Power transmission infrastructure daily plan risk level intelligent correction method and device, medium

Abstract

The application discloses a power transmission and transformation infrastructure daily plan risk level intelligent correction method and device, medium, and relates to the infrastructure daily plan management and control technical field.The method comprises the following steps: obtaining a to-be-corrected query text based on infrastructure daily plan data, and performing first retrieval on the to-be-corrected query text in a risk knowledge base to obtain an initial candidate knowledge set; encoding the to-be-corrected query text and the initial candidate knowledge set, and inputting the encoding result into a correction model to obtain a first intermediate correction result; based on the first intermediate correction result and the to-be-corrected query text, performing dynamic reordering on a candidate clause subset in the initial candidate knowledge set; inputting the candidate clauses with a candidate number in the dynamically reordered candidate knowledge set and the to-be-corrected query text into the correction model to obtain a second round correction result.The application realizes faster and more accurate intelligent correction of the risk level and risk number corresponding to the daily plan work content.

Description

Technical Field

[0001] This invention relates to the field of infrastructure daily plan management technology, specifically to a method, equipment, and medium for intelligently verifying the risk level of daily plans for power transmission and transformation infrastructure. Background Technology

[0002] During the daily planning and management of infrastructure projects, business personnel need to review and verify the risk levels and risk numbers corresponding to the "work content" in the daily plan, based on the "Regulations for Construction Safety Risk Management of Power Transmission and Transformation Projects," risk level tables, and relevant system requirements, to determine whether the risk levels reported in the plan are accurate. In existing technologies, to improve review efficiency, a search-based augmented generation method is typically used for auxiliary judgment. This method generally includes the following process: First, the work content text is obtained and encoded; then, candidate clauses related to the work content are retrieved from the risk knowledge base; next, the candidate clauses and the original work content are input into the generation model; finally, the generation model outputs the risk level judgment result or corresponding explanation.

[0003] Compared to purely manual review, the above-mentioned approach can reduce the workload of reviewing each clause of the regulations to some extent, and utilize a knowledge base to constrain the model output. However, in the specific scenario of verifying the risk level of daily infrastructure construction plans, the existing technology still has the following obvious shortcomings.

[0004] First, the accuracy of the initial search is limited. The work content in the daily plan is usually filled out by on-site personnel, containing a large amount of colloquial, abbreviated, and fragmented expressions, such as "tower assembly," "firefighting water system removal," and "attachment preparation." These expressions are not entirely consistent with the standard terminology in the regulations. Existing search enhancement generation methods typically perform the initial search directly based on the original input text. When there is a significant semantic deviation between the input expression and the standard clauses, the initial search may fail to accurately recall truly relevant risk clauses.

[0005] Second, the candidate clause ranking method is static and lacks dynamic correction capabilities. Existing methods typically rank candidate clauses once according to similarity scores after retrieval, and then select the top few as input for the generation model. In this process, the intermediate understanding results formed by the model's initial reasoning, such as key operational actions, process stages, potential risk directions, and clause preferences, do not have a reverse effect on adjusting the ranking of candidate clauses. Therefore, if there is a deviation in the initial ranking, the subsequent generated results are easily affected.

[0006] Third, a fixed number of candidate clauses makes it difficult to balance efficiency and accuracy. Existing methods typically pre-define a fixed number of candidate clauses, such as the first 5 or 10. If the number of candidate clauses is small, key risk factors may be overlooked; if the number of candidate clauses is large, more irrelevant clauses will be introduced, increasing the context length, prolonging model inference time, and potentially reducing the accuracy of the final judgment. Therefore, a fixed number of candidate inputs is difficult to adapt to workloads of varying complexity.

[0007] Fourth, the generation process is disconnected from the candidate selection process. In existing methods, although the generative model can form an intermediate understanding of the task during the initial generation process, the result of this understanding usually does not participate in the subsequent re-selection or re-ranking of candidate terms. As a result, the "generated semantic information" cannot be used to correct the previous retrieval bias, making it difficult to form a closed-loop optimization.

[0008] Fifth, there is insufficient utilization of business feedback for corrective actions. In practice, business departments can provide feedback on incorrectly judged samples and point out the correspondence between certain on-site descriptions and standard risk items. For example, they can indicate the highest risk level actually corresponding to a certain task and its risk level number in the basic risk level table. However, in existing solutions, this type of feedback usually only relies on human experience or is only patched up by supplementing the dictionary or modifying prompts. It fails to systematically participate in the subsequent retrieval, sorting, and generation processes, resulting in insufficient continuous optimization capabilities of the system.

[0009] Sixth, the response speed is slow. When the initial search is inaccurate, there are too many candidate terms, and the sorting cannot be dynamically corrected, the system needs to input a long context into the model for judgment, resulting in a long overall processing time, which makes it difficult to meet the requirements of business scenarios for rapid review of high-frequency daily plans.

