A Quality Inspection Method and System for Preliminary Design Reports of Reinforcement Projects for Dysfunctional Reservoirs

Abstract

This invention relates to the field of quality control technology in water conservancy engineering design, and discloses a method and system for quality inspection of preliminary design reports for the reinforcement and renovation of dilapidated reservoirs. This addresses the problems of traditional methods relying on subjective expert evaluation, difficulty in quantifying quality, insufficient indicator comprehensiveness, low efficiency, and susceptibility to subjective factors. In this invention, relevant historical data is first collected to construct a corpus and image library. After preprocessing, text and image features are extracted using deep learning and merged into an initial feature set. Simultaneously, a key calculation program model for water conservancy engineering is built, and an AHP method is used to establish a hierarchical initial evaluation system. Then, historical reports are used for verification and optimization to obtain a verified feature set and evaluation system. Finally, features of the report to be reviewed are extracted, and the accuracy of the calculated features is verified by the calculation program model. Scoring is then performed according to corresponding rules to complete the quality inspection. This invention is applicable to the quality inspection and evaluation of preliminary design reports for the reinforcement and renovation of dilapidated reservoirs.

Description

Technical Field

[0001] This invention relates to the field of quality control technology for water conservancy engineering design, specifically to a method and system for quality inspection of preliminary design reports for the reinforcement and renovation of dilapidated reservoirs. Background Technology

[0002] As a major country in water conservancy projects, my country has built over 90,000 reservoirs of various types. These reservoirs play an irreplaceable strategic role in flood control, irrigation, water supply, and power generation, and are crucial infrastructure for ensuring national water resource security, flood control security, and ecological security. However, with the increasing age of these reservoirs, due to multiple factors such as the technological level of their construction, long-term hydrological and meteorological influences, changes in geological conditions, and inadequate maintenance and management, many reservoirs have gradually developed problems such as dam leakage, structural aging, insufficient anti-sliding stability, and deterioration of spillway function. The number of reservoirs with defects is increasing year by year, and the safety hazards are becoming increasingly prominent. These reservoirs not only severely restrict the normal functioning of their own comprehensive benefits, but also pose a continuous and serious threat to the lives and property of people in downstream areas, important infrastructure, and the ecological environment.

[0003] Therefore, carrying out the reinforcement and remediation of dilapidated reservoirs has become a key task in my country's water conservancy industry. The preliminary design report for the reinforcement and remediation project is the core technical document of the remediation work. It serves as a crucial basis for guiding construction, ensuring reinforcement quality, and controlling project investment; its quality directly determines the implementation effect and safety benefits of the reinforcement project. If the design report contains missing content, incorrect data, logical contradictions, or does not comply with industry standards, it will lead to unreasonable reinforcement plans, inadequate engineering measures, and may even trigger secondary safety risks, failing to fundamentally eliminate the reservoir's hidden dangers.

[0004] Currently, the quality inspection of preliminary design reports for the reinforcement and renovation of dilapidated reservoirs in my country mainly relies on industry standards such as the "Regulations for the Preparation of Preliminary Design Reports for Water Conservancy and Hydropower Projects" and the "Standards for Quality Assessment of Preliminary Designs of Water Conservancy and Hydropower Projects," employing a traditional inspection model primarily based on expert review. This model has several limitations in practical application:

[0005] (1) The evaluation is mainly qualitative and the quality is difficult to quantify: It mainly relies on the professional knowledge and practical experience of experts to make subjective judgments, and lacks a unified and accurate quantitative evaluation standard, resulting in insufficient objectivity and comparability of the test results.

[0006] (2) The quantitative evaluation indicators are not comprehensive enough: the existing evaluation system focuses on a few core indicators such as the completeness of the report format and the rationality of the main technical parameters, and does not cover detailed dimensions such as the consistency of text logic, the standardization of image data and the accuracy of calculation process, making it difficult to fully reflect the overall quality of the report.

[0007] (3) Low inspection efficiency: As the renovation of dilapidated reservoirs is carried out on a large scale, the number of design reports to be inspected continues to increase. However, senior industry experts are scarce, and manual review requires a lot of time and manpower, which makes it difficult to meet the actual needs of efficient inspection.

[0008] (4) Subjectivity has a significant impact: Different experts have different knowledge structures and review experience, and their understanding and grasp of the evaluation criteria are inconsistent, which may lead to different review results for the same report, affecting the fairness and authority of the test.

