Intelligent compliance review method and system based on power engineering cost knowledge graph

Abstract

The application discloses an intelligent compliance review method and system based on a power engineering cost knowledge graph, and the method comprises the following steps: acquiring power engineering cost file data, and mapping to generate a knowledge graph subgraph; performing matrix representation of the topological structure of the subgraph to construct a hybrid Laplace matrix; extracting cost values to construct a spatial domain cost graph signal vector, performing graph Fourier transform on the hybrid Laplace matrix to generate a frequency domain graph signal vector, and identifying suspected abnormal candidate nodes; performing ontology semantic reverse query to generate a vertex adaptive dynamic cutoff frequency, constructing an adaptive polynomial filter matrix, performing secondary filtering operation on the spatial domain cost graph signal vector, and outputting a final high-frequency residual graph signal vector; comparing the high-frequency residual energy with a cost fluctuation tolerance threshold to identify abnormal cost nodes, backtracking associated neighbor nodes, and outputting a review report. The application can effectively identify cross-node collaborative deviation phenomena and screen abnormal cost projects with structural cost deviation.

Description

Technical Field

[0001] This invention relates to the field of power engineering data processing technology, specifically to an intelligent compliance review method and system based on a power engineering cost knowledge graph. Background Technology

[0002] Cost review of power engineering projects (especially substation projects) is a key link in controlling the investment cost of power grid construction and standardizing the use of project funds. In the construction process of power engineering projects such as substations, cost documents cover a large number of materials, equipment, construction procedures and related cost details. The compliance review aims to verify whether the actual submitted cost documents meet the standards and specifications, eliminate unreasonable project quantities, overestimation and other abnormal deviations, and ensure the reasonable allocation and use of power engineering construction funds.

[0003] In existing power engineering business practices, cost compliance review mainly relies on reviewers to complete the task using information-based cost software. The computer system usually performs condition matching and numerical comparison on each independent item in the cost document based on pre-set price thresholds or cost engines. When the price or calculated quantity of a single item in the document exceeds the normal benchmark range, the system triggers an over-limit alarm and outputs abnormal entries to assist reviewers in subsequent business auditing work.

[0004] However, existing compliance review technologies suffer from the following objective shortcomings: inconsistent review standards and low efficiency: different reviewers may have different judgment criteria for the same project, resulting in a lack of consistency and comparability in review results. Faced with a large number of complex cost documents, manual review is time-consuming and labor-intensive, making it difficult to meet the need for rapid response. Furthermore, existing technologies primarily fragment the engineering cost system into isolated data units for independent review, lacking the ability to globally represent the topological relationships between engineering cost business entities, leading to incomplete problem identification (difficulty in capturing deep logical contradictions): relying on manual experience and discrete rule matching easily overlooks potential risk points. In actual complex cost businesses such as power substations, there are often deep logical intersections between engineering quantities and material prices, and some abnormal cost deviations exhibit a "structural dispersion" characteristic, meaning that the amount of unbalanced cost composition is dispersed across multiple seemingly compliant entities. In related projects, traditional discrete rule matching engines can only perform local basic data feature comparisons and cannot capture the cost ratio break states that are discretely distributed among multiple cost nodes. Ultimately, it is difficult to fully detect such hidden structural cost deviations and distinguish between objective reasonable fluctuations and abnormal deviations. Existing technologies cannot effectively distinguish between reasonable cost fluctuations and structurally abnormal cost fluctuations caused by objective engineering environments (such as topography, geology, meteorological conditions, special construction techniques, etc.), which is prone to false alarms and lacks interpretability and knowledge accumulation. Some existing automated auditing tools can only output "pass / fail" or a single alarm conclusion. There is no traceable logical deduction chain between the judgment result and the underlying data indicators, which cannot provide clear judgment basis and problem location. In addition, the auditing experience and business rules of experts are scattered in individuals or documents, making it difficult to form a reusable and iterative knowledge system. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent compliance review method and system based on a knowledge graph of power engineering cost. This solves the problems of existing engineering cost compliance review technologies, such as the easy failure of discrete rule matching engines and the lack of clear judgment criteria and problem localization when facing deep logical contradictions and cross-node structural cost deviations.

[0006] The first aspect of this invention provides an intelligent compliance review method based on a knowledge graph of power engineering cost, comprising the following steps: The process involves acquiring power engineering cost document data, mapping this data to a pre-constructed benchmark power engineering cost knowledge graph, and instantiating a knowledge graph subgraph. The subgraph's topological structure is then matrix-represented, constructing a hybrid Laplace matrix for each subgraph. Actual engineering cost values ​​corresponding to the power engineering cost document data are extracted to construct a spatial domain cost graph signal vector. The hybrid Laplace matrix is ​​used to perform a graph Fourier transform on the spatial domain cost graph signal vector to generate a frequency domain graph signal vector, which is then used to identify potential anomaly candidate nodes. Finally, the process is repeated in the original text. In the knowledge graph of power engineering cost, a reverse semantic query of the ontology is performed on suspected abnormal candidate nodes to generate vertex adaptive dynamic cutoff frequencies. An adaptive polynomial filter matrix is ​​constructed based on the vertex adaptive dynamic cutoff frequencies. The adaptive polynomial filter matrix is ​​used to perform a second-order filtering operation on the spatial domain cost graph signal vector to output the final high-frequency residual graph signal vector. The high-frequency residual energy of the cost project nodes is calculated based on the final high-frequency residual graph signal vector. Abnormal cost nodes are identified by comparing the high-frequency residual energy with the cost fluctuation tolerance threshold, and the associated neighboring nodes are traced back to output an intelligent compliance review report.

[0007] In practice, the process of constructing the hybrid Laplacian matrix includes: classifying the edges in the knowledge graph subgraph according to the type of ontology association; extracting the connection strength weights on the edges of specific associations to construct multi-dimensional weight matrices; combining the weight matrices with the degree matrices of the corresponding dimensions to calculate the multi-dimensional normalized sub-Laplacian matrix; retrieving the preset cost impact factor weight parameters as weighting coefficients; multiplying the multi-dimensional normalized sub-Laplacian matrix by the corresponding weighting coefficients; and then performing a linear combination operation of summing the corresponding matrix elements on the multiple matrices obtained by multiplication to aggregate and generate the hybrid Laplacian matrix.

[0008] For the frequency domain conversion process of the cost map signal, specifically, an orthogonal decomposition operation is performed on the mixed Laplace matrix, which is decomposed into the form of an orthogonal matrix, a diagonal matrix, and the transpose of the orthogonal matrix multiplied together. The transpose of the orthogonal matrix is ​​used as the left multiplication matrix, and the spatial domain cost map signal vector is used as the right multiplication vector to perform matrix multiplication operation, thereby generating the frequency domain map signal vector.