[0010] Therefore, although existing search enhancement generation schemes can support the verification of risk levels of daily infrastructure construction plans to a certain extent, they still have problems such as inaccurate retrieval, static ranking, fixed candidates, insufficient utilization of feedback, and low overall timeliness. Summary of the Invention

[0011] This invention addresses the problems of inaccurate retrieval, static sorting, fixed candidates, insufficient feedback utilization, and low overall timeliness in existing infrastructure daily plan management schemes. It provides an intelligent verification method, equipment, and medium for risk levels of daily plans for power transmission and transformation infrastructure, enabling faster and more accurate intelligent verification of the risk levels and risk numbers corresponding to the daily plan work content.

[0012] The present invention is achieved through the following technical solution.

[0013] Firstly, a method for intelligently verifying the risk level of daily plans for power transmission and transformation infrastructure construction is provided, the method comprising:

[0014] The text to be checked is obtained based on the daily infrastructure construction plan data, and the text to be checked is searched for the first time in the risk knowledge base to obtain an initial candidate knowledge set.

[0015] The query text to be proofread and the initial candidate knowledge set are encoded, and the encoding results are input into the proofreading model to obtain the first intermediate proofreading result;

[0016] Based on the first intermediate proofreading result and the query text to be proofread, the subset of candidate clauses in the initial candidate knowledge set is dynamically reordered.

[0017] The candidate terms and the query text to be proofread are input into the proofreading model from the candidate knowledge set after dynamic reordering to obtain the second round of proofreading results.

[0018] In some embodiments, before obtaining the query text to be verified based on infrastructure daily plan data and performing an initial retrieval of the query text to be verified in a risk knowledge base, the method further includes: data acquisition and preprocessing, and constructing a risk knowledge base.

[0019] The data acquisition and preprocessing includes: acquiring daily infrastructure plan data and extracting text content, and performing text preprocessing on the text content to form a text to be checked and queried;

[0020] The construction of the risk knowledge base includes: constructing a knowledge base for risk level verification, wherein the risk knowledge base includes at least: clauses of the "Regulations for Construction Safety Risk Management of Power Transmission and Transformation Projects", a basic risk level table, risk level number mapping relationships, terminology mapping relationships, and historical manual correction samples.

[0021] In some embodiments, based on the first intermediate proofreading result and the query text to be proofread, the subset of candidate terms in the initial candidate knowledge set is dynamically reordered, including:

[0022] The dynamic reordering is modeled as a trajectory decision problem, and the action trajectory of the reorderer in a single calibration task is constructed:

[0023] ;

[0024] in, This indicates the sequence of actions for selecting and prioritizing terms. The parameter is Reordering strategy, Indicates the production environment. Indicates the current sample, Indicates an action;

[0025] The reordering strategy employs trajectory probabilities:

[0026] ;

[0027] in, Indicates the first The previous historical action sequence, Indicates the total number of steps. This indicates the t-th action.

[0028] In some embodiments, before inputting the candidate terms and the query text to be proofread from the dynamically reordered candidate knowledge set into the proofreading model to obtain the second round of proofreading results, the method further includes: determining the number of candidates, wherein determining the number of candidates includes cloning the target by the following behavior:

[0029] ;

[0030] in, The objective function representing behavioral cloning; Represents the mathematical expectation; Represents the expert example set Mid-sample triplet, This indicates the text to be checked. Represents the initial set of candidate knowledge. Indicates the number of candidate clauses. This indicates the number of target clauses corresponding to the preferred trajectory. This represents the collection of expert examples.

[0031] In some embodiments, the method further includes: using the results of the second round of verification as a feedback signal to optimize the dynamic reordering strategy.

[0032] Set exploration objectives:

[0033] ;

[0034] in, Indicates the objective function to be explored; This indicates the expectations for the environment and the sample; Represents the expected value of the policy trajectory; Indicates the current production environment Current sample and the current sorting trajectory The reward value obtained below;

[0035] The obtained strategy trajectories are sampled and sorted multiple times to select the optimal trajectory. and worst trajectory :

[0036] ;

[0037] in, This represents the value of the independent variable that makes the function reach its maximum value; Represents the sorting trajectory The sampled values;

[0038] Using a preference optimization objective, a reorderer for dynamic reordering is trained using the optimal and worst trajectories:

[0039] ;

[0040] in, Represent the objective function; Represents the expectation of a trajectory pair; Indicates the probability of preference; Indicates the degree of preference.

[0041] In some embodiments, the method further includes: after obtaining the second round of verification results, manually correcting the work content that is judged to be erroneous, determining the highest risk level and its risk number in the basic risk level table, storing the manually corrected result, and using it as a reward sample and a ranking preference sample.

