Incremental fine-tuning-based power transmission paper review model continuous optimization method

Abstract

The application provides a power transmission and transformation drawing review model continuous optimization method based on incremental fine-tuning, and belongs to the technical field of intelligent review and artificial intelligence continuous learning of power transmission and transformation projects. The difference labeling four-tuple modeling of the manual review behavior of drawing review is performed, the backflow sample pool is formed through three quality gate screening of qualification verification, consistency verification and diversity filtering, a specification article version difference and conflict resolution mechanism is established, high-value difficult samples are mined from three dimensions of model uncertainty, historical misjudgment rate and article complexity to construct an incremental training set, a LoRA low-rank adaptation method is used to perform parameter efficient fine-tuning on the multi-modal review model, and an EWC regular term is introduced to inhibit catastrophic forgetting, the history case problem mode is clustered and made explicit through the construction of the review experience knowledge graph, and the model performance tracking and closed-loop benefit quantification are completed through the construction of the gold standard verification set, and the model iteration optimization is automatically triggered relying on the event-driven mechanism.

Description

Technical Field

[0001] This invention relates to the fields of intelligent review and continuous learning of artificial intelligence in power transmission and transformation engineering, and in particular to a method for continuous optimization of power transmission and transformation drawing review models based on incremental fine-tuning. Background Technology

[0002] Design review of power transmission and transformation projects is a highly specialized activity that relies heavily on expert experience. An engineer with comprehensive review capabilities typically needs 5 to 10 years of project experience to develop a systematic understanding of various mandatory violation patterns. However, this experience, in the form of tacit knowledge, is attached to individuals and is difficult to standardize and pass on, nor can it be effectively preserved when experts retire or move on.

[0003] From a technical perspective, existing intelligent review systems generally face a common dilemma: the model remains static after deployment and cannot draw feedback from continuously accumulated review practices. This means that even if the system performs well during initial deployment, its review capabilities will inevitably decline as standards are updated, project types expand, and design styles evolve. This problem is particularly prominent in the field of power transmission and transformation engineering because the standard system in this field is frequently updated, and there are significant differences in engineering designs for different voltage levels and in different regions.

[0004] In the existing review process, manual review is an indispensable quality assurance step. The actions of reviewers in confirming, correcting, supplementing, and rejecting the model's output results are essentially precise calibrations of the model's reasoning capabilities. Each correction contains a wealth of knowledge: under what drawing features did the model misjudge, what kind of strong clauses' judgment logic is ambiguous, and how to refine the constraints in a specific engineering context.

[0005] However, existing systems treat these review actions merely as the final stage of the review process, without providing structured records or establishing connections with the model training data. A significant amount of high-value expert judgment information dissipates after the review task is completed, failing to translate into continuous improvement of system capabilities.

[0006] Mandatory provisions are not static. Taking the "Code for Fire Protection Design of Buildings" (GB50016) as an example, it has undergone several comprehensive revisions since its first publication in 1974. Each revision involves numerical adjustments to some mandatory provisions, changes in the scope of application, or the addition of new constraints. Under the current technical framework, updating the code requires a full retraining of the review model, which is costly and time-consuming, resulting in a long-term version lag between the model's knowledge and the current code.

[0007] A deeper problem lies in the fact that specification revision is not a simple text replacement, but involves structural changes to the constraint logic. Semantic conflicts may exist between the old and new versions of the provisions. If not handled properly, this can lead to contradictory constraints in the knowledge base, resulting in logical confusion in model reasoning. Existing systems lack a fine-grained management mechanism for provision version changes, making it impossible to achieve precise and targeted updates to the knowledge base when the specification is updated.

