsible for reading long sentences, while unidirectional models can only read half from left to right or from right to left, which is prone to long-range pairing loss. MHA is responsible for weighting and emphasizing keywords, giving higher weights to words that truly determine their attributes. CRF is responsible for making the boundaries legal, ensuring that the entity boundaries are completely aligned with the standard text.

[0020] Other features and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the green building structure in an embodiment of the present invention; Figure 2 This is a schematic diagram of the named entity recognition model structure in an embodiment of the present invention; Figure 3 This is a schematic diagram comparing the performance of named entity recognition with different parameters in an embodiment of the present invention; Figure 4 This is a schematic diagram of the GCN-PGRC model structure in an embodiment of the present invention; Figure 5 This is a flowchart of the algorithm for constructing a green building knowledge graph in an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating the green building knowledge graph in an embodiment of the present invention; Figure 7This is how the MaterialLevel attribute of the IfcSlab instance is represented in the ifcOWL example diagram in this embodiment of the invention; Figure 8 This is a residential building BIM model in an embodiment of the present invention. Detailed Implementation

[0022] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0023] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0024] Example 1 In one or more embodiments, a method for verifying the compliance of a green building model is disclosed, specifically including the following process: S101: Based on green building standard documents, construct domain knowledge ontology and sensor ontology respectively; achieve deep integration of domain knowledge ontology and sensor ontology through semantic mapping rules to form a unified green building ontology; In this embodiment, in order to unify the semantic expression in the field of green building and eliminate different data sources, a semantic relationship is established between building standard text, BIM model and sensor data by constructing a green building ontology. This allows the computer to understand what green building compliance requirements are, which building entities need to meet the requirements, and how to verify the requirements through data. This provides an abstract layer for the knowledge graph, that is, the structural definition of knowledge, to ensure that the entities and relationships extracted later can be organized and associated in an orderly manner.

[0025] It should be noted that sensor data refers to quantitative data related to building environmental performance collected in real time by various IoT devices deployed in green buildings. This data mainly includes environmental parameter data and resource and equipment operation data. Environmental parameter data includes indoor and outdoor temperature, humidity, PM2.5 concentration, concentration of harmful gases such as formaldehyde, light intensity, and noise decibel value. Resource and equipment operation data includes electricity and water consumption, ventilation system operation parameters, solar and wind energy utilization efficiency, and building component status. The data is used to determine whether the building meets the requirements of green building standards.

[0026] As a specific implementation method, the specific implementation process is as follows: S1011: Define the construction goals and scope; This embodiment focuses on the core requirements of green building compliance verification, taking the "China Green Building Evaluation Standard" (GB / T50378-2019) and IFC standards as the core basis, covering key dimensions such as building components, evaluation indicators, and sensor monitoring, avoiding redundancy caused by an overly broad scope, and ensuring that the ontology can support semantic alignment and interoperability of multi-source data.

[0027] At the same time, existing standard resources are reused, with core reuse of building entity definitions from the IFC (Industrial Foundations) standard, such as walls IfcWall and floors IfcSlab, to avoid redundant modeling. Meanwhile, the SOSA (Sensor, Observation, Sample, Actuator) ontology framework is referenced to ensure semantic compatibility of sensor data.

[0028] S1012: Construct the domain knowledge ontology (OntoDK) and the sensor ontology (OntoSen) respectively. In this embodiment, combined with Figure 1 The construction of the Domain Knowledge Sub-Ontology (OntoDK) first clarifies the domain scope of the ontology, focusing on the core concepts and relationships involved in green building to ensure its applicability and completeness. The Domain Knowledge Ontology (OntoDK) adopts a top-down approach. The top-level classes include five major categories: building type, evaluation criteria, building components, building structure, and evaluation role. Each top-level class has subclasses to form a hierarchical structure. At the same time, object attributes are designed, such as hasMaterial to associate building components with materials, and data attributes, such as numerical features like area and height. The core concepts of the IFC ontology are reused for extension to achieve deep integration of building design data and green building performance indicators. Table 1 gives an example of the OntoDK ontology conceptual framework.

[0029] Table 1. OntoDK Ontology Conceptual Framework

[0030] The OntoSen ontology is based on the SOSA framework and defines classes such as sensors, observed objects, observable attributes, and measurement results. It supports semantic modeling of real-time monitoring data such as temperature, humidity, and air quality. It associates observation behavior with measurement data through attributes such as hasResult. Table 2 gives an example of the OntoSen ontology conceptual framework.

[0031] Table 2 OntoSen ontology conceptual framework

[0032] In this embodiment, the core innovation of the Domain Knowledge Ontology (OntoDK) lies in its reuse and extension to enhance its interoperability and semantic expression capabilities within the BIM environment. OntoDK reuses core building entity classes from the IFC ontology, such as IfcWall and IfcSlab, and extends them with evaluation classes specific to green buildings, such as indoor air quality evaluation items and energy efficiency evaluation items. Through semantic mapping, it associates IFC building entities with green building standard requirements, such as binding IfcWall to wall insulation performance requirements.

