An industrial CAD drawing auditing rule optimization method and system based on self-evolution agent

Abstract

The present application relates to the technical field of CAD drawing auditing rule optimization, in particular to an industrial CAD drawing auditing rule optimization method and system based on self-evolution agent. The method comprises generating a structured feedback sample set based on obtained historical industrial CAD drawing auditing data; performing auditing deviation attribution and sample classification based on the structured feedback sample; constructing a drawing auditing reflection memory based on a large language model based on the attribution result; generating a candidate evolution strategy of drawing auditing rules and auditing strategies based on the drawing auditing reflection memory; and obtaining an effective evolution strategy through gate screening of the candidate evolution strategy. The present application forms a drawing auditing strategy self-evolution closed loop through candidate evolution strategy generation, verification screening and experience library version updating mechanism.

Description

Technical Field

[0001] This invention relates to the field of CAD drawing review rule optimization technology, and in particular to an industrial CAD drawing review rule optimization method and system based on a self-evolving intelligent agent. Background Technology

[0002] Existing industrial CAD drawing review methods typically rely on preset review rules or manually maintained review experience databases. While these methods can improve the efficiency of checking compliance issues such as layers, line types, and dimension formats to some extent, the review rules often require manual adjustment and supplementation when faced with different enterprise drafting standards, different professional drawing types, and constantly changing project specifications. For similar false alarms, omissions, or rule conflicts, the system struggles to automatically extract experience from historical review results and manual review comments, leading to the possibility of repeated erroneous judgments for similar drawings in subsequent reviews.

[0003] In recent years, large language models and intelligent agent technologies have demonstrated strong capabilities in task decomposition, tool invocation, semantic reasoning, and natural language interpretation, providing new technical pathways for industrial CAD drawing review. Through intelligent agent mechanisms, tasks such as drawing parsing, semantic recognition, rule review, conflict arbitration, and report generation can be assigned to different intelligent agents. However, existing intelligent agent review methods often focus on the completion of a single review task. After completing the review, the agent typically only outputs problem conclusions and modification suggestions, lacking a continuous accumulation and reflection mechanism for the review process trajectory, evidence usage, error causes, and human review feedback, making it difficult to form reusable review experience.

[0004] Furthermore, the causes of misjudgments in industrial CAD drawing review are diverse and scenario-dependent. For example, some false alarms may stem from overly strict rules or failure to recognize reserved interface descriptions in the drawings; some false alarms may arise from missing rules, omissions in cross-drawing references, or errors in semantic recognition of graphic elements; and some conflicting conclusions may arise from different agents using different rule interpretations for the same drawing object. If the system cannot attribute and classify these review errors, nor can it transform manual review results into rule corrections, prompt strategy optimizations, or tool call path adjustments, then the review agent will struggle to continuously optimize as project data and enterprise standards change.

[0005] Therefore, there is an urgent need for a self-evolving intelligent agent optimization technology solution for industrial CAD drawing review. This solution should be able to collect historical review trajectories, intelligent agent review conclusions, evidence primitives, manual review results, and false positive / false negative samples. It should perform attribution analysis and sample classification of review errors, and generate reusable reflective memories using a large language model. Simultaneously, it should be able to automatically generate candidate review rules, prompting strategies, tool call chains, and intelligent agent scheduling strategies based on the reflective memories. Furthermore, it should evaluate, filter, and roll back candidate strategies through validation samples, thereby achieving continuous optimization of industrial CAD drawing review rules and intelligent agent review strategies, improving the system's adaptability, accuracy, and maintainability in complex drawing scenarios. Summary of the Invention

[0006] To address the problems in existing industrial CAD drawing review processes, such as review rules relying on manual maintenance, lack of adaptive optimization capabilities in intelligent agent review strategies, difficulty in reusing historical review feedback, and recurring issues of similar false alarms and missed alarms, this invention provides an industrial CAD drawing review rule optimization method and system based on a self-evolving intelligent agent.

[0007] Firstly, the present invention provides a method for optimizing industrial CAD drawing review rules based on a self-evolving intelligent agent, which adopts the following technical solution: An optimization method for industrial CAD drawing review rules based on self-evolving intelligent agents includes: Obtain historical industrial CAD drawing review data; A set of structured feedback samples is generated based on the historical industrial CAD drawing review data obtained; Audit deviation attribution and sample classification based on structured feedback samples; Constructing a reflection and memory system for image appreciation based on attribution results and a large language model; Candidate evolutionary strategies for generating drawing review rules and approval strategies based on review reflection and memory; Effective evolutionary strategies are obtained by gating and screening candidate evolutionary strategies.

[0008] Secondly, an industrial CAD drawing review rule optimization system based on a self-evolving intelligent agent includes: The data acquisition module is configured to acquire historical industrial CAD drawing review data; The structured module is configured to generate a set of structured feedback samples based on the acquired historical industrial CAD drawing review data; The attribution module is configured to perform audit deviation attribution and sample classification based on structured feedback samples; The Reflective Memory module is configured to build a Reflective Memory of Image Review based on a large language model, based on attribution results. The evolution module is configured to generate candidate evolution strategies for review rules and review policies based on the review reflection memory. The filtering module is configured to obtain effective evolutionary strategies by gating the candidate evolutionary strategies.

[0009] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned method for optimizing industrial CAD drawing review rules based on a self-evolving intelligent agent.

[0010] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide the aforementioned method for optimizing industrial CAD drawing review rules based on a self-evolving intelligent agent.

[0011] In summary, the present invention has the following beneficial technical effects: Compared to existing industrial CAD drawing review methods, which suffer from technical bottlenecks such as reliance on manual maintenance of review rules, difficulty in reusing historical review feedback, recurring false positives and false negatives, and a lack of adaptive optimization capabilities in intelligent review strategies, this invention addresses the core need of "self-evolutionary optimization of industrial CAD drawing review rules and intelligent review strategies." It proposes an industrial CAD drawing review rule optimization method, system, device, and storage medium based on a self-evolutionary intelligent agent. This achieves a continuous optimization process from historical review data collection, review deviation attribution, construction of a large language model reflective memory, generation of candidate rules and strategies, verification and screening of candidate strategies, to version updates and rollbacks of the experience base. This effectively improves the adaptability, accuracy, and maintainability of the industrial CAD drawing review system.

[0012] First, this invention solves the problem of traditional audit systems only retaining the final report and making it difficult to trace the process of audit deviation formation by uniformly modeling historical industrial CAD drawing audit data, constructing structured feedback samples from drawing files, audit rules, audit process records, audit results, and manual review feedback. Furthermore, this invention establishes a correspondence between manual review opinions and audit objects, rule bases, evidence locations, and audit processes through a feedback alignment mechanism. This allows historical problems such as false alarms, omissions, grade deviations, and rule interpretation deviations to be transformed into calculable, attributable, and reusable optimized samples, providing a data foundation for subsequent drawing review experience extraction and rule strategy updates.

[0013] Secondly, this invention employs a deviation attribution and sample classification mechanism to probabilistically identify sources of deviation in structured feedback samples, such as missing rules, overly strict rules, rule conflicts, incorrect evidence location, incorrect object semantic matching, omissions in cross-drawing references, and inappropriate audit strategies. It also combines multi-label attribution and category prototype matching to retain the causes of complex deviations. This mechanism avoids the problem of coarse-grained corrections based solely on single manual labels, enabling the system to clearly identify the causes and sample sets corresponding to different types of audit deviations, thereby improving the targeting of subsequent reflective memory generation and strategy evolution.

[0014] Furthermore, this invention introduces a review reflection memory construction mechanism based on a large language model. It weights and semantically summarizes historical samples, manual feedback, and attribution results under different deviation types, generating a structured reflection memory that includes common causes of deviations, applicable drawing scenarios, types of objects involved, rule optimization hints, hint strategy optimization hints, tool call optimization hints, and scheduling strategy optimization hints. Simultaneously, this invention filters out low-consistency or high-risk reflection content through reflection memory quality assessment and index construction, ensuring that historical review experience is preserved in a searchable, matchable, and reusable form, providing a reliable basis for the generation of candidate review rules, hint strategies, tool call chains, and scheduling strategies.

[0015] Finally, this invention forms a self-evolving closed loop for drawing review strategies through candidate evolution strategy generation, verification and screening, and version-based update mechanisms for the experience base. On the one hand, the system can automatically generate candidate review rules, candidate prompt strategies, candidate tool call chains, and candidate review scheduling strategies based on high-quality reflective memory, and comprehensively evaluate their accuracy improvement, false alarm rate changes, false negative rate changes, location consistency, and rule conflict risks through verification samples. On the other hand, the system avoids writing unverified strategies directly into the experience base through gating screening, version merging, update effect monitoring, and anomaly rollback mechanisms, thereby reducing the risk of review performance degradation caused by erroneous evolution and enabling continuous optimization capabilities for industrial CAD drawing review rules and intelligent review strategies.

[0016] In the application of industrial CAD drawing review strategy optimization, the deviation attribution accuracy of the method of this invention is more than 12% higher than that of single-round large language model rule and prompt generation methods; the recurrence rate of similar false positives and the recurrence rate of similar false negatives are reduced by 6% and 6.8% respectively compared with the fixed rule base review method and the manual rule maintenance method. Therefore, this invention, while improving the adaptability of industrial CAD drawing review rules and the efficiency of strategy optimization, can effectively control the risks of writing erroneous strategies and version updates, providing a reliable technical solution for the continuous evolution and engineering deployment of intelligent industrial CAD drawing review systems. Attached Figure Description

[0017] Figure 1This is a schematic diagram of an industrial CAD drawing review rule optimization method based on a self-evolving intelligent agent according to Embodiment 1 of the present invention; Figure 2 This is a schematic diagram comparing the attribution accuracy of various methods in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram comparing the comprehensive review scores of various methods in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram comparing the rule coverage of each method in Embodiment 1 of the present invention; Figure 5 This is a schematic diagram comparing the false positive rates of various methods in Embodiment 1 of the present invention; Figure 6 This is a schematic diagram comparing the recurrence rates of similar missed detections in different methods according to Embodiment 1 of the present invention; Figure 7 This is a schematic diagram comparing the manual maintenance time of various methods in Embodiment 1 of the present invention. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to the accompanying drawings.