[0008] Historical review reports are stored in scattered PDF documents or Excel spreadsheets, lacking a unified structured presentation. The relationships between issues (co-occurrence patterns of similar issues in different projects, causal chains of specific design errors) cannot be systematically uncovered. This directly leads to two consequences: First, the same type of design error repeatedly appears in different projects, and the review system is unable to extract early warning patterns from historical cases; Secondly, review experience cannot be effectively transferred across projects, and the review quality of each new project still highly depends on the individual experience level of the reviewers involved in that review. Summary of the Invention

[0009] In view of this, the present invention provides a method for continuous optimization of the power transmission and transformation drawing review model based on incremental fine-tuning, which solves the technical problems of static solidification after deployment of existing intelligent review models for power transmission and transformation drawings, inability to structure and reuse the experience of manual review experts, synchronous lag between the updating of standard clauses and the model knowledge base, catastrophic forgetting during incremental training of the model, and difficulty in achieving cross-project experience transfer due to the fragmentation of historical review cases.

[0010] The technical solution proposed in this invention is: a continuous optimization method for a power transmission and transformation drawing review model based on incremental fine-tuning, comprising the following steps: S1 Review Behavior Structured Backflow: Captures manual review operations during the power transmission and transformation drawing review process, encodes them as difference label quadtuples and performs event classification processing, and forms an effective backflow data pool after being screened by three-level quality gating; S2 Standard Knowledge Base Dynamic Update: Establish a version history chain for mandatory clauses of power transmission and transformation projects, execute the clause version differential algorithm to extract change events, detect and resolve semantic conflicts between old and new clauses, and achieve precise and targeted updates of the standard knowledge base; S3 High-Value Training Sample Mining: Based on the three-dimensional difficult sample mining strategy, the comprehensive difficulty score of samples is calculated from the backflow data pool, and incremental training sets are obtained by filtering in descending order of the score. S4 Review Model Incremental Fine-tuning: The low-rank adaptation LoRA method is used to update the local parameters of the pre-trained multimodal review model. Elastic weights are introduced into the incremental fine-tuning loss function to consolidate the EWC regularization term and suppress catastrophic forgetting of the model. S5 Review Experience Knowledge Transfer: Construct a review experience knowledge graph, cluster and make explicit the problem points of historical review cases, generate a key concern list and inject it into the review context when reviewing new projects, and realize cross-project experience transfer; S6 Closed-Loop Quality Continuous Evaluation: Maintain a fixed gold standard validation set, evaluate model performance after each incremental training, quantify closed-loop benefit metrics, and trigger the next round of model optimization process based on the evaluation results.

[0011] Further, the difference-labeled quadruple mentioned in step S1 is represented as follows: , , , ,in: The original output of the model is a JSON object containing a problem description, trigger clauses, coordinates, and confidence score. To manually correct the results, record the field-level differences for each field in the model output; To correct the cause label, the value is an enumeration set {Misjudgment_Graphic misreading, Misjudgment_Textual comprehension deviation, Omission_Semantic occlusion, Omission_Cross-graph association, Coordinate offset, Severity misjudgment}; The sample quality score is calculated by weighting three dimensions: the reviewer's qualification level, the extent of the revision, and the difficulty of the clause.

[0012] Furthermore, the three-level quality gating described in step S1 specifically includes: Gate control 1: Verification of the qualifications of the reviewers; only data reviewed by personnel with registered electrical engineer qualifications or equivalent professional titles will be accepted. Gate control 2: Consistency verification. The Fleiss' Kappa coefficient is used to quantify the consistency of the review results of multiple reviewers of the same drawing. When the coefficient is lower than the preset threshold, the sample is transferred to the expert arbitration queue. Gating 3: Sample diversity filtering. The returned samples are clustered in the feature space, and only samples whose nearest neighbor distance to the existing training set exceeds the diversity threshold δdiv are retained.

[0013] Furthermore, the clause version difference algorithm described in step S2 specifically involves: performing a structured comparison of the clauses in the old and new versions of the specification, and extracting newly added events. Revision Event Three types of change events: deprecation event, invalidation event, and deprecation event. Semantic similarity is calculated using weighted similarity based on the components of each quintuple of clauses, with the weight of the judgment logic component set to 0.4.

[0014] Furthermore, the semantic conflict detection and resolution described in step S2 specifically involves: detecting three types of semantic conflicts: numerical conflicts, range conflicts, and obsolescence-citation conflicts; resolving conflicts according to the priority order of "new version priority > special standard priority > general standard > local standard" and recording conflict processing logs for manual review.