[0033] The core of the sensor sub-ontology (OntoSen) is to address the semantic gaps in sensor data. To solve this problem, semantic modeling injects semantic knowledge into the sensor monitoring process, thereby enabling effective interpretation of sensor data. The Sensor, Observation, Sample, and Actuator (SOSA) lightweight ontology is a standard framework for describing sensors, IoT devices, and their observation or actuation processes, allowing raw monitoring data such as 25℃ and 0.05mg / m³ to be mapped to specific green building evaluation indicators.

[0034] In this embodiment, the domain knowledge ontology (OntoDK) reuses the IFC standard and is geared towards static BIM data in the design phase; the sensor sub-ontology (OntoSen) reuses SOSA and is geared towards dynamic monitoring data in the operation phase. Both are first modeled according to their respective standards and then bridged through mapping rules, which not only ensures semantic consistency between design and monitoring but also avoids getting bogged down in cumbersome cross-domain coordination early on, solving the problem of difficulty in aligning heterogeneous cross-domain data.

[0035] When new types of sensors, such as photovoltaic monitoring, or new evaluation standards, such as near-zero energy consumption standards, need to be integrated, only horizontal expansion on the corresponding sub-ontology side is required, along with a small amount of additional mapping; the entire ontology will not be affected. Furthermore, both ontologies can be reused independently by other projects, improving the reusability of knowledge modules and demonstrating good scalability and reusability.

[0036] S1013: Merge the two sub-ontologies and perform semantic alignment to form a unified green building ontology (OntoGB). In the ontology construction process, merging the OntoDK and OntoSen ontology is a crucial step in building OntoGB. This embodiment achieves deep integration of OntoDK and OntoSen through semantic mapping rules. The rdfs:subClassOf rule is used to classify sensor devices into the building facilities subclass, the owl:equivalentClass rule clarifies the equivalence between environmental risk factors and sensor monitoring objects, and the owl:equivalentProp and rdfs:subPropertyOf rules associate the core attributes of the two ontology types, ensuring semantic consistency between building design data and sensor monitoring data. After merging, a unified green building ontology (OntoGB) is formed, and the consistency and completeness of the ontology are verified using a Pellet inference engine. Table 3 shows an example of the ontology merging rules.

[0037] Table 3 Examples of Ontology Merging Rules

[0038] This embodiment achieves deep semantic integration between the two ontologies through precise mapping rules, thereby promoting the effective combination of building design data and sensor monitoring data in green building evaluation. These mapping rules not only ensure semantic alignment between OntoDK and OntoSen at the conceptual and attribute levels, but also verify the consistency and completeness of the mapped ontologies through a description logic inference engine. Ultimately, the sub-ontologies were successfully merged into a unified OntoGB ontology, providing a complete semantic framework for the design, monitoring, and evaluation of green buildings.

[0039] S102: For the unstructured data of green building standard documents, entity recognition and relation extraction are performed to obtain entity-relation-attribute triple structure. Using the green building ontology as the semantic framework, a visualized green building knowledge graph is constructed.

[0040] In this embodiment, unstructured data from green building standard documents, semi-structured data from BIM models, and structured data from sensor monitoring are transformed into a unified entity-relationship-attribute triple structure to solve the problem of heterogeneous multi-source data. A visualized and queryable knowledge graph is constructed to clearly present the relationships between building components, attributes, standard requirements, and monitoring data, enabling computers to quickly locate which standards a building component must meet and which monitoring data corresponds to that standard. The knowledge graph's schema layer (or abstraction layer) is the constructed OntoGB ontology, while the instance layer contains standard specification text data, BIM model information, and sensor data, providing a knowledge base for SHACL rule verification and ensuring that rules can accurately match the corresponding building entities and attributes.

[0041] The specific implementation process is as follows: S1021: First, construct the dataset; This embodiment uses the "China Green Building Evaluation Standard" (GB / T50378-2019) as its core, integrates relevant building codes, removes duplicate content, breaks down complex text, and transforms tabular data into natural language clauses, ultimately forming a high-quality corpus containing 5,000 sentences. A BIO annotation scheme is adopted to annotate building components, attributes, and attribute values ​​in the text, clarifying entity categories and relationship types, providing labeled data for model training.

[0042] S1022: Utilize NLP to perform Named Entity Recognition (NER) and Relation Extraction (RE) on building code text data; Named Entity Recognition (NER) aims to automatically identify core entities such as building components, attributes, and standard requirements from standard text, addressing the question of which key objects exist. Texts in the field of green building typically contain a large number of professional terms and technical terms. The diversity and specialization of these terms place higher demands on the model's vocabulary understanding and semantic analysis capabilities. This embodiment utilizes the LERT-CNN-BiLSTM-MHA-CRF model, specifically optimized for professional terms in the field of green building, to extract seven types of entities from green building standard documents, including building components, building types, attributes, and attribute values. For example, for residential buildings, the wall has a formaldehyde concentration of 0.08 mg / m³.