[0019] Example 1 Reference Figure 1 This embodiment of an industrial CAD drawing review rule optimization method based on a self-evolving intelligent agent includes: (1) Historical review data and feedback sample construction module The historical drawing review data and feedback sample construction module receives historical data generated during the existing industrial CAD drawing review process and constructs a unified structured feedback sample from scattered drawing files, review rules, process records, review results, and manual review comments. Existing industrial CAD drawing review systems typically only retain the final review report or issue list, lacking a unified record of the correspondence between review objects, triggering rules, evidence locations, review processes, and manual review feedback. This makes it difficult to analyze the causes of false positives, omissions, and rule conflicts. Therefore, this module, through historical data modeling, review process coding, feedback alignment, and sample structuring, forms a set of feedback samples that can be used for subsequent deviation attribution and self-evolutionary optimization.

[0020] 1) Input of historical audit data To enable continuous optimization of drawing review rules and strategies based on historical samples, this module first receives historical industrial CAD drawing review data. This historical data can originate from manual review records, fixed-rule drawing review systems, image recognition-assisted drawing review systems, or other automated review systems; the specific implementation of the pre-review system is not limited. By uniformly representing historical drawing review data from different sources, a consistent data entry point can be provided for subsequent deviation attribution.

[0021] Let the historical industrial CAD drawing review data set be represented as: , in, This represents a collection of historical industrial CAD drawing review data. Indicates the first Historical audit data, This indicates the number of historical review records. Each historical review record consists of drawing files, rule sets, review process records, review results, and manual review feedback, and is represented as follows: , in, Indicates the first The industrial CAD drawing files corresponding to each historical data point. This represents the set of review rules used when reviewing the drawing. This represents the collection of records from the review process. This represents the set of audit results. This represents the set of feedback from manual review. This unified representation groups review data from different sources into the same data structure, providing input for subsequent review process coding.

[0022] 2) Feature coding of audit process and results After obtaining historical drawing review data, it is necessary to further characterize the review process and results. Deviations in industrial CAD drawing reviews are usually related to the reviewed object, rule triggering conditions, evidence location, and review conclusions. If this information is only stored in text or table format, it is difficult to use it directly for automatic attribution. Therefore, this step encodes the review process records and review results into computable vector representations.

[0023] For the The first in the historical drawing review data The audit process record is as follows: (The original record is:) , in, This indicates the drawing object being reviewed. This indicates the audit rule that was triggered or invoked. This indicates the condition for the rule to be judged. Indicates the location of evidence or evidence elements. This represents the preliminary review results. To convert this process record into a computable representation, a process feature vector is constructed: , in, This indicates the record of the review process. eigenvectors, , , , and These represent the encoding functions for objects, rules, conditions, evidence, and results, respectively. This represents the process feature dimension. This feature vector describes the relationship between the object, rules, evidence, and result in a single audit judgment.

[0024] For the The first in the historical data The audit result is shown in the original record as follows: , in, Indicates a problem identifier. Indicate the question type. Indicates the target object of the review. Indicates the basis of the rules, Indicates problem location information. Indicates the severity level of the problem. This indicates the confidence level of the audit results. To align the audit results with process records and feedback, they are encoded as a result feature vector: , in, Indicates the audit results eigenvectors, , , , and These are the coding functions representing the issue type, the audit target, the rule basis, the location information, and the severity level, respectively. This indicates the feature dimension of the review result. This feature is used to establish a correspondence with the feedback from manual review in the future.

[0025] 3) Manual review and feedback alignment After obtaining the characteristics of the audit process and audit results, it is necessary to align the manual review feedback with the corresponding audit results and process records. The manual review feedback contains information such as confirmed issues, false alarms, omissions, level adjustments, and modification suggestions, serving as a key monitoring signal for the system's self-evolution and optimization. If the feedback cannot be mapped to a specific audit object, triggering rule, or evidence location, the source of the problem cannot be determined. Therefore, this step uses feature matching to calculate the alignment relationship between the feedback and the audit results and audit process.

[0026] For the The first in the historical data The original record of the manual review feedback is as follows: , in, This indicates the corresponding drawing object or review issue. Indicates a review label. Indicate the reason for manual review. This indicates a suggestion for manual modification. To calculate its matching relationship with the review records, the feedback is encoded as a feedback feature vector: , in, Indicates manual review and feedback eigenvectors, This represents a text semantic encoding function. This represents the dimension of the feedback features. This vector is used for alignment calculations with the review process and review results.

[0027] To simultaneously consider the correspondence between human feedback, review results, and the review process, a ternary alignment score is defined: , in, Indicate feedback Audit Results and audit process records The ternary alignment score between them , and Let these represent the bilinear matching matrices for feedback-result, result-process, and feedback-process, respectively. Indicates the bias term. This represents the normalized activation function. This formula can simultaneously measure the consistency between feedback and review conclusions, review conclusions and the generation process, and feedback and the rule triggering process.

[0028] Based on the alignment score, records with scores exceeding a threshold are selected to form a feedback alignment set: , in, Indicates the first The set of feedback alignment records in the historical data. This represents the feedback alignment threshold. This set establishes a correspondence between human feedback and the review results and their formation process, providing a basis for constructing structured feedback samples.

[0029] 4) Structured feedback sample output After obtaining the feedback alignment set, the drawing files, rule sets, review process, review results, and manual review feedback need to be integrated into a unified structured feedback sample. This sample should be able to express the object, rule, evidence, conclusion, and manual feedback label of a review judgment, so that subsequent modules can determine whether the sample belongs to a false alarm, omission, level deviation, or rule conflict, and further analyze the reasons for its formation.

[0030] For aligned records Construct structured feedback samples: , in, Indicates by the first Historical data and the first A structured feedback sample generated by manual review and feedback. Represents industrial CAD drawing files. This represents a set of review rules. This indicates a record of the review process. Indicates the review results. This indicates manual review and feedback. , and These represent process characteristics, result characteristics, and feedback characteristics, respectively. This indicates the ternary alignment score. This structured sample fully preserves the correspondence between the reviewer's judgment and the human feedback.

[0031] The feedback sample set consists of structured feedback samples generated from all historical drawing review data: , in, Represents the set of structured feedback samples. This represents the quantity of historical audit data. This set unifies and consolidates drawing objects, rule triggers, audit conclusions, and manual review feedback from historical industrial CAD drawing audits into a calculable sample, providing input for the next module to perform audit deviation attribution and sample classification.

[0032] (2) Review of deviation attribution and sample classification module This module uses the structured feedback sample set output from the previous module as input to identify the causes of problems such as false alarms, missed alarms, grading bias, rule conflicts, and interpretation bias. It also categorizes the samples into bias types that can be used for subsequent reflective memory construction. Bias in industrial CAD drawing review is usually not caused by a single factor; it may originate from multiple aspects simultaneously, such as missing rules, overly strict rules, inaccurate evidence positioning, incorrect object semantic matching, omissions in cross-drawing references, or inappropriate review strategies. Therefore, this module employs a multi-label attribution mechanism and a category prototype matching mechanism to perform probabilistic attribution and classification of feedback samples.

[0033] 1) Feedback label analysis and bias sample screening In the structured feedback sample set, some samples were manually confirmed as correct, while others were false alarms, missed alarms, grade adjustments, or interpretation biases. To ensure that the subsequent attribution process focuses on the samples that need optimization, this step first parses the feedback tags and filters out the sample set with review biases.

[0034] For structured feedback samples From the feedback of manual review Extract feedback tags: , in, Indicates sample The corresponding feedback tags, This refers to the function for extracting feedback tags. The feedback tags include types such as confirmed issues, false alarms, missed alarms, level adjustments, and interpretation biases.

[0035] To filter samples that need to enter the deviation attribution process, a deviation-related label set is defined. And construct a set of biased samples: , in, This represents the set of samples for which bias attribution needs to be performed. This represents the set of bias-related labels, including false positives, false negatives, level adjustments, and explanatory biases. This set excludes confirmed samples that do not require correction, providing input for subsequent bias feature construction.

[0036] 2) Enhanced attribution bias characteristics After obtaining the set of deviation samples, it is necessary to extract features that reflect the source of the deviation. For industrial CAD drawing review, deviations may be related to rule text, review object, evidence location, semantics of human feedback, and the degree of process-outcome alignment. To enhance the interpretability of attribution, this step further constructs rule-evidence consistency features and feedback consistency features based on the process features, outcome features, and feedback features already present in the previous module.

[0037] For the biased sample First, its process features, outcome features, feedback features, and alignment scores are combined to form the basic attribution features: , in, Indicates biased samples Basic attribution characteristics , and These represent the process features, result features, and feedback features generated by the previous module, respectively. This indicates the score for tripartite alignment.

[0038] To further characterize the inconsistencies between rules, evidence, and feedback, a consistency feature vector is constructed: , in, Represents a consistent feature vector. This indicates the semantic matching degree between the review rules and the feedback reasons. This indicates the consistency between the location of the evidence and the feedback target. This indicates consistency between the audit target and the feedback target. This indicates the consistency between the initial review results and the feedback labels. This consistency feature is used to distinguish different sources of bias, such as rule bias, object bias, and evidence bias.

[0039] Based on the basic attribution features and consistency features, the enhanced biased attribution features are obtained: , in, Indicates biased samples Enhanced attribution features This represents the enhanced feature dimension. This feature vector will serve as input for subsequent multi-label attribution and prototype matching.

