Artificial intelligence-based power material procurement closed-loop review method and system

Abstract

The application discloses an artificial intelligence-based power material procurement closed-loop review method and system, belongs to the technical field of power material review, and extracts semantic role structures from procurement application texts, contract terms and performance materials by using a deep semantic analysis model to generate a three-dimensional semantic coupling tensor; a procurement behavior semantic time responsibility graph is constructed based on the tensor to reveal responsibility chains, sub-contracting paths and condition exemption logic; the graph is input into a counterfactual game model reasoning engine to generate a multi-round strategy confrontation graph and extract abnormal nodes; a review decision tensor is optimized by a supervised learning mechanism in combination with historical samples to output a behavior traceability graph, an intervention intensity vector and a rule deviation score; when the deviation exceeds a threshold, path reconstruction, term optimization and risk label pushing suggestions are proposed; the application can realize the structured modeling and dynamic review response of complex procurement semantics, and has strong risk identification ability, intelligent judgment ability and rule self-evolution ability.

Description

Technical Field

[0001] This invention relates to the field of power material review technology, specifically to a closed-loop review method and system for power material procurement based on artificial intelligence. Background Technology

[0002] The power industry's procurement systems generally employ a static, rule-driven process control framework. This framework relies heavily on traditional approval chains and manual acceptance mechanisms, lacking intelligent support for semantic understanding of procurement documents, dynamic correlation of contract terms, and identification of potential risks in performance. While some systems have introduced modules such as automatic price comparison and electronic approval, their core remains at the stage of logical verification driven by structured fields, making it difficult to perform semantic understanding and risk identification on unstructured procurement information (such as free-text application instructions, technical specification descriptions, invoices, and performance documentation).

[0003] Existing procurement risk control solutions often focus on judging static abnormal indicators, such as thresholds for duplicate purchase amounts and supplier rating warnings. They lack full-process modeling of the dynamic evolution of the procurement scenario, and in particular, it is difficult to link the initial intent, contract execution and subsequent performance to form a unified judgment system. This makes it easy to avoid abnormal procurement behavior and difficult to reconstruct the violation path.

[0004] Furthermore, there is currently no technology capable of dynamically coupling and modeling the contextual semantic nesting structure of procurement activities (such as indirect obligation chains between responsible parties, hidden conditions, and exemption clauses) with the behavioral time series. Existing methods cannot construct models of compliance expectation deviations, nor do they form closed-loop intervention mechanisms, and they lack a dynamic adjustment logic of "interpretation-game-reinforcement" for review strategies. Summary of the Invention

[0005] The purpose of this invention is to provide a closed-loop review method and system for power material procurement based on artificial intelligence, so as to solve the shortcomings of the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a closed-loop review method for power material procurement based on artificial intelligence, comprising:

[0007] By using a deep semantic parsing model, the semantic role structure in procurement application texts, contract terms, and performance materials is obtained, and a three-dimensional semantic coupling tensor containing compliance element vectors, responsibility mapping graphs, and condition triggering relationships is generated.

[0008] A semantic temporal responsibility graph for procurement behavior is constructed based on semantic coupling tensors. A dynamic expansion mechanism of polygonal graphs is used to represent the unexplicit but potentially existing responsibility chains, sub-delegation logic and condition exemption chains in the procurement path.

[0009] The graph is input into a dynamic inference engine based on an improved counterfactual game model to generate a multi-round strategy adversarial graph of potential violation paths and extract a set of anomalous nodes with context labels.

[0010] By combining typical adversarial sample sets from historical procurement scenarios, a policy alignment supervised learning mechanism is used to optimize the review decision tensor, and the output includes a behavior tracing probability graph, a responsible entity intervention intensity vector, and a rule deviation score decision triplet.

[0011] When any node deviates from the score and exceeds the preset threshold, a closed-loop review response instruction set is automatically generated, which includes a real-time restructuring strategy for the approval path, a set of clause adjustment suggestions, and a push notification of behavioral risk labels.

[0012] Preferably, the process of obtaining semantic role structure using a deep semantic parsing model includes:

[0013] A multi-scale semantic decomposition based on a self-attention mechanism is performed on the procurement application text to generate a syntactic nested structure tree containing procurement intent, behavioral subject, and constraints.

[0014] Multi-head cross-embedding encoding is performed on contract terms and performance materials to extract the sentence groups of rights and responsibilities, trigger clauses and additional condition chains, and to form a semantic association graph;

[0015] The syntactic nested structure tree and semantic association graph are input into a multimodal semantic coupling network that integrates graph convolutional networks and Transformer residual channels. The coupling weights are optimized through a semantic constraint consistency loss function, and the output is a three-dimensional semantic coupling tensor containing compliance element vectors, responsibility mapping paths and condition triggering relationships.

