Project risk monitoring method and system based on large language model

Abstract

The invention relates to the technical field of project risk management, in particular to a project risk monitoring method and system based on a large language model, and aims to guide a language model to complete risk identification in a professional context by analyzing a natural language supervision request of a user, identifying a task field, matching a corresponding knowledge graph and a rule base, generating a reasoning configuration set and guiding the language model to complete risk identification in a professional context. Through a multi-modal fusion mechanism, unstructured data such as contract texts, drawing images and progress logs are coded in a unified mode, context modeling and rule reasoning of cross-modal information are achieved in combination with a large language model guided by a strategy, hidden risks needing image-text linkage judgment are effectively recognized, the analysis capacity for complex semantic association is improved, and the method is suitable for large-scale popularization and application. And furthermore, through a reinforcement learning mechanism, a supervision sample is constructed according to user feedback, a reward signal is generated, language model strategy parameters are optimized in real time, and continuous evolution and self-adaptive updating of a risk monitoring model are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of project risk management, in particular to a project risk monitoring method and system based on a large language model. BACKGROUND

[0002] Risk management throughout the project has become a core concern in the field of engineering consulting, construction and supervision. Traditional project risk supervision mainly relies on manual experience, and potential problems are identified by manually checking contract texts and reviewing construction logs. This manual method has many shortcomings such as low efficiency, limited coverage, poor knowledge transfer ability, etc. Especially when faced with large-scale, multi-stage, cross-specialty engineering tasks, it is often difficult to discover compliance risks and quality risks in time, affecting project progress and quality safety.

[0003] The existing engineering supervision information system is mainly manually oriented, supplemented by a number of automated tools. However, these systems often use isolated model structures, such as text review and progress analysis as separate modules, lacking a unified multi-modal coordination mechanism. Specifically, when the system detects a delay in construction progress, it cannot automatically associate it with drawing change information and contract extension clauses for reasoning, resulting in the inability to identify implicit risks such as "whether non-compliance with the drawing has been agreed in the contract change". In addition, most systems rely on pre-set rules and lack a unified knowledge reasoning logic, making it difficult to flexibly identify risk paths based on context. In addition, most existing models are configured with static parameters, making it difficult to adjust based on user feedback during actual use, resulting in a decline in model effectiveness as the project progresses, and an inability to adapt to the dynamic evolution of risks. SUMMARY

[0004] To solve the above problems, the present application provides a project risk monitoring method and system based on a large language model.

[0005] To achieve the above purpose, the technical solution adopted by the present application is:

[0006] A project risk monitoring method based on a large language model, comprising the following steps:

[0007] S1, obtain a natural language supervision request input by a user, analyze its semantic features, determine the field to which the supervision task belongs, and match the knowledge graph and rule base corresponding to the field to generate a set of reasoning configurations related to the task;

[0008] S2, obtain non-structured data uploaded by the user, including contract texts, drawing images and progress logs, extract text entities, image structures and time sequence features respectively, and fuse to generate multi-modal encoding representations;

[0009] S3, input the pre-trained language model of the multi-modal encoding representation, and perform semantic analysis and rule inference in combination with the inference configuration set to generate a structured risk labeling result;

[0010] S4, generating a structured supervision report containing risk levels, associated regulations and modification suggestions based on the structured risk labeling result; constructing a feedback sample set with supervised labeling according to user feedback information to generate a reward value, and updating language model parameters through reinforcement learning.

[0011] Further, the S1 comprises the following steps:

[0012] S11, performing word segmentation, entity recognition and semantic feature extraction on the natural language supervision request input by the user to obtain a supervision target semantic vector;

[0013] S12, comparing the supervision target semantic vector with a plurality of preset domain semantic vector templates for similarity, and identifying the domain label to which the current task belongs;

[0014] S13, based on the domain label, retrieving corresponding subgraph structure data from the knowledge graph database, and extracting relevant rule items and hierarchical relationships according to the rule index associated with the subgraph to construct an inference configuration set.

[0015] Further, the S13 comprises the following steps:

[0016] Based on the identified domain label, locate the corresponding domain subgraph from the knowledge graph database, the domain subgraph contains semantic entity nodes and their semantic relationship edges related to the domain;

[0017] Performing semantic type screening on the semantic entity nodes in the domain subgraph, extracting a target node set having significant semantic correlation with the supervision target semantic vector, and generating a path weight graph according to the semantic edge types between the nodes;

[0018] According to the path weight graph, extract the rule index information associated with the target node, the rule index information includes constraint type, applicable condition and inference template identifier;

[0019] According to the extracted rule index information, retrieve the corresponding rule items from the rule library, and construct the rule applicable scope in combination with the hierarchical relationship information in the knowledge graph;

[0020] Encapsulating the target node set, path weight graph and rule item set as an inference configuration set.

[0021] Further, the S2 comprises the following steps:

[0022] S21, perform syntax analysis and entity recognition processing on the contract text uploaded by the user, extract the obligor, time agreement, penalty clause and technical parameters, and generate text entity features;

[0023] S22, perform OCR recognition and structure relationship extraction on the drawing image, generate image structure features containing component labeling and dimension information;

[0024] S23, perform time series change rate extraction and trend judgment processing on the progress log data, extract key nodes and abnormal change information, and generate time series feature representation;

[0025] S24, input the text entity features, image structure features and time series feature representation into the multi-modal fusion module, perform semantic space alignment, and generate multi-modal encoding representation.

