Construction map and cost map fusion method, device and system and storage medium

Abstract

The invention relates to a construction map and cost map fusion method, device and system and a storage medium, and the method comprises the steps: firstly providing a collaborative embedding method, generating initial semantic embedding through combining a TF-IKGF algorithm and a FastText model, fusing a topological structure of a map through employing a graph convolution network, and generating a node vector representation with depth semantics and a structure context; then, adopting a two-stage unsupervised entity alignment strategy of clustering coarse screening and intra-cluster fine alignment, reducing a matching range through K-Means clustering, then performing precise matching based on cosine similarity in a cluster, and efficiently identifying an aligned entity pair; and finally, integrating the construction knowledge into a cost system according to business logic through an automatic workflow containing manual verification, and completing deep integration of the knowledge graph. According to the method, a knowledge link of construction and cost can be automatically opened, a unified knowledge graph which is clear in logic and can serve technical and economic integrated analysis is constructed, and the intelligent application level of knowledge in the building field is remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of building informatization and knowledge management, and particularly to a construction graph and cost graph fusion method, device, system and storage medium. BACKGROUND

[0002] With the deep application of information technology in the architecture, engineering and construction (AEC) industry, how to effectively fuse the massive information throughout the project life cycle has become the key to improving the intelligent level of engineering management. In recent years, knowledge graph technology has been widely used in building knowledge bases in the construction field due to its strong semantic association and knowledge reasoning ability. For example, Chen et al. are committed to solving the alignment challenge of knowledge from different sources. They proposed a semi-supervised entity alignment method for heterogeneous building knowledge graphs, which effectively improved the alignment accuracy between different data sources by introducing a small amount of human annotation to guide model learning. Some research has also focused on automatically constructing knowledge graphs from specific documents. For example, Zhao et al. extracted key entities and event relationships from construction dispute texts to construct a knowledge graph to provide intelligent decision support for contract and legal issues in projects. In China, relevant research has also made significant progress. For example, Wang et al. began to explore a cross-domain knowledge graph entity alignment method based on graph neural networks, which provides a new way to break down the barriers between fields by learning the representation of entities in different knowledge systems. In addition, Liu et al. focused on knowledge graph construction for specific scenarios such as construction safety. They systematically analyzed the causes, effects and preventive measures of safety risks to form a professional knowledge base that can assist in risk early warning. These studies have promoted the process of structuring knowledge in the construction field and successfully formed numerous professional and vertical knowledge systems, such as the construction knowledge system centered on construction technology, technical specifications and quality control, and the cost knowledge system centered on engineering quantity list pricing specifications, quota sub-items and cost composition. The construction knowledge graph and cost knowledge graph handled by the present application are typical representatives of such professional knowledge bases that have a complete multi-level index hierarchy.

[0003] However, existing research mostly focuses on single-dimensional knowledge modeling, resulting in a long-term separation of the two core knowledge systems of construction and cost. Specifically, a knowledge graph describing construction technology follows the internal logic of technical processes and causal relationships, while a knowledge graph describing project cost is organized according to pricing specifications and cost composition. There is a natural "semantic gap" in concept granularity, description method, and organizational structure between the two. For example, a specific "high-strength bolt connection process" in construction knowledge may be abstracted as a sub-item under the "steel structure manufacturing and installation" quota in cost knowledge, and is associated with the corresponding labor, material, and mechanical costs. This asymmetric, many-to-many complex mapping relationship makes traditional association methods based on rules or surface text matching unsustainable. In addition, existing research rarely considers the context information of knowledge nodes in their graph topology when performing entity alignment, i.e., the complete semantics of a node is determined by its own attributes and neighbor nodes. Ignoring structural information and relying solely on text similarity for judgment can easily lead to ambiguity and mismatch in alignment.

[0004] In summary, although single-domain architectural knowledge graph research has begun to take shape, how to overcome the semantic and structural barriers between different knowledge systems and achieve efficient and accurate automated alignment and deep integration remains a weak link and a technical bottleneck that needs to be broken through. Project managers and technical personnel still need to rely on personal experience to manually compare the two knowledge systems, which is time-consuming and labor-intensive. The information silo problem seriously hinders the realization of higher-level applications such as cost and technology integration analysis and intelligent decision-making. Therefore, at the current stage, a new knowledge graph integration method that can understand the deep semantics of nodes and the topology of graphs is needed to automatically and intelligently bridge the knowledge link between construction and cost and build a unified architectural engineering knowledge model. SUMMARY

[0005] To solve the problems in the prior art, the purpose of the present application is to provide a construction graph and cost graph fusion method, device, system, and storage medium, which can automatically and intelligently bridge the knowledge link between construction and cost and build a unified architectural engineering knowledge model.

[0006] The specific goals are as follows: I. Constructing a field-adaptive initial semantic embedding: based on the FastText pre-training model, combining TF (term frequency) and IKGF (inverse knowledge graph frequency) statistical features, generate a weighted initial semantic embedding for each entity node. This goal aims to dynamically capture the core semantics of architectural engineering terms and represent the importance of different entities in their respective knowledge systems differently.