[0016] Preferably, the construction of the semantic temporal responsibility graph includes:

[0017] Based on the compliance element vector extracted from the three-dimensional semantic coupling tensor and the role interaction path, the semantic clustering context density method is used to filter the set of key behavioral nodes.

[0018] Dynamic semantic connectivity detection is performed on the key node set to construct a time-series directed graph, and a node temporal drift factor is introduced to construct a time-responsibility biaxial projection space.

[0019] By utilizing a dynamic expansion mechanism based on a polygonal graph structure, explicit role paths are completed into a composite behavioral graph that includes responsibility transfer chains, implicit delegation paths, and deferred trigger condition logic, outputting a procurement path graph structure that integrates semantic and temporal responsibilities.

[0020] Preferably, the generation of potential violation paths includes:

[0021] The semantic temporal responsibility graph is input into a counterfactual game reasoning engine that includes a strategy perturbation module and a role intention predictor.

[0022] By introducing the strategy perturbation factor and context adversarial tensor in path simulation, a multi-round evolution strategy adversarial graph is constructed to simulate the behavioral deviation path of the procurement entity under different constraints and incentives.

[0023] By using a temporal reinforcement scoring mechanism to label adversarial graphs with nodes exhibiting logical reverse loops and conditional mutations, a set of abnormal behavior nodes with contextual labels is extracted.

[0024] Preferred, policy alignment supervised learning mechanisms include:

[0025] By calling the typical adversarial sample set in the historical procurement scenario library, label mapping and path feature comparison are performed on the abnormal node set extracted from the multi-round strategy adversarial graph to construct a supervised aligned semantic comparison set;

[0026] The multi-channel attention supervision network, which integrates node behavior trajectories, contextual roles, and triggering logic inputs in the semantic temporal responsibility graph, performs cross-scenario policy transfer and adversarial residual optimization.

[0027] The output includes a ternary decision tensor consisting of a behavior tracing probability graph, a vector of the intensity of intervention by the responsible party, and a rule deviation score. This tensor serves as the basis for the intelligent review engine's judgment and is dynamically fed back to the inference engine to update the adversarial weights.

[0028] Preferably, the generation of the closed-loop response instruction set includes:

[0029] When any rule in the ternary decision tensor deviates from the score by more than a preset dynamic threshold range, the review response generation engine with policy graph adaptive capability is triggered.

[0030] The response generation engine generates approval path reconstruction strategies in real time based on the path backtracking chain in the behavior tracing probability graph and the influence index in the intervention intensity vector, including process adjustment, authority reallocation and objection node isolation operation instructions.

[0031] Simultaneously, a set of contract clause adjustment suggestions is generated, and key sentence elements marked with semantic role risk distribution charts are used for compliance replacement and enhanced logic prompts;

[0032] The output includes behavioral path labeling information with dynamic risk tags, and the abnormal policy path and response results are synchronously fed back to the review rule base for adaptive rule evolution and subsequent review iteration optimization.

[0033] This invention also provides an artificial intelligence-based closed-loop review system for power material procurement, comprising:

[0034] The semantic parsing module uses a deep semantic parsing model to obtain the semantic role structure in procurement application texts, contract terms, and performance materials, and generates a three-dimensional semantic coupling tensor containing compliance element vectors, responsibility mapping graphs, and condition triggering relationships.

[0035] The semantic graph construction module constructs a semantic temporal responsibility graph of procurement behavior based on semantic coupling tensors, and uses a multi-sided graph dynamic expansion mechanism to represent the unexplicit but potentially existing responsibility chains, sub-delegation logic and condition exemption chains in the procurement path;

[0036] The counterfactual reasoning module inputs the graph into a dynamic reasoning engine based on an improved counterfactual game model to generate a multi-round strategy adversarial graph of potential violation paths and extract a set of abnormal nodes with context labels.

[0037] The supervised learning optimization module combines typical adversarial sample sets from the historical procurement scenario library and uses a policy alignment supervised learning mechanism to optimize the review decision tensor, outputting a decision triplet that includes a behavior tracing probability graph, a vector of the responsible entity's intervention intensity, and a rule deviation score.

[0038] The closed-loop response generation module automatically generates a set of review closed-loop response instructions when any node deviates from the score and exceeds the preset threshold. This includes a real-time restructuring strategy for the approval path, a set of clause adjustment suggestions, and push notifications of behavioral risk labels.

[0039] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0040] 1. This invention, by introducing a deep semantic parsing model and a multimodal semantic coupling mechanism, enables cross-structure and cross-context semantic role identification and responsibility path reconstruction for application texts, contract terms, and performance materials involved in power material procurement. Compared to traditional static review methods based on rule or field matching, this invention can automatically uncover hidden responsibility chains, delegation paths, and condition exemption logic, achieving full-link structured modeling of procurement behavior and effectively improving the depth of anomaly identification and the semantic interpretation capability of review decisions.