[0043] In this embodiment, the LERT-CNN-BiLSTM-MHA-CRF model is used to extract target entities (such as building components, materials, and attributes) from green building standard documents. LERT is a domain knowledge injection module; LERT (Language-driven Encoder Representations from Transformers) is a pre-trained language model optimized for Chinese. Its input is the word sequence of the green building standard document, and its output is the contextual semantic representation of each word, forming word vectors. CNN is a convolutional neural network, BiLSTM is a bidirectional long short-term memory network, MHA is a multi-head attention mechanism, and CRF is a conditional random field. The green building domain text is preprocessed (word segmentation, stop word removal, and irrelevant character cleanup) to obtain a normalized word sequence. The word sequence is input into the LERT pre-trained model to generate deep semantic word vectors, capturing the contextual information of professional terms. The deep semantic word vectors are input into the CNN model to extract local phrase features, enhancing the recognition of domain terms. BiLSTM is used to bidirectionally model context dependencies, combined with a multi-head attention mechanism to capture long-distance semantic associations, and finally, CRF is used to optimize the entity label sequence.

[0044] Specifically, in combination Figure 2The core logic of entity extraction is text semantic encoding, local feature extraction, global feature modeling, and label sequence optimization. It consists of four parts: the representation layer (LERT), the CNN encoding layer, the recurrent network layer (BiLSTM+MHA), and the prediction layer (CRF). The representation layer (LERT) converts each word in the text into a vector containing contextual semantics. The semantic vector is a numerical vector that converts words, phrases, or sentences in the text into a computer-processable vector. The CNN encoding layer encodes the local features of the text, which can effectively capture key terms and phrase features in the text of green building, such as professional terms such as waterproof layer and step height. The recurrent network layer (BiLSTM+MHA) captures the long-distance dependencies in the text, such as the association between the building and the building envelope, and better handles the problem of long-distance semantic association. In the prediction layer, the CRF module effectively reduces label ambiguity by globally optimizing the label sequence, thereby significantly improving the overall accuracy of named entity recognition.

[0045] The specific implementation process of each layer is as follows: The representation layer uses LERT as a text encoder to generate deep semantic representations in the field of green building. LERT incorporates syntactic information and semantic features during the pre-training stage, giving it better generalization ability and contextual understanding when processing technical terms. The input text sequence is... , Representing the For each word, LERT computes a contextual representation of the word using a bidirectional Transformer: ; in, For word-level context representation, each Dimensions .

[0046] In the LERT computation process, each Transformer layer consists of self-attention, a feedforward network (FFN), and a layer normalization layer, and the computation method is as follows: Attention (Q, K, V) = softmax ( V; in, These are query, key, and value matrices, respectively. This refers to the dimension of the attention head. Ultimately, LERT generates word vectors. As input to downstream networks, it enhances the text semantic representation capabilities.

[0047] Building upon the contextual representation generated by LERT, the CNN encoding layer further extracts local features of the text. The core advantage of CNN lies in its ability to effectively capture phrase-level information, making the model's recognition of green building terminology more accurate. Assuming the input LERT word vector matrix is... ,in If the LERT word vector dimension is used, the CNN calculation process is as follows: ; in, For convolution kernel weights, Indicates the window size is Local segments, For bias terms, This outputs features for the CNN. This embodiment employs a multi-channel CNN, using convolutional kernels of different sizes. To extract features at different granularities, and obtain a global representation through max pooling: ; Finally, the feature matrix output by the CNN encoding layer As input to the BiLSTM layer, it enriches the local information of the text.

[0048] To model global information in the text, the recurrent network layer uses BiLSTM to model the sequence. BiLSTM computes hidden states through forward and backward LSTMs, capturing contextual information and thus improving the model's ability to understand the structure of long sentences. Given the feature matrix output by the CNN, the BiLSTM calculation process is as follows: ; ; ; in, and These are the hidden states calculated by the forward and backward LSTM, respectively. This is the concatenated BiLSTM representation. The hidden unit dimension of the BiLSTM layer is set to 256, and the final generated feature matrix dimension is... .

[0049] To further enhance the model's ability to model long-range dependencies, this embodiment introduces MHA. MHA captures information from different subspaces through multiple attention heads, thereby improving the model's ability to distinguish named entities. Given the feature matrix output by the BiLSTM... The MHA calculation process is as follows: ; ; in, For the linear transformation matrix of query, key, and value, For the number of attention heads, This is the final projection matrix. The addition of multi-head attention can effectively enhance the model's ability to model complex semantic relationships and improve the recognition accuracy of named entities.