[0026] Further, the S3 includes the following steps:

[0027] The multi-modal encoding representation and the reasoning configuration set are jointly input into a language model constructed based on the Transformer architecture, and the language model includes a subset of strategy parameters for guiding attention distribution and output determination;

[0028] The dynamic adjustment of the attention weight matrix and the position embedding vector is controlled by the subset of strategy parameters, the context modeling of the multi-modal semantic representation is completed, and the context encoding result focusing on the potential risk elements is obtained;

[0029] Based on the context encoding result, semantic matching and rule constraint verification are performed on the text entities, image structures and time series features contained therein, and the rule items in the reasoning configuration set are combined to output a structured risk labeling result.

[0030] Further, the structured risk labeling result is used to generate a structured supervision report containing risk levels, related regulations and modification suggestions, which includes:

[0031] Based on the structured risk labeling result, the risk type, field positioning and violation basis in each risk labeling are extracted, and an initial risk item set is constructed;

[0032] The initial risk item set is subjected to rule comparison and standard matching processing, based on the specification items and historical case information in the reasoning configuration set, the severity level of each risk is determined, and the corresponding regulation number and suggested revision text are supplemented, and an enhanced description result of the risk item is generated;

[0033] The enhanced description result is classified and sorted according to the risk level to generate a structured supervision report.

[0034] Furthermore, the step of constructing a feedback sample set with supervised annotations based on user feedback information to generate a reward value and updating language model parameters through reinforcement learning includes:

[0035] Based on the user adoption results of each risk recommendation in the structured supervision report, the original multimodal encoding representation, the risk annotation results output by the model, and the user feedback annotation are paired to construct a feedback sample set with supervision labels;

[0036] Calculate a corresponding reward value for each sample in the feedback sample set using a reward function, where the reward function is constructed based on semantic deviation, judgment consistency, and task completion between the model output and the user annotation;

[0037] Based on the reward value, a policy gradient algorithm is used to perform parameter optimization on a subset of policy parameters in the Transformer structure, and the parameter update amount is output;

[0038] Based on the parameter update amount, a subset of the strategy parameters in the Transformer structure is replaced and updated.

[0039] Furthermore, the formula of the reward function is as follows:

[0040]

[0041] Among them, R i is the reward value corresponding to the i-th feedback sample; y i Label the user feedback results for the i-th sample; The structured risk annotation results generated by the model for the corresponding sample; y i and Semantic matching degree; y i and The risk level deviation degree; α and β are hyperparameters used to adjust the ratio of positive incentives to negative penalties respectively.

[0042] Furthermore, the semantic matching degree is calculated by cosine similarity.

[0043] A project risk monitoring system based on a large language model, applied to any of the aforementioned project risk monitoring methods based on a large language model, comprising:

[0044] The request parsing module is used to obtain the natural language supervision request input by the user, parse its semantic features, determine the field to which the supervision task belongs, match it with the knowledge graph and rule base of the corresponding field, and generate a task-related reasoning configuration set;

[0045] A data analysis module is configured to obtain unstructured data uploaded by a user, including contract text, drawing images, and progress logs, extract text entities, image structures, and time sequence features, and generate a multi-modal encoding representation by fusion;

[0046] A risk monitoring module is configured to input the multi-modal encoding representation into a pre-trained language model, combine the inference configuration set, perform semantic analysis and rule inference, and generate a structured risk labeling result;

[0047] A self-optimization module is configured to generate a structured supervision report containing risk levels, associated regulations, and modification suggestions based on the structured risk labeling result, construct a feedback sample set with supervised labeling based on user feedback information to generate a reward value, and update language model parameters through reinforcement learning.

[0048] The present application has the following advantages: the present application analyzes the natural language supervision request input by the user, identifies the field to which the supervision task belongs, matches the knowledge graph and rule base of the corresponding field, constructs the inference configuration set, guides the language model to complete the task in the professional knowledge context, and solves the problem of poor knowledge transfer ability in the traditional method. By introducing a multi-modal fusion mechanism, unstructured data such as contract text, drawing images, and progress logs are uniformly encoded to form a multi-modal representation with consistent semantics. By guiding the Transformer architecture language model of the strategy parameter subset, context modeling and rule inference of cross-modal information are realized, and implicit risks such as "whether the construction delay is compliant" which require joint judgment of text and image semantics are effectively identified. Further, the present application does not rely on static rules, but uses a knowledge graph to construct a rule path associated with the task, and through a dynamic rule configuration set, the joint modeling of hierarchical relationship, semantic correlation, and logical dependence is realized, which significantly improves the flexibility and accuracy of risk identification, and overcomes the limitations of traditional systems that cannot adapt to multi-dimensional information fusion reasoning. Finally, in order to enhance the adaptability of the model in the actual engineering scene, the present application uses a reinforcement learning mechanism based on user feedback. By collecting the adoption of the user to the suggestion items in the structured supervision report, a supervised sample is constructed, and a reward signal is calculated, and the strategy parameter subset in the Transformer model is optimized in real time, so that the system can continuously evolve itself according to the supervision practice, and the problem of existing models that the effect decreases over time and cannot be dynamically optimized is solved. BRIEF DESCRIPTION OF DRAWINGS

[0049] Fig. 1 is a step flow chart of a project risk monitoring method based on a large language model in the present application.