[0042] The method further includes: after the manual correction, based on the number of clauses in the historical correct samples that ultimately contribute to obtaining the correct risk level and risk number, calculating the number of target clauses corresponding to the preferred trajectory. Update.

[0043] In some embodiments, the method further includes: evaluating the second round of proofreading results using the following comprehensive evaluation function:

[0044] ;

[0045] in, Indicates exact match reward; This represents a semantic similarity reward; Awards are given for the fluency of the text or the quality of the terms and conditions. Length penalty; This represents the overall quality score based on model evaluation; , , , , This represents the weighting coefficient.

[0046] Secondly, a smart verification device for the risk level of daily plans for power transmission and transformation infrastructure is provided, the device comprising:

[0047] The initial candidate knowledge set acquisition module is used to: acquire the query text to be checked based on the daily infrastructure plan data, and perform an initial search on the query text to be checked in the risk knowledge base to obtain the initial candidate knowledge set;

[0048] The intermediate proofreading result acquisition module is used to: encode the query text to be proofread and the initial candidate knowledge set, and input the encoding result into the proofreading model to obtain the first intermediate proofreading result;

[0049] The dynamic reordering module is used to: dynamically reorder the subset of candidate clauses in the initial candidate knowledge set based on the first intermediate proofreading result and the query text to be proofread;

[0050] The final proofreading result acquisition module is used to input the candidate terms of the candidate number in the dynamically reordered candidate knowledge set and the query text to be proofread into the proofreading model to obtain the second round of proofreading results.

[0051] Thirdly, a smart verification device for the risk level of daily plans for power transmission and transformation infrastructure is provided, the device comprising:

[0052] At least one processor;

[0053] At least one memory coupled to the at least one processor and storing instructions for execution by the at least one processor, the instructions implementing the method described in any of the above when executed by the at least one processor.

[0054] Fourthly, a computer-readable storage medium is provided, characterized in that the computer-readable storage medium stores instructions that, when executed by a computer, cause the computer to perform the method described in any of the preceding embodiments.

[0055] Compared with the prior art, the present invention has the following advantages and beneficial effects: the intelligent verification of the risk level of the daily construction plan based on dynamic feedback reordering retrieval enhancement generation can dynamically determine the priority and number of candidate clauses according to the initial generation results, and continuously optimize the matching process in combination with business correction feedback, thereby achieving faster and more accurate intelligent verification of the risk level and risk number corresponding to the daily plan work content. Attached Figure Description

[0056] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a flowchart of a method for intelligently verifying the risk level of daily plans for power transmission and transformation infrastructure according to an embodiment of the present invention.

[0058] Figure 2 This is an intelligent verification process for the risk level of daily plans for power transmission and transformation infrastructure, according to an embodiment of the present invention.

[0059] Figure 3 This is a schematic diagram comparing the intelligent verification method for daily risk levels of power transmission and transformation infrastructure construction of the present invention with existing methods.

[0060] Figure 4 This is a schematic diagram illustrating the working process of a dynamic reorderer according to an embodiment of the present invention.

[0061] Figure 5 This is a schematic diagram illustrating the composition of the comprehensive evaluation function according to an embodiment of the present invention.

[0062] Figure 6 This is a schematic diagram of a business correction feedback closed loop according to an embodiment of the present invention.

[0063] Figure 7 This is a schematic diagram of a typical work content verification process according to an embodiment of the present invention.

[0064] Figure 8 This is a structural block diagram of an intelligent calibration device for the daily risk level of power transmission and transformation infrastructure construction, according to an embodiment of the present invention.

[0065] Figure 9 This is a schematic diagram of a smart calibration device for the daily risk level of power transmission and transformation infrastructure construction, according to an embodiment of the present invention. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0067] To address the shortcomings of existing search-enhanced generation schemes in the scenario of verifying the risk level of daily construction plans, such as low accuracy in the first round of retrieval, static sorting of candidate clauses, fixed number of candidate clauses, inability to provide feedback on the generated results for optimized sorting, and difficulty in deeply integrating business correction information, this invention provides an intelligent verification scheme for the risk level of daily construction plans based on dynamic feedback reordering search-enhanced generation. This scheme can dynamically determine the priority and number of candidate clauses based on the initial generation results and continuously optimize the matching process by combining business correction feedback, thereby achieving faster and more accurate intelligent verification of the risk level and risk number corresponding to the daily plan work content.

[0068] On the one hand, the present invention provides an intelligent method for verifying the risk level of daily plans for power transmission and transformation infrastructure. Figure 1 This is a flowchart illustrating an intelligent verification method for the daily risk level of power transmission and transformation infrastructure construction according to an embodiment of the present invention. (Reference) Figure 1 The intelligent verification method for the risk level of the daily plan for power transmission and transformation infrastructure includes: S10 to S40.