[0007] II. Realize the deep node representation of the fusion topology: design and apply a multi-layer graph convolution network (GCN) with normalized adjacency matrix as input. The goal is to effectively integrate the context structure information of the node into its initial semantic embedding through the neighborhood aggregation mechanism of the graph neural network, so as to generate a final node representation with deep semantic and topological correlation, and improve the robustness and expression ability of the vector.

[0008] III. Propose an efficient unsupervised hierarchical alignment strategy: in order to balance the efficiency and accuracy of entity alignment, the invention aims to adopt a hierarchical strategy of "clustering coarse screening, fine matching within cluster". First, the K-Means clustering algorithm is used to quickly and unsupervisedly divide the final node embedding of the two graphs, so as to effectively reduce the search space of potential matching; secondly, the high threshold cosine similarity calculation is used for accurate matching within each cluster to ensure the high confidence of the alignment result.

[0009] IV. Establish a knowledge fusion workflow: design an automated process to fuse the verified aligned entities in the graph database. The goal is not only to realize the connection of nodes, but more importantly, to reconstruct the knowledge system according to the business logic that construction knowledge belongs to cost knowledge, and finally generate a clear structure of construction-cost integrated knowledge graph by creating new hierarchical relationships and deleting redundant nodes.

[0010] In order to achieve the above purposes, the invention adopts the following technical solutions: The fusion method of construction graph and cost graph comprises the following steps, Based on the pre-trained word vector model, combined with the statistical characteristics of the weight of the representative field term, the initial semantic embedding vector of each entity node in the construction graph and the cost graph is generated; Using graph neural network, combining the initial semantic embedding vector and the topological structure of the construction graph and the cost graph, learning and generating the final node embedding vector which integrates the structure information; Perform entity alignment, first cluster all final node embedding vectors by unsupervised clustering algorithm, then identify the aligned entity pairs from different graphs within each cluster based on similarity calculation; Based on the aligned entity pairs, fuse the construction graph and the cost graph in the graph database, establish new hierarchical relationships to generate an integrated knowledge graph.

[0011] Further, the statistical characteristics of the weight of the representative field term are the combination of word frequency and inverse knowledge graph frequency.

[0012] Further, the graph neural network is a multi-layer graph convolution network.

[0013] Further, the step of generating the final node embedding vectors comprises: taking the normalized adjacency matrix as input and applying a nonlinear activation function after each convolutional layer.

[0014] Further, the step of aligning entity pairs further comprises a preprocessing sub-step: identifying entities with the same name in the construction graph and the cost graph, and replacing their respective final node embedding vectors with a fusion vector, which is the arithmetic mean of the final node embedding vectors of the entities with the same name.

[0015] Further, the unsupervised clustering algorithm is the K-Means algorithm, and the similarity calculation within the cluster is the cosine similarity calculation.

[0016] Further, the fusion of the construction graph and the cost graph in the graph database comprises: creating a new node that inherits the properties of the entity nodes in the construction graph; establishing a hierarchical relationship between the new node and the corresponding entity nodes in the cost graph; and deleting the original entity nodes and their associated relationships in the construction graph.

[0017] The fusion device of the construction graph and the cost graph comprises: An initial semantic embedding vector generation module is configured to generate an initial semantic embedding vector for each entity node in the construction graph and the cost graph based on a pre-trained word vector model and in combination with statistical features representing the weights of domain-specific terms; A final node embedding vector generation module is configured to learn and generate final node embedding vectors that fuse structure information by using a graph neural network in combination with the initial semantic embedding vectors and the respective topological structures of the construction graph and the cost graph; An entity alignment execution module is configured to first cluster all the final node embedding vectors by using an unsupervised clustering algorithm, and then identify aligned entity pairs from different graphs within each cluster based on similarity calculation; A fusion module is configured to fuse the construction graph and the cost graph in a graph database based on the aligned entity pairs, and establish new hierarchical relationships to generate an integrated knowledge graph.

[0018] The fusion system of the construction graph and the cost graph comprises a processor and a memory, and the memory stores a computer program, and the processor executes the computer program to implement the fusion method described above.

[0019] The computer-readable storage medium stores a computer program, and the computer program is executed by the processor to implement the fusion method described above.

[0020] Overall, the present application has the following advantages: I. A collaborative embedding method for construction-cost field heterogeneous semantics is proposed: In view of the problem of knowledge alignment between construction knowledge description (such as process and flow) and cost knowledge description (such as bill of quantities), the application applies a TF-IKGF algorithm to quantify the weight of field terminology, and combines the topological structure information of the atlas with a graph convolution network (GCN). This method can effectively capture and quantify the entity connotation under two different discourse systems, and generate a highly adaptive and robust entity vector representation in the field.