[0009] In the second filtering stage, the mixed Laplace matrix is ​​normalized and mapped to generate a normalized mixed Laplace matrix. A recursive operation is used to calculate the polynomial approximation basis matrix sequence of the normalized mixed Laplace matrix. The vertex adaptive dynamic cutoff frequency is substituted into the polynomial coefficient calculation formula to independently calculate the expansion coefficients. According to the row index order of the cost project nodes in the mixed Laplace matrix, the expansion coefficients of the same order of all cost project nodes are arranged to generate the coefficient diagonal matrix corresponding to each order of polynomial. Finally, for each order of polynomial, matrix multiplication and weighted summation operations are performed sequentially on the coefficient diagonal matrix, the polynomial approximation basis matrix of the corresponding order, and the spatial domain cost map signal vector to calculate and generate the final high-frequency residual map signal vector.

[0010] The innovative principle of this invention lies in introducing graph Fourier transform and frequency domain analysis theory from the field of digital signal processing into the logical reasoning structure of knowledge graphs. This invention transforms the knowledge graph from discrete graph logical relationships into an algebraic state space reflecting business logic constraints. The engineering cost value is defined as a continuous graph signal attached to the topological nodes. In a compliant engineering cost system, the cost ratio between strongly correlated cost project nodes remains smooth, corresponding to a low-frequency cost signal in the topological frequency space. However, structural cost deviation behavior will break the smoothness of the physical topology, causing a break in the local cost ratio, corresponding to a high-frequency cost signal in the topological frequency space. The system performs high-pass filtering through matrix algebra operations, so that the cross-node collaborative deviation phenomenon hidden on multiple nodes is directly captured in the form of high-frequency residual energy oscillation.

[0011] Furthermore, to avoid misjudging legitimate cost fluctuations caused by objective engineering environments as abnormal deviations, this invention establishes a feedback connection channel from the business semantic domain to the mathematical frequency domain. The system initiates a reverse query in the benchmark power engineering cost knowledge graph for suspected abnormal candidate nodes, extracts environmental factor ontology tags and construction attribute ontology tags, converts textual semantic tags into numerical environmental impact weight values ​​to calculate the comprehensive environmental impact coefficient, and generates an adaptive dynamic cutoff frequency for vertices based on the comprehensive environmental impact coefficient. For cost project nodes with complex objective engineering environments, the filtering cutoff frequency is locally relaxed. Based on the aforementioned feedback connection channel, the system solidifies the ontological semantic facts of the knowledge graph into dynamic truncation parameters of the graph filter, realizing the cross-constraint between the pure mathematical graph signal filtering model and the underlying business rule logic, and accurately locking down structurally non-compliant cost projects.

[0012] A second aspect of this invention provides an intelligent compliance review system based on a knowledge graph of power engineering cost, comprising: The data access module is configured to acquire power engineering cost documents. The graph management module is configured to map power engineering cost document data to a pre-built benchmark power engineering cost knowledge graph and instantiate and generate knowledge graph subgraphs. The graph signal processing module is configured to perform a matrix representation of the topological structure of the knowledge graph subgraph to construct a hybrid Laplace matrix, extract the actual engineering cost value corresponding to the power engineering cost document data to construct a spatial domain cost graph signal vector, perform a graph Fourier transform on the spatial domain cost graph signal vector using the hybrid Laplace matrix to generate a frequency domain graph signal vector, and identify suspected abnormal candidate nodes based on the frequency domain graph signal vector. The semantic feedback control module is configured to perform ontology semantic reverse query on suspected abnormal candidate nodes in the benchmark power engineering cost knowledge graph and generate vertex adaptive dynamic cutoff frequency. The graph signal processing module is also configured to construct an adaptive polynomial filter matrix based on the vertex adaptive dynamic cutoff frequency, and use the adaptive polynomial filter matrix to perform a second-order filtering operation on the spatial domain cost graph signal vector to output the final high-frequency residual graph signal vector. The review output module is configured to calculate the high-frequency residual energy of the cost project nodes based on the final high-frequency residual map signal vector, perform abnormal cost node identification based on the comparison of high-frequency residual energy with the cost fluctuation tolerance threshold, trace back the related neighbor nodes, and output an intelligent compliance review report.

[0013] This invention provides an intelligent compliance review method and system based on a knowledge graph of power engineering cost. It has the following beneficial effects: 1. This invention utilizes a hybrid Laplace matrix to perform a graphical Fourier transform on the spatial domain cost map signal vector to generate a frequency domain graph signal vector, and then performs filtering on the frequency domain graph signal vector to retain high-frequency residual frequency domain signals. This realizes the transformation of discrete graph logical relationships into continuous frequency features in algebraic state space. By using high-frequency residual frequency domain signals, it identifies deep structural deviations and logical contradictions distributed across multiple cost project nodes, effectively screening out abnormal cost projects with structural cost imbalances.

[0014] 2. This invention extracts environmental factors and construction attribute ontology tags by performing ontology semantic reverse query on suspected abnormal candidate nodes in the knowledge graph, generates vertex adaptive dynamic cutoff frequencies to construct an adaptive multinomial filter matrix, and then performs a second-order filtering operation on the spatial domain cost graph signal vector to output the final high-frequency residual graph signal vector. This realizes the conversion of graph semantic tags into numerical filter parameters, adjusts the filtering boundary for cost project nodes with complex objective engineering environments, and avoids misjudging legitimate cost fluctuations caused by engineering environments as abnormal deviations.

[0015] 3. This invention calculates high-frequency residual energy and compares it with the cost fluctuation tolerance threshold generated based on variance analysis of historical compliant cost projects. For confirmed abnormal cost nodes, it performs a traversal operation along the ontology association edges to trace back the associated neighbor nodes. It then concatenates data indicators with ontology labels to generate a structured logical derivation chain. This enables the reconstruction of the split engineering cost structure within the graph topology space, establishes the correspondence between underlying data indicators and review results, and provides an interpretable basis containing computational data and engineering semantic facts for the final assembled intelligent compliance review report. Attached Figure Description

[0016] Figure 1 This is a flowchart of the intelligent compliance review method based on power engineering cost knowledge graph of the present invention; Figure 2 This is an architecture diagram of the intelligent compliance review system based on the knowledge graph of power engineering cost, as presented in this invention. Detailed Implementation

[0017] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] See attached document Figure 2 , Figure 2 This is an architecture diagram of an intelligent compliance review system based on a knowledge graph of power engineering cost according to an embodiment of the present invention. The intelligent compliance review system based on the knowledge graph of power engineering cost is deployed in a computing cluster or cloud server environment and includes a data access module, a graph management module, a graph signal processing module, a semantic feedback control module, and a review output module.