[0050] In named entity recognition tasks, MHA or BiLSTM alone may result in highly independent label predictions, failing to fully consider the correlation between sequence labels. Therefore, this embodiment uses CRF for sequence labeling optimization in the final prediction layer. CRF learns the label transition matrix. This modeling of the dependencies between labels before and after prediction improves the consistency of predictions. Given the output of the MHA layer... CRF calculates the tag sequence The scores are as follows: ; in, Indicates label Transferred to The score, This is the current label score output by MHA.

[0051] CRF computes the optimal label path using the Viterbi Algorithm and is trained using negative log-likelihood loss (NLL). ; in, This represents all possible label sequences. Through the CRF prediction layer, the model can more effectively perform named entity recognition, reduce false predictions, and improve overall accuracy.

[0052] To evaluate the performance of the LERT-CNN-BiLSTM-MHA-CRF model, experiments were conducted by varying the prompt length (prompt_len) and prompt dimension (prompt_dim), with values ​​of [64, 128, 256] and [500, 768, 1000], respectively. The prompt length refers to the maximum length of the text segment processed by the model in a single operation. By comparing the three prompt lengths of 64, 128, and 256, the optimal prompt length was determined for green building standard text, balancing key semantics with efficiency. The prompt dimension refers to the vector dimension of the virtual prompt token in the prompt template, which is the dimension of the feature vector used by the model to capture domain-specific semantics. A higher dimension results in a more expressive prompt vector, capable of carrying more professional semantics from the green building domain; however, excessively high dimensions increase the model parameter size and training difficulty. Experiments were conducted with values ​​of 500, 768, and 1000 to verify the impact of different feature expressive capabilities on domain entity recognition.

[0053] The results are as follows Figure 3 As shown, the 64-length template performed the worst, with all metrics (precision, recall, and F1 score) below 75%. The 128-length template performed better, while the 256-length template did not significantly outperform the 128-length template. The [256,768] configuration had an F1 score 3.95% lower than the [128,768] configuration. This indicates that 128 tokens are sufficient to capture the key information in incident reports, making [128,768] the optimal configuration for entity extraction, striking a balance between performance and efficiency while reducing the risk of overfitting or underfitting.

[0054] Relation extraction (RE) aims to extract semantic relationships between entities from green building-related texts, supporting the construction of green building knowledge graphs, addressing the question of what the relationships are between objects, and providing edges for triple construction; this embodiment uses the GCN-PGRC model, such as... Figure 4 As shown, GCN (Graph Convolutional Network) is added to the traditional PGRC model to enhance the ability to capture the contextual association of entity relationships. Node features are initialized through LERT, text graphs are constructed using TF-IDF and PMI, and then neighbor node information is aggregated through GCN to enhance entity representation. Subsequently, the subject and object of potential relationships are predicted, and effective relationship pairs are filtered through the global correspondence matrix to improve relationship matching accuracy.

[0055] Relation extraction in the field of green building, based on entity extraction (such as building envelope, curtain wall, stair tread height, 0.9 meters), further determines the semantic relationships between entity pairs. Combined with... Figure 4 The specific process is as follows: The input consists of entities and canonical text obtained from relation extraction. To obtain semantic information for the context, the overall text information is needed. Compared to the traditional PGRC relation extraction model, this embodiment introduces a Graph Convolutional Network (GCN) before the global correspondence step, forming the GCN-PGRC framework. The core consists of three modules: text graph construction, which transforms linear text into structured graph data with nodes and edges, using LERT semantic encoding to generate text feature vectors h and TF-IDF+PMI to calculate edge weights; relation-specific sequence labeling, divided into GCN and relation prediction, which aggregates neighbor node features through the Graph Convolutional Network (GCN), strengthens the semantic association between entities, and predicts relations, using Softmax classification and subject-object probability calculation to predict potential subject-object pairs and preliminary relation types; and global filtering, which filters invalid relation pairs using the global correspondence matrix and Sigmoid probability to ensure accurate alignment between entities and relations.

[0056] The specific processing steps for each layer are as follows: Constructing a text graph: A text graph transforms unstructured text into structured graph data in the form of nodes and edges. Its core function is to allow models to capture semantic relationships between words through the graph's topological structure. Defining nodes: Input text segmentation: splitting a sentence into a sequence of words. Each word corresponds to a node. LERT is used to generate a node feature vector h. The edge weights are calculated to represent the semantic association strength between two words. The calculation is performed by combining two statistical methods: TF-IDF (word importance) and PMI (point mutual information, word co-occurrence strength).