[0040] 3) Multi-label bias attribution calculation After obtaining the enhanced attribution features, it is necessary to determine the source of the bias corresponding to the sample. Since actual audit bias may have multiple causes, such as missing rules and omissions in cross-drawing references, which may lead to false negatives, and overly strict rules and incorrect object semantic matching, this step adopts multi-label probabilistic attribution instead of single-category attribution.

[0041] Let the set of deviation cause categories be represented as: , in, This represents the set of categories of causes of deviation. Indicates the first Reasons for class bias This indicates the number of categories of deviation causes. These causes include missing rules, overly strict rules, rule conflicts, incorrect evidence location, incorrect object semantic matching, omissions in cross-drawing references, and inappropriate audit strategies.

[0042] To calculate the probability that a sample belongs to each cause of bias, a non-linear multi-label mapping is performed on the enhanced attribution features: , in, The hidden layer attribution representation of biased samples. This represents the first mapping weight matrix. This represents the bias vector. This represents a non-linear activation function. Based on this hidden layer representation, the multi-label bias probability is obtained: , in, This represents the probability vector of the causes of the deviation. This represents the second mapping weight matrix. This represents the bias vector. element-wise Function. The first element in a vector. Each component This indicates that the sample is a cause of bias. The probability of.

[0043] To enhance semantic consistency between attribution results and preset bias categories, this module further introduces category prototype matching. Let the... The prototype vector of the cause of class bias is The matching probability between a sample and each category prototype is expressed as: , in, Indicates sample With the The matching probability of the prototype for the cause of class deviation. Represents cosine similarity. This represents the temperature coefficient. Combining the multi-label probability and prototype matching probability, the fused attribution score is obtained: , in, Indicates the reasons for sample bias Fusion attribution score, This represents the fusion weight between the multi-label probability and the prototype matching probability. This fusion score is used to determine the final set of causes of deviation.

[0044] 4) Attribution results classification output After obtaining the fusion attribution score, each biased sample needs to be assigned to one or more categories of biased cause. This classification result does not simply provide a label, but retains the original sample, feedback label, augmenting features, and probabilities of each biased cause, which are then used by the subsequent large language model to generate more targeted reflective memories.

[0045] Based on the fused attribution score, the sample is defined. The set of reasons for the deviation: , in, Indicates sample The corresponding set of reasons for deviation, This represents the attribution threshold for bias. If the combined attribution score of all bias causes is below the threshold, the category with the highest score is selected as the primary bias cause to ensure that each biased sample has at least one attribution result.

[0046] To unify the representation of samples, labels, and attribution results, a biased attribution record is defined: , in, Indicates biased samples The corresponding attribution records, Indicates feedback label, This indicates enhanced attribution features. This represents a multi-label bias probability vector. This represents the prototype matching probability vector. Represents the fused attribution score vector. This represents the set of causes of deviation.

[0047] The attribution records are divided according to the category of the cause of the deviation, resulting in the first... The sample subset corresponding to the cause of class bias: , in, This indicates a cause of deviation. The attribution sample subset. The audit bias attribution results set consists of sample subsets from all categories: , in, This represents the set of attribution and sample classification results for review deviations. This set clarifies the historical samples, feedback labels, attribution probabilities, and feature representations corresponding to different deviation causes, providing input for the next module to construct a review reflection memory based on a large language model.

[0048] (3) Image review reflection and memory construction module based on large language model This module takes the attribution results of the review deviations output by the previous module as input, and transforms historical samples, manual feedback, and attribution results under different deviation types into reusable review reflection memories. Industrial CAD drawing review deviations have a clear scenario dependency, such as misjudgments of reserved interfaces, overly strict layer rules, omissions in cross-drawing numbering, and evidence location offsets. If these are only saved as discrete samples, subsequent systems will find it difficult to directly reuse the experience gained. Therefore, this module transforms historical review deviations into reflective memories that can be used for the evolution of rules, prompts, toolchains, and scheduling strategies through sample weighting, large language model reflection generation, reflection quality assessment, and memory index construction.

[0049] 1) Weighted organization of attribution samples In the previous module, the biased samples were divided into multiple subsets according to the category of bias cause. To enable the large language model to extract stable patterns from similar biased samples, the samples need to be weighted. Different samples have varying degrees of representativeness; samples with high attribution scores, clear human feedback, and complete evidence are more suitable as the core basis for reflection generation. Therefore, this step calculates sample weights based on the fused attribution score and feedback alignment strength, and constructs the input context for the large language model.

[0050] For the Attribution sample subset corresponding to class bias causes First, calculate the sample Representative score for this category: , in, Indicates attribution samples For the first Representative scores of the causes of class bias Indicates the reasons for sample bias Fusion attribution score, This indicates the feedback alignment score. This represents the completeness score of the human review feedback. This representativeness score is used to control for the contribution of different samples to the generated reflections.

[0051] Based on the representativeness score, for the first Normalized weighting of biased samples: , in, Indicates sample In the Weights in class bias reflection generation This represents the sample weight normalized temperature coefficient. The higher the weight, the more suitable the sample is as a representative sample for this type of bias.

[0052] To construct a reflexive input that is readable by a large language model, the sample content is organized based on sample weights to obtain a weighted reflexive context: , in, Indicates the first The weighted reflective context corresponding to the cause of the deviation includes sample weights, structured feedback samples, a set of deviation causes, and feedback labels. This context will serve as input for the large language model to generate reflective memories of image review.

[0053] 2) Reviewing drawings, reflecting on them, and generating memory. After obtaining the weighted reflective context, a large language model is needed to transform similar deviation samples into reusable review reflection memories. These reflection memories should not merely be restates of individual error samples, but rather summarize the common causes, applicable scenarios, triggering conditions, rule correction directions, prompt strategy optimization directions, and tool call optimization directions for this type of deviation. In this way, historical review deviations can be transformed into the knowledge base for subsequent self-evolving strategy generation.

[0054] To constrain the large language model to generate structured reflections from the perspective of image review rule optimization, the first definition is... Input generated from reflection on the causes of class biases: , in, Indicates the first Reflecting on the causes of class biases to generate input. This indicates a weighted reflection on the context. Indicates the category of the cause of the deviation. This represents a template for generating reflection instructions. This template is used to request the model to output common causes of errors, applicable drawing scenarios, rule optimization suggestions, drawing review strategy suggestions, and tool call optimization suggestions.

[0055] Based on this input, a large language model is invoked to generate reflective text: , in, Indicates the first Reflective texts corresponding to the causes of class deviations This represents the function call used for the large language model generated by reflection. To enable the reflected text to be directly used by subsequent policy evolution modules, it is structured into reflection memory units: , in, Indicates the first The corresponding review and reflection memory unit for the causes of deviations. Indicate the common causes of deviations. Indicates the applicable drawing scenario. Indicates the type of drawing object involved. This indicates a rule optimization suggestion. This indicates an optimization of the suggestion strategy. This indicates a tool call optimization suggestion. This indicates a suggestion for optimizing the review or scheduling strategy. This structured reflective memory will serve as the basis for generating subsequent candidate evolutionary strategies.

[0056] 3) Reflective memory quality assessment After generating reflective memory units, it is necessary to evaluate their reliability, feasibility, and consistency with historical samples. Since the content generated by large language models may suffer from overgeneralization, insufficient evidence, or unenforceable optimization suggestions, failure to screen these units could lead to incorrect updates to subsequent rules or prompting strategies. Therefore, this step scores reflective memories based on four aspects: sample consistency, feasibility, coverage, and conflict risk.

[0057] To assess the consistency between reflective memory and the input sample, a reflective memory quality score is defined: , in, Reflection and memory Quality rating A score indicating the consistency between the content of the reflection and the weighted context of the reflection. This represents the executability score of rules, prompts, tools, and scheduling optimization prompts from reflected memory. Reflective memory on attribution sample subset Coverage score, This indicates a risk score for reflective memory introducing false generalizations or rule conflicts. to Let represent the scoring weight, and satisfy: , in, Indicates the first The weights of each rating component are determined. A higher quality score indicates that the reflective memory is more consistent with historical sample evidence and is more suitable for subsequent self-evolutionary strategy generation processes.

[0058] Filtering effective reflective memory based on quality scoring thresholds: , in, This represents the set of reflective memories from the review of drawings that passed the quality screening process. This represents the threshold for reflective memory quality. This screening process prevents low-quality or high-risk large language model reflections from entering the subsequent evolutionary process.

[0059] 4) Reflective memory vectorization and indexed output After obtaining a valid set of reflective memories, these memories need to be converted into searchable and matchable representations. Subsequent modules, when generating candidate rules, prompting strategies, tool call chains, and scheduling strategies, need to retrieve similar reflective memories based on the current deviation type, drawing scenario, and review object. Therefore, this step involves vectorizing and indexing the reflective memories.

[0060] For reflective memory units Define a memory vector representation: , in, Reflection and memory The vector representation of , This represents a reflective memory encoding function. This represents the dimension of the memory vector. This vector is used for subsequent similar memory retrieval and candidate strategy generation.

[0061] To enable reflective memories to be invoked by subsequent modules, a reflective memory index is constructed: , in, This represents a set of memory indexes for reviewing drawings and reflecting on their use. This represents a structured reflective memory unit. This represents the corresponding memory vector. This indicates a score for the quality of reflected memory. Finally, this module outputs: , in, This represents the output of the review and reflection memory building module. This output will serve as the input for the subsequent self-evolutionary generation module of review rules and intelligent review strategies, used to generate candidate review rules, prompt strategies, tool call chains, and scheduling strategies.

[0062] (4) Self-evolutionary generation module for drawing review rules and intelligent review strategies This module takes the review reflection memory output from the previous module as input and uses it to generate candidate review rules, prompt strategies, tool call chains, and review scheduling strategies based on the reflection memory accumulated from historical review deviations. If rules and strategies in industrial CAD drawing review rely on manual maintenance for a long time, problems such as lagging rule updates, prompt templates that cannot adapt to new scenarios, fixed tool call paths, and insufficient generalization ability of review strategies can easily arise. Therefore, this module achieves the self-evolution of review rules and intelligent review strategies through reflection memory retrieval, candidate rule generation, prompt and tool chain generation, scheduling strategy generation, and unified encapsulation of candidate strategies.