[0015] Furthermore, the three dimensions of the three-dimensional hard sample mining strategy described in step S3 are: Dimension 1: Model uncertainty; Dimension Two: Historical Misjudgment Rate, which is the ratio of the number of historical misjudgments to the total number of reviews for each type of drawing feature combination; Dimension 3: Clause complexity, measured by the number of variables involved in the clause judgment logic and the depth of the logical hierarchy; The comprehensive difficulty score H(xi) is obtained by weighting and fusing the scores from the three dimensions, and the top K samples are used to form the incremental training set.

[0016] Furthermore, the LoRA-based incremental fine-tuning described in step S4 specifically includes: In the pre-trained Qwen-VL model, low-rank adaptation matrices are inserted into the visual encoder cross-attention layer and the language decoder self-attention layer, respectively. The original pre-trained weights are frozen, and only the parameters of the low-rank decomposition matrix are updated. The rank r is set to 16. The incremental fine-tuning loss function is Where Ltask is the training loss for the current incremental task, and Fi is the diagonal element of the Fisher information matrix for parameter θi. λ represents the parameter values ​​after training on the old task, and λ is the forgetting inhibition strength hyperparameter.

[0017] Furthermore, the incremental training in step S4 adopts an event-driven triggering mechanism, with triggering conditions including: reflow sample accumulation triggering: when the number of new samples passing quality gating in the reflow data pool exceeds 500; specification update triggering: when the specification knowledge base executes the clause revision event Emod, targeted incremental training is triggered for the relevant drawing types involving the revised clauses; performance degradation triggering: when the model's strong clause recognition F1 value on the gold standard validation set is lower than 95% of the baseline value for two consecutive evaluations.

[0018] Furthermore, step S5 specifically includes: defining the entity types and relationship types of the review experience knowledge graph, where entities include problem types, specification clauses, drawing types, and project categories, and relationships include triggering, co-occurrence, and mutual exclusion; using a hierarchical clustering algorithm based on BERT embedding to perform semantic clustering of historical problem points, generating standardized problem pattern nodes and knowledge cards; when reviewing new projects, retrieving historical high-frequency problem patterns that match the current project category, voltage level, and drawing type, generating a list of key concerns, and dynamically injecting it into the context layer of the review Prompt.

[0019] Furthermore, the evaluation metrics mentioned in step S6 include: model performance metrics: precision, recall, and F1 score for strong bar identification, which are subdivided and statistically analyzed by strong bar type to generate a capability radar chart; and closed-loop benefit quantification metrics: knowledge accumulation rate, model iteration benefit, standard synchronization timeliness, and experience transfer efficiency, which are displayed in the form of time-series curves.

[0020] The continuous optimization method for the power transmission and transformation drawing review model based on incremental fine-tuning provided by this invention has the following beneficial effects: Transforming tacit expert experience into accumulable structured knowledge: By capturing manual review behavior at the event level through differential labeling quadruples, a structured feedback channel for expert judgment information in the field of power transmission and transformation engineering review was established for the first time, so that the knowledge value of each manual review behavior can be fully utilized.

[0021] Achieving precise and targeted synchronization of specification updates: The clause version differential management mechanism downgrades the specification update response from full retraining to targeted incremental updates, significantly shortening the time cycle from specification update to system capability synchronization and eliminating the risk of version lag.

[0022] Effectively suppressing catastrophic forgetting in continuous learning: The combined application of LoRA and EWC regularization achieves continuous improvement of the review model with extremely low parameter overhead. At the same time, the Fisher information matrix constraint protects existing knowledge from being overwritten by new tasks, thus solving the core technical contradiction of continuous learning in engineering review scenarios.

[0023] Building a knowledge infrastructure for cross-project experience transfer: The review experience knowledge graph transforms fragmented historical cases into structured and related knowledge, supports the dynamic generation of key concern lists, and enables new project reviews to systematically inherit historical experience, fundamentally changing the current situation where review quality depends on personal experience. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments will be briefly described below.

[0025] The accompanying drawings described below are only related to some embodiments of the invention and are not intended to limit the invention.