[0069] Figure 2 This is an intelligent verification process for the risk level of daily plans for power transmission and transformation infrastructure, according to an embodiment of the present invention. Figure 3 This is a schematic diagram comparing the intelligent risk level verification method for daily plans of power transmission and transformation infrastructure of the present invention with existing methods. The following references... Figures 1 to 3 The present invention provides a detailed description of the intelligent verification method for the risk level of daily plans for power transmission and transformation infrastructure.

[0070] In S10, the query text to be checked is obtained based on the daily infrastructure plan data, and the query text to be checked is searched for the first time in the risk knowledge base to obtain an initial candidate knowledge set.

[0071] Before acquiring the query text to be verified based on the daily construction plan data and performing an initial search on the risk knowledge base, this intelligent verification method for the risk level of daily construction plans for power transmission and transformation also includes: data acquisition and preprocessing, and constructing a risk knowledge base. Data acquisition and preprocessing includes: acquiring daily construction plan data and extracting text content, and performing text preprocessing on the text content to form the query text to be verified. Constructing the risk knowledge base includes: building a knowledge base for risk level verification, wherein the risk knowledge base includes at least: clauses of the "Regulations for Construction Safety Risk Management of Power Transmission and Transformation Projects", a basic risk level table, risk level number mapping relationships, terminology mapping relationships, and historical manual correction samples.

[0072] Specifically, the daily infrastructure construction plan data is obtained, and fields such as work content, work specialty, equipment object, voltage level, planned date, and original risk level (text content) are extracted. The text content undergoes preprocessing, including removing invalid symbols, standardizing terminology, splitting action phrases, normalizing synonyms, and mapping non-standard expressions to form the text to be checked and queried. .

[0073] Build a knowledge base for risk level verification The knowledge base should include at least the clauses of the "Regulations for Construction Safety Risk Management of Power Transmission and Transformation Projects", a basic risk level table, risk level number mapping relationships, terminology mapping relationships, and historical manual correction samples. The knowledge base content should preferably be structured and segmented at the clause, risk level, number, and terminology levels to facilitate subsequent retrieval and matching.

[0074] Text to be checked Input the search module into the knowledge base The first search is performed to obtain the first (initial) candidate knowledge set. This process can be represented as:

[0075] ;

[0076] in, Indicates the search engine. This represents the initial set of candidate knowledge.

[0077] In some implementations, the first Candidate knowledge units With query text The basic correlation can be expressed as:

[0078] ;

[0079] in, This can be a vector similarity, keyword relevance, or other relevance scoring function. The relevance score calculated based on the above formula can be compared with a predetermined threshold to obtain an initial candidate knowledge set.

[0080] In S20, the query text to be proofread and the initial candidate knowledge set are encoded, and the encoding results are input into the proofreading model to obtain the first intermediate proofreading result.

[0081] Specifically, for the query text The terms in the initial candidate knowledge set are encoded to obtain a vector representation:

[0082] ;

[0083] Then, the query text and candidate terms are input into the generation (proofreading) model to obtain the first intermediate proofreading result:

[0084] ;

[0085] in, This refers to the intermediate verification results generated by the generative model, which include at least one or more of the following: candidate risk level tendency, candidate risk number tendency, key operation terms, process stage information, and clause preference information. Indicates the output variable; Indicates candidate clauses.

[0086] In some implementations, the conditional generation probability of the generative model can be further expressed as:

[0087] ;

[0088] The generation loss can be expressed as:

[0089]

[0090] In S30, based on the first intermediate proofreading result and the query text to be proofread, the subset of candidate terms in the initial candidate knowledge set is dynamically reordered.

[0091] This invention no longer uses fixed Top-K terms directly as the final input. Instead, after the initial generation, the reordering algorithm uses the query text to determine the final input. Initial candidate knowledge set In addition to the feedback information carried by the first intermediate proofreading results, a dynamically rearranged candidate subset is output. This subset embodies both "which terms to select" and "how many terms to retain," and can be represented as:

[0092] ;

[0093] in, The parameter is Dynamic reorderer, ⊆ This represents a subset of candidate terms after dynamic rearrangement.

[0094] In the context of this invention, the input to the reorderer includes not only the original query text. It also includes feedback information carried by the intermediate verification results generated initially, such as key action terms, process stages, risk level tendencies, and clause preferences. In other words, the reorderer is a reorderer with generation feedback constraints, and its goal is to select a subset of clauses that is more helpful in determining the final risk level and risk number.