[0021] II. Design an entity alignment strategy that adapts to the many-to-one mapping relationship in the construction field: To solve the complex mapping problem that multiple specific knowledge points on the construction side may correspond to a single quota item on the cost side, the application proposes a two-stage alignment method of "clustering coarse screening and intra-cluster fine matching". This method efficiently aggregates construction and cost entities with potential associations into the same matching space through unsupervised clustering, and then performs accurate matching through intra-cluster similarity calculation. This strategy not only significantly improves the efficiency and accuracy of entity alignment in large-scale and high-complexity building knowledge graphs, but also provides a clear candidate set for subsequent manual review.

[0022] III. An automated knowledge fusion workflow driven by engineering business logic is constructed: The application not only provides an alignment algorithm, but more importantly, establishes an end-to-end automated workflow from alignment to fusion. The fusion operation in this process is not simply node merging, but according to the core engineering logic that "construction knowledge is the specification and supplement of cost items", the knowledge system is reconstructed in the graph database, thereby generating a new knowledge graph with clear logic, knowledge value-added, and direct service to construction scheme for technical and economic integrated analysis. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 A fusion method flowchart of the construction graph and the cost graph provided by an embodiment of the application; Figure 2 A fusion system logic framework diagram of the construction graph and the cost graph provided by an embodiment of the application; Figure 3 A node representation learning process schematic diagram of the fusion method and system of the construction graph and the cost graph provided by an embodiment of the application; Figure 4 A TF-IKGF algorithm principle schematic diagram of the fusion method and system of the construction graph and the cost graph provided by an embodiment of the application; Figure 5 A graph convolution network (GCN) neighborhood aggregation schematic diagram of the fusion method and system of the construction graph and the cost graph provided by an embodiment of the application; Figure 6A flowchart illustrating the two-stage entity alignment strategy of the method and system for fusing construction drawings and cost estimates, as provided in an embodiment of the present invention. Figure 7 This is a schematic diagram illustrating the knowledge fusion process of a method and system for fusing construction drawings and cost estimates, provided in an embodiment of the present invention.

[0024] Figure 8 This is a structural block diagram of a device for integrating construction drawings and cost estimates, provided in one embodiment of the present invention.

[0025] Figure 9 This is a structural block diagram of a system for integrating construction drawings and cost estimates, provided in one embodiment of the present invention. Detailed Implementation

[0026] The present invention will now be described in further detail.

[0027] This invention proposes a method, apparatus, system, and storage medium for integrating construction drawings and cost estimates, aiming to provide a solution to the problems of fragmented knowledge, disconnect between construction technology and cost quota information, and low efficiency of knowledge utilization in the current construction engineering field. The overall research framework for this task is as follows: Figure 2 As shown. The specific implementation of this invention will elaborate on the entire process from the collection and processing of heterogeneous data sources, to the initial construction of two independent domain knowledge graphs, and finally to the high-quality fusion of the two graphs through an innovative entity alignment method that integrates deep learning and domain knowledge. The following will combine... Figure 1 and Figure 2 Please provide a detailed explanation.

[0028] Example 1: This embodiment provides a method for integrating construction drawings and cost estimates.

[0029] I. Data Preparation and Construction of Multi-level Indicator Knowledge Graph like Figure 1 As shown, step S100 in this embodiment involves constructing two independent, high-quality domain knowledge graphs: a construction engineering knowledge graph (hereinafter referred to as the "construction graph") and a construction engineering cost knowledge graph (hereinafter referred to as the "cost graph"). This process encompasses data collection, structured preprocessing, and the design and instantiation of the graph schema based on a multi-level indicator system.

[0030] 1. Data Collection and Sources The data in this embodiment mainly comes from two core knowledge systems in the field of construction engineering: Construction Knowledge Domain: This embodiment processes source data on construction knowledge, including information on construction techniques and operating procedures. The data is primarily collected from authoritative building construction codes, construction method standards, technical guidelines, and professional textbooks. This data describes the "how-to" question, encompassing specific construction techniques, operating procedures, key quality control points, material requirements, and machinery selection, among other technical knowledge.

[0031] Cost Knowledge Domain: This embodiment processes cost knowledge source data containing information such as quota items and quantity calculation rules. The data is primarily collected from national or local construction engineering budget quotas, bill of quantities pricing specifications, and related cost composition documents. This data describes the question of "how much money is spent," including economic knowledge such as the division of engineering projects, work content, quantity calculation rules, and pricing units.

[0032] The data collected in this embodiment are of various forms, mainly semi-structured text, which needs to be preprocessed before it can be used for map construction.

[0033] 2. Establishment of a multi-level indicator system and data structuring In order to systematically express complex architectural engineering knowledge, this embodiment introduces a multi-level index system to structure the collected data, which lays a solid foundation for subsequent map construction.

[0034] The multi-level indicator system of construction drawings: Referring to the construction knowledge source data table used in this invention, the knowledge in the construction drawings is organized into a hierarchical structure. For example, the top-level node can be "concrete engineering," under which there can be sub-nodes such as "formwork engineering," "reinforcement engineering," and "concrete preparation and pouring." Each sub-node can be further subdivided, such as "formwork engineering," which includes more specific knowledge nodes such as "formwork design," "formwork installation and removal," and "various types of formwork (such as wall formwork, beam formwork)." This hierarchical indicator system allows complex construction knowledge to be clearly expressed in a "general-specific" structure. In the source data, the "node data" table defines these entities, and the "relationship data" table connects these hierarchical and related entities through relationships such as "contains," forming a set of triples (Subject, Predicate, Object).