[0050] Fig. 2 is a step flow chart of constructing a feedback sample set with supervised labeling based on user feedback information to generate a reward value, and updating language model parameters through reinforcement learning in the present application. DETAILED DESCRIPTION

[0051] See also Figs. 1-2 As shown, the present invention relates to a project risk monitoring method based on a large language model, comprising the following steps:

[0052] S1. Obtain the natural language supervision request input by the user, parse its semantic features, determine the domain to which the supervision task belongs, match it with the knowledge graph and rule base of the corresponding domain, and generate a task-related reasoning configuration set;

[0053] S2. Obtain unstructured data uploaded by users, including contract text, drawing images, and progress logs, extract text entities, image structure, and temporal features, and fuse them to generate a multimodal encoding representation;

[0054] S3. The pre-trained language model of the multimodal encoding representation input is combined with the reasoning configuration set to perform semantic parsing and rule reasoning to generate a structured risk annotation result;

[0055] S4. Based on the structured risk labeling results, a structured supervision report is generated that includes risk levels, related regulations, and modification suggestions. Based on user feedback information, a feedback sample set with supervised labeling is constructed to generate reward values, and the language model parameters are updated through reinforcement learning.

[0056] In some embodiments, the user inputs a natural language supervision request through the system interface: "Please check if there are any violations of the drawing requirements or non-compliance with the delay procedures in the current construction progress". The system first performs word segmentation and semantic vectorization processing on the request, extracts keywords such as "construction progress", "drawing requirements", "delay procedures", etc., and identifies the task domain as "urban rail transit-structure engineering supervision" by combining the pre-set domain semantic vector template. The system further retrieves the corresponding sub-graph nodes in the knowledge graph in this domain, extracts high-relevance semantic entities such as "drawing change approval process", "construction plan node", "contract delay clause", etc., and generates a configuration set for reasoning by combining semantic paths and rule indexes, clearly specifying the rules and logic such as drawing legality constraints, progress deviation thresholds, and delay procedure conditions that need to be checked. Subsequently, the user uploads the contract file PDF, the latest construction drawing scan, and the progress daily report form filled out by the construction unit (in Excel format). The system performs syntactic analysis on the contract PDF, extracts the agreed milestone nodes, penalty clauses, and delay application conditions; performs OCR recognition and structural line extraction on the drawing image, extracts wall numbering, reinforcement arrangement parameters, and modification annotations; and performs time series analysis on the daily report form, extracts the key node construction progress, construction interruption explanation, and fill-in timestamps for the past three weeks. After aligning the above information in the semantic space, it is fused into a unified multi-modal encoding representation, reflecting the cross-modal semantic association of "drawing structural requirements", "contractually agreed plans", and "actual construction performance". In the analysis stage, the system inputs the fused encoding representation and the aforementioned reasoning configuration set into a language model based on the Transformer architecture, which particularly introduces a subset of strategy parameters to guide the attention mechanism to focus on key fields such as structural differences, progress deviations, and violation constraints. For example, when processing the daily report form containing "wall C12 not constructed", the strategy parameter control model automatically focuses on the "C12 wall layout" in the drawing image and the relevant construction period agreement content in the contract file, and automatically identifies that the delay of this item is not explicitly stated in the contract. Finally, the system generates a structured risk annotation result: "wall C12 construction delay without specified delay basis, which may constitute a breach of contract". Further based on this risk annotation, a structured supervision report is automatically generated, pointing out the corresponding legal provisions of the risk and providing a suggestion text: "suggest supplementing the delay application process materials or revising the progress plan". After viewing the report, the user adopts the suggestion, and the system accordingly forms a triple sample of the current multi-modal input, model output, and user feedback annotation, calculates the reward value (e.g., based on semantic consistency and user acceptance), and fine-tunes the strategy parameter subset in the model through policy gradient.Different from the prior art which only processes single-modal information or is based on static rule matching, the scheme introduces a multi-modal semantic fusion and strategy-guided language model collaborative mechanism to realize cross-modal and cross-logical level information reasoning, solves the problem of multi-source information coupling such as "change of drawings without agreed process" which involves contracts, drawings and construction behaviors. At the same time, through the adaptive optimization mechanism driven by user feedback, the model can continuously learn and update the reasoning mode.

[0057] Further, the S1 comprises the following steps:

[0058] S11, performing word segmentation, entity recognition and semantic feature extraction on the natural language supervision request input by the user to obtain a supervision target semantic vector;

[0059] S12, comparing the supervision target semantic vector with a plurality of preset domain semantic vector templates to identify a domain label to which the current task belongs;

[0060] S13, based on the domain label, retrieving corresponding subgraph structure data from a knowledge graph database, and extracting relevant rule items and upper and lower relationship according to the rule index associated with the subgraph to construct a reasoning configuration set.