[0063] 1) Reflective memory retrieval and evolutionary context construction In the previous module, the system has already formed a set of reflection memories for image review and its index. To avoid indiscriminately calling all reflection memories, this step first selects reflection memories suitable for the current evolutionary round based on their quality, memory vector similarity, and relevance to the applicable scenario, and then constructs an evolutionary context. This allows the candidate strategy generation process to focus on high-quality, highly relevant historical bias experiences.

[0064] Let the reflective memory index output by the previous module be: , in, This represents a set of memory indexes for reviewing drawings and reflecting on their use. Indicates the first A reflective memory unit, The vector representation of this memory. This indicates the quality score of the reflective memory. This represents the amount of reflective memory. To describe the current experience base for reviewing drawings that needs optimization, the current experience base is defined as: , in, This refers to the previous drawing review experience database. This indicates that there is an existing set of review rules. This indicates that there is already a set of suggestion strategies. This indicates an existing set of tool call chains. This indicates that there is already a set of audit scheduling policies. This indicates the version information of the experience library.

[0065] To measure the correlation between reflective memory and the current optimization needs of the experience base, we first construct the experience base state vector: , in, This represents the state vector of the current drawing review experience base. This represents the experience base state encoding function. Based on this state vector and the reflective memory vector, the evolutionary relevance of reflective memory is calculated: , in, Indicates the first The evolutionary correlation of reflective memory. Represents cosine similarity. This represents the temperature coefficient. This correlation considers both the quality of reflective memory and its fit with the current state of the experience base.

[0066] Reflective memories are filtered based on evolutionary relevance to form an evolutionary context: , in, This represents the set of reflective contexts used for this round of self-evolution. This represents the memory relevance threshold. This context set will serve as input for the generation of subsequent candidate rules, cue strategies, tool call chains, and scheduling strategies.

[0067] 2) Generation of candidate review rules After obtaining the evolutionary context, the rule optimization suggestions in the reflective memory need to be transformed into executable candidate review rules. Industrial CAD drawing review rules typically include elements such as applicable objects, triggering conditions, evidence requirements, judgment thresholds, and severity levels. If only natural language suggestions are generated, they cannot be directly entered into the review system. Therefore, this step extracts the rule modification intent from the reflective memory and generates structured candidate review rules.

[0068] In order to generate candidate rules based on reflective memory, firstly, for each reflective memory... Extracting the optimized vector based on the rules: , in, Indicates the first Each reflective memory corresponds to a rule optimization vector. This indicates the rule suggestion encoding function. This indicates a rule optimization suggestion. Indicates the applicable drawing scenario. This indicates the type of drawing object involved. This vector is used to indicate the direction of the current reflective memory's rule update.

[0069] To generate structured candidate rules, candidate rule parameters are calculated based on the rule optimization vector and the existing rule set: , in, This represents the candidate rule parameter vector. , The rules represent the generation of the weight matrix. , This represents the bias vector. This represents a non-linear activation function. This parameter vector is used to generate the triggering conditions, threshold parameters, and applicable scope of candidate rules.

[0070] Define candidate review rules based on candidate rule parameters: , in, Indicates memory through reflection The generated candidate review rules, Indicates the candidate rule identifier. Indicates the review area to which the rule belongs. Indicates the applicable drawing object. Indicates the triggering condition. Indicates the candidate rule parameters. This indicates the default severity level. This candidate rule will proceed to the subsequent validation and filtering module instead of being directly written to the official experience base.

[0071] 3) Prompt strategy and tool call chain generation After generating candidate review rules, it is also necessary to simultaneously generate matching prompt strategies, tool call chains, and review scheduling strategies. For industrial CAD drawing review, the rules themselves often cannot cover all judgment details. The system also needs to constrain how the model interprets drawing objects, how it reads evidence, and how it avoids misjudgments through prompt strategies, and determine whether to call element query, geometric calculation, text retrieval, or cross-drawing matching tools through tool call chains. Therefore, this step generates corresponding candidate strategies based on the prompt optimization directions, tool optimization directions, and scheduling optimization directions from the reflection and memory.

[0072] To generate candidate cue policies, the cue optimization cue from reflective memory is first encoded into a cue policy vector: , in, This indicates the suggestion policy optimization vector. This indicates the prompt encoding function. This indicates an optimization of the suggestion strategy. Indicate the common causes of deviations. This indicates the applicable drawing scenario. Candidate prompt template parameters are generated based on this vector: , in, Indicates the candidate suggestion strategy parameters. , This indicates that the suggestion strategy generates a weight matrix. , This represents the bias vector. From this, we obtain the candidate suggestion strategy: , in, Indicates the candidate suggestion strategy, Indicates the prompt policy identifier, This indicates the content of the generated prompt template. This prompt policy is used to constrain the methods of problem judgment, evidence reading, and conclusion interpretation during subsequent review processes.

[0073] To generate candidate tool call chains, tool optimization hints from reflective memory are encoded as toolchain state vectors: , in, Indicates the first The initial state vector of each candidate tool call chain. This indicates the tooltip encoding function. This indicates tool call optimization suggestions. Let the set of available tools be... Then the first in the candidate tool call chain The probability of selecting a step tool is: , in, Indicates the first The candidate tool call chain is in the... Step selection tool The probability, Indicates the first Step toolchain state vector, Indicates the selection of the weight matrix in the tool. This represents the bias vector. To ensure the toolchain forms a continuous sequence of calls, the next state is updated based on the selected tool: , in, Indicates the first Step toolchain state vector, This represents the gated loop update function. Indicates the first Embedded representation of the selected tools at each step. Based on the tool selection results at each step, the candidate tool call chain is obtained: , in, Indicates the call chain of candidate tools. Indicates the first The tool called by the step This indicates the length of the tool call chain. This tool call chain guides subsequent review processes on how to sequentially invoke tools such as element lookup, geometric calculation, text retrieval, or cross-drawing matching.

[0074] To supplement the self-evolution capability at the review strategy level, candidate review scheduling strategies are further generated based on scheduling optimization hints from reflective memory: , in, Represents the scheduling policy optimization vector. This indicates the scheduling hint encoding function. This vector indicates a suggestion for optimizing the review or scheduling strategy. Candidate review and scheduling strategies are generated based on this vector. , in, This indicates the candidate review and scheduling strategy. The scheduling strategy generates a weight matrix. This represents the bias vector. This scheduling strategy is used to determine the priority order of rules, prompting strategies, and tool call chains, or the selection of audit paths, in a specific drawing scenario.

[0075] 4) Output of the candidate evolutionary strategy set After generating candidate rules, suggestion strategies, tool call chains, and scheduling strategies, these elements need to be uniformly encapsulated into a candidate evolution strategy. Since a single self-evolutionary update may involve rule conditions, suggestion templates, toolchains, and scheduling strategies simultaneously, if these parts are stored separately, subsequent verification modules will find it difficult to evaluate the overall effect. Therefore, this step organizes different types of candidate update items into a unified candidate evolution strategy and calculates the generation confidence score.

[0076] To unify the expression of rule, hint, toolchain, and scheduling policy updates, candidate evolutionary policies are defined: , in, Indicates memory through reflection The generated candidate evolutionary strategies Indicates the candidate strategy identifier. Indicates the candidate review rules. Indicates the candidate suggestion strategy, Indicates the call chain of candidate tools. This indicates the candidate review and scheduling strategy. This indicates the confidence level for candidate strategy generation. Candidate review and scheduling strategy. This describes the preferred audit path or combination of audit strategies for a specific drawing object or deviation scenario.

[0077] To evaluate the reliability of candidate evolutionary strategies during the generation phase, generation confidence is calculated: , in, This represents the confidence level for generating candidate strategies. The score indicates the consistency between candidate rules and reflective memory. The score indicates the consistency between candidate cueing strategies and reflective memory. The score indicates the consistency between the candidate tool call chain and reflective memory. This indicates the consistency score between candidate scheduling strategies and reflective memory. This represents the conflict risk score between the candidate rule and the existing rule set. to This indicates the generation of confidence weights.

[0078] The candidate evolutionary strategy set consists of all candidate strategies that meet the generation confidence requirements. , in, Represents the set of candidate evolutionary strategies. This represents the confidence threshold for generation. This set will serve as input to the subsequent candidate evolution strategy verification and gating selection modules to further verify whether the candidate strategies can truly reduce the risks of false positives, false negatives, or rule conflicts.

[0079] (5) Candidate evolution strategy verification and gating screening module This module uses the candidate evolutionary strategy set output by the previous module as its main input and calls the aforementioned structured feedback sample set and bias attribution result set to construct a validation sample set for evaluating the effectiveness and safety of the candidate strategies. Since the candidate rules, suggestion strategies, tool call chains, and scheduling strategies generated by the large language model may have risks such as overfitting, rule conflicts, or insufficient generalization, directly writing them into the experience base without validation may lead to a decrease in subsequent review performance. Therefore, this module implements admission control for candidate evolutionary strategies through validation sample construction, candidate strategy replay testing, comprehensive benefit scoring, and gating screening mechanisms.

[0080] 1) Construction of the validation sample set To evaluate whether the candidate evolutionary strategies have a practical improvement effect, a validation sample set covering different deviation types and drawing scenarios needs to be constructed. The validation sample set should include both deviation samples targeted by the candidate strategy and normal samples in adjacent scenarios to determine whether the strategy introduces new misjudgments while correcting the original problems. Therefore, this step constructs the validation sample set from structured feedback samples and deviation attribution results.