[0035] The cost estimation map employs a multi-level index system: Referring to the cost knowledge source data table used in this invention, the cost estimation map also utilizes multi-level indicators. Its system is closely integrated with the coding system of the bill of quantities pricing specification. For example, a top-level node could be "cast-in-place concrete engineering," under which, according to project coding rules, are sub-nodes such as "cast-in-place concrete foundation," "cast-in-place concrete column," and "cast-in-place concrete beam." Each sub-node is associated with specific attributes such as "project characteristics," "project name," "unit of measurement," and "quantity calculation rules." For instance, the node "cast-in-place concrete column" is associated with project characteristics such as "column shape (rectangular column, irregular column)," "concrete strength grade," and "concrete type," as well as its quantity calculation rule "calculated by volume according to the dimensions shown in the design drawings..." This structure tightly binds the pricing rules to the engineering entity.

[0036] Through the aforementioned multi-level indicator system, this embodiment transforms unstructured text knowledge into structured triplet data, providing high-quality data input for subsequent automated graph construction.

[0037] 3. Database instantiation of the graph In this embodiment, Neo4j is used as the graph database storage platform, and the processed construction triplet data and cost triplet data are imported into Neo4j to create two independent knowledge graphs.

[0038] Construction GraphSG: Nodes are assigned a unified "GraphSG" label and additional labels representing their specific types. For example, the source data contains the entity "concrete pouring," which contains specific construction points, such as entity A: "Non-self-compacting concrete must be poured and compacted in layers. The thickness of each layer depends on the vibration method." These entities and their hierarchical relationships are converted into triples, such as (concrete pouring, contains, entity A). This invention loads all such triples into a graph database (Neo4j) to form an initial construction knowledge graph.

[0039] GraphZJ (Concrete Cost Knowledge Graph): Nodes are assigned a unified "GraphZJ" label and additional labels representing their list items. For example, the source data contains the entity "Concrete Pouring and Vibration," which includes specific pricing or process requirements, such as entity B: "Requirements for Concrete Pouring and Vibration." These entities are also converted into triples and loaded into the graph database, forming an independent cost knowledge graph that coexists with the construction drawings.

[0040] Thus, this embodiment has yielded two independent domain knowledge graphs, each with its own internal knowledge structure, but which are not yet connected to each other. These graphs contain entities A and B that need to be aligned. This lays the foundation for subsequent entity alignment and knowledge fusion.

[0041] II. Node Representation Learning As Figure 1 shown, in step S200, in order to achieve a deep understanding and matching of heterogeneous knowledge, the present invention first needs to convert each entity node representing specific knowledge in the graph into a machine-computable and information-rich mathematical vector. This process is node representation learning. This embodiment proposes a two-stage representation method that integrates domain features and structural context, as Figure 3 shown: First, through text preprocessing and the TF-IKGF algorithm, combined with a general word vector model, "initial semantic embeddings" are generated that can accurately capture the core connotations of professional terms in the construction industry. Subsequently, the graph convolutional network (GCN) is used to further incorporate the neighborhood structure information of the nodes into this embedding, and finally, "enhanced node embeddings" with both deep semantics and topological relevance are output, laying a solid foundation for subsequent high-precision entity alignment.

[0042] 1. Text Preprocessing Since the entity names in the graph are Chinese natural language texts, standard natural language processing (NLP) preprocessing is first required.

[0043] Chinese Word Segmentation: The "jieba" word segmentation library is used to segment the names (Name attributes) of all entity nodes in the two graphs. "Jieba" can effectively process Chinese texts and segment them into a series of tokens.

[0044] Stop Word Filtering: After word segmentation, in order to reduce noise, computational complexity, and highlight the core semantics, this embodiment uses a predefined stop word list to filter out those commonly occurring but meaningless words.

[0045] Taking the name of entity A as an example, the token sequence obtained after word segmentation is: ["for", "non", "self-compacting", "concrete", "must", "layered", "pouring", ",", "layered", "tamping", ".",...]. After removing stop words (such as "for", "must", "the"), the core semantic tokens are retained: ["non", "self-compacting", "concrete", "layered", "pouring", "layered", "tamping",...]. Similarly, the core tokens obtained after processing the name of entity B are: ["concrete", "pouring", "vibrating", "requirements"].

[0046] 2. Hybrid Weight Semantic Representation: Integrating Domain Knowledge and General Semantics Traditional entity alignment methods often represent entities only using the average of word vectors, ignoring the importance differences of different words in a specific domain knowledge system. In this embodiment, a hybrid weight representation method is adopted to generate an initial embedding vector for each entity that contains both general semantics and reflects domain importance.

[0047] First is the basic semantic embedding: the FastText model.