[0061] In some embodiments, natural language processing is first performed on the text in step S11, specifically including word segmentation, named entity recognition (NER) and context semantic feature extraction. The word segmentation stage divides the input sentence into word items such as "check / current / pile foundation construction / whether / satisfy / design requirements / and / consistent / standard process"; the NER model identifies "pile foundation construction" as an engineering component entity, "design requirements" as a technical parameter entity, and "standard process" as a process compliance term. Subsequently, the system extracts the semantic vector of the whole sentence based on the BERT model, and constructs the supervision target semantic vector reflecting the supervision intention. In step S12, the system predefines a plurality of semantic vector templates in the engineering field, covering standard task intention representations such as "foundation engineering supervision", "structure acceptance supervision", "progress compliance review", etc. The cosine similarity is calculated between the supervision target semantic vector and all templates in turn, and the domain label corresponding to the highest similarity is selected. In this example, the target semantic vector has the highest similarity with the "foundation engineering supervision" template, so it is identified that the current task belongs to this field. After entering step S13, based on the identified domain label "foundation engineering supervision", the corresponding subgraph structure is located from the knowledge graph library. The subgraph includes semantic nodes such as pile foundation type, design parameter, construction technology, acceptance process, and semantic edges between them (such as "require to comply", "should be executed", "belong to the Xth step in the process", etc.). The system uses graph neural networks (such as GCN) to propagate and aggregate the features of the nodes in the subgraph, while performing semantic type screening to filter out entity nodes with high relevance to "pile foundation construction", such as "concrete bored pile", "pore inspection process", "design reinforcement parameter", etc. Then, based on the edge weights between these nodes, a path weight graph is constructed, and the key path is identified using the shortest path search and path attention mechanism to extract the rule index most relevant to the supervision target. These indexes may include rule item numbers such as "pile foundation design drawings should be consistent with detection reports", "poured construction needs to record pouring parameters simultaneously", etc. Finally, the system retrieves the complete rule provisions from the rule library according to the nodes and rule indexes involved in the above path, and limits the scope of application of the rules in combination with the superior-inferior relationship between entities in the graph (such as "pile foundation construction" is a subtask of "foundation engineering"). All extracted information is encapsulated into a structured reasoning configuration set, including the target supervision semantic vector, the graph subgraph structure, the rule path graph and the rule item set, providing context and rule support for the subsequent Transformer language model reasoning stage.

[0062] Further, the S13 comprises the following steps:

[0063] Based on the identified domain label, a corresponding domain subgraph is located from the knowledge graph library, which contains semantic entity nodes and semantic relationship edges related to the domain;

[0064] performing semantic type screening on semantic entity nodes in the field subgraph, extracting a target node set having significant semantic correlation with a supervision target semantic vector, and generating a path weight graph according to semantic edge types between the nodes;

[0065] extracting rule index information associated with the target node according to the path weight graph, the rule index information including a constraint type, an applicable condition, and an inference template identifier;

[0066] retrieving a corresponding rule item from a rule library according to the extracted rule index information, and constructing a rule applicable range in combination with upper-lower relationship information in the knowledge graph;

[0067] encapsulating the target node set, the path weight graph, and the rule item set as an inference configuration set.

[0068] In some embodiments, first, based on the identified domain label "foundation engineering", the system calls the structured stored engineering knowledge graph from the knowledge graph library, and locates the domain subgraph corresponding to the label. The subgraph is represented by a graph database (such as Neo4j) structure, containing semantic entity nodes such as foundation construction process, acceptance specification, design parameter, common risk type, and semantic edges such as "belongs to process", "requirement to meet", "associated acceptance index". Each node has attribute embedding (such as node type, weight label, upper and lower level, etc.), and the semantic edge marks the edge type and semantic directionality. Next, the system performs a semantic type screening operation. By constructing a supervision target semantic vector (generated by the BERT model in S11), the vector space matching mechanism (such as cosine similarity) is used to perform similarity calculation on the embedding vector of each entity node in the graph. The most relevant entity nodes to the supervision target are screened out to form a target node set. For example, if the supervision target is "foundation steel cage installation quality review", the target nodes may include "steel cage positioning error range", "steel lap length specification", "pouring sequence influence factor", etc. Based on the target node set, the system constructs a path weight graph. The graph takes the target node as the center, traverses its adjacent nodes, and assigns edge weights according to the semantic edge type. The edge weight design considers the logical dependence strength of the edge (such as "must follow" with high weight) and the prior importance index configured by the domain expert. Through graph traversal algorithms (such as weighted breadth-first search or attention mechanism based path aggregation algorithm), the paths that have logical dependence on the target node in the semantic sense are identified in the path weight graph, and the rule index information on all paths is extracted. Each rule index contains constraint type (such as numerical threshold, process sequence), applicable condition (such as project type, stage), and its corresponding reasoning template identifier (such as "foundation steel inspection-process type rule"). Subsequently, the system retrieves complete rule items from the rule library according to the above rule index information. The rule library stores structured rule content in the form of "rule ID + template variable", for example: "Rule ID 03-145: When the steel cage length > 12m, segmented hoisting and segmented positioning acceptance should be used." The system aligns the retrieved rule items with the upper and lower relationships of the entity nodes in the knowledge graph, eliminates rules that are not related to the current project stage and subtask, and constructs a rule applicable range with stage applicability. Tree relationship filters or pruning algorithms based on parent node semantic type can be used in this process to improve matching accuracy. Finally, the system encapsulates the target node set, path weight graph, and filtered rule item set into a structured reasoning configuration set. The configuration set is input into the subsequent language model reasoning process in the form of JSON format or tensor dictionary, as context structure data to guide attention distribution, rule verification, and semantic explanation, thereby realizing the deep integration of semantic driven and knowledge reasoning.