[0041] 2. This invention integrates improved counterfactual game reasoning and supervised learning mechanisms. It can not only generate multi-round strategy adversarial graphs to dynamically simulate potential violation paths, but also output a ternary decision tensor containing a behavior source probability graph, an intervention intensity vector, and a rule deviation score. This triggers the review response engine to generate path reconstruction and clause optimization suggestions in real time, achieving a shift from "static judgment" to "intelligent intervention." Through the labeling of abnormal paths and the self-feedback updating of the rule base, this invention constructs a closed-loop review system with adaptive learning capabilities, significantly improving the accuracy, real-time performance, and system evolution capabilities of procurement risk control. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0043] Figure 1 This is a mind map of the method of the present invention.

[0044] Figure 2 This is a mind map of the system modules of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of 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, 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.

[0046] Example 1, please refer to Figure 1 As shown in this embodiment, the closed-loop review method for power material procurement based on artificial intelligence includes:

[0047] This paper utilizes a deep semantic parsing model to obtain the semantic role structure in procurement application texts, contract terms, and performance materials, generating a three-dimensional semantic coupling tensor containing compliance element vectors, responsibility mapping graphs, and condition triggering relationships. Based on the semantic coupling tensor, a semantic temporal responsibility graph of procurement behavior is constructed, and a polygonal graph dynamic expansion mechanism is used to represent the implicit but potential responsibility chains, sub-delegation logic, and condition exemption chains in the procurement path. The graph is input into a dynamic inference engine based on an improved counterfactual game model to generate a multi-round strategy adversarial graph of potential violation paths, and extracts a set of abnormal nodes with context labels. Combining a typical adversarial sample set from a historical procurement scenario library, a strategy alignment supervised learning mechanism is used to optimize the review decision tensor, outputting a decision triplet including a behavior tracing probability graph, a responsible entity intervention intensity vector, and a rule deviation score. When the deviation score of any node exceeds a preset threshold, a review closed-loop response instruction set is automatically generated, including a real-time reconstruction strategy for the approval path, a set of clause adjustment suggestions, and a push of behavior risk labels.

[0048] In this embodiment of the invention, the deep semantic parsing model is used to extract semantic role structures with contextual features from procurement-related unstructured text, and to establish explicit and implicit semantic graphs of behavioral paths, responsibility boundaries and condition triggering mechanisms in the text, providing basic input for subsequent review and reasoning.

[0049] First, the procurement application text is semantically deconstructed using a multi-scale semantic decomposition algorithm based on a self-attention mechanism. This algorithm constructs a multi-layer context-aware representation of the word sequence in the text input, capturing key components such as subjective preferences, the subject of the action, and the target constraints in the procurement intent.

[0050] The multi-scale semantic decomposition module extracts feature representations at different semantic levels by setting multi-level attention windows (e.g., within a sentence, within a paragraph, and throughout the entire text), and simultaneously establishes a syntactic nested structure tree. This structure tree explicitly encodes syntactic dependency relations such as subject-verb-object relationships, subordinate adverbs, and negative qualifiers. Nodes represent semantic units, and edges represent structural dependencies, thereby accurately depicting the logical organization between intent and constraints in a procurement application.

[0051] For example, for text like "It is planned to purchase three SVG devices, with a delivery date no later than December and local supply preferred", the structure tree uses "purchase" as the root node, "three SVG devices" as the object node, and "delivery date", "no later than December" and "preferred" as time constraints and preference modifiers as child nodes, effectively extracting the primary skeleton of the semantic execution graph.

[0052] Secondly, the contract terms and performance materials (such as acceptance records, invoice images, and delivery confirmation documents) are processed in parallel. Multi-head cross-embedding is used to embed semantic entities at the character, word, and phrase levels, and to construct a group of sentences for dividing rights and responsibilities, trigger clauses, and chains of additional conditions.

[0053] The cross-embedding network includes parallel attention channels and interaction residual blocks, which can capture semantic co-occurrence relationships in the context of multi-source documents. For example, it can establish a mapping between the payment responsibility and performance deviation between "Party B shall pay the final payment within 10 working days after the equipment installation and acceptance" and "The invoice amount is 2.4 million yuan and the payment progress is 80%", thus forming a semantic association graph.

[0054] The nodes in this graph represent semantic units of sentence groups or clauses, and the edges represent responsibility mappings or conditional dependencies. The graph explicitly marks semantic roles such as "responsible subject," "time constraint," and "dependency condition" and their structured expressions.