[0009] In summary, traditional methods for quality inspection of preliminary design reports for the reinforcement and renovation of dilapidated reservoirs are no longer adequate to meet the demands for refined and efficient quality control in water conservancy projects under the new circumstances. There is an urgent need for an intelligent inspection technology that can overcome reliance on manual methods, achieve quantitative evaluation, and improve the comprehensiveness and accuracy of inspections, thereby ensuring the high-quality implementation of reinforcement and renovation projects for dilapidated reservoirs. Summary of the Invention

[0010] The technical problem to be solved by this invention is to provide a method and system for quality inspection of preliminary design reports for the reinforcement and renovation of dilapidated reservoirs, which solves the problems of traditional solutions relying on subjective evaluation by experts, difficulty in quantifying quality, insufficient indicators, low efficiency, and susceptibility to subjective factors.

[0011] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0012] On the one hand, this invention provides a method for quality inspection of preliminary design reports for the reinforcement and renovation of dilapidated reservoirs, comprising the following steps:

[0013] S1. Obtain relevant historical data and construct a corpus and image library;

[0014] S2. After preprocessing the corpus and image database, text feature sets are generated separately using deep learning. and image feature set The initial feature set is formed by merging these features. ;

[0015] S3. Construct a calculation program model based on the calculation content related to water conservancy projects in the corpus. ;in, The name of the calculation program. For the calculation formula, For calculating parameters, For the calculation results, The number of input parameters. The number of calculation results, For the computation function of the computation program model;

[0016] S4. Determine the hierarchical structure model of quality evaluation indicators, use the Analytic Hierarchy Process (AHP) to determine the weights of indicators at each level, and construct the initial evaluation system. ;

[0017] S5. Based on the initial feature set The feature set of the preliminary design report for the reinforcement and renovation project of dangerous reservoirs was extracted from the corpus and image database. The inspection process is executed, wherein the computational features are calculated using the computational program model constructed in step S3. Accuracy verification was performed; based on the initial evaluation system. Conduct scoring to validate the initial feature set. To improve the quality and obtain a verified feature set. and verified evaluation system ;

[0018] S6. Based on validated feature sets Extract the feature set to be tested from the preliminary design report of the dangerous reservoir reinforcement project to be reviewed. The inspection process is executed, wherein the computational features are calculated using the computational program model constructed in step S3. Accuracy verification was performed, and the results were based on a validated evaluation system. The review report is evaluated and scored.

[0019] Furthermore, in step S1, the relevant historical data includes the preliminary design report of the reinforcement project for the dilapidated reservoir before review, the preliminary design report of the reinforcement project for the dilapidated reservoir after review and approval, the design specifications for the reinforcement project for the dilapidated reservoir, the dam registration data, the national water conservancy census data, and the statistical yearbook of the region where the dilapidated reservoir is located.

[0020] Furthermore, in step S2, the preprocessing of the corpus and image database includes:

[0021] Tables and images were removed from the corpus, leaving only plain text to form a text dataset. The plain text dataset was then cleaned to remove unnecessary characters, spaces, and punctuation marks. Word segmentation technology was used to divide the continuous text into semantically meaningful lexical units, and part-of-speech tagging was performed to obtain tagged text.

[0022] The initial images in the image library are uniformly converted to JPG format. First, all JPG images are categorized and archived using title bar recognition to create a first-level image directory. Based on the first-level image directory, the images are sequentially annotated with Labels and segmented using YOLO. The categories are further refined by combining the segmented instance names to create a second-level image directory. Based on the segmented instances in the second-level image directory, BIM modeling is performed using Revit to generate a reservoir holographic library. The holographic library is then sliced ​​to obtain more detailed instance images, and a third-level image directory is created based on these sliced ​​images. All images in the first-level, second-level, and third-level image directories are combined to form an instance image set.

[0023] Furthermore, in step S2, a text feature set is generated through deep learning. The methods include:

[0024] Set training parameters, import the labeled text into the Bert pre-training layer and BiLSTM network layer in sequence to generate a bidirectional hidden layer text vector sequence, and then connect the bidirectional hidden layer text vector sequence into the CRF model to construct a feature function sequence. Combine the label transition probability constraint to select the optimal label sequence.

[0025] Based on the consecutive "B-entity type" and "I-entity type" labels in the optimal label sequence, the entity name e and the relationship r between entities are extracted.

[0026] Import the extracted entity name e and the relationship r between entities into the Neo4j graph database, create the corresponding nodes and relationships using Cypher statements, generate a knowledge graph and store it.

[0027] By integrating the extracted entity names e, inter-entity relationships r, and the generated knowledge graph, a text feature set is formed. .