[0057] TF-IDF is an indicator that measures the importance of a word in a text. The formula is: ; in, Indicator Frequency of occurrence in the current text Indicator Its rarity within the entire green building corpus; PMI is an indicator that measures whether two terms frequently appear together; the formula is: ; in, Indicator The probability of occurrence in the corpus; Indicator and The probability of them occurring simultaneously; Combining TF-IDF and PMI yields two terms. and Boundary weights between: ; Finally, an adjacency matrix A is generated. The adjacency matrix is ​​a structured representation of the text graph, and the elements in the matrix... Corresponding node and Boundary weights between: like (If there is a correlation), then ; like (No relation) .

[0058] Relation-specific sequence labeling is a key step in the GCN-PGRC framework. Its core objective is to classify relationships between entities based on their representations and contextual interactions within the sequence. First, a graph convolutional network is used to update the initial entity representations to fully explore the semantic connections between entities. Specifically, GCN uses a normalized adjacency matrix to aggregate entity features, enabling each node to integrate information from its neighbors, thereby improving the model's ability to perceive complex relationships. The update formula is as follows: ; Where H represents the initial node feature matrix, A is the adjacency matrix, and D is the degree matrix. σ is the learnable weight matrix of the i-th layer of GCN, where σ represents the non-linear activation function. After processing by GCN, the representation of an entity not only includes its own information but also incorporates the contextual features of its neighboring nodes.

[0059] After updating the entity representations in GCN, the next step is to predict the subject and object of the potential relationship. The model first calculates the prediction distributions of the subject and object, using the following formula: ; ; in, It is the representation of entity i after GCN processing. Representing contextual information or a representation of another entity j. and These are the learnable weight matrices used for subject and object prediction, respectively. and This is the corresponding bias term. The role of the Softmax function is to map the model's predicted values ​​to a probability distribution so that the model can classify the subject and object, thereby obtaining all possible subjects and objects.

[0060] Finally, after obtaining the possible subjects and objects, a global correspondence matrix is ​​introduced to filter the correct subject-object pairs. The global correspondence matrix can be learned synchronously with the latent relation prediction task because it is independent of specific relation types. Specifically, all possible subject-object pairs are enumerated and filtered using the rich features extracted by GCN. Then, the corresponding scores in the global correspondence matrix are checked; if they exceed a set threshold, the relation pair is retained; otherwise, it is filtered. After GCN processing, the formula for calculating the global correspondence is as follows: ; ; in, and Let i and j represent the updated features of subject i and object j after GCN processing, respectively. The learnable weights are the corresponding global matrix. It is the corresponding bias term. This is a non-linear activation function. This method can fully utilize the entity relationship information extracted by GCN, improve the filtering accuracy of subject-object relationship pairs, and ensure more accurate final prediction results.

[0061] Ultimately, the output is a structured triplet, such as wall - includes - waterproof layer, formaldehyde concentration - comparison relationship - ≤0.08mg / m³, residential building - subordinate relationship - indoor air quality requirements.

[0062] S1023: Construction of the core framework of knowledge graph; This embodiment utilizes named entity recognition technology to extract predefined entities in the green building domain from relevant text, such as building materials, energy-saving technologies, and environmental certifications. Relation extraction technology is used to classify the identified entities, and relationships between entities are established based on predefined relation templates, thus forming a triple structure in the green building domain. Finally, the generated triple data is imported into the Neo4j graph database using Cypher statements to construct the core framework of a visualized green building knowledge graph. Figure 5 This paper presents the algorithm for constructing a green building knowledge graph using Cypher statements. The final result is a comprehensive and structurally complete green building knowledge graph, as shown below. Figure 6 As shown, this provides users with more accurate and comprehensive green building knowledge services.

[0063] This embodiment presents a visualized knowledge graph covering entities and relationships such as building components, materials, and performance parameters. It achieves a structured presentation and retrieval of green building knowledge, providing a unified knowledge carrier for subsequent compliance verification. The entire process utilizes NLP technology to automate the conversion from text to knowledge, breaking down the barriers between unstructured text and structured data, and providing efficient and accurate knowledge support for green building compliance verification.

[0064] S103: Convert the semi-structured data of the BIM model and the structured data of sensor monitoring into a unified RDF format, and store the RDF format data in the knowledge base under the green building ontology.

[0065] In this embodiment, both BIM information and sensor data are converted into RDF (Resource Description Framework) format to achieve unification; the essence of RDF format is a triple represented by subject-predicate-object.

[0066] BIM data is stored according to the IFC standard and contains structured information such as building components, attributes, and spatial relationships. During conversion, these core elements must first be extracted and converted to RDF format using the IFC-to-RDF conversion method. The IFC-to-RDF conversion method extracts core information about building components, such as the geometric attributes, material composition, and spatial relationships of walls, floors, and HVAC systems, ensuring that each building element is accurately aligned with the concepts in the Green Building Ontology (OntoGB). For example, an IfcWall instance is mapped to the Wall class in the ontology, while retaining its fire resistance, thermal insulation, and other attributes.