[0081] Let the set of structured feedback samples be... The set of bias attribution results is To ensure that the validation samples cover different types of bias, a validation sample set is constructed: , in, This represents the validation sample set. Represents the sample extraction function. Indicates the first The attribution sample subset corresponding to the cause of class bias This indicates the number of samples drawn from this category. Represents the set of unbiased samples. This indicates the number of normal samples drawn. The validation set includes both biased and normal samples, used to evaluate the error-correcting ability and side-effect risk of the candidate strategy.

[0082] To ensure that validation samples have balanced weights across different classes, sample weights are defined as follows: , in, Indicates verification sample The weight, Indicates the category to which the sample belongs. This indicates the number of samples of this category in the validation set. This weight is used in subsequent performance metric calculations to avoid evaluation bias caused by class imbalance.

[0083] 2) Candidate strategy replay test After constructing the validation sample set, it is necessary to test the audit performance before and after the application of the candidate strategies. This process does not directly modify the official experience base; instead, in the validation environment, candidate evolution strategies are temporarily overlaid onto the old experience base to form a candidate experience base, and the audit process is replayed on the validation samples. By comparing the outputs of the old experience base and the candidate experience base, it can be determined whether the candidate strategies truly improve audit performance.

[0084] For candidate evolutionary strategies Construct a temporary candidate experience base: , in, Indicates the superposition of candidate strategies The subsequent temporary candidate experience pool, This indicates a strategy merging operation. This operation is only used during the validation phase and does not directly modify the official experience base.

[0085] To compare the effectiveness of candidate strategies, validation samples were used. The following calculations are performed separately for the review outputs of the old experience base and the candidate experience base: , in, This indicates that the old experience library is used to validate samples. The audit output on the website This represents the audit output of the candidate experience base on the same validation sample. This represents the audit replay function in the verification environment. The output can include audit results such as issue type, location, rule basis, and severity level.

[0086] 3) Calculation of comprehensive benefit score After completing the replay testing of candidate strategies, it is necessary to comprehensively evaluate their suitability for inclusion in the formal experience base based on multiple indicators. For map review rule optimization, simply improving the accuracy of a single type of sample is insufficient; it is also necessary to avoid increasing false positives, false negatives, rule conflicts, and review costs. Therefore, this step calculates the comprehensive benefit score of candidate strategies in terms of review output consistency, false positive rate, false negative rate, location consistency, rule coverage, and strategy risk.

[0087] To accommodate complex review outputs, a review output scoring function is defined: , in, Indicates the audit output With manual verification of labels Consistency score between them Indicates the consistency score of the problem type. This indicates the consistency score in positioning. The rules are based on consistency scoring. Indicates the consistency score of severity level. This indicates the weight of the score.

[0088] Based on the weights of the validation samples, candidate strategies are calculated. The corresponding review output score improvement: , in, This indicates the change in the review output score caused by the candidate strategy. Indicates verification sample The corresponding manual verification label. If... A value greater than zero indicates that the candidate strategy improves the consistency between the audit results and the manual review results compared to the old experience base.

[0089] Similarly, define the changes in false alarm rate and false negative rate: , , in, This indicates the change in false alarm rate caused by the candidate strategy. This indicates the change in the false positive rate. For both of these indicators, smaller values ​​are better; negative values ​​indicate that the candidate strategy has reduced the false positive or false negative rate.

[0090] To comprehensively measure the effectiveness and risk of candidate strategies, a comprehensive benefit score is defined: , in, Represent candidate strategies The overall benefit score, This indicates the change in the accuracy of primitive positioning. This indicates a change in rule coverage. This indicates the risk score of candidate strategies introducing rule conflicts or overfitting. to This indicates the weighting of the benefit score. This score considers both performance improvement and risk constraints, and is used for subsequent gating and screening.

[0091] 4) Gating and candidate strategy output After obtaining the overall benefit score, it is necessary to determine whether the candidate evolutionary strategies meet the conditions for inclusion in the experience base. To prevent low-return or high-risk strategies from entering the formal experience base, this step adopts a gating screening mechanism, requiring candidate strategies to simultaneously meet the conditions of improved accuracy, no worsening of false positives and false negatives, controllable conflict risk, and satisfactory overall benefit.

[0092] Define the gating function for the candidate policy: , in, Represent candidate strategies The gating results This indicates that the threshold for raising the review output score has been increased. Indicates the threshold for changes in false alarm rate. Indicates the threshold for changes in the false negative rate. Indicates the risk threshold. This represents the threshold for comprehensive benefits. Only when... Only when the candidate strategy is validated is it considered to have passed the validation process.

[0093] Based on the gating results, the set of effective evolutionary strategies after filtering is obtained: , in, This represents the set of candidate evolutionary strategies that have passed the validation screening. The strategies in this set have demonstrated effectiveness on the validation samples without introducing significant side effects, and will serve as input for subsequent version updates and rollback modules of the map review experience base.

[0094] (6) Version update and rollback module for the drawing review experience library This module takes the set of valid evolutionary strategies output by the previous module as input, and writes the verified candidate strategies into the drawing review experience base. It also manages the update process, monitors its effectiveness, and handles rollback for anomalies. Since drawing review rules, prompting strategies, tool call chains, and scheduling strategies directly affect subsequent drawing review results, the lack of version control and rollback mechanisms could lead to increased review errors if updated strategies perform poorly in new scenarios. Therefore, this module achieves stable self-evolution of the drawing review experience base through strategy merging, version updates, online effectiveness monitoring, and rollback determination.

[0095] 1) Effective strategy merging and experience item construction After obtaining a set of effective evolutionary strategies, the candidate rules, hint strategies, tool call chains, and scheduling strategies from different strategies need to be merged into experience items that can be written into the experience base. Since different candidate strategies may originate from different bias types, direct appending may cause duplication or conflict. Therefore, this step merges effective strategies while retaining their verification performance and source memory.

[0096] For effective strategies Build experience items: , in, Indicates an effective strategy The constructed drawing review experience items, This indicates the reflective memory corresponding to the strategy. This indicates the overall benefit score. , and These represent the changes in accuracy, false positive rate, and false negative rate during the validation phase, respectively. This indicates the update time. This experience item records the strategy content, source reflections, and verification results.

[0097] The set of empirical terms is formed from all effective strategies: , in, This represents the set of valid experience items to be written to the experience library. This set provides the specific update content for subsequent version updates of the experience library.

[0098] 2) Version updates of the drawing review experience database After obtaining a valid set of experience items, they need to be merged into the current drawing review experience library, and a new version of the experience library needs to be generated. To ensure the traceability of the update process, this step does not directly overwrite the old experience library, but generates a new version based on the old version, and records the rules, prompting strategies, toolchains and scheduling strategies for adding, modifying and replacing.

[0099] Let the current experience base version be The corresponding experience base is Versioned merging is performed based on a set of valid experience items: , in, This indicates a candidate new version experience library. This indicates a versioned merge function. This refers to the old version of the experience library. This represents the set of valid experience items. To make version updates traceable, a version difference record is defined: , in, This indicates the new or modified content in the candidate version compared to the old version. This difference record is used for subsequent effect monitoring and rollback operations when necessary.

[0100] To avoid conflicts between multiple experience items for the same rule or strategy, a combined weight is calculated based on the benefit scores of the experience items: , in, Representing experience items Update weighting in version merge This indicates an update to the temperature coefficient. A higher update weight indicates that the empirical item performed better during the validation phase and has a higher priority during conflict merging.

[0101] 3) Update effect monitoring and stability assessment After generating the candidate new version experience base, its effectiveness needs to be monitored to determine whether its performance is stable in subsequent review tasks or monitoring samples. Even if the candidate strategy has passed the validation set screening, it may still produce side effects under new drawing types or project specifications. Therefore, this step evaluates the accuracy, false positive rate, false negative rate, rule conflict rate, and execution cost of the new version experience base through online monitoring indicators.

[0102] Let the monitoring sample set be Calculate performance vectors under the old version and candidate new version experience bases respectively: , , in, This represents the monitoring performance vector of the old version of the experience base. This represents the monitoring performance vector of the candidate new version experience base. Score This indicates the overall score output by the review process. FPR Indicates the false alarm rate. FNR Indicates the false negative rate. Conflict Indicates the rule conflict rate. Cost This indicates the execution cost.

[0103] To determine whether a candidate new version is overall superior to the old version, a stability gain score is defined: , in, This indicates the stability gain score of the candidate new version's experience base relative to the old version. to This indicates the weight of the monitoring indicator. If the score is positive and exceeds the threshold, it means that the candidate new version is better than the old version in terms of overall performance.

[0104] Define a version acceptance flag based on stability benefit scores: , in, This indicates the acceptance flag for the candidate new version. This indicates the threshold for version update benefits. When... If the candidate experience base is accepted as the new version, then the rollback process will begin.

[0105] 4) Experience base rollback and final output After the stability assessment is completed, if the candidate new version fails to reach the update benefit threshold, or if the false positive rate, false negative rate, or rule conflict rate in the monitoring samples are abnormally high, it needs to be rolled back to the previous stable version. This mechanism can prevent the self-evolution process from introducing erroneous rules or unstable strategies, thus providing a safe boundary for experience base updates.

[0106] Based on the version acceptance identifier, define the final experience base output: , in, This represents the final output of the drawing review experience database. This indicates a candidate new version experience library. This represents the old version's experience base. The piecewise function outputs the new version if the candidate new version passes the stability check; otherwise, it retains the old version.

[0107] To record the reasons for rollback and update status, a version log is generated: , in, Indicates the first This experience base update log records the old version number, new version number, version differences, stability gains, acceptance flag, and performance vectors between the old and new versions. Finally, this module outputs: , in, This output represents the update and rollback module for the drawing review experience base. Through this output, the system can write verified rules, prompting strategies, tool call chains, and scheduling strategies into the experience base while maintaining version traceability and rollback capability, thus achieving continuous self-evolution and optimization of industrial CAD drawing review rules and intelligent review strategies.