[0055] Finally, the constructed syntactic nested structure tree and semantic association graph are input into a multimodal semantic coupling network. This network integrates graph convolutional network (GCN) and Transformer residual branch structure to perform fusion modeling of semantic nesting, behavioral paths, responsibility mapping, and conditional triggering chains between graph structures and sequence structures.

[0056] The multimodal coupling network contains an alignment mapping module and a graph order cross-attention unit, which can perform dimensional consistency mapping on the structure tree and graph, compressing them into a fixed-structure three-dimensional semantic coupling tensor. The first dimension of the tensor represents the semantic role type (such as intention, obligation, constraint, incentive), the second dimension represents the time and context trajectory dimension, and the third dimension represents the semantic edge weights and correlations between entities.

[0057] During the training phase, the coupled network is optimized using a semantic constraint consistency loss function, which is defined as follows: ;in, These represent the embedding vectors of semantic roles in the structure tree and semantic graph, respectively. Let λ be the KL divergence function and λ be the regularization factor. represents the semantic consistency loss function, used to measure the degree of difference between the syntactic structure tree and the semantic association graph in semantic embedding representation; N represents the total number of semantic units participating in the loss calculation, usually corresponding to the number of semantic role nodes; This represents the role classification probability distribution vector of the i-th semantic unit in the syntax tree structure, usually output by the Softmax function; This represents the role classification probability distribution vector of the i-th semantic unit in the semantic graph structure;

[0058] The final output three-dimensional semantic coupling tensor integrates semantic elements, role relationships, and condition triggering mechanisms from multiple source texts, providing high-quality structured semantic support for the subsequent construction of semantic temporal responsibility graphs and game reasoning.

[0059] To achieve dynamic reconstruction of complex behavioral paths in procurement review, this invention, based on the acquisition of a three-dimensional semantic coupling tensor, further constructs a semantic-time-responsibility graph structure to reveal potential responsibility transfer and condition constraint paths across documents, time points, and entities.

[0060] First, compliance element vectors and role interaction path sequences are extracted from the 3D semantic coupling tensor, and potential key behavior nodes are identified using semantic clustering-context density method. The method embeds vectors in the semantic space... Perform context density estimation: ;in: This represents the context density score of the i-th semantic unit; σ represents the semantic neighborhood set of the i-th unit; σ represents the Gaussian kernel bandwidth parameter, which controls the sensitivity of density calculation.

[0061] By using density peak judgment and semantic dispersion filtering, behavioral nodes that are semantically concentrated but geographically dispersed in the context are retained as the core node set for subsequent construction of the time responsibility graph.

[0062] Based on the key nodes, extract the time stamp of each node in the source document (such as application time, contract signing time, invoice issuance time, etc.) to construct a directed time series graph. The direction of the edges represents the causal order of events from first to last.

[0063] Introducing node timing drift factor This is used to model the contextual resilience representation of nodes under different time reference frames: ;in: This represents the current semantic topic distribution vector of node i; This represents the initial semantic topic distribution of node i; Indicates the time interval between nodes.

[0064] Based on the timeline, a responsibility axis is constructed, forming a time-responsibility dual-axis projection space through the role assignment chain and responsibility semantic tags in the semantic graph. The position of the node in this space is as follows: ;in: This indicates the sequence position of node i on the time axis; This represents the weight center of node i in the responsibility role space.

[0065] Finally, in the biaxial space, a composite behavioral graph is constructed using a dynamic expansion mechanism of polygon graphs. This mechanism allows for dynamic adjustment of edge types, including:

[0066] Explicit role path edge (explicit instruction or contract role);

[0067] Transfer of responsibility chain (e.g., "Party B entrusts Party C to perform the contract");

[0068] Implicit delegate path edges (not explicitly stated in the text but presumed to exist in the behavior);

[0069] Conditional deferred logic edge (e.g., "Payment terms: Equipment acceptance and no breach of contract dispute").

[0070] Each edge type employs a different edge weight update mechanism, and a semantic entropy adjustment function is dynamically introduced. Control structure diagram complexity: ;in: This represents the weight of edge (i,j) under the specified type; Indicates semantic similarity score; α represents the semantic entropy function based on the local structure of the subgraph; α and β represent the control factors used to balance similarity and entropy complexity.

[0071] The final output graph structure is the procurement path graph with semantic-temporal responsibility fusion, which serves as the input basis for subsequent counterfactual game reasoning and supervised decision optimization.

[0072] In this invention, in order to identify possible violations, evasion and implicit responsibility transfer behaviors in the procurement of power materials, based on the constructed semantic-time-responsibility graph structure, counterfactual game modeling and strategy adversarial graph simulation are further introduced to generate potential violation paths and locate abnormal nodes.