[0095] To characterize the process of the reorderer progressively selecting a subset of terms, dynamic reordering can be modeled as a trajectory decision problem. The action trajectory of the reorderer in a single proofreading task can be represented as:

[0096] ;

[0097] in, This indicates the sequence of actions for selecting and prioritizing terms. The parameter is Reordering strategy, Indicates the production environment. Indicates the current sample, This represents an action. Its trajectory probability can be expanded as follows:

[0098] ;

[0099] in, Indicates the first Following previous historical actions, Indicates the total number of steps. This indicates the t-th action.

[0100] In S40, the candidate terms and the query text to be proofread are input into the proofreading model from the candidate knowledge set after dynamic reordering to obtain the second round of proofreading results.

[0101] In this invention, the number of candidate terms in the second round of input is not fixed. Instead, the system dynamically decides how many terms to retain through policy learning. The behavioral cloning objective can be represented as:

[0102] ;

[0103] in, The objective function representing behavioral cloning; Represents the mathematical expectation; Represents the expert example set Mid-sample triplet, This indicates the text to be checked. Represents the initial set of candidate knowledge. Indicates the number of candidate clauses. This indicates the number of target clauses corresponding to the preferred trajectory. This represents the collection of expert examples.

[0104] In the business scenario of this invention This can be determined by the number of clauses in the historical correct sample that ultimately helped to obtain the correct risk level and risk number. Therefore, fewer clauses can be selected for simple tasks, and more clauses can be selected for complex tasks.

[0105] In some embodiments, the method further includes: using the quality of the second proofreading result as a feedback signal to optimize the dynamic reordering strategy, the exploration objective of which can be expressed as:

[0106] ;

[0107] in, Indicates the objective function to be explored; This indicates the expectations for the environment and the sample; Represents the expected value of the policy trajectory; Indicates the current production environment Current sample and the current sorting trajectory The reward value obtained below.

[0108] In some implementations, the optimal trajectory is selected by sampling and sorting the trajectories multiple times. and worst trajectory :

[0109] ;

[0110] in, This represents the value of the independent variable that makes the function reach its maximum value; Represents the sorting trajectory The sampled values.

[0111] Train a dynamic reorderer using a preference optimization objective:

[0112] ;

[0113] in, Represent the objective function; Represents the expectation of a trajectory pair; Indicates the probability of preference; This indicates the degree of preference. The expression reflects the principle of giving higher probability to high-quality sorted trajectories than low-quality sorted trajectories.

[0114] In some embodiments, the method further includes: after obtaining the second verification result, manually correcting the highest risk level of the work content that was judged to be incorrect and its risk number in the basic risk level table, storing the manually corrected result, and using it as a reward sample and a ranking preference sample. Additionally, after manual correction, the number of target clauses corresponding to the preferred trajectory can be determined based on the number of clauses in the historical correct samples that ultimately contribute to obtaining the correct risk level and risk number. The system is then updated. Specifically, when a business department discovers an error in the system's judgment, it provides feedback on the erroneous sample, indicating the highest risk level corresponding to the work content and its risk number in the basic risk level table. The system writes this corrective relationship into the historical sample library for use in constructing higher-quality reward samples and ranking preference samples, thereby continuously improving the performance of the dynamic reorderer.

[0115] In some embodiments, the method further includes: evaluating the second proofreading result using the following comprehensive evaluation function.

[0116] In this invention, the final reward is no longer based solely on the correctness of ordinary questions and answers, but is constructed by combining the risk level verification task. Its comprehensive evaluation function can be expressed as:

[0117] ;

[0118] in, Indicates exact match reward; This represents a semantic similarity reward; Awards are given for the fluency of the text or the quality of the terms and conditions. Length penalty; This represents the overall quality score based on model evaluation; , , , , This represents the weighting coefficient.

[0119] In the risk level verification scenario of this invention, the above items can be specifically defined as follows:

[0120] An exact match can be defined as whether the risk level and risk number are exactly the same as the standard answer:

[0121] ;

[0122] in, This represents an exact matching function; This indicates the standard answer; This indicates the model's prediction results.

[0123] The semantic similarity term can be defined as the degree of semantic closeness between the basis terms provided by the system and the standard basis terms:

[0124] ;

[0125] in, This represents the semantic similarity function.

[0126] Text fluency can be represented as the degree of consistency between the output specification and the standard specification:

[0127] ;

[0128] in, This represents a function for text fluency.

[0129] The length penalty term can be expressed as:

[0130] ;

[0131] in, This represents the function for the length penalty term.

[0132] The comprehensive evaluation items can be represented as:

[0133] ;

[0134] in, This indicates a prompt template used for evaluation. Indicates system output, This indicates the standard answer.

[0135] In the context of this invention, it is also possible to further... It is divided into two components: "risk level matching" and "risk number matching" to better meet business needs.