[0048] This embodiment uses the FastText pre-trained word vector model proposed by Facebook AI Research in 2016 as the source of basic semantics. Compared with models such as Word2Vec and GloVe that rely on complete vocabulary for learning, the core innovation of FastText lies in its subword embedding mechanism. It treats each word as a set of character n-grams and learns vector representations for these n-grams. The final vector of a word is the sum of the vectors of all its character n-grams. For example, for the word "concrete", in addition to the vector of the whole word, it will also learn the vectors of subwords such as "混", "凝", "土", "混凝", "凝土", etc.

[0049] The advantage of this mechanism is reflected in professional fields such as construction engineering, where new terms and compound words (such as "post-cast strip", "keyed board", "self-compacting concrete") emerge in an endless stream. For these out-of-vocabulary (OOV) words that may not appear in the general corpus, traditional models cannot generate effective vector representations. However, FastText can synthesize a high-quality vector representation for any out-of-vocabulary word through the known subword vectors it consists of, which greatly enhances the model's representation ability and generalization for domain-specific terms. In this embodiment, each preprocessed token will be mapped to a 300-dimensional FastText vector.

[0050] Secondly is the domain knowledge weight: the TF-IKGF algorithm.

[0051] To measure the importance of a token in the current knowledge domain, the invention does not adopt the general TF-IDF, but uses a TF-IKGF (Term Frequency - Inverse Knowledge Graph Frequency) algorithm that is more suitable for the knowledge graph scenario. The TF-IDF algorithm was originally designed for linear document collections. When directly applied to a knowledge graph with a complex topological structure, the concept of "document" becomes blurred, making it difficult to fully utilize the structural information of the graph.

[0052] Therefore, the TF-IKGF algorithm adopted in this embodiment makes an adaptive improvement to this: Term Frequency (TF): This part is calculated in accordance with the standard TF and is used to measure the local importance of a term in a single entity name. The higher the frequency of a term appearing in an entity name, the higher its TF value. In entity A, "layering" appears twice, indicating that "layering" is a core concept describing this construction point. The larger its TF value, the better. The calculation formula is: ; in, It is a word element t In entity d The number of times the name appears, the denominator is the entity. d The total number of all morphemes in the name.

[0053] Inverse Knowledge Graph Frequency (IKGF): This is the core algorithm applied in this invention to adapt to the graph structure. In this embodiment, each triplet in the knowledge graph is treated as a mini-document because it encapsulates an atomic knowledge fact: if a term appears frequently in numerous triplets throughout the entire knowledge base (containing both construction and cost graphs), it indicates high prevalence and low distinguishability, and its importance should be correspondingly suppressed. The formula for calculating IKGF is as follows: ; in, df(t) represents the total number of triples in both the construction drawing and cost drawing datasets, and df(t) is the number of triples containing the term t. Logarithmic smoothing and the increment operation are used to prevent the denominator from being zero and to moderate the numerical scale. This design effectively suppresses the weight of terms that are very common in both datasets (such as "concrete" and "engineering"), while increasing the weight of keywords that only appear under specific topics (such as "slipform" and "list").

[0054] Finally, weighted vectorization is used to generate the initial entity embedding.

[0055] Combining general semantics and domain weights, the initial semantic embedding vector V for each entity node E is... E The weighted average is calculated using the following formula:

[0056] in, t It is a physical entity E morphemes in the name, It is its FastText word vector. α and β These are hyperparameters used to adjust the relative importance of TF and IKGF (in this embodiment, they are set experimentally). α=1000, β =0.4), ϵ is a small smoothing factor (such as 0.1) to avoid weights being zero.

[0057] like Figure 4 As shown, this method yields the initial semantic embedding vectors of entities A and B. and These two vectors retain the rich semantics of the general language model while highlighting the professional characteristics of the construction engineering field through TF-IKGF weights.

[0058] 3. Structural and semantic fusion: Graph Convolutional Networks (GCN) An entity's semantic information alone is insufficient; its neighboring nodes in the graph (i.e., its context structure) also contain rich alignment cues. This embodiment employs a Graph Convolutional Network (GCN) to fuse the initial semantic embeddings of nodes with the graph's topological information, generating more discriminative "enhanced node embeddings." GCN, as a cutting-edge graph representation learning technique, can learn feature representations of nodes in the graph end-to-end.

[0059] like Figure 5 As shown, the core idea of ​​GCN originates from graph signal processing and message passing mechanisms. It treats each node in the graph as a signal, aggregating neighborhood information through convolution operations on the graph to update the node's own representation. The propagation rule of a single-layer GCN can be formalized as: ; in, It is an adjacency matrix with self-connections, and I is the identity matrix. yes The degree matrix. H (l) It is the first l The node feature matrix of the layer, W (l) It is the trainable weight matrix of this layer. σ It is a non-linear activation function (such as ReLU). The essence of this formula is that the feature of a node in the next layer is a weighted average of its own features and the features of all its neighboring nodes in the current layer, with the weights determined by the graph structure, and then subjected to a linear transformation. W (l) () and nonlinear activation.