[0073] The counterfactual game reasoning engine consists of a strategy perturbation module, a role intention predictor, and a path evolution simulation unit. This engine uses a semantic temporal responsibility graph as input to construct a game decision space within a multi-agent, dynamic context.

[0074] The role intent predictor, based on the responsibility label sequence, semantic tensor, and historical behavior label set of nodes in the graph, uses a conditional variational coding model (CVAE) to predict the set of actions that each node may take in a specific context. and output the intention distribution vector. .

[0075] For example, for a contract signing node, the intention distribution may be "standard performance", "with avoidance clause", "introduction of temporary supplier", etc., and the prediction results are used to drive the simulation of subsequent path disturbances.

[0076] To simulate the potential evolutionary paths of agent behavior under different institutional conditions and constraint relaxation scenarios, the engine constructs a strategy perturbation factor. Context-adversarial tensors To generate multi-round policy graphs : ;in: Let represent the policy perturbation vector of node i in the k-th round. This represents a Gaussian perturbation with the current node's policy covariance as a parameter; The context adversarial tensor represents the condition for a specific context. (e.g., contract delays, relaxed regulations) simulated strategy deformation factors; γ is the disturbance amplitude adjustment factor; This represents the policy adversarial graph generated in the kth round after policy perturbation.

[0077] Each round of strategy diagrams is constructed through a counterfactual path backtracking mechanism: given the behavior of a certain role on the real path, its alternative behavioral trajectory under different incentive, punishment or responsibility structures is simulated to form a multi-branch adversarial path tree structure.

[0078] In all strategy graphs In this study, a time-series reinforcement scoring mechanism is used to jointly score the rationality and risk significance of the strategy shift at each node: ;in: Indicates the magnitude of the strategy offset; This refers to the context-sensitive weight metric, which measures the current node's ability to induce violations in a specific context. It represents the logical inverse loop closing factor, and the probability that a node is located in a closable condition chain or an abnormal path loop. This is the weighting coefficient, which can be adjusted according to the strategy complexity.

[0079] For rating For node groups that exceed the dynamic risk threshold θrisk, further contextual labels (such as "abnormal sub-delegation", "excessive contract performance", "double pricing path", etc.) are added as high-priority trigger points in the review system and included in the abnormal behavior node set for subsequent supervised learning model optimization.

[0080] After identifying potential violation paths, this invention introduces a policy alignment supervised learning mechanism to perform in-depth judgment and attribution of abnormal behavior nodes. This mechanism uses historical procurement scenarios as the basis for supervision and combines multimodal graph representation learning methods to train a high-dimensional decision tensor that can explain the procurement behavior deviation mechanism, and realizes dynamic adjustment of the adversarial inference model.

[0081] First, a set of typical adversarial samples that have been manually reviewed and labeled in the historical procurement scenario library is called up. This set of samples contains the strategy paths, responsibility chains and semantic tags of known violations or circumvention behaviors, forming a "labeled behavior semantic graph set".

[0082] The set of anomalous nodes extracted from the policy adversarial graph is semantically embedded and compared with the behavioral paths in the adversarial sample set, and then labeled. A supervised comparison set is constructed using a graph alignment algorithm. Its definition is: ;in: The policy graph generated for the i-th simulation; This is the j-th reference sample image; This represents the node alignment label matrix between the two sets, recording the semantic and responsibility matching degrees of the corresponding behavioral units. This comparison set is used to supervise the training of the neural network, achieving pattern transfer from the policy offset graph to high-confidence violation judgments.

[0083] To effectively integrate multi-dimensional data input from the semantic-temporal-responsibility graph, the behavioral trajectories, role participation tensors, and trigger logic vectors of abnormal nodes are input into a multi-channel attention-supervised network. Its structure includes:

[0084] Behavioral trajectory encoder: LSTM encoding is performed on the node time series and state transition sequence to form a behavioral context trajectory tensor;

[0085] Role-based channel: The semantic role interaction matrix is ​​used as input to the graph convolution to generate a responsibility propagation tensor;

[0086] Logical triggering channel: Model the dependency coupling strength on the conditional triggering chain using the causal convolution module.

[0087] The three-channel outputs are fused by the cross-attention module to form a node behavior description vector. The ultimate training objective of the network is: Where: CE represents the cross-entropy loss function; This represents the actual label and the predicted label of a node; Predictive and reference behavior embeddings; η semantic embedding alignment regularization factor.

[0088] After the network training is completed, the output is a decision tensor triple containing three types of high-dimensional indicators. : This represents a probability graph for behavioral tracing, recording the possible behavioral evolution or tracing avoidance relationships between each pair of nodes. This represents the intervention intensity vector of the responsible party, measuring the urgency of needing to review the response or issue an early warning intervention at a specific node; The rule deviation score measures the degree to which the current node's behavior deviates from the reference rule template.