[0028] Furthermore, in step S2, an image feature set is generated through deep learning. The methods include:

[0029] The example image set is divided into training, validation, and test sets. A graph feature library is built using YOLO learning, and the resulting numerical vector set is input into the ResNet50 model and stored. The hierarchical image resources of the example image set are integrated with the numerical vector set output by the ResNet50 model to form an image feature set. .

[0030] Furthermore, in step S2, the generated initial feature set Represented as:

[0031] ;

[0032] in, This represents the total number of feature terms in the initial feature set. Indicates the first One feature term;

[0033] ;

[0034] Where e is the entity name, extracted using named entity recognition technology, and is in the form of Chinese characters or English words;

[0035] r represents the relationships between entities, which are extracted using relation extraction techniques.

[0036] V is the corresponding standard feature value extracted from the entity name e using regularization, and its form includes words, numbers, numerical ranges, knowledge graphs, or vector sets;

[0037] t represents the feature type, which can be numerical, computational, word / vector, or knowledge graph type.

[0038] Address is the location information corresponding to this feature in the relevant data;

[0039] RuleV is the test rule, which uses the difference ratio, numerical range comparison, calculation result threshold comparison, cosine similarity comparison or comprehensive comparison.

[0040] ScoreM is the scoring standard, which calculates the score based on the feature type t and the test rule RuleV, and the score range is limited to [0,1].

[0041] Furthermore, in step S3, the constructed computational program model covers:

[0042] Design storm calculation, design peak flow calculation, design flood total volume calculation, design flood hydrograph derivation, design flood rationality analysis, phased flood calculation, sediment calculation, dead water level verification, reservoir flood regulation calculation, comprehensive safety assessment, current dam crest elevation verification, dam seepage analysis, dam anti-sliding stability analysis, spillway hydraulic verification, earth-rock cofferdam design, investment estimate calculation and economic evaluation calculation.

[0043] Furthermore, in step S4, the hierarchical structure model includes: target layer A, criterion layer B, indicator layer C, and sub-indicator layer D;

[0044] Target layer A refers to the quality of the preliminary design report for the reinforcement and renovation project of the dilapidated reservoir;

[0045] The criteria layer B includes primary evaluation indicators such as comprehensiveness, standardization, logic, and accuracy. Comprehensiveness includes secondary evaluation indicators such as engineering characteristics and basic information about the project area. Standardization includes secondary evaluation indicators such as qualification management, report format, and cost information. Logic includes secondary evaluation indicators such as scheme design, reinforcement measures, construction schemes, demolition design, environmental protection design, soil and water conservation design, and construction drawing design. Accuracy includes secondary evaluation indicators such as design storm calculation, design peak flow calculation, design flood total volume calculation, design flood hydrograph derivation, design flood rationality analysis, phased flood calculation, sediment calculation, dead water level verification, reservoir flood regulation calculation, comprehensive safety assessment, current dam crest elevation verification, dam seepage analysis, dam anti-sliding stability analysis, spillway hydraulic verification, earth-rock cofferdam design, investment estimate calculation, and economic evaluation calculation.

[0046] Indicator layer C has There are 1 primary evaluation indicator, and each primary evaluation indicator is defined as follows: ;

[0047] Sub-index layer D has Sub-indicators Sub-indicators are derived from the initial sub-indicator set. The initial sub-index set By extracting the initial feature set All feature data items Construct the corresponding entity name.

[0048] Furthermore, in step S4, the AHP method is used to determine the weights of the indicators at each level, including:

[0049] Target layer A: Target layer A has only one indicator, with a weight set to 1;

[0050] Quasi-testing layer B: This refers to the primary evaluation indicators within quasi-testing layer B. Industry experts used a nine-scale method to compare and score the indicators pairwise, using a formula... calculate The final score, of which, For the first One expert on the indicators The score, The number of experts who participated in the scoring;

[0051] Based on the final score, the eigenvalues ​​are calculated using the product method. After vector normalization, solving for the maximum eigenvalue, and random consistency detection, the primary evaluation index is obtained. weight ;

[0052] Indicator Layer C: This refers to each primary evaluation indicator within Indicator Layer C. Industry experts used a nine-scale method to compare and score the indicators pairwise, using a formula... calculate The final score, of which, For the first One expert on the indicators The score, The number of experts who participated in the scoring;

[0053] Based on the final score, the eigenvalues ​​are calculated using the product method. After vector normalization, solving for the maximum eigenvalue, and random consistency detection, the primary evaluation index is obtained. weight ;

[0054] Sub-index layer D: The weight of each sub-index in sub-index layer D is set to 1.