[0060] Secondly, adjacency matrix construction and normalization are performed. Based on the triplet relationships in the construction and cost source data, this embodiment constructs adjacency matrix A for the construction map and cost map respectively.SG and A ZJ If a relationship exists between node i and node j, then the corresponding position A(i,j) in the matrix is ​​1; otherwise, it is 0. Since the graph is directed, this embodiment uses a simplified normalization method suitable for directed graphs: ,in D It is the out-degree matrix of the nodes. This row normalization method can be understood as each node passing its "information" evenly to all its outgoing neighbors, effectively preventing the scale explosion or vanishing problem of node features during the aggregation process.

[0061] Finally, GCN-enhanced node embedding is performed, and this invention constructs a deep model containing multiple GCN layers. The initial entity embedding matrix E, calculated using mixed weights, is... initial As input to GCN (i.e., H) (0) And it is fed into a multi-layered GCN model along with the adjacency matrix.

[0062] In the first convolutional layer, the vector V of entity A EA Generate the first hidden representation H EA (1) Next, entity A in the second convolutional layer, H EA (1) This further aggregates the features of its two-hop neighbors. Finally, the matrix containing entities is propagated through multiple layers of GCN and subjected to nonlinear activation (ReLU) to output the final enhanced node embedding matrix E. final .

[0063] It is worth noting that as the number of GCN layers increases, an over-smoothing problem may occur, where the representations of all nodes tend to converge to the same value, thus losing discriminative power. To mitigate this problem, this invention employs a residual-like connection in the GCN model, accumulating the output of each convolutional layer into the final result. The resulting enhanced node embedding is V. EA ′ This design not only allows the model to utilize both shallow (focusing more on local neighborhoods) and deep (focusing more on global structure) information simultaneously, but also effectively preserves the individual characteristics of nodes (including entity A). Similarly, the enhanced embedding V of entity B... EB ′ It also incorporates the structural information from the cost map. After GCN processing, the vector of each node contains its original semantics as well as the structural and semantic information of its multi-hop neighbors, greatly enhancing its expressive power.

[0064] III. Cross-graph entity alignment likeFigure 1 As shown, in step S300, the present invention enters the cross-graph entity alignment stage, the core task of which is to accurately identify entity pairs that refer to the same concept or have strong correlations between two independent knowledge graphs. To balance the efficiency of large-scale graph processing with the accuracy of alignment results, this embodiment adopts a two-stage alignment strategy from coarse to fine. This strategy first uses an unsupervised clustering algorithm to embed all nodes in a vector space for rapid grouping, efficiently generating candidate matching sets; then, within each candidate set (cluster), fine-grained matching is performed through high-threshold similarity calculation, thereby filtering out highly reliable aligned entity pairs without relying on any manually labeled data.

[0065] 1. Semi-supervised clustering and candidate entity pair selection After obtaining high-quality node embeddings, this embodiment adopts a coarse-to-fine strategy to find aligned entities.

[0066] First, perform entity vector fusion with the same name (weakly supervised guidance).

[0067] Before performing fully unsupervised clustering, this embodiment introduces a simple heuristic rule as weak supervision. This rule is based on strong prior knowledge: if two entities have the exact same name in construction drawings and cost estimates, they are highly likely to refer to the same concept. Therefore, this embodiment first identifies all pairs of entities with the same name, and then arithmetically averages their respective enhanced embedding vectors generated by GCN, replacing their original vectors with this average vector. This step is an effective bootstrapping strategy, which forcibly narrows the distance between these highly aligned entities in the vector space, providing key "anchors" for subsequent clustering algorithms, enabling the clustering results to better group semantically related but not exactly identical entities together.

[0068] Secondly, candidate entity generation is based on clustering.

[0069] To avoid brute-force comparisons of all possible entity pairs (whose complexity is the product of the number of entities in the two graphs), this invention employs clustering as an efficient candidate generation (or "blocking") strategy, embedding all enhanced nodes of the two graphs after name-matching fusion (including V... EA ′ and V EB ′The data is then merged and partitioned within the same high-dimensional vector space using the unsupervised K-Means clustering algorithm. The goal of K-Means is to find K clusters and minimize the Within-Cluster Sum of Squares (WCSS) of all points within each cluster to the cluster centroid. ; in, C i It is the first i Clusters, x It is the embedding vector of a node in the cluster. μ i It is a cluster C i The algorithm iteratively performs two steps, "assignment" and "update," until convergence. Since the core semantics of entities A and B both point to "concrete pouring and vibration," their enhanced embedding vectors are spatially highly similar. Therefore, entities A and B have a very high probability of being assigned to the same cluster. This embodiment further assigns all nodes to... K Within each cluster, coarse-grained grouping of entities is achieved. Entities within the same cluster have high semantic and structural similarity, which greatly reduces the search space for subsequent precise matching. The choice of K value can be assisted by methods such as the Elbow Method; in this embodiment, a suitable K value was experimentally determined.

[0070] 2. Fine-grained alignment within clusters Within each cluster, the present invention performs a final step of fine alignment.