[0089] The ternary tensor also serves as the basis for judgment in the intelligent review system and is fed back to the counterfactual reasoning engine to reconstruct the weights of the policy perturbation magnitude and the simulated priority path, thereby realizing a closed-loop optimization mechanism of game-learning.

[0090] This invention constructs a review response engine based on the ternary decision tensor output by the intelligent review system. This engine has the capabilities of dynamic judgment, adaptive reconstruction, and risk closed-loop handling. After detecting rule deviation behavior, it can promptly propose path adjustment, contract optimization, and system rule update strategies to complete the closed-loop control from risk identification to strategy intervention.

[0091] When the rule of any node in the ternary decision tensor deviates from the score Exceeding the system's set dynamic risk threshold That is, satisfying: ;in: , This represents the mean and standard deviation of all deviation scores within the current period; κ represents the deviation tolerance coefficient, set in the range of 1.5 to 2.0. This indicates a risk threshold that is dynamically adjusted based on the actual distribution. This triggers the "Strategy Graph Adaptive Review Response Generation Engine," initializing the response construction sequence.

[0092] The response engine reads the backtracking path chain in the behavior tracing probability graph Ptrace and extracts the set of predecessor nodes Pipre that may lead to abnormal behavior through depth-first path tracing. This is combined with the influence index of the corresponding node in the intervention intensity vector Vintervene. Priority of computational stage reconstruction: ;in: This represents the in-degree of node j, measuring the degree to which it is dependent on multiple paths. represents the betweenness centrality of a node in the path graph; e is the adjustment coefficient.

[0093] Based on priority, output the following path reconstruction strategies:

[0094] Adjustments to the approval process: Adjusting or replacing the approval roles or order corresponding to high-risk nodes;

[0095] Permission reallocation: reauthorize key nodes or add a review mechanism;

[0096] Dispute node isolation: Temporarily freeze or report and mark nodes with abnormal multi-path associations.

[0097] For contractual behavior nodes deemed to deviate from the rules, the system extracts their semantic role risk distribution map, marks key sentence elements (such as time limits, delegation authority, and acceptance conditions), and outputs compliance optimization suggestions, divided into two categories:

[0098] Compliance replacement recommendation: Replace easily circumventable clauses, such as "can be executed by any third party" with "limited to registered qualified contractors";

[0099] Enhance logic hints: Add examples or scope definitions to clauses containing ambiguous expressions, such as “reasonable time” → “no more than 15 calendar days”.

[0100] All suggestions are output in the form of a structured set of proposed amendments, which can be manually selected by the approver for adoption.

[0101] Finally, the system outputs a set of behavioral path tagging information containing dynamic risk labels. Each path node is tagged with the following:

[0102] Risk level (high / medium / low);

[0103] Deviation types (overlapping responsibilities, path avoidance, conditional loopholes);

[0104] Recommended actions (refactoring, replacement, locking, review delay, etc.).

[0105] Simultaneously, the aforementioned abnormal strategy paths and response outputs, including the behavior path marker information set and the structured clause modification suggestion set, are synchronously written into the review rule base as input samples for the next cycle update of the review engine, thereby realizing the self-evolution of the rule knowledge base and the iterative optimization of the model.

[0106] Example 2: To verify the applicability and effectiveness of the present invention in the review of material procurement in the power industry, a sample scenario was taken from the 2023 framework procurement of power distribution equipment by a provincial company of the State Grid Corporation of China. A total of 12,460 unstructured documents covering bidding documents, winning contracts, acceptance reports and invoices were collected to build a semantic sample training library and a behavior label supervision set.

[0107] The document types include: purchase requisition forms (3,200 copies), purchase contracts (PDF, 2,700 copies), performance invoice images (3,800 copies), and scanned copies of acceptance records (2,760 copies).

[0108] Tag set construction method: Through comparison with historical monitoring records, 57 types of known violation paths and 23 typical avoidance strategy templates were manually labeled;

[0109] System deployment environment: NVIDIA A100 GPU×2, 256GB memory, using the PyTorch framework to build the review model, with semantic tensor dimensions of (128×128×32).

[0110] In the actual test, 104 bidding and procurement projects in Q2 of 2024 (involving a total of 1,129 contract paths) were selected, and "system identification accuracy, avoidance path identification capability, and suggested clause adoption rate" were used as evaluation indicators.