[0136] Compared with existing search enhancement generation schemes, this invention achieves the following technical effects through a technical process of "initial search - initial generation - dynamic feedback reordering - dynamic candidate subset output - feedback optimization based on result quality - regeneration".

[0137] (1) This invention no longer uses fixed Top-K terms to generate the candidate subset directly, but instead uses a dynamic reordering machine to output the rearranged candidate subset. This allows for dynamic decision-making regarding "which clauses to select and how many clauses to retain" based on different work content.

[0138] (2) This invention utilizes intermediate understanding information in the initial generated results to reverse optimize the candidate clause ranking, which can correct the initial retrieval deviation and improve the matching accuracy between the work content and the standard risk clause, rather than being difficult to correct once the first round of recall deviation is obtained, as is the case with the existing static solution.

[0139] (3) This invention enables the dynamic reorderer to learn "what clause order and number of clauses are more conducive to the correct output of the final risk level and risk number" through behavioral cloning and reward / preference-based strategy optimization, thereby improving the optimization objective from simple clause relevance to the quality of the final proofreading result.

[0140] (4) This invention incorporates "whether the risk level is correct, whether the risk number is correct, whether the basis clause is reasonable, and whether the output is concise" into the optimization process through a multi-dimensional evaluation function, so that the system no longer only pursues "like the relevant clauses", but pursues "being able to correctly complete the risk level verification".

[0141] (5. This invention incorporates the manual correction results provided by the business department into the dynamic reordering optimization closed loop, enabling the system to gradually absorb the mapping relationship between "on-site description - standard risk item" and enhance its adaptability to non-standard and colloquial descriptions.

[0142] Therefore, this invention can achieve higher proofreading accuracy, better processing efficiency, and stronger continuous optimization capabilities compared to existing retrieval enhancement generation schemes in the scenario of intelligent proofreading of risk levels in daily infrastructure construction plans.

[0143] The present invention will be described below through specific application implementations.

[0144] Implementation Case 1: Risk Level Verification in Non-Standard Expression Scenarios

[0145] like Figure 2 , Figure 3 , Figure 4 and Figure 7As shown, the work content in a certain infrastructure daily plan is: "Tower assembly, dismantling of fire hydrant systems, and preparation for accessory installation," and the original risk level was reported as Level 2. The system first preprocesses the work content and generates a query. Then, according to the formula:

[0146] ;

[0147] The first candidate knowledge set was retrieved from the risk knowledge base. Then, according to the formula:

[0148] ;

[0149] Initial results have been obtained. These results indicate that the current work focuses more on risk clauses related to tower assembly and dismantling.

[0150] Then, the dynamic reorderer is invoked, according to the formula:

[0151] ;

[0152] Output the rearranged candidate subsets Furthermore, clauses that are literally related to "preparation" but contribute little to the final risk level assessment should be reduced, while clauses directly related to "angle steel tower assembly" and "module dismantling" should be given higher priority.

[0153] Ultimately, the system is based on The output structured verification results are as follows: The highest risk level is Level 3 risk, the corresponding risk number is a certain number, the matching basis clause is the corresponding entry in the basic risk level table, the original reported Level 2 risk is inconsistent with the system's judgment, and it is recommended to correct it to Level 3 risk.

[0154] Implementation Case 2: Rapid Proofreading in Standard Expression Scenarios

[0155] like Figure 2 and Figure 3 As shown, a certain work item is described as "installation of conductor and ground wire accessories," a description that closely resembles standard terminology in the knowledge base. The system can obtain a high-quality candidate knowledge set after its initial retrieval, and the risk level tendency of the initial generated results is clear. The dynamic reordering machine outputs a small subset of candidate knowledge. This reduces the number of clauses required for the second round of input. The system then quickly outputs the risk level, risk number, and applicable clauses, confirming consistency with the original information. This implementation case illustrates that for work content with standardized wording and low ambiguity, this invention can achieve faster processing speeds through dynamic subset output.

[0156] Implementation Case 3: Manual Correction and Feedback Participation in Subsequent Optimization

[0157] like Figure 5 and Figure 6 As shown, the original text of a certain work item was "fire hydrant removal." During the initial system run, it failed to accurately match the correct risk number. After review by the business department, it was confirmed that it actually corresponded to a specific standard risk item, and the highest risk level and corresponding risk number were provided. The system then included this sample in the preference optimization dataset, and in subsequent training, it constructed the following optimal / worst trajectories:

[0158] ;

[0159] And the following preference optimization objectives:

[0160] ;

[0161] This makes it easier to retain clauses corresponding to the correct risk number in the candidate subset when identical or similar expressions appear in the future, thereby improving the matching accuracy of similar samples. This implementation case illustrates that this invention can achieve continuous optimization through business feedback, rather than just remaining at the level of a single proofreading result.