[0019] The data access module includes a relational database interface, a spreadsheet parsing engine, and an unstructured text parsing unit based on a named entity recognition model. This module receives and structures the submitted power engineering cost documents. The graph management module uses a distributed graph database platform to store the benchmark power engineering cost knowledge graph, responsible for storing and maintaining engineering materials, process nodes, and the topological connections between nodes. The graph signal processing module embeds a matrix operation unit and an eigenvalue decomposition acceleration operator, specifically handling graph Fourier transforms and filtering operations for high-dimensional sparse matrices. The semantic feedback control module integrates a graph query language parser, configured to initiate reverse queries to the graph database for environmental attributes and construction objective condition ontology tags, and converts the retrieved textual semantic tags into numerical adjustment parameters for controlling the filter. The review output module includes a numerical comparator and a report generation engine, used to compare high-frequency residual energy with preset thresholds and output the final compliance review report.

[0020] See attached document Figure 1 , Figure 1 This is a flowchart of an intelligent compliance review method based on a knowledge graph of power engineering cost according to an embodiment of the present invention. The intelligent compliance review method is executed by the processing device in the above-mentioned intelligent compliance review system, and specifically includes the following steps: S1. Cost document access and benchmark graph subgraph instantiation: The system obtains the actual submitted power engineering cost document data through the data access module. The cost document data covers specific engineering quantities, material unit prices, and construction-related costs. The system maps the parsed cost data to a pre-constructed benchmark power engineering cost knowledge graph, extracts a knowledge graph subgraph containing all relevant materials, processes, and relationships of this project. The benchmark power engineering cost knowledge graph uses nodes to represent specific cost project entities and edges to represent the ontology relationships between nodes, assigning quantified connection strength weights to the edges. The system thus completes the graph structure instantiation for this review project.

[0021] S2, Graph Topology Matrixing and Normalized Laplace Matrix Construction: The system uses the graph signal processing module to perform matrixing representation of the topological structure of the extracted subgraphs, calculates the degree matrix and weight matrix of the subgraphs respectively, and combines the degree matrix and weight matrix to construct a normalized Laplace matrix, and further aggregates them into a hybrid Laplace matrix. In this way, the discrete graph logical relationship is transformed into an algebraic state space that can perform continuous signal frequency analysis, thereby providing a low-level structural conservation mapping at the physical structure and business association levels.

[0022] S3, Cost Mapping and Graph Fourier Transform: The system defines the actual total cost data or cost deviation data of each node in the submitted documents as a continuous graph signal vector, and maps the continuous graph signal vector to the corresponding nodes of the subgraph. The system performs eigenvalue decomposition on the hybrid Laplace matrix to extract eigenvalues ​​reflecting the topological spatial characteristics and the corresponding eigenvector orthogonal matrix. The system uses the eigenvector orthogonal matrix as the spatial basis of the graph Fourier transform, performs matrix multiplication, and transforms the cost map signal in the spatial domain to the frequency domain to generate a frequency domain graph signal.

[0023] S4, semantic feedback-driven dynamic frequency band adaptive local filtering: The system configures an initial global high-pass filter function to perform preliminary filtering on the frequency domain graph signal to extract high-frequency residual signals, and performs graph Fourier inverse transform to map the signal back to the spatial domain. This identifies a set of suspected abnormal candidate nodes with residual energy exceeding the limit. The system triggers the semantic feedback control module to initiate a reverse query in the knowledge graph for the aforementioned suspected abnormal candidate nodes, extracting the environmental factors and construction attribute ontology tags bound to the corresponding nodes. The system constructs a numerical mapping function to convert the queried ontology tags into cutoff frequency offsets specific to the corresponding local area. Based on the generated frequency offsets, the system constructs a vertex adaptive polynomial filter matrix, relaxes the cutoff frequency for nodes with extreme objective environmental tags, performs secondary filtering with frequency offset adjustment, and outputs the final high-frequency residual graph signal.

[0024] S5, Anomaly Location and Explainable Review Report Output: The review output module calculates the residual energy of each node in signal transmission based on the final high-frequency residual graph signal from the secondary filtering output. The system sets the expected cost fluctuation tolerance threshold. When the residual energy of a node is greater than the cost fluctuation tolerance threshold, the system determines the corresponding node as a high-confidence abnormal cost node. Based on the topological edge direction in the subgraph, the system traces back the associated neighboring nodes that caused the high-frequency oscillation of the abnormal cost node. The system integrates the abnormal cost node data, the associated set data of structural deviations, and the extracted high objective environmental factor data to output the compliance review analysis results.

[0025] Step S1, cost document integration and benchmark map sub-map instantiation, specifically includes the following sub-steps: S101, Cost Data Access and Parsing: The system establishes a data connection channel for the power engineering cost system through the data access module. The data access module receives cost documents submitted for review from multiple sources. The received file types include detailed cost documents exported from relational databases, quantities of engineering work in Building Information Modeling (BIM) format using Extensible Markup Language (XML), and construction contract text in Portable Document format. For structured data in spreadsheet format, the system calls the spreadsheet parsing engine to extract material names, specifications, quantities, and comprehensive unit prices from the cells. For unstructured text data, the system uses an unstructured text parsing unit built based on a named entity recognition model to perform text segmentation and entity boundary recognition operations, extracting a set of feature entities with engineering business attributes.

[0026] S102, Entity Vectorization Encoding and Alignment Mapping in the Cost Graph: The system acquires the extracted feature entity set and inputs it into a pre-trained domain word vector encoding model. The domain word vector encoding model extracts the character-level features and contextual semantic features of the feature entities, generating fixed-length entity semantic feature vectors. The system performs vector space comparison between the entity semantic feature vectors and the pre-stored benchmark node feature vectors in the benchmark power engineering cost knowledge graph. The system calculates the cosine similarity between the entity semantic feature vectors and the benchmark node feature vectors. When the cosine similarity is greater than the preset alignment threshold, the system determines that the feature entity and the corresponding benchmark graph node are successfully matched, completing the ontology mapping from the actual engineering cost item to the physical node of the standard knowledge graph, and aggregating all successfully matched benchmark graph nodes into a subgraph node set.