[0026] In the attached diagram: Figure 1 The overall architecture diagram of the knowledge closed loop of this invention is shown. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the described embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0028] Terminology Explanation: Knowledge loop review: The correction information generated by manual review, the clause change information generated by standard updates, and the problem pattern information accumulated from historical review cases are fed back to the standard knowledge base and reasoning model in a structured manner, forming a continuous iterative mechanism of "review-review-feedback-optimization-re-review".

[0029] Difference-labeled quadruples: Basic data units for structured coding of manual review processes, expressed in the form of... ,in The original output of the model. This is a result of manual correction. To correct the reason label, Score the quality of the sample.

[0030] Incremental Fine-tuning: A training strategy that updates local parameters of a pre-trained model using newly labeled samples without retraining all parameters. This invention employs the Low-Rank Adaptation (LoRA) method, updating only the low-rank decomposition matrix of the model's attention layer, thus preserving existing knowledge while incorporating new review experience.

[0031] Catastrophic forgetting refers to the phenomenon where, when a neural network learns a new task, its ability to perform previously mastered tasks rapidly declines due to parameter overwriting. In the continuous optimization of review models, if the incremental training strategy is inappropriate, the model may lose its ability to recognize existing strong points while absorbing experience from reviewing new projects.

[0032] Hard samples are drawing samples with low model inference confidence, high historical misclassification rates, or involving multiple intersecting constraints. Hard samples contribute significantly more to improving model capabilities than ordinary samples and are the preferred choice for constructing incremental training datasets.

[0033] RegulationVersionDiff: This function performs a structured comparison of the content changes of the same regulation in different versions of the specification, extracts three types of change events: added constraints, revised constraints, and abolished constraints, and triggers targeted updates to the knowledge base.

[0034] Review Experience Knowledge Graph: A semantic network that uses a graph structure to express the relationships between issue points, trigger clauses, drawing types, and project categories in historical review cases. Nodes represent entities (issue types, clauses, equipment types), and edges represent relationships (triggers, co-occurrences, mutual exclusions), supporting cross-project review experience transfer and issue pattern reasoning.

[0035] LoRA (Low-Rank Adaptation): A parameter-efficient fine-tuning method that connects a low-rank decomposition matrix in parallel with the original weight matrix. (in , , This allows for targeted enhancement of model capabilities with minimal trainable parameters.

[0036] To address the aforementioned technical problems, this invention proposes a knowledge closed-loop accumulation and continuous model optimization method for power transmission and transformation engineering design review. The overall technical process of this invention is as follows: Figure 1 As shown, it comprises four core modules: a structured feedback module for review behavior, a dynamic update module for the standard knowledge base, an incremental optimization module for the examination model, and a knowledge graph module for examination experience. These four modules together form a complete knowledge loop, connecting with the examination output of the preceding patent (multimodal mandatory clause examination method).

[0037] Review behavior structured reflow module: Difference-labeled quadruple modeling: The manual review interface captures each action of the reviewer at the event level and encodes the review action as a difference label quadruple: ; The components are defined as follows: Q_i^{model}: The raw output of the model, which is a complete JSON object containing the problem description, trigger clauses, coordinates, and confidence score; Q_i^{human}: Manual correction result, recording the modifications made by the reviewers to the above fields (field-level differences); R_i: Correction reason label, the value is an enumeration set \mathcal{R} = {Misjudgment_Graphic misreading, Misjudgment_Textual comprehension deviation, Omission_Semantic occlusion, Omission_Cross-graph association, Coordinate offset, Severity misjudgment}; S_i: Sample quality score, calculated by weighting three dimensions: reviewer qualification level, scope of revision, and difficulty of the clause. ; in The weights are configured by the system administrator based on the project type.