[0108] Experimental verification To verify the effectiveness and reliability of the proposed industrial CAD drawing review rule optimization method based on self-evolving intelligent agents in terms of review feedback reuse, deviation attribution, rule and prompt strategy optimization, tool call chain adjustment, and stable updating of the review experience base, an industrial CAD drawing review feedback and strategy evolution dataset was constructed. This dataset originates from typical industrial scenarios such as mechanical manufacturing, piping technology, electrical control, equipment layout, and engineering construction. It includes DWG format drawings, historical review rules, review process records, review results, manual review feedback, rectification records, and different versions of the review experience base, covering typical samples under different drawing types, professional categories, project stages, and sources of review deviations.

[0109] ① Feedback samples for mechanical drawings mainly include issues such as missing dimensions, incorrect hole positions, incomplete title blocks, and non-standard layer usage, used to verify the ability to optimize basic rules; ② Feedback samples for piping and process drawings mainly include issues such as inconsistent pipeline numbers, missing interface connections, incorrect valve directions, and inconsistent cross-drawing references, used to verify the ability to attribute semantic deviations in complex engineering projects; ③ Feedback samples for electrical and control drawings mainly include issues such as conflicting terminal numbers, inconsistent circuit annotations, missing text descriptions, and incorrect adaptation of professional rules, used to verify the self-evolving optimization capabilities of prompting strategies, tool call chains, and audit scheduling strategies. All feedback samples have undergone manual review, resulting in annotation data that includes deviation type, evidence object, rule source, manual processing opinions, and final confirmation results.

[0110] The historical drawing review data in the dataset was first processed through feedback sample structuring and then divided into three task levels based on the source of review deviations: ① Rule deviation samples, including missing rules, overly strict rules, rule conflicts, and incorrect rule application scope; ② Evidence and semantic deviation samples, including errors in element positioning, object semantic matching, misunderstanding of text annotations, and omissions in cross-drawing references; ③ Strategy execution deviation samples, including insufficient prompt constraints, unreasonable tool call order, redundant review paths, and incompatible scheduling strategies. A total of 1,500 industrial CAD drawing samples, 4,600 historical review records, and 3,200 manual review feedback records were constructed. Among them, 1,000 drawings and corresponding feedback were used for drawing review experience initialization and candidate strategy generation, and 500 drawings and corresponding feedback were used for verification and testing. The test samples covered different drawing sizes, feedback densities, deviation types, and rule version statuses.

[0111] To comprehensively evaluate system performance, four comparison methods were selected: ① Manual-RuleUpdate: a method of manually maintaining the rule base, where reviewers manually adjust rules and review instructions based on feedback; ② Static-RuleCAD: a fixed rule base review method, which only completes drawing review based on preset rules and does not adaptively update based on historical feedback; ③ LLM-PromptGen: a single-round large language model rule and prompt generation method, which generates rules or prompt suggestions based on feedback text, but does not perform deviation attribution, candidate verification, or version rollback; ④ AutoEvo-NoRollback: an automatic update method without rollback, which can generate candidate optimization strategies based on feedback, but lacks gating and version stability control; ⑤ The method of this invention achieves continuous optimization of the industrial CAD drawing review experience base through structured review feedback data, attribution of review deviations, generation of large language model reflective memory, evolution of candidate rules / prompts / toolchains / scheduling strategies, verification and screening, and version rollback mechanisms.

[0112] All methods were evaluated using the same historical feedback samples, the same initial rule base, and a unified test data partition. To verify the system's stability in real-world engineering applications, three types of feedback complexity were set: ① Low-complexity feedback, mainly stemming from basic specification issues such as layers, line types, title blocks, scale, and dimension styles; ② Medium-complexity feedback, involving object recognition, annotation association, partial views, block referencing, and rule adaptation issues; ③ High-complexity feedback, involving cross-drawing references, multi-disciplinary rule conflicts, historical version differences, and the superposition of multiple types of deviations.

[0113] The following evaluation metrics are used: ① Deviation Attribution Accuracy (DAA) measures the system's accuracy in identifying false positives, false negatives, and sources of policy execution deviations; ② Audit Overall Score (Score) comprehensively evaluates the overall effectiveness of the optimized strategy in terms of issue type, location, rule basis, and severity level judgment; ③ Rule Coverage Rate (RCR) measures the updated experience base's coverage of typical audit issues and feedback scenarios; ④ False Positive Recurrence Rate (RFPR) measures the proportion of false positives that have been reported that reappear in subsequent audits; ⑤ False Negative Recurrence Rate (RFNR) measures the proportion of missed reports that have been reported that are repeatedly missed in subsequent audits; ⑥ Manual Maintenance Time (M-Time) assesses the manual intervention time required for each round of experience base updates.

[0114] Experimental results show that the method of this invention exhibits good overall performance in the self-evolution task of industrial CAD drawing review strategies. Manual-RuleUpdate relies on human experience for rule adjustment and can handle some explicit feedback, but its update efficiency is low, with a deviation attribution accuracy of 63.5% and an average maintenance time of 86.4 minutes per round. Static-RuleCAD lacks feedback learning ability, with a rule coverage of 70.9%, and similar false positive and false negative recurrence rates of 16.2% and 18.7%, respectively. LLM-PromptGen can generate some rules and suggestions based on feedback text, improving the overall review score to 0.801, but due to the lack of candidate validation and rollback control, the strategy stability is insufficient. AutoEvo-NoRollback can automatically generate candidate optimization strategies with a rule coverage of 84.6%, but it is prone to introducing new rule conflicts under high-complexity feedback conditions. In comparison, the method of this invention achieves a bias attribution accuracy of 91.7%, a comprehensive audit score of 0.903, a rule coverage rate of 92.8%, reduces the false positive rate of similar cases to 6.1%, reduces the false negative rate of similar cases to 6.8%, and reduces manual maintenance time to 24.6 minutes / round, which is significantly better than the comparative methods such as Manual-RuleUpdate, Static-RuleCAD, LLM-PromptGen, and AutoEvo-NoRollback.

[0115] Under high-complexity feedback testing conditions, the method of this invention maintains good stability. Because this invention structurally aligns review process records, issue results, human feedback, and evidence objects, and controls the risk of experience base updates through deviation attribution, large language model reflective memory, candidate strategy verification, and version rollback mechanisms, it can reduce invalid updates and error propagation in scenarios involving cross-drawing references, multi-disciplinary rule conflicts, and complex deviations. Test results show that the overall review score reduction of the method of this invention in high-complexity feedback samples is less than 4.9%, while the reductions for LLM-PromptGen and AutoEvo-NoRollback are approximately 11.6% and 8.9%, respectively, indicating that the method of this invention has stronger robustness and strategy stability in complex industrial CAD drawing review feedback environments.

[0116] In summary, the experimental results verify the effectiveness of the method of this invention in structuring review feedback, attributing review deviations, generating reflective memory from a large language model, screening candidate evolutionary strategies, and updating the experience base version. This method not only reduces the probability of recurring similar false positives and false negatives, but also reduces the cost of manual rule maintenance and improves the continuous optimization capabilities of review rules, prompting strategies, tool call chains, and intelligent review scheduling strategies. It provides an efficient, stable, and rollbackable technical solution for the long-term operation of industrial CAD drawing review systems, enterprise standard adaptation, and the accumulation of review experience.

[0117] Table 1. Comparison of data from different methods under six major indicators. In performance evaluation, Deviation Attribution Accuracy (DAA), Audit Overall Score (Score), and Rule Coverage Rate (RCR) are positive indicators. DAA measures the system's accuracy in identifying false positives, false negatives, and sources of policy execution deviations. Score comprehensively evaluates the overall performance of the optimized audit strategy in terms of problem type, location, rule basis, and severity level judgment. RCR measures the coverage of the updated audit experience base for typical audit problems and feedback scenarios. False Positive Recurrence Rate (RFPR) and False Negative Recurrence Rate (RFNR) are negative indicators, measuring the proportion of previously reported false positives and false negatives that recur in subsequent audits. Manual Maintenance Time (M-Time) is a negative indicator, measuring the time required for manual intervention in each round of audit experience base updates.

[0118] The experimental results are shown in Table 1. Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 , Figure 7As shown in the figure, in the task of optimizing rules and strategies for industrial CAD drawing review, the manual rule base maintenance method can revise some explicit rules based on review feedback, but it relies on the experience of the reviewers, has a long update cycle, and is difficult to systematically reuse historical feedback; the fixed rule base review method has certain applicability to stable drafting specifications, but lacks feedback learning ability and is difficult to adapt to changes in project stages, enterprise specifications, and drawing types; the single-round large language model rule and prompt generation method can generate some optimization suggestions based on feedback text, but lacks deviation attribution, candidate verification, and version rollback mechanisms, and is prone to introducing unstable strategies; the automatic update method without rollback can achieve a certain degree of automatic strategy iteration, but may produce rule conflicts or accumulation of erroneous strategies in complex feedback scenarios. In contrast, the method of this invention, by introducing structured review feedback data, review deviation attribution, large language model reflective memory, candidate rule / prompt / toolchain / scheduling strategy evolution, and verification screening and rollback mechanisms, realizes a complete closed loop from historical feedback collection, deviation cause identification, strategy candidate generation to stable update of the experience base, and has better adaptive optimization capabilities and version stability in complex industrial CAD drawing review environments.

[0119] The results show that the deviation attribution accuracy (DAA) of the method of this invention reaches 91.7%, which is 12.3% higher than the single-round large language model rule and prompt generation method; the overall review score reaches 0.903, and the rule coverage rate (RCR) reaches 92.8%; the false positive rate (RFPR) and false negative rate (RFNR) of the same type are reduced to 6.1% and 6.8% respectively, which are significantly lower than the manual rule base maintenance method, the fixed rule base review method, and the single-round large language model rule and prompt generation method; the manual maintenance time required for each round of experience base update is reduced to 24.6 minutes, which is about 71.5% less than the manual rule base maintenance method. These results demonstrate that the method of this invention has significant advantages in industrial CAD drawing review feedback reuse, deviation attribution, rule and prompt strategy optimization, toolchain adjustment, and experience base versioning updates, verifying its practical application potential in the continuous optimization and engineering deployment of intelligent industrial CAD drawing review systems.