[0027] S103: Based on the extraction of relationships between mapped nodes and the instantiation of knowledge graph subgraphs, after completing the alignment mapping of all cost document entities, the system activates the successfully matched benchmark graph nodes in the distributed graph database. Starting from the activated benchmark graph nodes, the system performs a depth-first traversal operation in the benchmark power engineering cost knowledge graph. The system extracts the ontology relationship edges between the activated nodes. The ontology relationship edges include physical connection edges reflecting the connection characteristics of engineering materials and cost allocation edges reflecting the composition of financial quotas. The system aggregates the extracted ontology relationship edges into a subgraph edge set. The system simultaneously extracts the pre-stored connection strength weights on the relationship edges and combines the connection strength weights into a subgraph weight matrix. The system integrates the subgraph node set, the subgraph edge set, and the subgraph weight matrix to instantiate and generate a knowledge graph subgraph reflecting the topological structure characteristics of the submitted project. The knowledge graph subgraph represents the local business relationship logic of the submitted project in a computer-readable data structure form, providing underlying structural support for the graph signal processing module.

[0028] Step S2, the matrix transformation of the graph topology and the construction of the normalized Laplacian matrix, specifically includes the following sub-steps: S201. Based on heterogeneous ontology association relationships, a multidimensional weight matrix is ​​extracted. The system traverses the subgraph edge set generated in the previous steps and classifies the edges in the subgraph edge set according to the ontology association relationship type. The ontology association relationship types are divided into physical connection relationship, cost sharing relationship, and material substitution relationship. For each ontology association relationship type, the system extracts the connection strength weight on the specific association relationship edge and constructs the physical connection weight matrix, cost sharing weight matrix, and material substitution weight matrix respectively. The system calculates the node degree for each dimension of the weight matrix. The node degree is obtained by performing a numerical summation operation on the connection strength weight of the target node in the same dimension. The system constructs degree matrices corresponding to different association relationship types. The degree matrix is ​​set as a diagonal matrix. The main diagonal elements of the degree matrix are the sum of the degree values ​​of the corresponding nodes, and the off-diagonal elements of the degree matrix are uniformly set to zero.

[0029] S202, Constructing the Normalized Sub-Laplace Matrix: The system utilizes the matrix operation unit in the graph signal processing module to perform matrix algebra operations on the physical connection weight matrix, cost allocation weight matrix, and material substitution weight matrix, respectively, in conjunction with the degree matrix of the corresponding dimension. Based on the normalization operation logic, the system calculates the physical connection sub-Laplace matrix, cost allocation sub-Laplace matrix, and material substitution sub-Laplace matrix corresponding to different ontology association types. The system uses the following basic formula to calculate the normalized Laplace matrix: In the formula, Represents the normalized Laplace matrix. This represents the identity matrix with the same dimensions as the degree matrix. Degree matrix, Let represent the diagonal matrix obtained by inverting the degree matrix and taking the square root. The system represents the weight matrix and performs independent calculations of the sub-Laplace matrices of each dimension by repeatedly calling the basic formula.

[0030] S203, based on heterogeneous semantic feature aggregation, generates a hybrid Laplace matrix. The system retrieves the cost impact factor weight parameters preset in the benchmark power engineering cost knowledge graph. The cost impact factor weight parameters are used to characterize the degree of influence of different ontology relationship types in the formation of the total project cost. The system configures the cost impact factor weight parameters as weighting coefficients and performs tensor weighted collapse operation on the physical connection sub-Laplace matrix, cost allocation sub-Laplace matrix, and material substitution sub-Laplace matrix. The specific collapse operation logic is as follows: the system multiplies the sub-Laplace matrix of each dimension by the corresponding weighting coefficient, and then performs a linear combination operation of summing the corresponding matrix elements on the multiple matrices obtained by multiplication to generate a hybrid Laplace matrix that maps the semantic features of specific engineering business. The hybrid Laplace matrix completes the transformation of discrete graph logical relations into continuous algebraic state space, so that when the intelligent compliance review system performs graph signal frequency analysis, it is subject to the cross constraints of engineering physical topology, financial allocation logic, and material substitution logic, blocking structural cost deviation and item-buying behavior that circumvents single physical dimension review at the underlying data structure level.

[0031] Step S3, cost map signal mapping and graph Fourier transform, specifically includes the following sub-steps: S301, Extract node cost feature values. The system traverses the node set in the knowledge graph subgraph. For each cost item node in the node set, the system extracts the corresponding actual project cost value from the actual submitted power engineering cost documents. The actual project cost value includes the actual total submitted cost obtained by multiplying the comprehensive unit price of material settlement by the actual reported project quantity. The system simultaneously retrieves the standard benchmark cost pre-bound to the corresponding cost item node from the benchmark power engineering cost knowledge graph. The standard benchmark cost serves as a benchmark comparison value for measuring cost fluctuations.

[0032] S302, calculate the cost deviation rate signal scalar. The system uses the actual submitted total cost minus the standard benchmark cost to obtain the absolute cost difference. The system divides the absolute cost difference by the standard benchmark cost to calculate the cost deviation rate value. The system configures the cost deviation rate value as the graph signal scalar of the corresponding cost item node. By using the cost deviation rate value to replace the absolute cost value to construct the graph signal, the system eliminates the problem of graph signal distribution imbalance caused by excessive differences in the absolute value of materials, so that the graph signal truly reflects the abnormal fluctuation state of the cost data.

[0033] S303, Construct the spatial domain cost graph signal vector. The system reads the hybrid Laplacian matrix generated in the previous steps, extracts the row index order of the nodes in the hybrid Laplacian matrix, and performs an alignment operation on all graph signal scalars in the node set according to the row index order. The system concatenates all the aligned graph signal scalars to generate a one-dimensional continuous signal column vector. The one-dimensional continuous signal column vector constitutes the spatial domain cost graph signal vector. Each dimension element in the spatial domain cost graph signal vector forms an algebraic mapping relationship with the spatial topological position of the nodes in the hybrid Laplacian matrix, providing the input data structure for the graph signal processing module to perform frequency domain conversion operations.

[0034] S304, Eigenvalue Decomposition of the Mixed Laplacian Matrix: The system inputs the mixed Laplacian matrix into the eigenvalue decomposition acceleration operator within the graph signal processing module. The eigenvalue decomposition acceleration operator performs orthogonal decomposition operations on the mixed Laplacian matrix, decomposing it into the mathematical form of an orthogonal matrix, a diagonal matrix, and the product of the transpose of the orthogonal matrix. The elements on the diagonal of the diagonal matrix are composed of the eigenvalues ​​of the mixed Laplacian matrix, arranged in ascending order of value, and the eigenvalues ​​correspond to graph frequencies in the topological space. The orthogonal matrix is ​​composed of the eigenvectors corresponding to the eigenvalues, and each column vector in the orthogonal matrix corresponds to an independent graph frequency component.