[0038] Review behavior incident classification and handling: Based on the type of operation performed by the reviewers, the review behavior is divided into five categories, and each category triggers different return processing logic: Confirm The model is correct and no modifications are needed. Positive sample enhancement, update confidence calibration parameters Correct The problem exists, but its description / location is incorrect. Generate corrected sample pairs and add them to the incremental training queue. Supplement The model missed some cases, so new problem points were manually added. Generate missed negative samples, triggering hard sample labeling. Reject The model misjudged the issue; the problem point is invalid. Generate misjudged negative samples to trigger clause interpretation correction. Downgrade The severity was overestimated. Update the severity classifier training data Return data quality gating: Not all review actions are directly fed into the training dataset. The reflow module uses a three-level quality gating system: Gate control 1: The reviewer qualification verification only accepts the review data of the reviewers who have the qualifications of registered electrical engineers or equivalent professional titles, so as to prevent low-quality annotations from polluting the training set.

[0039] Gate Control 2: Consistency Verification. Consistency is verified by checking the results of multiple reviewers of the same drawing, using Fleiss' Kappa coefficient to quantify consistency. ; when If the sample is deemed ambiguous, it will be transferred to the expert arbitration queue and temporarily excluded from the training set.

[0040] Gating 3: Sample diversity filtering performs feature space clustering on the returned samples, calculates the nearest neighbor distance between the new sample and the existing training set, and retains only samples whose distance exceeds the diversity threshold Δ_{div}, thus avoiding the risk of overfitting caused by excessive accumulation of homogeneous samples in the training set.

[0041] Standardized knowledge base dynamic update module: Differential management of clause versions: The specification knowledge base employs a version control mechanism, with each entry maintaining a complete version history chain: ; When a new version of the specification is released, the system executes a clause version differential algorithm to perform a structured comparison of the clauses in the old and new versions, extracting three types of change events: New event E_{add}: A constraint clause that appears in the new version but does not exist in the old version; Revision event E_{mod}: Clauses with the same clause number but with changed constraints, further subdivided into four subcategories: numerical adjustment, scope expansion, scope contraction, and logical modification; Deprecation event E_{del}: A constraint clause that existed in the old version but was deleted in the new version.

[0042] The core of the difference algorithm is the calculation of semantic similarity of the text, which adopts a weighted similarity based on the components of the quintuple: ; Where \alpha_k represents the weight of each component, with the judgment logic component L having the highest weight (\alpha_L = 0.4), as it directly determines the review conclusion.

[0043] Conflict detection and resolution: After the specification is updated, semantic conflicts may arise between the old and new provisions. The conflict detection module checks all provision pairs in the knowledge base. Perform constraint consistency checks to identify the following three types of conflicts: Type 1: Numerical Conflict Two clauses give different constraint thresholds for the same parameter, such as the new clause requiring fire separation distance. The old version of the clause required When the two coexist, a conflict arises.

[0044] Type 2: Scope Conflict The applicable scopes of two clauses, P_i and P_j, overlap, but the constraint logics L_i and L_j contradict each other.

[0045] Type 3: Repeal - Reference Conflict with a Certain Clause Citing repealed provisions This leads to a broken reference chain.

[0046] Conflict resolution strategies are implemented according to priority: new versions first > specific standards first > general standards > local standards. Resolution results are recorded in the conflict handling log for manual review and confirmation.

[0047] Dynamic evaluation of knowledge base coverage: Knowledge base coverage is defined as the ratio of the number of structured, modeled strong rules in the current knowledge base to the total number of currently valid strong rules. ; The system performs a coverage assessment quarterly, identifying areas with coverage below a threshold. Based on the applicant's professional area, knowledge base supplementation tasks are automatically generated and pushed to the knowledge engineer's work queue.

[0048] Review the incremental optimization module of the model: Hard sample mining strategy: The data efficiency of incremental training depends on the quality and diversity of the training samples. This invention designs a three-dimensional hard sample mining strategy to prioritize the selection of high-value training samples from the backflow data pool: Dimension 1: Model Uncertainty. The confidence level of the model output is used as a measure of uncertainty. Samples with low confidence levels correspond to the knowledge boundary region of the model and have higher training value. \text{Uncertainty}(x_i) = 1 - \max_c P(y=c \mid x_i) Dimension Two: Historical Misjudgment Rate Statistics. Historical misjudgment rates are calculated for each type of drawing feature combination (professional category × drawing type × clause type). Samples corresponding to feature combinations with high misjudgment rates are prioritized for selection. ; Dimension 3: Clause complexity is measured by the constraint complexity (number of variables and depth of logic hierarchy) of the judgment logic L_i in the clause quintuple. Samples corresponding to complex clauses are given priority.