[0071] Cosine similarity calculation: For each cluster, including the clusters containing entities A and B, this embodiment calculates the cosine similarity between all nodes from the construction drawings and all nodes from the cost drawings within that cluster. The calculated similarity is... V EA ′ and V EB ′ Cosine similarity value: ; Cosine similarity measures the degree of proximity of two vectors in a direction. It is not sensitive to the absolute magnitude of the vectors and is very suitable for similarity judgment in high-dimensional semantic spaces.

[0072] Threshold filtering and matching pair generation: This invention sets a relatively high similarity threshold (e.g., 0.86 or higher). For each node in the construction drawing, within its cluster, it searches for the node in the cost drawing with the highest cosine similarity, and this similarity value is greater than the preset threshold. If found, this pair of entities ((construction node, cost node)) is output as a candidate alignment pair. That is, if Sim(E A E B If the value is higher than the preset threshold, the system determines that entity A and entity B constitute a valid aligned entity pair.

[0073] like Figure 6 The alignment strategy process shown above, through the above steps, automatically generates a high-confidence candidate alignment list including entity A and entity B.

[0074] IV. Validation of Knowledge Alignment and Integration with Knowledge Graphs like Figure 1 As shown, in step S400, the final stage of the method of the present invention is to finally confirm the knowledge alignment relationships generated in the preceding steps to construct a physically unified construction-cost integrated knowledge graph. To ensure the quality and practicality of the final result, this stage includes two core steps: a crucial human-machine collaborative verification process and automated database fusion implementation. First, a human reviews the candidate alignment list generated by the machine to eliminate potential incorrect matches and confirm the correct alignment relationships. Subsequently, as... Figure 7 The knowledge fusion process shown involves the automated fusion module performing atomic "create-link-delete" operations in the graph database based on this verified final list. This structured integration of independent construction knowledge nodes into the corresponding cost knowledge system completes the deep fusion of the two knowledge graphs.

[0075] 1. Manual verification While the aforementioned automated methods can efficiently discover a large number of potential entity alignments, 100% accuracy is difficult to guarantee due to the complexity of natural language and the specificity of domain knowledge. To ensure that the quality of the final fused knowledge graph meets production application standards, this invention incorporates a crucial human-in-the-loop verification step between automated alignment and database fusion: a human reviews the machine-generated candidate alignment list, including entities A and B, eliminating incorrect matches (e.g., entities that, although textually similar, actually refer to different construction stages or pricing rules), and confirming correct matches. This step ensures that the data input into the next fusion process is highly accurate. The final alignment list, after verification, is used to drive subsequent fusion operations.

[0076] This is the final step of the method of the present invention. The embodiment will be executed by an automated fusion module, which is responsible for physically fusing the verified alignment relationships in the graph database to ultimately form a unified knowledge graph that connects construction and cost.

[0077] 2. Establish database connection The fusion module first establishes a secure connection with the running Neo4j database instance.

[0078] 3. Traverse the alignment pairs and perform blending. The fusion module iterates through the manually validated set of entity pairs. For each aligned entity pair (containing entity A and entity B), the following atomic transaction is executed: Locating the parent node: In the cost estimate diagram, find the corresponding "parent" node based on its name. This node will serve as the input point for the merged knowledge.

[0079] Locating child nodes: In the construction drawings, find the corresponding "child" node based on its name.

[0080] Create a new node and migrate information: If both parent and child nodes are found, the system will create a new node in the database. The attributes of this new node (name, description, etc.) are completely copied from the construction node to be merged, such as entity A, but its label is specially set to GraphSGinZJ to identify its source and merging status.

[0081] Establish a new relationship: Between the parent node (cost node), such as entity B, and this newly created GraphSGinZJ node, establish a "contains" (or other custom) relationship. In this way, the previously independent construction knowledge is linked into the cost graph system as a detailed supplement to the cost knowledge.

[0082] Deleting old nodes: To avoid information redundancy and keep the graph clean, after successfully creating new nodes and relationships, the system will delete the original merged node, such as entity A, and all its associated relationships from the construction graph.

[0083] After execution, the previously independent construction point entity A was successfully and structurally integrated into the cost knowledge system, becoming a supplement to the specific technical details of the cost requirement entity B. This atomic operation of "create-link-delete" ensures the smooth transfer and integration of knowledge without disrupting the overall structure of the knowledge map.

[0084] 4. Establish top-level connections After all entity pairs have been merged, the merging module performs a final operation: creating a high-level association, such as SGcomplement (Construction Knowledge Supplement), between the root node of the construction drawing and the root node of the cost drawing. This establishes a connection between the two original knowledge systems at a macro level, allowing users to easily navigate to the knowledge of the other from the entry point of either drawing.

[0085] Through the above specific implementation methods, this invention fully demonstrates how to construct a unified knowledge graph that deeply integrates construction technology and engineering cost knowledge from scratch. This method effectively solves the problem of domain knowledge silos through innovative TF-IKGF algorithms, GCN-enhanced semantic structure representation, semi-supervised clustering alignment, and rigorous database fusion strategies. It provides a solid data foundation and technical path for realizing advanced applications in the construction engineering field, such as intelligent knowledge querying, intelligent question answering, cost prediction, and construction scheme optimization.