[0055] Furthermore, in step S4, the initial quality evaluation system for the preliminary design report of the reinforcement project for dilapidated reservoirs is constructed from the indicator names and corresponding indicator weights of the target layer A, criterion layer B, indicator layer C, and sub-indicator layer D. .

[0056] Furthermore, step S5 specifically includes:

[0057] S51. Extract the feature data items and sub-indicators to be tested:

[0058] Initial feature set based on the preliminary design report of the reinforcement project for dilapidated reservoirs From the approved and pre-approval preliminary design reports of the dangerous reservoir reinforcement projects, extract the characteristic data items to be tested and record them as follows: Then extract the entity names to form the sub-indicators to be tested. ;

[0059] S52. Clarify the calculation formula for the actual score of sub-indicators:

[0060] The formula for calculating the actual score of a sub-indicator is as follows: ;

[0061] in, for The test result value, ; This is the function for calculating the actual score of the sub-indicator. For the feature data item being tested ; for eigenvalues; For feature type; Initial feature set for the preliminary design report of the reinforcement project for dilapidated reservoirs Feature data items in ; for eigenvalues; To verify the rules; As the scoring criteria;

[0062] S53. Match the corresponding test rules according to the feature type and calculate the actual score of the sub-indicator. :

[0063] Traverse the tested sub-indicator set For each sub-indicator in the set, the name of the currently tested sub-indicator will be compared with the initial sub-indicator set. All sub-indicator names are compared one by one in the initial sub-indicator set; if the name of the currently tested sub-indicator is not in the initial sub-indicator set... If a match is found, the actual score of the tested sub-indicator is determined. ;

[0064] If the names match, then... and Comparison, and when comparing, according to Feature type selection test rules:

[0065] like Feature types For numerical data, the difference ratio is used in RuleV1.1 mode to determine whether the numerical values ​​meet the requirements of the design specifications for the reinforcement and renovation of dilapidated reservoirs. Internally, or through the RuleV1.2 pattern for feature types. Determine whether it falls within the range of values ​​by comparison;

[0066] like Feature types For calculation-related cases, the calculation results are compared and judged using the RuleV3 mode.

[0067] like Feature types For word / vector classes, cosine similarity comparison is used in RuleV4 mode;

[0068] like Feature types For knowledge graphs, a comprehensive comparison method is used through the RuleV5 pattern;

[0069] S54. Calculate the actual scores of the indicator layer C, criterion layer B, and target layer A in sequence:

[0070] C in the indicator layer C j Actual score The calculation method is as follows: ;

[0071] In criterion layer B Actual score The calculation method is as follows: ; This is a first-level evaluation indicator in indicator layer C. The weights;

[0072] Calculate the actual score of target layer A The calculation method is as follows: ; This is the first-level evaluation indicator in criterion layer B. The weights;

[0073] The actual score of target layer A As the quality score result of the inspection report;

[0074] S55. Verify the correctness of the initial evaluation system, define the quality standards, and output the verified feature set. and verified evaluation system :

[0075] For both the preliminary design reports for the reinforcement and renovation of dilapidated reservoirs before review and the preliminary design reports for the reinforcement and renovation of dilapidated reservoirs after review and approval, the following procedures shall be performed:

[0076] Based on the method in step S51, the tested feature data items of the two types of reports are extracted. and extract The entity names in the test constitute the feature set. ;

[0077] Based on the method in steps S52-S54, the tested feature set Each sub-indicator in the report is sequentially subjected to name matching, rule adaptation, and hierarchical score calculation to obtain the quality score results for the two types of reports.

[0078] Integrate the quality scores from the reviewed and approved reports to form a benchmark score set; statistically analyze the score distribution range of this benchmark score set to verify the initial evaluation system. The rationality;

[0079] The quality score results of the pre-review report are compared with the score range of the above benchmark score set to determine the quality standard threshold of the preliminary design report for the reinforcement and renovation project of the dangerous reservoir.

[0080] The initial feature set is selected based on the quality standard threshold. The feature set that meets the standard is obtained from the preliminary design report of the reinforcement project of the dilapidated reservoir. The optimized preliminary design report for the reinforcement and renovation project of the dilapidated reservoir has been simultaneously output, and its quality has been verified and evaluated. .

[0081] Furthermore, in step S53, the calculation method for the RuleV1.1 mode is as follows:

[0082] ;

[0083] in , for The eigenvalues ​​of are denoted as . ; for The eigenvalues ​​of are denoted as . ;

[0084] The numerical values ​​are those required by the design specifications for the reinforcement and renovation of dilapidated reservoirs.