[0120] Example 2 This embodiment provides an industrial CAD drawing review rule optimization system based on a self-evolving intelligent agent.

[0121] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device, the aforementioned method for optimizing industrial CAD drawing review rules based on a self-evolving intelligent agent.

[0122] A terminal device includes a processor and a computer-readable storage medium, the processor being configured to implement various instructions; the computer-readable storage medium being configured to store multiple instructions adapted for loading and execution by the processor of the aforementioned method for optimizing industrial CAD drawing review rules based on a self-evolving intelligent agent.

[0123] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for optimizing industrial CAD drawing review rules based on self-evolving intelligent agents, characterized in that, include: Obtain historical industrial CAD drawing review data; A set of structured feedback samples is generated based on the historical industrial CAD drawing review data obtained; Audit deviation attribution and sample classification based on structured feedback samples; Constructing a reflection and memory system for image appreciation based on attribution results and a large language model; Candidate evolutionary strategies for generating drawing review rules and approval strategies based on review reflection and memory; Effective evolutionary strategies are obtained by gating and screening candidate evolutionary strategies.

2. The method for optimizing industrial CAD drawing review rules based on a self-evolving intelligent agent according to claim 1, characterized in that, The process of generating a structured feedback sample set based on the acquired historical industrial CAD drawing review data includes, whereby the historical industrial CAD drawing review data set is represented as follows: in, This represents a collection of historical industrial CAD drawing review data. Indicates the first Historical audit data, This indicates the number of historical review data entries. Each historical review data entry is represented as: in, Indicates the first The industrial CAD drawing files corresponding to each historical data point. This represents the set of review rules used when reviewing the drawing. This represents a collection of records from the review process. This represents the set of audit results. Represents the set of manual review feedback; for the first... The first in the historical drawing review data The audit process record is as follows: (The original record is:) in, This indicates the drawing object being reviewed. This indicates the audit rule that was triggered or invoked. This indicates the condition for the rule to be judged. Indicates the location of evidence or evidence elements. To represent the preliminary review results, and in order to convert the process records into a computable representation, a process feature vector is constructed: in, This indicates the record of the review process. eigenvectors, , , , and These represent the encoding functions for objects, rules, conditions, evidence, and results, respectively. Represents the process feature dimension; for the first The first in the historical data The audit result is shown in the original record as follows: in, Indicates a problem identifier. Indicate the question type. Indicates the target object of the review. Indicates the basis of the rules, Indicates problem location information. Indicates the severity level of the problem. To represent the confidence level of the audit results, and to align the audit results with process records and feedback, they are encoded as a result feature vector: in, Indicates the audit results eigenvectors, , , , and These are the coding functions representing the issue type, the audit target, the rule basis, the location information, and the severity level, respectively. This indicates the feature dimension of the audit result; for the first... The first in the historical data The original record of the manual review feedback is as follows: in, This indicates the corresponding drawing object or review issue. Indicates a review label. Indicate the reason for manual review. This represents a suggestion for manual modification. To calculate its matching relationship with the review record, the feedback is encoded as a feedback feature vector: in, Indicates manual review and feedback eigenvectors, This represents a text semantic encoding function. This represents the feedback feature dimension; to simultaneously consider the correspondence between human feedback, review results, and the review process, a ternary alignment score is defined: in, Indicate feedback Audit Results and audit process records The ternary alignment score between them , and Let these represent the bilinear matching matrices for feedback-result, result-process, and feedback-process, respectively. Indicates the bias term. This represents the normalized activation function; based on the alignment score, records with scores exceeding a threshold are selected to form a feedback alignment set. , in, Indicates the first The set of feedback alignment records in the historical data. This indicates the feedback alignment threshold; ultimately, for the aligned record... Construct structured feedback samples: in, Indicates by the first Historical data and the first A structured feedback sample generated by manual review and feedback. Represents industrial CAD drawing files. This represents a set of review rules. This indicates a record of the review process. Indicates the review results. This indicates manual review and feedback. , and These represent process characteristics, result characteristics, and feedback characteristics, respectively. The score represents the ternary alignment score; the feedback sample set consists of structured feedback samples generated from all historical drawing review data. in, Represents the set of structured feedback samples. This indicates the number of historical audit data.

3. The method for optimizing industrial CAD drawing review rules based on a self-evolving intelligent agent according to claim 2, characterized in that, The attribution of audit deviations and sample classification based on structured feedback samples include the following: Feedback from manual review Extract feedback tags: in, Indicates sample The corresponding feedback tags, This represents the feedback label extraction function; to filter samples that need to enter the bias attribution process, a set of bias-related labels is defined. And construct a set of biased samples: in, This represents the set of samples for which bias attribution needs to be performed. Represents the set of labels related to deviation; for deviation samples First, its process features, outcome features, feedback features, and alignment scores are combined to form the basic attribution features: in, Indicates biased samples Basic attribution characteristics , and These represent the process features, result features, and feedback features generated by the previous module, respectively. The ternary alignment score is represented; to characterize the inconsistency between rules, evidence, and feedback, a consistency feature vector is constructed: in, Represents a consistent feature vector. This indicates the semantic matching degree between the review rules and the feedback reasons. This indicates the consistency between the location of the evidence and the feedback target. This indicates the consistency between the audit target and the feedback target. This indicates the consistency between the preliminary review results and the feedback labels; based on the basic attribution features and consistency features, the enhanced bias attribution features are obtained: in, Indicates biased samples Enhanced attribution features This represents the enhanced feature dimensions; then, the source of bias corresponding to the sample is determined, using multi-label probabilistic attribution. Let the set of bias cause categories be represented as: in, This represents the set of categories of causes of deviation. Indicates the first Reasons for class bias This represents the number of categories of causes of bias. To calculate the probability that a sample belongs to each cause of bias, a non-linear multi-label mapping is performed on the enhanced attribution features: in, The hidden layer attribution representation for biased samples. This represents the first mapping weight matrix. This represents the bias vector. Representing a nonlinear activation function, the multi-label bias probability is obtained based on the hidden layer representation: in, This represents the probability vector of the causes of the deviation. This represents the second mapping weight matrix. This represents the bias vector. element-wise Function, the first in a vector Each component This indicates that the sample is a cause of bias. To enhance semantic consistency between attribution results and preset bias categories, a category prototype matching is introduced, where the probability is given by the first category. The prototype vector of the cause of class bias is The matching probability between a sample and each category prototype is expressed as: in, Indicates sample With the The matching probability of the prototype for the cause of class deviation. Represents cosine similarity. The temperature coefficient is used to combine the multi-label probability and the prototype matching probability to obtain the fused attribution score: in, Indicates the reasons for sample bias Fusion attribution score, The fusion weights represent the ratio between multi-label probabilities and prototype matching probabilities; based on the fusion attribution score, the sample is defined. The set of reasons for the deviation: in, Indicates sample The corresponding set of reasons for deviation, This represents the bias attribution threshold; to unify the representation of samples, labels, and attribution results, a bias attribution record is defined: , in, Indicates biased samples The corresponding attribution records, Indicates feedback label, This indicates enhanced attribution features. This represents a multi-label bias probability vector. This represents the prototype matching probability vector. Represents the fused attribution score vector. This represents the set of causes of deviation.

4. The method for optimizing industrial CAD drawing review rules based on a self-evolving intelligent agent according to claim 3, characterized in that, The construction of image-reflective memory based on a large language model, based on attribution results, includes calculating sample weights according to the fused attribution score and feedback alignment strength, and constructing the input context of the large language model. Attribution sample subset corresponding to class bias causes First, calculate the sample Representative score for this category: in, Indicates attribution samples For the first Representative scores of the causes of class bias. Indicates the reasons for sample bias Fusion attribution score, This indicates the feedback alignment score. This represents the completeness score of the manual review feedback; based on the representativeness score, the score is used for the first... Normalized weighting of biased samples: in, Indicates sample In the Weights in class bias reflection generation This represents the normalized temperature coefficient of the sample weights; to construct a reflexive input that can be read by a large language model, the sample content is organized based on the sample weights to obtain a weighted reflexive context: in, Indicates the first The weighted reflective context corresponding to the cause of class bias; to constrain the large language model to generate structured reflection from the perspective of image review rule optimization, the first... Input generated from reflection on the causes of class biases: in, Indicates the first Reflecting on the causes of class biases to generate input. This indicates a weighted reflection on the context. Indicates the category of the cause of the deviation. This represents a template for generating reflection instructions; based on this input, a large language model is invoked to generate reflection text. in, Indicates the first Reflective texts corresponding to the causes of class deviations This represents the function call used for the large language model generated by reflection. To enable the reflected text to be directly used by the subsequent policy evolution module, it is structured into reflection memory units: in, Indicates the first The corresponding review and reflection memory unit for the causes of deviations. Indicate the common causes of deviations. Indicates the applicable drawing scenario. Indicates the type of drawing object involved. This indicates a rule optimization suggestion. This indicates an optimization of the suggestion strategy. This indicates a tool call optimization suggestion. This indicates a suggestion for optimizing the review or scheduling strategy.