[0111] Abnormal behavior identification accuracy 91.3% 62.4% +28.9% Path-level avoidance recognition rate 87.5% 41.8% +45.7% Contract terms optimization adoption rate 76.2% Not supported —— Automatic feedback closed-loop trigger rate 94.1% Not supported ——

[0112] In addition, the system successfully identified 14 previously undetected "sub-delegation + risk clause avoidance" combination paths in the sample, involving 3 suppliers and 2 approvers; 6 of these paths were confirmed as risky procurement behaviors, proving that the method of the present invention has the ability to detect risks in practice.

[0113] The C08 procurement project includes a total of 180 "10kV cable branch boxes";

[0114] The original contract contained the phrase "may be delivered by Party B or its designated third party";

[0115] The system identified that the node deviated from the score by 0.73 (the threshold is 0.52), and output a strategy reconstruction suggestion of "limiting the main body of supplier registration and filing", and successfully triggered the replacement of "Approver B";

[0116] The final feedback log indicates that the system's suggested contract optimizations were adopted by the salesperson, and after manual review, it was confirmed that there was potential avoidance behavior but no actual violation had occurred.

[0117] The above experiments confirm the following conclusions:

[0118] This invention can effectively integrate unstructured semantic analysis, role intent modeling, and strategy path simulation to accurately identify complex avoidance paths;

[0119] The system demonstrates significantly better performance than traditional rule-based logic systems in detecting compliance deviations.

[0120] The constructed closed-loop response mechanism transforms risk identification from "post-event judgment" to "process intervention," enabling substantial intelligent risk control capabilities.

[0121] Example 3, please refer to Figure 2 As shown in this embodiment, the artificial intelligence-based closed-loop review system for power material procurement includes:

[0122] The semantic parsing module uses a deep semantic parsing model to obtain the semantic role structure in procurement application texts, contract terms, and performance materials, and generates a three-dimensional semantic coupling tensor containing compliance element vectors, responsibility mapping graphs, and condition triggering relationships.

[0123] The semantic graph construction module constructs a semantic temporal responsibility graph of procurement behavior based on semantic coupling tensors, and uses a multi-sided graph dynamic expansion mechanism to represent the unexplicit but potentially existing responsibility chains, sub-delegation logic and condition exemption chains in the procurement path;

[0124] The counterfactual reasoning module inputs the graph into a dynamic reasoning engine based on an improved counterfactual game model to generate a multi-round strategy adversarial graph of potential violation paths and extract a set of abnormal nodes with context labels.

[0125] The supervised learning optimization module combines typical adversarial sample sets from the historical procurement scenario library and uses a policy alignment supervised learning mechanism to optimize the review decision tensor, outputting a decision triplet that includes a behavior tracing probability graph, a vector of the responsible entity's intervention intensity, and a rule deviation score.

[0126] The closed-loop response generation module automatically generates a set of review closed-loop response instructions when any node deviates from the score and exceeds the preset threshold. This includes a real-time restructuring strategy for the approval path, a set of clause adjustment suggestions, and push notifications of behavioral risk labels.

[0127] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A closed-loop review method for power material procurement based on artificial intelligence, characterized by: include: By using a deep semantic parsing model, the semantic role structure in procurement application texts, contract terms, and performance materials is obtained, and a three-dimensional semantic coupling tensor containing compliance element vectors, responsibility mapping graphs, and condition triggering relationships is generated. Based on semantic coupling tensors, a semantic temporal responsibility graph of procurement behavior is constructed. A dynamic expansion mechanism of polygonal graphs is used to represent the unexplicit but potentially existing responsibility chains, sub-delegation logic and condition exemption chains in the procurement path. The graph is input into a dynamic inference engine based on an improved counterfactual game model to generate a multi-round strategy adversarial graph of potential violation paths and extract a set of anomalous nodes with context labels. By combining typical adversarial sample sets from historical procurement scenarios, a policy alignment supervised learning mechanism is used to optimize the review decision tensor, and the output includes a behavior tracing probability graph, a responsible entity intervention intensity vector, and a rule deviation score decision triplet. When any node deviates from the score and exceeds the preset threshold, a closed-loop review response instruction set is automatically generated, which includes a real-time restructuring strategy for the approval path, a set of clause adjustment suggestions, and a push notification of behavioral risk labels.

2. The closed-loop review method for power material procurement based on artificial intelligence according to claim 1, characterized in that: The process of obtaining semantic role structure by deep semantic parsing models includes: A multi-scale semantic decomposition based on a self-attention mechanism is performed on the procurement application text to generate a nested syntactic structure tree containing procurement intent, behavioral subject, and constraints. Multi-head cross-embedding encoding is performed on contract terms and performance materials to extract the sentence groups of rights and responsibilities, trigger clauses and additional condition chains, and to form a semantic association graph; The syntactic nested structure tree and semantic association graph are input into a multimodal semantic coupling network that integrates graph convolutional networks and Transformer residual channels. The coupling weights are optimized through a semantic constraint consistency loss function, and the output is a three-dimensional semantic coupling tensor containing compliance element vectors, responsibility mapping paths and condition triggering relationships.