[0069] Further, the S2 comprises the following steps:

[0070] S21, perform syntax analysis and entity recognition processing on the contract text uploaded by the user, extract the obligor, time agreement, penalty clause and technical parameters, and generate text entity features;

[0071] S22, perform OCR recognition and structure relationship extraction on the drawing image, generate image structure features containing component labeling and dimension information;

[0072] S23, perform time series change rate extraction and trend judgment processing on the progress log data, extract key nodes and abnormal change information, and generate time series feature representation;

[0073] S24, input the text entity features, image structure features and time series feature representation into the multi-modal fusion module, perform semantic space alignment, and generate multi-modal encoding representation by fusion.

[0074] In some embodiments, first, for the user-uploaded contract text, a Chinese pre-trained language model based on the Transformer structure (such as RoBERTa-wwm-ext) is called to perform sentence segmentation and syntax analysis on the text, and combined with the named entity recognition module (NER), the obligation subject (such as “general contractor”) in the contract, the time agreement (such as “the completion time shall not be later than × years × months”), the penalty clause (such as “one day delay, fine × yuan”) and the technical parameters (such as “bored pile diameter ≥1000mm”) are extracted. In order to enhance the degree of semantic separation, the syntactic dependency tree is introduced, the extracted key entities and their modifier structures are bound into phrase segments, and the context semantic graph is constructed, and finally the text entity feature tensor with entity type, context position index and parameter value field is generated. Next, for the uploaded image data of the drawing, a graph-text joint recognition model based on deep convolutional network (such as CRAFT+LayoutLM) is used for OCR recognition, and at the same time, the structure information such as component number, position label, size line segment and legend description in the image is extracted combined with spatial position embedding. Through the image preprocessing steps (such as edge detection and tilt correction) to improve the accuracy of structure segmentation, the graph neural network (GNN) modeling is performed on the relative spatial structure in the image, and the image structure feature tensor containing component ID, spatial position vector, connection relationship and labeled text is generated. For example, “pile foundation A area φ1000 pile, spacing 3.2m” is identified from the drawing, and is converted into component nodes and attribute description in the graph structure. For the user-provided construction progress log data, input the multi-layer bidirectional gated recurrent network (Bi-GRU) model in time series format, extract the change rate of key indicators (such as concrete pouring amount, completed pile foundation number) in different time windows. Introduce the sliding window mechanism and the first-order difference operator to calculate the local trend value, and based on the set fluctuation threshold, automatically identify the abnormal change nodes such as “lag”, “stop work” or “jump progress”. Through the time attention mechanism, the importance of each node in the overall progress is weighted, and the time series feature vector sequence containing key time points, event types and trend factors is generated. In the feature fusion stage, based on the multi-channel Transformer architecture, the text channel, image channel and time channel are set, and the three types of features are encoded respectively. Through the unified embedding space alignment method, each type of feature is mapped to a unified dimensional vector space; based on the cross-modal attention mechanism (Cross-modal Attention), the dependence relationship of each modality to other modalities in the current context is calculated, and the fusion attention matrix is generated. In order to enhance the semantic consistency, the semantic alignment loss function is added to constrain the distance between the representation vectors of the same semantic entities in different modalities to be minimized. The final output of the multi-modal encoding representation is in the form of a tensor, containing the joint representation semantic chain between “supervision target-contract clause-drawing component-construction progress”, having cross-modal context interaction capability, and serving as the input of the language model reasoning in S3.

[0075] Further, the S3 comprises the following steps:

[0076] The multi-modal encoding representation is inputted into a language model based on the Transformer architecture, which contains a subset of strategy parameters for guiding attention distribution and output determination;

[0077] The dynamic adjustment of attention weight matrix and position embedding vector is controlled by the subset of strategy parameters, completing the context modeling of multi-modal semantic representation, and obtaining the context encoding result focusing on potential risk elements;

[0078] Based on the context encoding result, semantic matching and rule constraint verification are performed on the text entities, image structures and time sequence features contained therein, combined with the rule items in the inference configuration set, and the structured risk labeling result is outputted.