[0162] On the other hand, the present invention provides an intelligent calibration device for the risk level of daily plans for power transmission and transformation infrastructure. Figure 8 This is a structural block diagram of an intelligent verification device for the daily risk level of power transmission and transformation infrastructure construction, according to an embodiment of the present invention. (Reference) Figure 8 The device includes: an initial candidate knowledge set acquisition module, an intermediate proofreading result acquisition module, a dynamic reordering module, and a final proofreading result acquisition module.

[0163] The initial candidate knowledge set acquisition module is used to: acquire the query text to be checked based on the daily infrastructure plan data, and perform an initial search on the query text to be checked in the risk knowledge base to obtain the initial candidate knowledge set.

[0164] The intermediate proofreading result acquisition module is used to encode the query text to be proofread and the initial candidate knowledge set, and input the encoding result into the proofreading model to obtain the first intermediate proofreading result.

[0165] The dynamic reordering module is used to dynamically reorder a subset of candidate terms in the initial candidate knowledge set based on the first intermediate proofreading result and the query text to be proofread.

[0166] The final proofreading result acquisition module is used to input the candidate terms and the query text to be proofread from the candidate knowledge set after dynamic reordering into the proofreading model to obtain the second proofreading result.

[0167] For other details regarding the intelligent verification equipment for the daily risk level of power transmission and transformation infrastructure construction, please refer to the previous description of the intelligent verification method for the daily risk level of power transmission and transformation infrastructure construction, which will not be repeated here.

[0168] In implementing the functions of the integrated modules described above in hardware, this embodiment of the invention provides a structure for the intelligent verification device for the daily risk level of power transmission and transformation infrastructure construction involved in the above embodiments. Figure 9 This is a schematic diagram of a smart verification device for the daily risk level of power transmission and transformation infrastructure construction, according to an embodiment of the present invention. (Reference) Figure 9 The intelligent verification device for the daily risk level of power transmission and transformation infrastructure construction includes: at least one processor; and at least one memory. The at least one memory is coupled to the at least one processor and stores instructions for execution by the at least one processor, which, when executed by the at least one processor, implement the methods described above.

[0169] A processor can be a set of logic blocks, modules, and circuits that implement or execute the various exemplary logic blocks, modules, and circuits described in connection with embodiments of the present invention. The processor can be a central processing unit, a general-purpose processor, a digital signal processor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in connection with embodiments of the present invention. A processor can also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a digital signal processor and a microprocessor, etc.

[0170] The memory may be read-only memory (ROM) or other types of static storage devices that can store static information and instructions, random access memory (RAM) or other types of dynamic storage devices that can store information and instructions, electrically erasable programmable read-only memory (EEPROM), disk storage media or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer, but is not limited thereto.

[0171] In one implementation, the memory can exist independently of the processor. The memory can be connected to the processor via a bus and used to store instructions or program code. When the processor calls and executes the instructions or program code stored in the memory, it can implement the method provided in the embodiments of the present invention. In another implementation, the memory can also be integrated with the processor.

[0172] On the other hand, the present invention also provides a computer-readable storage medium (e.g., a non-transitory computer-readable storage medium) storing computer program instructions that, when executed on a computer, cause the computer to perform the method as described in any of the above embodiments.

[0173] Exemplary examples show that the aforementioned computer-readable storage media may include, but are not limited to: magnetic storage devices (e.g., hard disks, floppy disks, or magnetic tapes), optical discs (e.g., compact disks (CDs), digital versatile disks (DVDs), etc.), smart cards, and flash memory devices (e.g., erasable programmable read-only memory (EPROMs), cards, sticks, or key drives, etc.). The various computer-readable storage media described in this invention may represent one or more devices and / or other machine-readable storage media for storing information. The term "machine-readable storage medium" may include, but is not limited to, wireless channels and various other media capable of storing, containing, and / or carrying instructions and / or data.

[0174] This invention provides a computer program that, when run on a computer, causes the computer to perform the method of any of the above embodiments.

[0175] This invention provides a computer program product containing instructions that, when run on a computer, cause the computer to perform the method of any of the above embodiments.

[0176] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligently verifying the risk level of daily plans for power transmission and transformation infrastructure construction, characterized in that, The method includes: The text to be checked is obtained based on the daily infrastructure construction plan data, and the text to be checked is searched for the first time in the risk knowledge base to obtain an initial candidate knowledge set. The query text to be proofread and the initial candidate knowledge set are encoded, and the encoding results are input into the proofreading model to obtain the first intermediate proofreading result; Based on the first intermediate proofreading result and the query text to be proofread, the subset of candidate clauses in the initial candidate knowledge set is dynamically reordered. The candidate terms and the query text to be proofread are input into the proofreading model from the candidate knowledge set after dynamic reordering to obtain the second proofreading result.