[0035] S305, Frequency domain transformation of cost map signal: The system extracts the transpose of the orthogonal matrix and the spatial domain cost map signal vector constructed in the previous steps. The graph signal processing module uses the transpose of the orthogonal matrix as the left multiplication matrix and the spatial domain cost map signal vector in the form of a one-dimensional continuous signal column vector as the right multiplication vector, and performs matrix multiplication operation. The output result of the matrix multiplication operation is defined as the frequency domain graph signal vector. The system uses matrix multiplication operation to complete the graph Fourier transform of the cost map signal vector.

[0036] S306, Frequency Domain Component Business Logic Mapping: The system divides the frequency spectrum of elements in the frequency domain graph signal vector according to the magnitude of the eigenvalues. The system maps the frequency domain graph signal components corresponding to the eigenvalues ​​with smaller values ​​to low-frequency cost signals. The low-frequency cost signals indicate that the cost ratio between strongly correlated cost project nodes remains smooth and conforms to the physical ratio specified in the historical standard cost quota. The system maps the frequency domain graph signal components corresponding to the eigenvalues ​​with larger values ​​to high-frequency cost signals. The high-frequency cost signals indicate that the cost ratio between strongly correlated cost project nodes breaks or changes abruptly. The system uses graph Fourier transform to convert the discrete local cost deviation values ​​distributed in different spatial nodes into continuous frequency domain components in the global topological frequency space. Based on the continuous frequency domain components in the topological frequency space, the system enables the structural cost deviation fraud behavior hidden in multiple nodes to be directly captured in the frequency domain in the form of high-frequency energy oscillation.

[0037] Step S4, semantic feedback-driven dynamic frequency band adaptive local filtering, specifically includes the following sub-steps: S401, Set the initial global high-pass filter. The system configures the initial global high-pass filter function and sets the initial global cutoff frequency in the initial global high-pass filter function. The system inputs the diagonal matrix composed of eigenvalues ​​into the initial global high-pass filter function. For eigenvalues ​​below the initial global cutoff frequency, the initial global high-pass filter function outputs the first coefficient value, which is configured to attenuate low-frequency cost signals. For eigenvalues ​​above the initial global cutoff frequency, the initial global high-pass filter function outputs the second coefficient value, which is configured to retain high-frequency cost signals. The system arranges the first coefficient value and the second coefficient value according to the position of the diagonal elements of the diagonal matrix to construct the initial global filter diagonal matrix.

[0038] S402, extract the high-frequency residual frequency domain signal. The graph signal processing module extracts the frequency domain graph signal vector generated in the previous steps. The graph signal processing module uses the initial global filtering diagonal matrix as the left multiplication matrix and the frequency domain graph signal vector as the right multiplication vector to perform matrix multiplication. The output result of the matrix multiplication operation is defined as the high-frequency residual frequency domain signal vector. The system uses matrix multiplication to filter out low-frequency components that reflect normal engineering business logic and retains high-frequency components caused by structural cost camouflage in the topology frequency space.

[0039] S403, Execute the graph Fourier inverse transform. The system extracts the orthogonal matrix obtained by eigenvalue decomposition in the previous steps. The graph signal processing module uses the orthogonal matrix as the left multiplication matrix and the high-frequency residual frequency domain signal vector as the right multiplication vector to perform matrix multiplication. The matrix multiplication operation reverses the mapping of the high-frequency residual signal from the topological frequency space back to the discrete physical cost node space. The output result of the matrix multiplication operation is defined as the initial spatial domain residual signal vector.

[0040] S404, Identifying Suspected Anomaly Candidate Nodes: The system traverses all component elements in the initial spatial domain residual signal vector, calculates the absolute value of each component element, configures the initial residual discrimination threshold in the review output module, and compares the absolute value with the initial residual discrimination threshold sequentially. When the absolute value of a component element is greater than the initial residual discrimination threshold, the system determines that the cost project node corresponding to the component element has high-frequency energy abnormal oscillation. The system extracts and aggregates the cost project nodes with high-frequency energy abnormal oscillation into a set of suspected anomaly candidate nodes. The set of suspected anomaly candidate nodes provides the target query range for the semantic feedback control module.

[0041] S405, Execute the target node ontology semantic reverse query. For each cost project node in the suspected abnormal candidate node set, the system extracts the unique entity identifier of the cost project node. The system concatenates the unique entity identifier into the preset graph database query template to generate the target graph query statement. The semantic feedback control module sends the target graph query statement to the distributed graph database platform. Based on the target graph query statement, the distributed graph database platform searches for attribute nodes that have direct edge connections with the current cost project node in the benchmark power engineering cost knowledge graph. The system extracts the environmental factor ontology tags and construction attribute ontology tags recorded in the attribute nodes. The environmental factor ontology tags include terrain and geological state tags and meteorological condition tags. The construction attribute ontology tags include construction operation efficiency reduction tags and special efficiency reduction process requirement tags. By generating standardized graph query statements, the system obtains the status data of the cost project node in the real engineering environment from the attribute dimension of the knowledge graph.

[0042] S406, Mapping Calculation of Comprehensive Environmental Impact Coefficient: The system has a pre-set semantic penalty mapping dictionary in the semantic feedback control module. The semantic penalty mapping dictionary establishes a deterministic mapping relationship between various ontology tag strings and environmental impact weight values. The system queries the semantic penalty mapping dictionary and converts the extracted environmental factor ontology tags and construction attribute ontology tags into numerical environmental impact weight values. When a single cost project node is associated with multiple ontology tags, the system performs a numerical accumulation operation on the multiple environmental impact weight values ​​corresponding to the multiple ontology tags to calculate the comprehensive environmental impact coefficient of the current cost project node. The system converts discrete text-type semantic tags into aggregated numerical adjustment parameters through numerical accumulation operations, establishing a feedback connection channel from the business semantic domain to the mathematical frequency domain.