5. The method for optimizing industrial CAD drawing review rules based on a self-evolving intelligent agent according to claim 4, characterized in that, The construction of image review reflective memory based on a large language model based on attribution results also includes defining a reflective memory quality score to assess the consistency between reflective memory and input samples: in, Reflection and memory Quality rating A score indicating the consistency between the content of the reflection and the weighted context of the reflection. This represents the executability score of rules, prompts, tools, and scheduling optimization prompts from reflected memory. Reflective memory on attribution sample subset Coverage score, This indicates a risk score for reflective memory introducing false generalizations or rule conflicts. to Let represent the scoring weight, and satisfy: in, Indicates the first Weights of each rating component; effective reflective memory is selected based on quality rating thresholds: in, This represents the set of reflective memories from the review of drawings that passed the quality screening process. The threshold for reflective memory quality is represented; then, reflective memory is vectorized and indexed for each reflective memory unit. Define a memory vector representation: in, Reflection and memory The vector representation of , This represents a reflective memory encoding function. Representing the dimension of the memory vector; to enable reflective memory to be invoked by subsequent modules, a reflective memory index is constructed: in, This represents a set of memory indexes for reviewing drawings and reflecting on the process. This represents a structured reflective memory unit. This represents the corresponding memory vector. The final output represents the quality score of the reflected memory. in, This represents the output of the drawing review, reflection, and memory building module.

6. The method for optimizing industrial CAD drawing review rules based on a self-evolving intelligent agent according to claim 5, characterized in that, The candidate evolutionary strategy for generating review rules and review strategies based on review reflection memory includes selecting reflection memories suitable for the current evolutionary round based on reflection memory quality, memory vector similarity, and relevance to the applicable scenario, and constructing an evolutionary context. The reflection memory index is set as follows: in, This represents a set of memory indexes for reviewing drawings and reflecting on the process. Indicates the first A reflective memory unit, The vector representation of this memory. This indicates the quality score of the reflective memory. This represents the number of reflected memories and defines the current experience base as: in, This refers to the previous drawing review experience database. This indicates that there is an existing set of review rules. This indicates that there is already a set of suggestion strategies. This indicates an existing set of tool call chains. This indicates that there is already a set of audit scheduling policies. This represents the version information of the experience base; to measure the correlation between reflective memory and the current optimization needs of the experience base, an experience base state vector is constructed: in, This represents the state vector of the current drawing review experience base. The experience base state encoding function is used to calculate the evolutionary relevance of reflective memory based on the state vector and reflective memory vector. in, Indicates the first The evolutionary correlation of reflective memory. Represents cosine similarity. Representing temperature coefficient; filtering reflective memories based on evolutionary relevance to form evolutionary context: in, This represents the set of reflective contexts used for this round of self-evolution. This represents the threshold for memory relevance; it is used to generate candidate rules based on reflective memory for each reflective memory. Extracting rule-optimized vectors: in, Indicates the first Each reflective memory corresponds to a rule optimization vector. This indicates the rule suggestion encoding function. This indicates a rule optimization suggestion. Indicates the applicable drawing scenario. Indicates the type of drawing object involved; to generate structured candidate rules, candidate rule parameters are calculated based on the rule optimization vector and the existing rule set: in, This represents the candidate rule parameter vector. , The rules represent the generation of the weight matrix. , This represents the bias vector. This represents a non-linear activation function; finally, based on the candidate rule parameters, candidate review rules are defined: in, Indicates memory through reflection The generated candidate review rules, Indicates the candidate rule identifier. Indicates the review area to which the rule belongs. Indicates the applicable drawing object. Indicates the triggering condition. Indicates the candidate rule parameters. This indicates the default severity level.

7. The method for optimizing industrial CAD drawing review rules based on a self-evolving intelligent agent according to claim 6, characterized in that, The candidate evolutionary strategy for generating review rules and review strategies based on review reflection memory also includes, in order to generate candidate suggestion strategies, first encoding the suggestions optimized from the reflection memory into suggestion strategy vectors: in, This indicates the suggestion policy optimization vector. This indicates the prompt encoding function. This indicates an optimization of the suggestion strategy. Indicate the common causes of deviations. Indicates the applicable drawing scenario, and generates candidate prompt template parameters based on vectors: in, Indicates the candidate suggestion strategy parameters. , This indicates that the suggestion strategy generates a weight matrix. , The bias vector is used to derive the candidate suggestion strategy: in, Indicates the candidate suggestion strategy, Indicates the prompt policy identifier, This represents the content of the generated suggestion template; to generate candidate tool call chains, the tool optimization suggestions from reflective memory are encoded into a toolchain state vector: in, Indicates the first The initial state vector of each candidate tool call chain. This indicates the tooltip encoding function. This indicates tool call optimization suggestions, assuming the set of available tools is . Then the first in the candidate tool call chain The probability of selecting a step tool is: in, Indicates the first The candidate tool call chain is in the... Step selection tool The probability, Indicates the first Step toolchain state vector, Indicates the selection of the weight matrix in the tool. This represents the bias vector; to ensure the toolchain forms a continuous call sequence, the next state is updated based on the selected tool: ,in, Indicates the first Step toolchain state vector, This represents the gated loop update function. Indicates the first The embedded representation of the selected tools at each step, based on the tool selection results at each step, yields the candidate tool call chain: in, Indicates the call chain of candidate tools. Indicates the first The tool called by the step Indicates the length of the tool call chain; to supplement the self-evolution capability at the review strategy level, candidate review scheduling strategies are generated based on scheduling optimization hints from reflective memory: , in, Represents the scheduling policy optimization vector. This indicates the scheduling hint encoding function. This vector indicates a suggestion for optimizing the review or scheduling strategy, and a candidate review and scheduling strategy is generated based on this vector: in, This indicates the candidate review and scheduling strategy. The scheduling strategy generates a weight matrix. Representing the bias vector, and finally, to unify the expression of rule, hint, toolchain, and scheduling policy updates, we define candidate evolutionary policies: in, Indicates memory through reflection The generated candidate evolutionary strategies Indicates the candidate strategy identifier. Indicates the candidate review rules. Indicates the candidate suggestion strategy, Indicates the call chain of candidate tools. This indicates the candidate review and scheduling strategy. The generation confidence score represents the confidence score of the candidate strategy. To evaluate the reliability of the candidate evolutionary strategy during the generation phase, the generation confidence score is calculated as follows: in, This represents the confidence level for generating candidate strategies. The score indicates the consistency between candidate rules and reflective memory. The score indicates the consistency between candidate cueing strategies and reflective memory. The score indicates the consistency between the candidate tool call chain and reflective memory. This indicates the consistency score between candidate scheduling strategies and reflective memory. This represents the conflict risk score between the candidate rule and the existing rule set. to This indicates the generation of confidence weights.

8. The method for optimizing industrial CAD drawing review rules based on a self-evolving intelligent agent according to claim 7, characterized in that, The process of obtaining effective evolutionary strategies through gating screening of candidate evolutionary strategies includes constructing a validation sample set from structured feedback samples and bias attribution results, whereby the structured feedback sample set is denoted as . The set of bias attribution results is To ensure that the validation samples cover different types of bias, a validation sample set is constructed: in, This represents the validation sample set. Represents the sample extraction function. Indicates the first The attribution sample subset corresponding to the cause of class bias This indicates the number of samples drawn from this category. Represents the set of unbiased samples. This represents the number of normal samples drawn; to ensure that the validation samples have balanced weights across different categories, sample weights are defined: in, Indicates verification sample The weight, Indicates the category to which the sample belongs. This indicates the number of samples of that category in the validation set; for candidate evolutionary strategies Construct a temporary candidate experience base: in, Indicates the superposition of candidate strategies The subsequent temporary candidate experience pool, This indicates a strategy merging operation; to compare the effectiveness of candidate strategies, validation samples are used. The following calculations are performed separately for the review outputs of the old experience base and the candidate experience base: in, This indicates that the old experience base is used to validate samples. The audit output on the website This represents the audit output of the candidate experience base on the same validation sample. This represents the audit playback function in the verification environment.

9. The method for optimizing industrial CAD drawing review rules based on a self-evolving intelligent agent according to claim 8, characterized in that, The process of obtaining effective evolutionary strategies through gating screening of candidate evolutionary strategies also includes calculating the comprehensive benefit score of the candidate strategies after completing the replay test; to adapt to complex review output, a review output scoring function is defined: in, Indicates the audit output With manual verification of labels Consistency score between them Indicates the consistency score of the problem type. This indicates the consistency score in positioning. The rules are based on consistency scoring. Indicates the consistency score of severity level. Represent the scoring weights; calculate candidate strategies based on the validation sample weights. The corresponding review output score improvement: in, This indicates the change in the review output score caused by the candidate strategy. Indicates verification sample The corresponding manual verification labels; and then the changes in false alarm rate and false negative rate are defined: , in, This indicates the change in false alarm rate caused by the candidate strategy. This indicates the change in the false negative rate; to comprehensively measure the effectiveness and risk of candidate strategies, a comprehensive benefit score is defined: in, Represent candidate strategies The overall benefit score, This indicates the change in the accuracy of primitive positioning. This indicates a change in rule coverage. This indicates the risk score of candidate strategies introducing rule conflicts or overfitting. to The weights for performance evaluation are represented; finally, a gating mechanism is used to define the gating function for candidate strategies: in, Represent candidate strategies The gating results This indicates that the threshold for raising the review output score has been increased. Indicates the threshold for changes in false alarm rate. Indicates the threshold for changes in the false negative rate. Indicates the risk threshold. The threshold represents the overall benefit; based on the gating results, the set of effective evolutionary strategies after screening is obtained: in, This represents the set of candidate evolutionary strategies that have passed the validation screening.

10. An industrial CAD drawing review rule optimization system based on a self-evolving intelligent agent, executing the industrial CAD drawing review rule optimization method based on a self-evolving intelligent agent as described in claim 1, characterized in that, include: The data acquisition module is configured to acquire historical industrial CAD drawing review data; The structured module is configured to generate a set of structured feedback samples based on the acquired historical industrial CAD drawing review data; The attribution module is configured to perform audit deviation attribution and sample classification based on structured feedback samples; The Reflective Memory module is configured to build a Reflective Memory of Image Review based on a large language model, based on attribution results. The evolution module is configured to generate candidate evolution strategies for review rules and review policies based on the review reflection memory. The filtering module is configured to obtain effective evolutionary strategies by gating the candidate evolutionary strategies.

Images