2. The method according to claim 1, characterized in that, Before obtaining the query text to be verified based on the daily infrastructure construction plan data and performing an initial retrieval of the query text in the risk knowledge base, the method further includes: data acquisition and preprocessing, and constructing a risk knowledge base. The data acquisition and preprocessing includes: acquiring daily infrastructure plan data and extracting text content, and performing text preprocessing on the text content to form a text to be checked and queried; The construction of the risk knowledge base includes: constructing a knowledge base for risk level verification, wherein the risk knowledge base includes at least: clauses of the "Regulations for Construction Safety Risk Management of Power Transmission and Transformation Projects", a basic risk level table, risk level number mapping relationships, terminology mapping relationships, and historical manual correction samples.

3. The method according to claim 1 or 2, characterized in that, Based on the first intermediate proofreading result and the query text to be proofread, the candidate clause subset in the initial candidate knowledge set is dynamically reordered, including: The dynamic reordering is modeled as a trajectory decision problem, and the action trajectory of the reorderer in a single calibration task is constructed: ； in, This indicates the sequence of actions for selecting and prioritizing terms. The parameter is Reordering strategy, Indicates the production environment. Indicates the current sample, Indicates an action; The reordering strategy employs trajectory probabilities: ； in, Indicates the first The previous historical action sequence, Indicates the total number of steps. This indicates the t-th action.

4. The method according to claim 3, characterized in that, Before inputting the candidate terms and the query text to be proofread from the dynamically reordered candidate knowledge set into the proofreading model to obtain the second proofreading result, the method further includes: determining the number of candidates, wherein determining the number of candidates includes cloning the target through the following actions: ； in, The objective function representing behavioral cloning; Represents the mathematical expectation; Represents the expert example set Mid-sample triplet, This indicates the text to be checked. Represents the initial set of candidate knowledge. Indicates the number of candidate clauses. This indicates the number of target clauses corresponding to the preferred trajectory. This represents the collection of expert examples.

5. The method according to claim 4, characterized in that, The method further includes: using the second proofreading result as a feedback signal to optimize the dynamic reordering strategy. Set exploration objectives: ； in, Indicates the objective function to be explored; This indicates the expectations for the environment and the sample; Represents the expected value of the policy trajectory; Indicates the current production environment Current sample and the current sorting trajectory The reward value obtained below; The obtained strategy trajectories are sampled and sorted multiple times to select the optimal trajectory. and worst trajectory : ； in, This represents the value of the independent variable that makes the function reach its maximum value; Represents the sorting trajectory The sampled values; Using a preference optimization objective, a reorderer for dynamic reordering is trained using the optimal and worst trajectories: ； in, Represent the objective function; Represents the expectation of a trajectory pair; Indicates the probability of preference; Indicates the degree of preference.

6. The method according to claim 4, characterized in that, The method further includes: after obtaining the second proofreading result, manually correcting the work content that is judged to be wrong, determining the highest risk level and its risk number in the basic risk level table, storing the manually corrected result, and using it as a reward sample and a ranking preference sample. The method further includes: after the manual correction, based on the number of clauses in the historical correct samples that ultimately contribute to obtaining the correct risk level and risk number, calculating the number of target clauses corresponding to the preferred trajectory. Update.

7. The method according to claim 1, characterized in that, The method further includes: evaluating the second proofreading result using the following comprehensive evaluation function: ； in, Indicates exact match reward; This represents a semantic similarity reward; Awards are given for the fluency of the text or the quality of the terms and conditions. Length penalty; This represents the overall quality score based on model evaluation; , , , , This represents the weighting coefficient.

8. A smart verification device for the daily risk level of power transmission and transformation infrastructure construction plans, characterized in that, The device includes: The initial candidate knowledge set acquisition module is used to: acquire the query text to be checked based on the daily infrastructure plan data, and perform an initial search on the query text to be checked in the risk knowledge base to obtain the initial candidate knowledge set; The intermediate proofreading result acquisition module is used to: encode the query text to be proofread and the initial candidate knowledge set, and input the encoding result into the proofreading model to obtain the first intermediate proofreading result; The dynamic reordering module is used to: dynamically reorder the subset of candidate clauses in the initial candidate knowledge set based on the first intermediate proofreading result and the query text to be proofread; The final proofreading result acquisition module is used to input the candidate terms of the candidate number in the dynamically reordered candidate knowledge set and the query text to be proofread into the proofreading model to obtain the second proofreading result.

9. A smart verification device for the daily risk level of power transmission and transformation infrastructure construction, characterized in that, The device includes: At least one processor; At least one memory coupled to the at least one processor and storing instructions for execution by the at least one processor, the instructions implementing the method of any one of claims 1 to 7 when executed by the at least one processor.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 7.

Images