[0043] S407, Generate Vertex Adaptive Dynamic Cutoff Frequency: The system retrieves the initial global cutoff frequency set by the initial global high-pass filter. The system multiplies the comprehensive environmental influence coefficient with the system's preset frequency offset adjustment constant to calculate the cutoff frequency offset of the current cost project node. The system adds the initial global cutoff frequency and the cutoff frequency offset to generate the dynamic cutoff frequency of the current cost project node. The system configures the dynamic cutoff frequency as the cutoff parameter for constructing the vertex adaptive local filter in subsequent steps. By injecting the cutoff frequency offset, the system moves the filter cutoff boundary towards the high-frequency direction for topological regions with complex engineering environment labels, relaxing the frequency review scale of specific cost project nodes and avoiding misjudging high-frequency oscillations caused by the engineering environment as structural abnormal deviations.

[0044] S408, construct Chebyshev polynomial approximation basis, the system extracts the mixed Laplace matrix and the spatial domain cost map signal vector constructed in the previous steps, the system sets the expansion order of the Chebyshev polynomial, the system normalizes the mixed Laplace matrix to a specified interval based on the largest eigenvalue of the mixed Laplace matrix, generates a normalized mixed Laplace matrix, the system uses recursive operation to calculate each basis matrix of the normalized mixed Laplace matrix, and generates a polynomial approximation basis matrix sequence covering low to high order.

[0045] S409, Fusion Vertex Adaptive Dynamic Cutoff Frequency: The system obtains the dynamic cutoff frequency generated for each cost project node in the aforementioned steps. The system substitutes the dynamic cutoff frequency into the polynomial coefficient calculation formula of the Chebyshev polynomial filter and independently calculates the Chebyshev polynomial expansion coefficients for each cost project node. Based on the row index order of the cost project node in the mixed Laplace matrix, the system arranges the Chebyshev polynomial expansion coefficients of the same order of all cost project nodes into a diagonal matrix, generating the coefficient diagonal matrix corresponding to each order polynomial. The coefficient diagonal matrix solidifies the local semantic feedback parameters of each node into a matrix operation structure.

[0046] S410 performs a second-order spatial-frequency domain joint filtering. The system extracts the polynomial approximation basis matrix sequence, the coefficient diagonal matrix, and the spatial domain cost map signal vector. The system uses the following vertex adaptive Chebyshev polynomial second-order filtering formula to calculate the final high-frequency residual signal: In the formula, This represents the final high-frequency residual image signal vector output after adaptive secondary filtering. Denotes the order of the Chebyshev polynomial expansion. Indicates by the first The coefficient diagonal matrix formed by expanding all the nodal coefficients of a polynomial of order 1. The first normalized mixture Laplace matrix represents the first normalized mixture matrix. Chebyshev polynomial matrix of order, The input spatial domain cost map signal vector is represented by the system. By executing the vertex adaptive Chebyshev polynomial quadratic filtering formula, the system dynamically relaxes the spectral boundary for local nodes with complex engineering environment labels, filters out high-frequency fluctuations caused by legitimate engineering environments, and outputs a high-frequency residual signal that truly reflects structural anomaly deviations.

[0047] Step S5, anomaly location and output of an interpretable review report, specifically includes the following sub-steps: S501, calculate the high-frequency residual energy of the node. The review output module extracts the final high-frequency residual map signal vector from the secondary filtering output of the previous steps. The system traverses the final high-frequency residual map signal vector and extracts the component elements corresponding to each cost project node. The system performs absolute value calculation for each component element and squares the absolute value calculation result to obtain the high-frequency residual energy of the corresponding cost project node. The system amplifies the numerical difference of abnormal high-frequency residual components through the square operation, and converts the residual signal scalar in a continuously fluctuating state into an energy parameter that measures the deviation of a single cost project node.

[0048] S502, Configure the cost fluctuation tolerance threshold. The system retrieves historical standard review archive data from the data access module. The system filters out historical compliant cost projects of the same type as the project under review from the historical standard review archive data. The system calculates the benchmark residual energy of the same type of historical compliant cost projects according to the aforementioned normalized Laplace matrix construction steps and graphical Fourier transform steps. The system aggregates the benchmark residual energy of all the same type of historical compliant cost projects, performs variance analysis using numerical calculation software, and calculates the statistical standard deviation of the benchmark residual energy. The system multiplies the statistical standard deviation by the system's preset proportional constant to generate the cost fluctuation tolerance threshold corresponding to the project under review. By introducing historical standard review archive data to perform variance analysis, the system replaces the fixed values ​​set by manual experience and establishes a numerical comparison benchmark that strictly conforms to the rules of power engineering cost quotas.

[0049] S503, Execute Abnormal Cost Node Identification. The system inputs the high-frequency residual energy of all cost project nodes and the cost fluctuation tolerance threshold into the numerical comparator inside the review output module. The numerical comparator performs numerical comparison calculations. When the high-frequency residual energy of a specific cost project node is less than or equal to the cost fluctuation tolerance threshold, the review output module determines that the specific cost project node is within the compliant fluctuation range. When the high-frequency residual energy of a specific cost project node is greater than the cost fluctuation tolerance threshold, the review output module determines that the specific cost project node is a high-confidence abnormal cost node. The system extracts the unique entity identifier of the high-confidence abnormal cost node and aggregates it to generate a list of high-confidence abnormal cost nodes. The list of high-confidence abnormal cost nodes excludes cost increases caused by changes in the objective environment and accurately identifies illegal cost projects that disrupt the smoothness of the physical topology by splitting cost document entries.

[0050] S504, Topology Backtracking and Abnormal Item Location: The system extracts unique entity identifiers from the list of high-confidence abnormal cost nodes. In the benchmark power engineering cost knowledge graph, the system locates the abnormal source node corresponding to the unique entity identifier. The system performs a breadth-first traversal operation along the ontology association edges of the abnormal source node. The system sets a topology backtracking hop count threshold and extracts neighbor nodes whose topological distance to the abnormal source node is within the threshold range. The system calculates the actual ratio of the actual submitted total cost of the abnormal source node to the actual submitted total cost of its neighbor nodes. The system extracts the pre-stored benchmark cost ratios for the abnormal source node and its corresponding neighbor nodes in the benchmark power engineering cost knowledge graph. The system compares the actual ratio with the benchmark cost ratio. When the actual ratio deviates from the benchmark cost ratio by more than the system tolerance range, the system determines that the neighbor node causing the deviation is an abnormal item node. Through breadth-first traversal and cost ratio comparison operations, the system reconstructs the split engineering cost association structure within the physical space of the knowledge graph.