[0049] The three-dimensional scores are weighted and fused to generate a comprehensive difficulty score H(x_i). The top K samples, sorted in descending order of score, constitute the incremental training set for this iteration. ; LoRA-based incremental fine-tuning strategy: This invention employs the LoRA method to incrementally fine-tune the Qwen-VL model, inserting low-rank adaptation matrices into the cross-attention layer of the visual encoder and the self-attention layer of the language decoder, respectively. ; in For frozen pre-trained weights, , ,rank Setting it to 16 results in approximately 0.3% of the number of trainable parameters compared to full fine-tuning.

[0050] To suppress catastrophic forgetting, this invention introduces an Elastic Weight Consolidation (EWC) regularization term into the incremental fine-tuning loss function: ; in This is the training loss for the current incremental task. For parameters The diagonal elements of the Fisher information matrix (measure the importance of this parameter to older tasks). These are the parameter values ​​after training the old task was completed. This represents the hyperparameter for the strength of forgetting inhibition.

[0051] Incremental training trigger mechanism: Incremental training does not use fixed-period triggering, but is based on the following three types of event-driven triggering: Event 1: Recycled Sample Accumulation Trigger: When the number of new samples that have passed quality gating in the recycled data pool exceeds the threshold N_{trigger} = 500, the incremental training task is automatically triggered; Event 2: Specification Update Trigger: After the knowledge base executes the clause revision event E_{mod}, targeted incremental training is triggered for the relevant drawing types involving the revised clauses; Event 3: Performance degradation trigger: When the model’s F1 score for strong bar recognition on the validation set is lower than 95% of the baseline value for two consecutive evaluations, an emergency incremental training task is triggered.

[0052] Review experience knowledge graph module: Knowledge graph entity and relation definitions: Review experience knowledge graph Includes the following entity types and relationship types: Entity type \mathcal{E}: Problem Pattern Insufficient fire separation distance for main transformer Frequency of occurrence, average severity Standard Clauses GB50229-5.1.3 Version number, constraint type Equipment type Main transformer, GIS Voltage level, capacity range Drawing type Floor plan, main wiring diagram Professional category, design stage Engineering Category 220kV substation, 110kV substation Voltage level, construction scale Relation type \mathcal{R}_{graph}: Triggers Problem Pattern Standard Clauses This issue violates the clause. Co-occurs Problem Pattern Problem Pattern Both types of problems appear simultaneously on the same drawing. Involves Problem Pattern Equipment type The issue pertains to this device. Highly prevalent in (prevalent_in) Problem Pattern Engineering Category This problem occurs frequently in this type of project. Associated with Problem Pattern Drawing type This problem mainly occurs in this type of drawing. Problem pattern clustering and explicitation: Semantic clustering is performed on the issue points in historical review cases, grouping semantically similar issue points into issue pattern nodes. Clustering employs a hierarchical clustering algorithm based on BERT embeddings, using the semantic vectors of the issue description text as input and Ward distance as the clustering criterion. ; After manual review, the clustering results are entered into the database under standardized problem pattern names. The frequency of occurrence, distribution of engineering categories involved, and set of related clauses for each problem pattern are statistically analyzed to form a problem pattern knowledge card that can be consulted by reviewers.

[0053] Cross-project experience transfer: When a new project is initiated for review, the system retrieves historical high-frequency problem patterns from the knowledge graph that match the current project category, voltage level, and drawing type, generates a "key focus list," and dynamically injects it into the review prompt's context layer to guide the model to prioritize the review of historically frequent problems. ; This mechanism transforms historical review experience from tacit knowledge into actionable reasoning priors, enabling experience transfer across projects.

[0054] Closed-loop quality assessment module: Continuous tracking of model performance: The system maintains a fixed Gold Standard Set, containing sample drawings consistently annotated by multiple senior review engineers, covering all strength bar types. After each incremental training iteration, the following metrics are evaluated on the validation set: ; Each indicator is subdivided and statistically analyzed according to the type of strong clause, generating a capability radar chart to intuitively present the distribution and changing trend of the model's capabilities across different clause types.