[0067] Sensor data is stored in a relational database, converted using an RDB-to-RDF conversion method. The original sensor data is stored in relational database tables, with real-time monitoring data such as temperature, humidity, air quality, and energy consumption conforming to the W3C RDB2RDF standard. This embodiment uses a D2RQ tool to generate mapping rules, establishing semantic relationships between database fields and ontology entities, and converting sensor readings to RDF format. For example, temperature data is mapped to the `hasTemperature` attribute of the ontology and associated with the corresponding building space instance, ensuring the correspondence between real-time monitoring data and building physical entities.

[0068] The BIM information and sensor data are converted into triple relationships in RDF format and used as instance layers, and then incorporated into the knowledge base under the ontology of the green building domain.

[0069] S104: Define SHACL rules based on the green building knowledge graph, including numerical constraints, logical constraints, and instance semantic mapping rules; based on the SHACL rules, perform green building compliance checks on RDF format data in the knowledge base.

[0070] After acquiring structured building data and standard information, the key to achieving compliance verification lies in how to automatically validate the building data according to standard requirements. Shape Constraint Language (SHACL), as an RDF data validation language, can define the structure and constraints of data, enabling automated validation of structured data, and has broad application prospects in the field of Architecture, Engineering, and Construction (AEC). SHACL's core advantage lies in its powerful constraint expression capabilities, supporting various types of constraints such as data type constraints, value range constraints, logical constraints, and custom SPARQL query constraints, flexibly addressing complex semantic validation needs.

[0071] However, most existing SHACL rules are manually defined, making it difficult to adapt to the dynamic updates of green building standards, and they lack support for verification of multiple constraint combinations in complex scenarios.

[0072] Based on this, the core objective of this embodiment is to transform the constraints of green building standards into rules that can be automatically executed, to accurately verify building data, and to ensure that building design, material selection and environmental performance meet the requirements of the "China Green Building Evaluation Standard" (GB / T50378-2019), thus solving the problems of low efficiency and strong subjectivity of traditional manual verification.

[0073] Specifically, the textual requirements of the "China Green Building Evaluation Standard" will be transformed into machine-executable SHACL structured constraints to build a rule-based foundation for compliance verification.

[0074] Numerical constraints target quantifiable requirements, such as room area, harmful gas concentration, and step height. Regular expressions are used to identify qualifiers such as "not less than" and "not exceeding", which are then converted into SHACL numerical range constraints. For example, a step height ≤ 175mm corresponds to sh:maxInclusive175.

[0075] Logical constraints focus on qualitative and relational requirements, extracting expressions such as "must have" and "should be avoided," and transforming them into constraints such as sh:minCount (required attribute) and sh:disjoint (disallow association). For example, the site should avoid storing hazardous chemicals, corresponding to the disjoint constraint.

[0076] Instance semantic mapping achieves efficient association between entities, attributes, and IFC model elements in the knowledge graph by constructing a matching dictionary between Chinese terms and the ifcOWL ontology. For example, "wall" and "wall structure" are uniformly mapped to ifc:IfcWall, and "moisture barrier" is mapped to ifc:IfcMaterial.

[0077] For quantitative comparisons and directive terms, predefined mapping rules are used for direct conversion, ensuring accurate alignment between building data and SHACL rules. At the attribute mapping level, the focus is on the attribute sets of building components in the BIM model, such as IfcPropertySet, associating standard requirements with component attributes. For example, the material requirements for the damp-proof layer are bound to the Material attribute in the Pset_SlabCommon attribute set of IfcSlab, achieving precise positioning from rules to data.

[0078] Figure 7 This describes how the MaterialLevel attribute of the IfcSlab instance is represented in the ifcOWL example graph, ensuring that constraints are applied precisely to the target data, ultimately forming a SHACL rule base stored in .ttl format.

[0079] This embodiment establishes a fully automated SHACL rule definition method. First, at the semantic deconstruction layer, a pre-trained LERT-CNN-BiLSTM-MHA-CRF named entity recognition model and GCN-PGRC relation extraction model are used to perform end-to-end parsing of the text standard specifications. This extracts entity-attribute-value-qualifier quadruples and triplet relationships in one go. Even if the standard text is revised, the latest semantic elements can be obtained in seconds simply by re-reasoning, without manual re-annotation. At the template transformation layer, common qualifying phrases such as mandatory, lower bound, upper bound, prohibited, recommended, and optional are abstracted into parameterizable SHACL fragments. Template variables are then populated in real-time with specific values ​​or concepts from the triples, achieving batch rule generation with fixed templates and dynamic parameters. At the rule activation layer, the system utilizes the sh:shapeGraph mechanism of SHACL-Advanced, dynamically injecting the newly generated shape graph into the TopBraid-Composer verification engine via the SPARQL LOAD instruction, while simultaneously marking the old version as deprecated, directly completing the replacement. The entire process, from publishing the standard text to deploying the executable rules, takes less than one second on average, solving the problems of manual writing, version delays, and high maintenance costs.