[0051] S505, Review data aggregation and construction of interpretable logical chains. The review output module aggregates the cost deviation rate value of the abnormal source node, the high-frequency residual energy of the abnormal source node, the cost data of the associated abnormal package node, and the objective environmental factor ontology label extracted in the previous steps. The review output module maps the high-frequency residual energy value of the abnormal source node to the cost over-limit alarm judgment item. The review output module maps the actual ratio of the abnormal source node to the associated abnormal package node to the topology ratio breakage data item. The review output module maps the objective environmental factor ontology label to the semantic environment description item. The review output module performs data text splicing on the cost over-limit alarm judgment item, the topology ratio breakage data item, and the semantic environment description item according to the preset audit logic rules to generate a structured logical deduction chain. The structured logical deduction chain establishes the correspondence between the underlying data indicators and the high-level abnormal results.

[0052] S506, an interpretable review report assembly output, includes a report generation engine that retrieves a preset compliance review electronic report template. The engine aggregates the review data and structured logical derivation chain, filling them into the corresponding text fields of the compliance review electronic report template according to the template field mapping rules. The engine extracts the graphical coordinates and associated edge data of the anomaly source node and related anomaly set nodes, calls the underlying graphical rendering interface to generate a local anomaly topology visualization component, embeds this component into the compliance review electronic report template, and outputs the final intelligent compliance review report. This intelligent compliance review report provides compliance review evidence including underlying computational data, algorithmic judgment logic, and engineering semantic facts.

[0053] For unstructured text parsing, knowledge graph entity cosine similarity calculation, graph database language query operations, and report data rendering and export operations in multi-source cost data access, those skilled in the art can use conventional natural language recognition algorithm models, graph database native query statement systems, and front-end visualization page rendering frameworks to implement these operations. The underlying data interaction code and rendering logic are well-known technologies in this field and will not be elaborated here.

[0054] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent compliance review method based on a knowledge graph of power engineering cost, characterized in that, Includes the following steps: Obtain power engineering cost document data, map the power engineering cost document data to a pre-constructed benchmark power engineering cost knowledge graph, and instantiate and generate a knowledge graph subgraph; The knowledge graph subgraphs are represented by matrix representation of their topological structure, and a hybrid Laplace matrix corresponding to the knowledge graph subgraphs is constructed. Extract the actual project cost value corresponding to the power engineering cost document data to construct a spatial domain cost map signal vector; The hybrid Laplace matrix is ​​used to perform a graphical Fourier transform on the spatial domain cost map signal vector to generate a frequency domain graphical signal vector; And based on the frequency domain graph signal vector, suspected abnormal candidate nodes are identified; In the benchmark power engineering cost knowledge graph, perform an ontology semantic reverse query on the suspected abnormal candidate nodes to generate vertex adaptive dynamic cutoff frequency; An adaptive polynomial filter matrix is ​​constructed based on the vertex adaptive dynamic cutoff frequency. The adaptive polynomial filter matrix is ​​used to perform a second-order filtering operation on the spatial domain cost map signal vector to output the final high-frequency residual map signal vector. Based on the final high-frequency residual map signal vector, the high-frequency residual energy of the cost project node is calculated. Based on the comparison of the high-frequency residual energy with the cost fluctuation tolerance threshold, abnormal cost nodes are identified, and related neighbor nodes are traced back to output an intelligent compliance review report.

2. The intelligent compliance review method based on power engineering cost knowledge graph as described in claim 1, characterized in that, The power engineering cost document data is mapped to a pre-constructed benchmark power engineering cost knowledge graph, and knowledge graph subgraphs are instantiated and generated, including: The power engineering cost document data is parsed and entity features are extracted to generate a feature entity set, which is then mapped to the benchmark power engineering cost knowledge graph. When the ontology mapping is complete, the successfully matched baseline graph nodes are aggregated into a set of subgraph nodes; Starting from the activated baseline graph node, perform graph traversal operation, extract the ontology association edges and aggregate them into a subgraph edge set, and simultaneously extract the connection strength weight combination on the ontology association edges into a subgraph weight matrix; The knowledge graph subgraph is instantiated by integrating the subgraph node set, the subgraph edge set, and the subgraph weight matrix.

3. The intelligent compliance review method based on power engineering cost knowledge graph as described in claim 1, characterized in that, The step of performing a matrix representation of the topological structure of the knowledge graph subgraph and constructing a hybrid Laplacian matrix corresponding to the knowledge graph subgraph includes: The edges in the knowledge graph subgraph are classified according to the ontology association type, and the connection strength weights on the edges of specific associations are extracted to construct multi-dimensional weight matrices. The multi-dimensional normalized sub-Laplace matrix is ​​obtained by combining the weight matrix with the degree matrix of the corresponding dimension; The preset cost impact factor weight parameters are retrieved as weighting coefficients. The normalized sub-Laplace matrix of the multidimensional matrix is ​​multiplied by the corresponding weighting coefficients. Then, the multiple matrices obtained by multiplication are combined by performing a linear combination operation of summing the corresponding matrix elements to generate the mixed Laplace matrix.

4. The intelligent compliance review method based on power engineering cost knowledge graph as described in claim 1, characterized in that, The step of extracting the actual project cost value corresponding to the power engineering cost document data to construct a spatial domain cost map signal vector includes: Calculate the cost deviation rate between the actual project cost and the standard benchmark cost, and configure the cost deviation rate as a graphical signal scalar. Extract the row index order of the nodes in the hybrid Laplacian matrix, perform alignment and arrangement operations on all the graph signal scalars according to the row index order, and splice them to generate the spatial domain cost graph signal vector in the form of a one-dimensional continuous signal column vector.

5. The intelligent compliance review method based on power engineering cost knowledge graph as described in claim 1, characterized in that, The step of performing initial global filtering based on the frequency domain graph signal vector to identify suspected abnormal candidate nodes includes performing a graph Fourier transform on the spatial domain cost map signal vector using the mixed Laplace matrix to generate a frequency domain graph signal vector, and identifying suspected abnormal candidate nodes based on the frequency domain graph signal vector, including: Perform an orthogonal decomposition operation on the mixed Laplacian matrix to decompose the mixed Laplacian matrix into the form of an orthogonal matrix, a diagonal matrix, and the product of the transpose of the orthogonal matrix, wherein the diagonal matrix is ​​composed of eigenvalues; The transpose of the orthogonal matrix is ​​used as the left multiplication matrix, and the spatial domain cost map signal vector is used as the right multiplication vector. Matrix multiplication is performed to generate the frequency domain map signal vector. The diagonal matrix is ​​input into the set initial global high-pass filter function, and the first coefficient value for attenuating the low-frequency cost signal and the second coefficient value for retaining the high-frequency cost signal are output. The initial global filter diagonal matrix is ​​constructed according to the arrangement of the diagonal elements of the diagonal matrix. The initial global filter diagonal matrix is ​​used as the left multiplication matrix, and the frequency domain signal vector is used as the right multiplication vector. Matrix multiplication is then performed to generate a high-frequency residual frequency domain signal vector. The orthogonal matrix is ​​used as the left multiplication matrix, and the high-frequency residual frequency domain signal vector is used as the right multiplication vector to perform matrix multiplication operations, and the initial spatial domain residual signal vector is generated by inverse mapping. When the absolute value of the component elements in the initial spatial domain residual signal vector is greater than the initial residual discrimination threshold, the corresponding cost project nodes are extracted and aggregated into the suspected abnormal candidate node set.