[0055] Closed-loop return quantification: The actual benefits of the closed-loop mechanism are quantified and evaluated using the following indicators: Knowledge accumulation rate: The number of effective backflow samples passing through quality gating per unit time; Model iteration benefit: The improvement in F1 score after each incremental training step, denoted as Delta F_1; Specification synchronization timeliness: The average time interval between specification updates being published to the knowledge base and completing the update; Experience transfer effectiveness: Focus on the proportion of problem patterns in the checklist that are actually triggered in the current project.

[0056] The above indicators are continuously displayed in the management dashboard in the form of time-series curves, supporting the system operations team in evaluating the effectiveness of the closed-loop mechanism and optimizing its parameters.

[0057] The following points should be noted in this article: 1. The accompanying drawings of the embodiments of the present invention only involve the structures involved in the embodiments of the present invention; other structures can refer to general designs.

[0058] 2. Where there is no conflict, the embodiments of the present invention and the features thereof can be combined with each other to obtain new embodiments.

[0059] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for continuous optimization of a power transmission and transformation drawing review model based on incremental fine-tuning, characterized in that, Includes the following steps: S1 Review Behavior Structured Backflow: Captures manual review operations during the power transmission and transformation drawing review process, encodes them as difference label quadtuples and performs event classification processing, and forms an effective backflow data pool after being screened by three-level quality gating; S2 Standard Knowledge Base Dynamic Update: Establish a version history chain for mandatory clauses of power transmission and transformation projects, execute the clause version differential algorithm to extract change events, detect and resolve semantic conflicts between old and new clauses, and achieve precise and targeted updates of the standard knowledge base; S3 High-Value Training Sample Mining: Based on the three-dimensional difficult sample mining strategy, the comprehensive difficulty score of samples is calculated from the backflow data pool, and incremental training sets are obtained by filtering in descending order of the score. S4 Review Model Incremental Fine-tuning: The low-rank adaptation LoRA method is used to update the local parameters of the pre-trained multimodal review model. Elastic weights are introduced into the incremental fine-tuning loss function to consolidate the EWC regularization term and suppress catastrophic forgetting of the model. S5 Review Experience Knowledge Transfer: Construct a review experience knowledge graph, cluster and make explicit the problem points of historical review cases, generate a key concern list and inject it into the review context when reviewing new projects, and realize cross-project experience transfer; S6 Closed-Loop Quality Continuous Evaluation: Maintain a fixed gold standard validation set, evaluate model performance after each incremental training, quantify closed-loop benefit metrics, and trigger the next round of model optimization process based on the evaluation results.

2. The method for continuous optimization of the power transmission and transformation drawing review model based on incremental fine-tuning according to claim 1, characterized in that, The difference-labeled quadruple in step S1 is represented as follows: , , , ,in: The original output of the model is a JSON object containing a problem description, trigger clauses, coordinates, and confidence score. To manually correct the results, record the field-level differences for each field in the model output; To correct the cause label, the value is an enumeration set {Misjudgment_Graphic misreading, Misjudgment_Textual comprehension deviation, Omission_Semantic occlusion, Omission_Cross-graph association, Coordinate offset, Severity misjudgment}; The sample quality score is calculated by weighting three dimensions: the reviewer's qualification level, the extent of the revision, and the difficulty of the clause.

3. The method for continuous optimization of the power transmission and transformation drawing review model based on incremental fine-tuning according to claim 1, characterized in that, The three-level quality gating mentioned in step S1 specifically includes: Gate control 1: Verification of the qualifications of the reviewers; only data reviewed by personnel with registered electrical engineer qualifications or equivalent professional titles will be accepted. Gate control 2: Consistency verification. The Fleiss' Kappa coefficient is used to quantify the consistency of the review results of multiple reviewers of the same drawing. When the coefficient is lower than the preset threshold, the sample is transferred to the expert arbitration queue. Gating 3: Sample diversity filtering. The returned samples are clustered in the feature space, and only samples whose nearest neighbor distance to the existing training set exceeds the diversity threshold δdiv are retained.