[0080] Finally, based on the SHACL rules, a green building compliance check was performed on the RDF format data in the knowledge base. Specifically, the defined SHACL rules were applied to the RDF format BIM and sensor data. Figure 8 This is a BIM model used for compliance verification, outputting compliance conclusions and details through automated verification. First, prepare and associate OntoGB ontology RDF instance data, and load the .ttl format SHACL rule library into tools such as TopBraidComposerFE. The tool iterates through each building instance, matches the corresponding constraints, and determines whether the rules are met.

[0081] The output results are divided into compliance and non-compliance lists. Each item clearly indicates the inspection object and the corresponding standard rules. Compliance is marked with attribute values ​​that meet the standards, while non-compliance is explained in detail for the reasons, such as the floor slab not being associated with the moisture-proof layer attribute. This provides clear guidance for rectification. At the same time, it supports the generation of compliance reports, replacing manual review and improving efficiency and objectivity.

[0082] In this context, building instances refer to actual building components (such as walls and stairs), sensor devices, and their specific information (such as wall thickness and formaldehyde concentration measured by the sensor) in BIM. Each instance has a unique identifier and is the specific object for SHACL rule verification. For example, a 200mm thick concrete exterior wall in a residential building, a staircase with a step height of 160mm, and a formaldehyde sensor deployed in room 502 that measured a formaldehyde concentration of 0.06mg / m³, etc. These instances will clearly indicate which category they belong to (e.g., wall, formaldehyde sensor) and their specific attribute values ​​(e.g., thickness, concentration). The tool will compare each instance against the SHACL rules (exterior wall thickness ≥ 200mm) and "formaldehyde concentration ≤ 0.08mg / m³" to determine compliance.

[0083] This embodiment can define SHACL rules for different dimensions such as site safety, building materials, and spatial parameters. For example, site safety verification requires constraining buildings to avoid geological disaster areas and hazardous chemical storage areas, and automated inspection can be achieved through Boolean constraints of attributes such as hasGeotechnicalElement and hasHazardousChemicalStorage.

[0084] For example, for verifying the properties of building components, such as checking whether IfcWall and IfcSlab contain waterproofing material, the association constraints between entities and materials are defined through SHACL rules, and components that do not meet the requirements are automatically filtered out. For numerical properties, such as the step height not exceeding 0.18m, the RiserHeight attribute value of IfcStairFlight is extracted by combining the numerical range constraints of SHACL with SPARQL queries and compared with the standard to generate a compliance report.

[0085] The entire process is fully automated and requires no human intervention, which ensures the objectivity and consistency of the verification results and increases the verification efficiency several times over, effectively supporting the large-scale application of green building compliance review.

[0086] Example 2 In one or more embodiments, a green building model compliance verification system is disclosed, comprising: The ontology construction module is configured to construct a domain knowledge ontology and a sensor ontology based on green building standard documents; and to achieve deep integration of the domain knowledge ontology and the sensor ontology through semantic mapping rules to form a unified green building ontology. The knowledge graph construction module is configured to perform entity recognition and relation extraction on the unstructured data of green building standard documents to obtain entity-relation-attribute triple structure, and construct a visualized green building knowledge graph using the green building ontology as a semantic framework. The data conversion module is configured to convert semi-structured data from BIM models and structured data from sensor monitoring into a unified RDF format, and store the RDF format data in the knowledge base under the green building entity. The compliance check module is configured to define SHACL rules based on the green building knowledge graph, including numerical constraints, logical constraints, and instance semantic mapping rules; and to perform green building compliance checks on RDF format data in the knowledge base based on the SHACL rules.

[0087] It should be noted that the specific implementation methods of the above modules are exactly the same as those in Example 1, and will not be described in detail again.

[0088] Example 3 In one or more embodiments, a terminal device is disclosed, comprising a processor and a memory, wherein the processor is used to implement instructions; and the memory is used to store multiple instructions adapted to be loaded by the processor and executed by the green building model compliance verification method described in Embodiment 1.

[0089] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0090] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0091] In the implementation process, each step of the above method can be completed by the integrated logic circuits in the processor hardware or by software instructions.

[0092] Example 4 In one or more embodiments, a computer-readable storage medium is disclosed, wherein a plurality of instructions are stored, the instructions being adapted to be loaded by a processor of a terminal device and executed by the green building model compliance verification method described in Embodiment 1.