6. The intelligent compliance review method based on power engineering cost knowledge graph as described in claim 1, characterized in that, The step of performing an ontology semantic reverse query on the suspected abnormal candidate nodes in the benchmark power engineering cost knowledge graph to generate an adaptive dynamic cutoff frequency for the vertex includes: Extract the unique entity identifier of the current cost project node from the set of suspected abnormal candidate nodes, and concatenate them to generate a target graph query statement; In the benchmark power engineering cost knowledge graph, retrieve attribute nodes that have direct edge connections with the current cost project node, and extract the environmental factor ontology tags and construction attribute ontology tags recorded in the attribute nodes; The semantic penalty mapping dictionary is used to convert the environmental factor ontology tags and the construction attribute ontology tags into numerical environmental impact weight values. A numerical accumulation operation is performed on multiple environmental impact weight values ​​to calculate the comprehensive environmental impact coefficient. The cutoff frequency offset is calculated by multiplying the comprehensive environmental impact coefficient by the frequency offset adjustment ratio constant. The initial global cutoff frequency is then added to the cutoff frequency offset to generate the adaptive dynamic cutoff frequency of the corresponding cost project node.

7. The intelligent compliance review method based on power engineering cost knowledge graph as described in claim 1, characterized in that, The process of constructing an adaptive polynomial filter matrix based on the vertex adaptive dynamic cutoff frequency, and using the adaptive polynomial filter matrix to perform a quadratic filtering operation on the spatial domain cost map signal vector to output the final high-frequency residual map signal vector includes: The mixed Laplacian matrix is ​​normalized and mapped to generate a normalized mixed Laplacian matrix. The polynomial approximation basis matrix sequence of the normalized mixed Laplacian matrix is ​​calculated using a recursive operation. The vertex adaptive dynamic cutoff frequency is substituted into the polynomial coefficient calculation formula to independently calculate the expansion coefficients. Based on the row index order of the cost project nodes in the hybrid Laplace matrix, the expansion coefficients of the same order of all cost project nodes are arranged to generate the coefficient diagonal matrix corresponding to each order of polynomial. Perform vertex adaptive Chebyshev polynomial quadratic filtering operation. For each order polynomial, perform matrix multiplication and weighted summation operations sequentially on the coefficient diagonal matrix, the polynomial approximation basis matrix of the corresponding order, and the spatial domain cost map signal vector to calculate and generate the final high-frequency residual map signal vector.

8. The intelligent compliance review method based on power engineering cost knowledge graph as described in claim 1, characterized in that, The process of calculating the high-frequency residual energy of cost project nodes based on the final high-frequency residual map signal vector, and performing abnormal cost node discrimination based on comparing the high-frequency residual energy with the cost fluctuation tolerance threshold, includes: For each component element in the final high-frequency residual map signal vector, the absolute value is calculated, and the result of the absolute value calculation is squared to obtain the high-frequency residual energy. Historical compliant cost projects of the same type as the project under review are selected from the historical standard review archive data to calculate the benchmark residual energy. An analysis of variance is performed to calculate the statistical standard deviation of the benchmark residual energy, and the cost fluctuation tolerance threshold is generated by combining it with the proportionality constant. When the high-frequency residual energy of a specific cost project node exceeds the cost fluctuation tolerance threshold, the specific cost project node is determined to be an abnormal cost node, and an abnormal cost node list is generated.

9. The intelligent compliance review method based on power engineering cost knowledge graph according to claim 1, characterized in that, The process of tracing back related neighbor nodes and outputting an intelligent compliance review report includes: A traversal operation is performed along the ontology association edge of the abnormal source node. When the actual ratio of the abnormal source node to the neighbor node deviates from the pre-stored benchmark cost ratio and exceeds the system tolerance range, the neighbor node that caused the deviation is determined to be an associated abnormal set of nodes, wherein the abnormal source node is a node in the abnormal cost node list. The high-frequency residual energy is mapped to a cost over-limit alarm judgment item, the actual ratio is mapped to a topological structure ratio fracture data item, the ontology tag is mapped to a semantic environment description item, and data text is spliced ​​according to preset audit logic rules to generate a structured logic deduction chain. Based on the structured logic derivation chain and the local anomaly topology visualization component generated by the graphics rendering interface, the intelligent compliance review report is assembled and generated according to the compliance review electronic report template mapping rules.

10. An intelligent compliance review system based on a knowledge graph of power engineering cost, characterized in that: include: The data access module is configured to acquire power engineering cost documents. The graph management module is configured to map the power engineering cost document data to a pre-built benchmark power engineering cost knowledge graph and instantiate and generate knowledge graph subgraphs. The graph signal processing module is configured to perform a matrix representation of the topological structure of the knowledge graph subgraph to construct a hybrid Laplace matrix, extract the actual engineering cost value corresponding to the power engineering cost document data to construct a spatial domain cost graph signal vector, perform a graph Fourier transform on the spatial domain cost graph signal vector using the hybrid Laplace matrix to generate a frequency domain graph signal vector, and perform global filtering inverse mapping on the frequency domain graph signal vector to generate an initial spatial domain residual signal vector. The semantic feedback control module is configured to perform an ontology semantic reverse query on the suspected abnormal candidate nodes in the benchmark power engineering cost knowledge graph to generate an adaptive dynamic cutoff frequency for the vertex. The graph signal processing module is further configured to construct an adaptive polynomial filter matrix based on the vertex adaptive dynamic cutoff frequency, and use the adaptive polynomial filter matrix to perform a second-order filtering operation on the spatial domain cost graph signal vector to output the final high-frequency residual graph signal vector. The review output module is configured to calculate the high-frequency residual energy of the cost project nodes based on the final high-frequency residual map signal vector, perform abnormal cost node discrimination based on comparing the high-frequency residual energy with the cost fluctuation tolerance threshold, trace back the associated neighbor nodes, and output an intelligent compliance review report.

Images