4. The method for continuous optimization of the power transmission and transformation drawing review model based on incremental fine-tuning according to claim 1, characterized in that, The specific steps of the clause version difference algorithm in step S2 are as follows: perform a structured comparison of the clauses in the old and new versions of the specification, and extract newly added events. Revision Event Three types of change events: deprecation event, invalidation event, and deprecation event. Semantic similarity is calculated using weighted similarity based on the components of each quintuple of clauses, with the weight of the judgment logic component set to 0.

4.

5. The method for continuous optimization of the power transmission and transformation drawing review model based on incremental fine-tuning according to claim 1, characterized in that, The semantic conflict detection and resolution described in step S2 specifically involves: detecting three types of semantic conflicts: numerical conflicts, range conflicts, and obsolescence-reference conflicts; resolving conflicts according to the priority of "new version first > special standard first > general standard > local standard" and recording conflict processing logs for manual review.

6. The method for continuous optimization of the power transmission and transformation drawing review model based on incremental fine-tuning according to claim 1, characterized in that, The three dimensions of the three-dimensional hard sample mining strategy described in step S3 are: Dimension 1: Model uncertainty; Dimension Two: Historical Misjudgment Rate, which is the ratio of the number of historical misjudgments to the total number of reviews for each type of drawing feature combination; Dimension 3: Clause complexity, measured by the number of variables involved in the clause judgment logic and the depth of the logical hierarchy; The comprehensive difficulty score H(xi) is obtained by weighting and fusing the scores from the three dimensions, and the top K samples are used to form the incremental training set.

7. The method for continuous optimization of the power transmission and transformation drawing review model based on incremental fine-tuning according to claim 1, characterized in that, The incremental fine-tuning based on LoRA mentioned in step S4 specifically refers to: In the pre-trained Qwen-VL model, low-rank adaptation matrices are inserted into the visual encoder cross-attention layer and the language decoder self-attention layer, respectively. The original pre-trained weights are frozen, and only the parameters of the low-rank decomposition matrix are updated. The rank r is set to 16. The incremental fine-tuning loss function is Where Ltask is the training loss for the current incremental task, and Fi is the diagonal element of the Fisher information matrix for parameter θi. λ represents the parameter values ​​after training on the old task, and λ is the hyperparameter for forgetting inhibition strength.

8. The method for continuous optimization of the power transmission and transformation drawing review model based on incremental fine-tuning according to claim 1, characterized in that, In step S4, incremental training adopts an event-driven triggering mechanism. The triggering conditions include: reflow sample accumulation trigger: when the number of new samples that have passed the quality gating in the reflow data pool exceeds 500; specification update trigger: when the specification knowledge base executes the clause revision event Emod, targeted incremental training is triggered for the relevant drawing types involving the revised clauses; performance degradation trigger: when the model's strong clause recognition F1 value on the gold standard validation set is lower than 95% of the baseline value for two consecutive evaluations.

9. The method for continuous optimization of the power transmission and transformation drawing review model based on incremental fine-tuning according to claim 1, characterized in that, Step S5 specifically includes: defining the entity types and relationship types of the review experience knowledge graph. Entities include problem types, specification clauses, drawing types, and project categories. Relationships include triggering, co-occurrence, and mutual exclusion. A hierarchical clustering algorithm based on BERT embedding is used to perform semantic clustering of historical problem points to generate standardized problem pattern nodes and knowledge cards. When reviewing new projects, historical high-frequency problem patterns that match the current project category, voltage level, and drawing type are retrieved, a list of key concerns is generated, and dynamically injected into the context layer of the review Prompt.

10. The method for continuous optimization of the power transmission and transformation drawing review model based on incremental fine-tuning according to claim 1, characterized in that, The evaluation metrics mentioned in step S6 include: model performance metrics: precision, recall, and F1 score for strong bar identification, which are subdivided and statistically analyzed by strong bar type and a capability radar chart is generated; closed-loop benefit quantification metrics: knowledge accumulation rate, model iteration benefit, standard synchronization timeliness, and experience transfer efficiency, which are displayed in the form of time series curves.

Images