[0093] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for verifying the compliance of a green building model, characterized in that, include: Based on green building standard documents, domain knowledge ontology and sensor ontology are constructed respectively; By using semantic mapping rules, a deep fusion of domain knowledge ontology and sensor ontology is achieved, forming a unified green building ontology; For the unstructured data of green building standard documents, entity recognition and relation extraction are performed to obtain entity-relation-attribute triple structure. Using the green building ontology as a semantic framework, a visualized green building knowledge graph is constructed. The semi-structured data of the BIM model and the structured data of the sensor monitoring are converted into a unified RDF format, and the RDF format data is stored in the knowledge base under the green building ontology; Based on the green building knowledge graph, SHACL rules are defined, including numerical constraints, logical constraints, and instance semantic mapping rules; based on the SHACL rules, green building compliance checks are performed on RDF format data in the knowledge base.

2. The green building model compliance verification method as described in claim 1, characterized in that, The domain knowledge ontology is constructed in a top-down manner. The top-level classes include five categories: building type, evaluation criteria, building components, building structure, and evaluation role. Each top-level class has several subclasses, forming a hierarchical structure.

3. The green building model compliance verification method as described in claim 1, characterized in that, The sensor ontology is based on the SOSA framework, defining sensor, observation object, observable attribute and measurement result class. Semantic knowledge is injected into the sensor monitoring process through semantic modeling of real-time monitoring data, and the observation behavior and measurement data are associated through the hasResult attribute.

4. The green building model compliance verification method as described in claim 1, characterized in that, Deep fusion of domain knowledge ontology and sensor ontology is achieved through semantic mapping rules, specifically including: The rdfs:subClassOf rule is used to classify sensor devices into the building facilities subclass. The owl:equivalentClass rule is used to clarify the equivalence relationship between environmental risk factors and sensor monitoring objects. The core attributes of the two types of ontology are associated through the owl:equivalentProp and rdfs:subPropertyOf rules to ensure the semantic consistency between building design data and sensor monitoring data.

5. The green building model compliance verification method as described in claim 1, characterized in that, Entity identification is performed on the unstructured data of green building standard documents as follows: The LERT-CNN-BiLSTM-MHA-CRF model is used to extract target entities from green building standard documents. The green building standard documents are input into the LERT pre-trained model to generate deep semantic word vectors, which capture the contextual information of professional terms. The deep semantic word vectors are input into the CNN model to extract local phrase features and enhance the recognition of domain terms. The context dependency is modeled bidirectionally by BiLSTM and combined with a multi-head attention mechanism to capture long-distance semantic associations. Finally, the entity label sequence is optimized by CRF.

6. The method for verifying the compliance of a green building model as described in claim 1, characterized in that, Relation extraction is performed on the unstructured data of green building standard documents, specifically as follows: LERT semantic encoding is used to generate text feature vectors, and TF-IDF and PMI are used to calculate edge weights to construct a text graph. A graph convolutional network (GCN) is used to aggregate the features of neighbor nodes, strengthen the semantic association between entities and perform relationship prediction. Softmax classification and subject-object probability calculation are used to predict potential subject-object pairs and preliminary relationship types. A global correspondence matrix is ​​introduced to filter out correct subject-object relationship pairs and invalid relationship pairs to ensure accurate alignment of entities and relationships.

7. The green building model compliance verification method as described in claim 1, characterized in that, The instance semantic mapping rules establish a matching dictionary between Chinese terms and ifcOWL ontology to associate entities, attributes, and BIM model elements in the knowledge graph. Load the defined SHACL rules, traverse each building instance in the BIM model, match the corresponding constraints, determine whether the rules are met, and output a list of compliance and non-compliance. The list clearly indicates the inspection object, the corresponding standard rule, whether the attribute value is compliant, and the reason for non-compliance.

8. A green building model compliance verification system, characterized in that, include: The ontology construction module is configured to construct a domain knowledge ontology and a sensor ontology based on green building standard documents; and to achieve deep integration of the domain knowledge ontology and the sensor ontology through semantic mapping rules to form a unified green building ontology. The knowledge graph construction module is configured to perform entity recognition and relation extraction on the unstructured data of green building standard documents to obtain entity-relation-attribute triple structure, and construct a visualized green building knowledge graph using the green building ontology as a semantic framework. The data conversion module is configured to convert semi-structured data from BIM models and structured data from sensor monitoring into a unified RDF format, and store the RDF format data in the knowledge base under the green building entity. The compliance check module is configured to define SHACL rules based on the green building knowledge graph, including numerical constraints, logical constraints, and instance semantic mapping rules; and to perform green building compliance checks on RDF format data in the knowledge base based on the SHACL rules.

9. A terminal device comprising a processor and a memory, the processor for implementing instructions; the memory for storing multiple instructions, characterized in that, The instructions are adapted to be loaded by a processor and executed as the green building model compliance verification method according to any one of claims 1-7.

10. A computer-readable storage medium storing a plurality of instructions, characterized in that, The instructions are adapted to be loaded by the processor of a terminal device and executed by the green building model compliance verification method according to any one of claims 1-7.

Images