[0079] In some embodiments, first, in the S2 step, the key clauses in the contract text (such as the installation error requirement of ±5 cm), the component size annotations in the drawing image (such as the offset between the equipment foundation annotation size and the actual installation coordinates), and the records of equipment installation completion in the progress log are extracted. These unstructured information from text, image and time sequence are fused and encoded into a unified multi-modal semantic representation. Subsequently, the above multi-modal encoding representation is inputted into the constructed language model together with the inference configuration set generated in the S1 step. The model is based on an improved Transformer structure, and a set of strategy parameters is specially introduced to guide the attention distribution and the judgment process, so that the model can focus more on the fragments related to potential risks when understanding the context. For example, when the model notices that the size of the equipment foundation in the drawing is inconsistent with the offset data in the construction record, and there is no clause in the contract that allows the offset to exceed the range, the model will significantly enhance the attention weight of these information fragments. In the context modeling process, the model not only considers the information dependency relationship within the same modality (such as the logical association between multiple clauses in the contract), but also establishes a collaborative understanding relationship between cross-modal through the guide parameters. For example, the size information on the drawing will be aligned and judged with the deviation limit condition in the contract clause, and the construction status information in the progress log will be checked with the time node agreed in the contract, forming a unified context logical link. Based on the modeling results, the system further utilizes the rule items in the inference configuration set to perform conditional matching and logical verification on the multi-modal context representation. For example: the model identifies that the component offset exceeds the standard, the construction is completed, the drawing is not updated, and there is no change clause in the contract, and matches the rule item “equipment offset without record is illegal”, thereby outputting the structured risk labeling: “the component A foundation offset exceeds the contract specified range, the drawing is not updated, and the construction site is recommended to be reviewed and technical records are supplemented.”

[0080] Furthermore, generating a structured supervision report including risk levels, associated regulations, and modification suggestions based on the structured risk annotation results includes:

[0081] Based on the structured risk annotation results, the risk type, field location, and violation basis in each risk annotation are extracted to construct an initial risk entry set;

[0082] Performing rule comparison and standard matching processing on the initial risk item set, determining the severity level of each risk based on the regulatory items and historical case information in the reasoning configuration set, and supplementing the corresponding regulatory number and suggested revision text to generate an enhanced description result of the risk item;

[0083] The enhanced description results are classified and sorted according to risk levels to generate a structured supervision report.

[0084] In some embodiments, risk factor extraction and semantic enhancement are first performed. Specifically, the system accurately parses and structuredly represents key fields within each risk annotation entry, including risk type (e.g., schedule delay, design discrepancy, contract breach), field location (e.g., clause number, construction node number, drawing component number), and the basis for the violation (e.g., contract clause, industry standard, or construction drawing requirements), to construct a risk semantic vector set. Subsequently, based on the pre-set rule entries in the inference configuration, the industry standard semantic graph, and the historical case database, a semantic similarity metric algorithm (e.g., cosine similarity based on the BERT embedding space) is used to perform rule matching and standard mapping. During this process, the system not only determines the direct correspondence between risk types and standard entries, but also assesses the severity of risk entries through the hierarchical relationships and semantic edge weights in the path-weighted graph. The system then outputs a level label (e.g., low risk, medium risk, high risk), along with the potentially affected construction phases and relevant responsible parties. For each classified risk entry, the system further retrieves the corresponding regulatory entry index information and generates suggested correction text based on the task context. The text is constructed by a pre-trained language model combined with a semantic template generation module to ensure the compliance of the recommended content.

[0085] Furthermore, the step of constructing a feedback sample set with supervised annotations based on user feedback information to generate a reward value and updating language model parameters through reinforcement learning includes:

[0086] Based on the user adoption results of each risk recommendation in the structured supervision report, the original multimodal encoding representation, the risk annotation results output by the model, and the user feedback annotation are paired to construct a feedback sample set with supervision labels;

[0087] According to each sample in the feedback sample set, a corresponding reward value is calculated by a reward function, the reward function is constructed based on semantic deviation between model output and user annotation, judgment consistency and task completion degree;

[0088] Based on the reward value, a parameter optimization is performed on a policy parameter subset in the Transformer structure by using a policy gradient algorithm, and a parameter update amount is output;

[0089] Based on the parameter update amount, the policy parameter subset in the Transformer structure is replaced and updated.

[0090] It should be noted that the structured supervision report-based user feedback information is used to construct a supervised sample set required by reinforcement learning, and the policy parameter subset in the Transformer model is updated accordingly. Specifically, first, the user's processing behavior for each risk suggestion item in the structured supervision report is tracked, including whether to adopt the suggestion, whether to modify the suggestion text, whether to report to the supervision system, etc. The system pairs these user behavior signals with the original input multi-modal encoding representation, the model generated risk annotation result and the final user revised annotation result, thereby constructing a feedback sample set with supervised labels. For each feedback sample, the system introduces a designed reward function to quantitatively evaluate the model behavior. The reward function not only considers the similarity between the model output annotation and the user annotation in the semantic space (for example, by calculating the cosine similarity after BERT embedding), but also quantifies the risk level consistency (for example, if the model judges as high risk but the user adopts as low risk, it is punished) and the task completion degree (for example, if the user completely accepts the model output, a higher positive reward is given). Based on the obtained reward value sequence, the system uses a policy gradient algorithm (Policy Gradient) to perform gradient update on the policy parameter subset in the language model responsible for attention control and output guidance. The specific operation is as follows: by calculating the gradient contribution of each policy parameter under the current policy, combining the positive and negative rewards obtained by sampling, updating the gradient direction and step length corresponding to it, and generating a parameter update amount. The system then uses the parameter update amount to perform replacement update on the original policy parameters, realizing the continuous self-optimization of the language model driven by user interaction feedback.