3. The closed-loop review method for power material procurement based on artificial intelligence as described in claim 1, characterized in that: The construction of the semantic temporal responsibility graph includes: Based on the compliance element vector extracted from the three-dimensional semantic coupling tensor and the role interaction path, the semantic clustering context density method is used to filter the set of key behavioral nodes. Dynamic semantic connectivity detection is performed on the key node set to construct a time-series directed graph, and a node temporal drift factor is introduced to construct a time-responsibility biaxial projection space. By utilizing a dynamic expansion mechanism based on a polygonal graph structure, explicit role paths are completed into a composite behavioral graph that includes responsibility transfer chains, implicit delegation paths, and deferred trigger condition logic, outputting a procurement path graph structure that integrates semantic and temporal responsibilities.

4. The closed-loop review method for power material procurement based on artificial intelligence according to claim 1, characterized in that: The generation of potential violation paths includes: The semantic temporal responsibility graph is input into a counterfactual game reasoning engine that includes a strategy perturbation module and a role intention predictor. By introducing the strategy perturbation factor and context adversarial tensor in path simulation, a multi-round evolution strategy adversarial graph is constructed to simulate the behavioral deviation path of the procurement entity under different constraints and incentives. By using a temporal reinforcement scoring mechanism to label adversarial graphs with nodes exhibiting logical reverse loops and conditional mutations, a set of abnormal behavior nodes with contextual labels is extracted.

5. The closed-loop review method for power material procurement based on artificial intelligence according to claim 1, characterized in that: Policy alignment supervised learning mechanisms include: By calling the typical adversarial sample set in the historical procurement scenario library, the abnormal node set extracted from the multi-round strategy adversarial graph is labeled and compared with path features to construct a supervised aligned semantic comparison set; The multi-channel attention supervision network, which integrates node behavior trajectories, contextual roles, and triggering logic inputs in the semantic temporal responsibility graph, performs cross-scenario policy transfer and adversarial residual optimization. The output includes a ternary decision tensor consisting of a behavior tracing probability graph, a vector of the intensity of intervention by the responsible party, and a rule deviation score. This tensor serves as the basis for the intelligent review engine's judgment and is dynamically fed back to the inference engine to update the adversarial weights.

6. The closed-loop review method for power material procurement based on artificial intelligence according to claim 1, characterized in that: The generation of the closed-loop response instruction set includes: When any rule in the ternary decision tensor deviates from the score by more than a preset dynamic threshold range, the review response generation engine with policy graph adaptive capability is triggered. The response generation engine generates approval path reconstruction strategies in real time based on the path backtracking chain in the behavior tracing probability graph and the influence index in the intervention intensity vector, including process adjustment, authority reallocation and objection node isolation operation instructions. Simultaneously, a set of contract clause adjustment suggestions is generated, and key sentence elements marked with semantic role risk distribution charts are used for compliance replacement and enhanced logic prompts; The output includes behavioral path labeling information with dynamic risk tags, and the abnormal policy path and response results are synchronously fed back to the review rule base for adaptive rule evolution and subsequent review iteration optimization.

7. An AI-based closed-loop review system for power material procurement, used to implement the AI-based closed-loop review method for power material procurement as described in any one of claims 1-6, characterized in that: include: The semantic parsing module uses a deep semantic parsing model to obtain the semantic role structure in procurement application texts, contract terms, and performance materials, and generates a three-dimensional semantic coupling tensor containing compliance element vectors, responsibility mapping graphs, and condition triggering relationships. The semantic graph construction module constructs a semantic temporal responsibility graph of procurement behavior based on semantic coupling tensors, and uses a multi-sided graph dynamic expansion mechanism to represent the unexplicit but potentially existing responsibility chains, sub-delegation logic and condition exemption chains in the procurement path; The counterfactual reasoning module inputs the graph into a dynamic reasoning engine based on an improved counterfactual game model to generate a multi-round strategy adversarial graph of potential violation paths and extract a set of abnormal nodes with context labels. The supervised learning optimization module combines typical adversarial sample sets from the historical procurement scenario library and uses a policy alignment supervised learning mechanism to optimize the review decision tensor, outputting a decision triplet that includes a behavior tracing probability graph, a vector of the responsible entity's intervention intensity, and a rule deviation score. The closed-loop response generation module automatically generates a set of review closed-loop response instructions when any node deviates from the score and exceeds the preset threshold. This includes a real-time restructuring strategy for the approval path, a set of clause adjustment suggestions, and push notifications of behavioral risk labels.

Images