[0086] Example 2: like Figure 8 As shown, this embodiment provides a device for fusing construction drawings and cost estimates, including: The initial semantic embedding vector generation module 501 is used to generate an initial semantic embedding vector for each entity node in the construction drawing and cost drawing based on a pre-trained word vector model and combined with statistical features representing the weights of domain terms. The terminal node embedding vector generation module 502 is used to learn and generate a terminal node embedding vector that integrates structural information by using a graph neural network, combining the initial semantic embedding vector and the respective topological structures of the construction drawing and the cost drawing. The entity alignment module 503 is used to first cluster all the final node embedding vectors using an unsupervised clustering algorithm, and then identify aligned entity pairs from different graphs within each cluster based on similarity calculation. The fusion module 504 is used to merge construction drawings and cost drawings in the graph database based on aligned entity pairs, and establish new hierarchical relationships to generate an integrated knowledge graph.

[0087] The specific implementation of each module in this embodiment can be found in Embodiment 1 above, and will not be repeated here. It should be noted that the device provided in this embodiment is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure can be divided into different functional modules to complete all or part of the functions described above.

[0088] Example 3: like Figure 9As shown, this embodiment provides a system for integrating construction drawings and cost estimates, including a processor 602, a memory, an input device 603, a display 604, and a network interface 605 connected via a system bus 601. The processor 602 provides computing and control capabilities. The memory includes a non-volatile storage medium 606 and internal memory 607. The non-volatile storage medium 606 stores an operating system, computer programs, and a database. The internal memory 607 provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium 606. When the computer programs are executed by the processor 602, the aforementioned racing car assisted driving method is implemented.

[0089] Example 4: This embodiment provides a storage medium, which is a computer-readable storage medium, storing a computer program that, when executed by a processor, implements the above-described racing car assisted driving method.

[0090] The storage medium described in this embodiment can be a disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), USB flash drive, portable hard drive, etc.

[0091] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for fusing construction drawings and cost drawings, characterized in that: The method comprises the following steps, Based on the pre-trained word vector model, combined with the statistical characteristics representing the weight of the field terms, an initial semantic embedding vector is generated for each entity node in the construction graph and the cost graph; Using a graph neural network, combined with the initial semantic embedding vector and the topological structure of the construction graph and the cost graph, a final node embedding vector is learned and generated, which integrates the structure information; Performing entity alignment, first clustering all final node embedding vectors by unsupervised clustering algorithm, then identifying aligned entity pairs from different graphs within each cluster based on similarity calculation; Based on the aligned entity pairs, the construction graph and the cost graph are integrated in the graph database to establish a new hierarchical relationship to generate an integrated knowledge graph.

2. The fusion method of claim 1, wherein: The statistical characteristics representing the weight of the field terms are the combination of word frequency and inverse knowledge graph frequency.

3. The fusion method of claim 1, wherein: The graph neural network is a multi-layer graph convolutional network.

4. The fusion method of claim 1, wherein: The step of generating the final node embedding vector includes: taking the normalized adjacency matrix as input, and applying a nonlinear activation function after each convolutional layer.

5. The fusion method of claim 1, wherein: The aligned entity pair step also includes a preprocessing sub-step: identifying entities with the same name in the construction graph and the cost graph, and replacing their respective final node embedding vectors with a fusion vector, which is the arithmetic mean of the final node embedding vectors of the entities with the same name.

6. The fusion method of claim 1, wherein: The unsupervised clustering algorithm is the K-Means algorithm, and the similarity calculation within the cluster is the cosine similarity calculation.

7. The fusion method of claim 1, wherein: Integrating the construction graph and the cost graph in the graph database includes: creating a new node that inherits the entity node attributes in the construction graph; establishing a hierarchical relationship between the new node and the corresponding entity node in the cost graph; deleting the original entity nodes and their associated relationships in the construction graph.

8. A fusion device for implementing the fusion method according to any one of claims 1 to 7, characterized in that, It comprises: An initial semantic embedding vector generation module for generating an initial semantic embedding vector for each entity node in the construction graph and the cost graph based on a pre-trained word vector model combined with statistical characteristics representing the weight of the field terms; A final node embedding vector generation module for learning and generating a final node embedding vector that integrates structure information using a graph neural network combined with the initial semantic embedding vector and the topological structure of the construction graph and the cost graph; An entity alignment execution module for first clustering all final node embedding vectors by unsupervised clustering algorithm, then identifying aligned entity pairs from different graphs within each cluster based on similarity calculation; A fusion module for integrating the construction graph and the cost graph in the graph database based on the aligned entity pairs to establish a new hierarchical relationship to generate an integrated knowledge graph.

9. A system for fusing construction drawings and cost drawings, characterized by, It comprises a processor and a memory, and the memory stores a computer program, and the processor executes the computer program to realize the integration method of any one of claims 1-7.

10. A computer readable storage medium, characterized in that, The storage medium stores a computer program, and the computer program is executed by the processor to realize the integration method of any one of claims 1-7.

Images