[0091] Further, the formula of the reward function is as follows:

[0092]

[0093] Wherein, R i is the reward value corresponding to the i-th feedback sample; y i is the user feedback annotation result in the i-th sample; is the structured risk annotation result generated by the model in the corresponding sample; is y i and Semantic matching degree; y i and The risk level deviation degree; α and β are hyperparameters used to adjust the ratio of positive incentives to negative penalties respectively.

[0094] Specifically, this reward function constructs a comprehensive evaluation mechanism for model output quality by jointly considering semantic matching and consistency in risk level judgment. A higher semantic match indicates greater consistency between the model and the user's language in describing risk, resulting in positive incentives. A greater risk level deviation indicates that the model's judgment of risk severity deviates from the user's perception, imposing corresponding penalties. By flexibly adjusting α and β, the emphasis on language consistency and risk judgment accuracy can be adapted to different scenarios.

[0095] Furthermore, the semantic matching degree is calculated by cosine similarity.

[0096] It should be noted that during the user feedback annotation stage, the structured risk annotation results generated by the model are automatically As user annotation y i The initial value of If the user does not modify the annotation result, the system regards it as a complete adoption of the model inference result, assigns the maximum semantic matching degree of 1 to the sample, and then guides the model to strengthen the behavior pattern. If the user modifies the annotation result, the system inputs the model annotation result and the user's final annotation into a unified semantic encoder (such as BERT or the Transformer encoding module shared with the backbone model) to obtain a semantic vector, and then calculates the cosine similarity between the two vectors to obtain the semantic matching degree. Measures the degree of semantic consistency between model inference results and user feedback. The closer this value is to 1, the more consistent the model is with user perception in terms of risk type, field extraction, or contextual judgment. In addition to semantic matching, the system also assesses the deviation between the risk level output by the model and the risk level annotated by the user. The absolute difference between the two is calculated as the risk level deviation, which measures the accuracy of the model in judging risk severity.

[0097] The present invention also includes a project risk monitoring system based on a large language model, which is applied to any of the aforementioned project risk monitoring methods based on a large language model, comprising:

[0098] The request parsing module is used to obtain the natural language supervision request input by the user, parse its semantic features, determine the field to which the supervision task belongs, match it with the knowledge graph and rule base of the corresponding field, and generate a task-related reasoning configuration set;

[0099] a data analysis module configured to obtain unstructured data uploaded by a user, including contract text, drawing images, and progress logs, extract text entities, image structures, and time sequence features, and generate a multi-modal encoding representation by fusion;

[0100] a risk monitoring module configured to input the multi-modal encoding representation into a pre-trained language model, combine the inference configuration set, perform semantic analysis and rule inference, and generate a structured risk labeling result;

[0101] a self-optimization module configured to generate a structured supervision report containing risk levels, associated regulations, and modification suggestions based on the structured risk labeling result, construct a feedback sample set with supervised labeling based on user feedback information to generate a reward value, and update language model parameters through reinforcement learning.

[0102] The above embodiments merely describe the preferred embodiments of the present application, and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements to the technical solutions of the present application made by ordinary engineering technicians in the art shall fall within the protection scope determined by the claims of the present application.

Claims

1. A project risk monitoring method based on a large language model, characterized in that: The following steps are involved: S1. Obtain the natural language supervision request input by the user, parse its semantic features, determine the domain to which the supervision task belongs, match it with the knowledge graph and rule base of the corresponding domain, and generate a task-related reasoning configuration set; S2. Obtain unstructured data uploaded by users, including contract text, drawing images, and progress logs, extract text entities, image structure, and temporal features, and fuse them to generate a multimodal encoding representation; S3. The pre-trained language model of the multimodal encoding representation input is combined with the reasoning configuration set to perform semantic parsing and rule reasoning to generate a structured risk annotation result; S4. Based on the structured risk labeling results, a structured supervision report is generated that includes risk levels, related regulations, and modification suggestions. Based on user feedback information, a feedback sample set with supervised labeling is constructed to generate reward values, and the language model parameters are updated through reinforcement learning.

2. A project risk monitoring method based on a large language model according to claim 1, characterized in that: Said S1 comprises the following steps: S11. Perform word segmentation, entity recognition, and semantic feature extraction on the natural language supervision request input by the user to obtain a supervision target semantic vector; S12, performing a similarity comparison between the supervision target semantic vector and a plurality of preset domain semantic vector templates to identify the domain label to which the current task belongs; S13. Based on the domain label, the corresponding subgraph structure data is retrieved from the knowledge graph library, and according to the rule index associated with the subgraph, the relevant rule entries and hierarchical relationships are extracted to construct an inference configuration set.

3. A project risk monitoring method based on a large language model according to claim 2, characterized in that: The S13 includes the following steps: Based on the identified domain label, locate the corresponding domain subgraph from the knowledge graph library, where the domain subgraph contains semantic entity nodes and semantic relationship edges related to the domain; Performing semantic type screening on the semantic entity nodes in the domain subgraph, extracting a set of target nodes that have significant semantic relevance to the supervision target semantic vector, and generating a path weight graph based on the semantic edge types between the nodes; Extracting rule index information associated with the target node according to the path weight graph, wherein the rule index information includes a constraint type, an applicable condition, and an inference template identifier; According to the extracted rule index information, the corresponding rule entry is retrieved from the rule base, and the rule application scope is constructed by combining the upper and lower relationship information in the knowledge graph; The target node set, the path weight graph, and the rule entry set are encapsulated as an inference configuration set.

4. A project risk monitoring method based on a large language model according to claim 1, characterized in that: The S2 comprises the following steps: S21. Perform syntactic analysis and entity recognition on the contract text uploaded by the user, extract the obligor, time agreement, penalty clauses, and technical parameters, and generate text entity features; S22, performing OCR recognition and structural relationship extraction on the drawing image to generate image structural features containing component annotation and dimension information; S23. Perform time series change rate extraction and trend judgment processing on the progress log data, extract key nodes and abnormal change information, and generate time series feature representation; S24. Input the text entity features, image structure features and temporal feature representations into a multimodal fusion module, perform semantic space alignment, and fuse them to generate a multimodal coding representation.

5. The project risk monitoring method based on a large language model according to claim 1 is characterized in that: The S3 includes the following steps: Inputting the multimodal encoded representation and the inference configuration set into a language model built based on the Transformer architecture, wherein the language model includes a subset of strategy parameters for guiding attention distribution and output decision; By controlling the dynamic adjustment of the attention weight matrix and position embedding vector through a subset of strategy parameters, we can complete the context modeling of multimodal semantic representation and obtain context encoding results that focus on potential risk factors. Based on the context encoding results, semantic matching and rule constraint verification are performed on the text entities, image structures and temporal features contained therein, and the structured risk labeling results are output in combination with the rule entries in the reasoning configuration set.

6. A project risk monitoring method based on a large language model according to claim 5, characterized in that: The structured supervision report generated based on the structured risk labeling results, which includes risk levels, related regulations, and modification suggestions, includes: Based on the structured risk annotation results, the risk type, field location, and violation basis in each risk annotation are extracted to construct an initial risk entry set; Performing rule comparison and standard matching processing on the initial risk item set, determining the severity level of each risk based on the regulatory items and historical case information in the reasoning configuration set, and supplementing the corresponding regulatory number and suggested revision text to generate an enhanced description result of the risk item; The enhanced description results are classified and sorted according to risk levels to generate a structured supervision report.

7. The project risk monitoring method based on a large language model according to claim 5 is characterized in that: The process of constructing a supervised and labeled feedback sample set based on user feedback information to generate a reward value and updating the language model parameters through reinforcement learning includes: Based on the user adoption results of each risk recommendation in the structured supervision report, the original multimodal encoding representation, the risk annotation results output by the model, and the user feedback annotation are paired to construct a feedback sample set with supervision labels; Calculate a corresponding reward value for each sample in the feedback sample set using a reward function, where the reward function is constructed based on semantic deviation, judgment consistency, and task completion between the model output and the user annotation; Based on the reward value, a policy gradient algorithm is used to perform parameter optimization on a subset of policy parameters in the Transformer structure, and the parameter update amount is output; Based on the parameter update amount, a subset of the strategy parameters in the Transformer structure is replaced and updated.

8. A project risk monitoring method based on a large language model according to claim 7, characterized in that: The formula of the reward function is as follows: Among them, R i is the reward value corresponding to the i-th feedback sample; y i Label the user feedback results for the i-th sample; The structured risk annotation results generated by the model for the corresponding sample; y i and Semantic matching degree; y i and The risk level deviation degree; α and β are hyperparameters used to adjust the ratio of positive incentives to negative penalties respectively.

9. A project risk monitoring method based on a large language model according to claim 8, characterized in that: The semantic matching degree is calculated by cosine similarity.

10. A project risk monitoring system based on a large language model, applied to a project risk monitoring method based on a large language model according to any one of claims 1 to 9, characterized in that: include: The request parsing module is used to obtain the natural language supervision request input by the user, parse its semantic features, determine the field to which the supervision task belongs, match it with the knowledge graph and rule base of the corresponding field, and generate a task-related reasoning configuration set; The data parsing module is used to obtain unstructured data uploaded by users, including contract text, drawing images, and progress logs, extract text entities, image structure, and temporal features, and fuse them to generate a multimodal encoding representation; A risk monitoring module, configured to input the multimodal encoding representation into a pre-trained language model, and in combination with the inference configuration set, perform semantic parsing and rule reasoning to generate a structured risk annotation result; The self-optimization module is used to generate a structured supervision report containing risk levels, related regulations and modification suggestions based on the structured risk labeling results; construct a feedback sample set with supervised annotations based on user feedback information to generate reward values, and update the language model parameters through reinforcement learning.

Images