[0085] The calculation method for the RuleV1.2 mode is as follows:

[0086] ;

[0087] The calculation method for the RuleV3 mode is as follows:

[0088] From the feature data items being examined eigenvalues Extract the set of calculation results to be tested and the set of calculated parameters to be tested ;Will Recorded as ;

[0089] Will Substitute into the calculation program model In the middle, the recalculation result is obtained. , recorded as ;

[0090] From the initial feature set Feature data items eigenvalues Extract the standard calculation result set , recorded as ;

[0091] Calculate relative error and : , ;

[0092] Determine relative error , Does it exceed the set threshold? ;

[0093] ;

[0094] The calculation method for the RuleV4 mode is as follows:

[0095] The set of features to be examined in the preliminary design report for the reinforcement and renovation project of dilapidated reservoirs Initial feature set of the preliminary design report for the reinforcement and renovation project of dilapidated reservoirs The eigenvalues ​​in the data are compared and tested, specifically:

[0096] calculate eigenvalues and eigenvalues Cosine similarity between them:

[0097] ;

[0098] in: This is the dot product operation for vectors; , They are respectively and The modulus length;

[0099] ;

[0100] The calculation method for the RuleV5 mode is as follows:

[0101] Perform entity alignment checks on the knowledge graph, including checking the completeness of tag similarity, attribute similarity, and relation similarity; the calculation formula is:

[0102] ;

[0103] in, , , These are the weighting coefficients; For tag similarity, For attribute similarity, For relational similarity;

[0104] Among them, label similarity The cosine similarity is calculated using word vectors, and the formula is as follows:

[0105] ;

[0106] in, for entity tags The corresponding word vectors; for entity tags The corresponding word vectors;

[0107] Attribute similarity The calculation uses normalized Euclidean distance for numerical attributes and word vector similarity weighted aggregation for text attributes. The formula is as follows:

[0108] ;

[0109] in, The number of common attributes, For the first Similarity of attributes;

[0110] Relationship similarity The calculation formula is:

[0111] ;

[0112] in, Eigenvalues A collection of relation types; Eigenvalues A set of relation types;

[0113] for Adjacent nodes connected along relation r; for Adjacent nodes connected along relation r; This is the similarity function.

[0114] On the other hand, the present invention also provides a quality inspection system for the preliminary design report of a dangerous reservoir reinforcement project, which is used to implement the above inspection method. The system includes: a data collection module, a data preprocessing module, a data transmission module, a model generation module, an evaluation system module, an inspection module, a storage module, and a publishing output module.

[0115] The data collection module is used to acquire historical data related to the preliminary design report of the dangerous reservoir reinforcement project, and to build a corpus and an image library;

[0116] The data preprocessing module is used to receive the corpus and image library output by the data collection module, and perform classification preprocessing according to the two data formats of text and images respectively;

[0117] The data transmission module is used to establish a data interaction channel between modules using a transmission protocol;

[0118] The model generation module is used to receive preprocessed text data and image data, and generate text feature sets respectively through deep learning. and image feature set The initial feature set is obtained after merging. Based on the calculation content related to water conservancy projects in the corpus, a calculation program model was constructed. ;in, The name of the calculation program. For the calculation formula, For calculating parameters, For the calculation results, The number of input parameters. The number of calculation results, For the computation function of the computation program model;

[0119] The evaluation system module is used to determine the hierarchical structure model of quality evaluation indicators, and uses the Expert Hierarchy Process (AHP) method to determine the weights of indicators at each level, thus constructing the initial evaluation system. ;

[0120] The verification module is based on the initial feature set. Extract the tested feature set from the corpus and image database. Call the computational program model Accuracy verification of computational features; based on the initial evaluation system. Conduct scoring to validate the initial feature set. To improve the quality and obtain a verified feature set. and verified evaluation system It is also used based on verified feature sets. Extract the feature set to be tested from the preliminary design report of the dangerous reservoir reinforcement project to be reviewed. Call the calculation program model again Accuracy verification is performed on computational features, based on a validated evaluation system. The review reports are evaluated and scored to obtain the report quality inspection results;

[0121] The storage module is used to perform data storage and provide support for data retrieval by various modules;

[0122] The output module is used to output the report quality inspection results to the user via screen output or file output.

[0123] The beneficial effects of this invention are:

[0124] (1) Enhance the comprehensiveness of the inspection and cover the core dimensions of the report:

[0125] This invention utilizes deep learning technology to simultaneously extract textual and image features from design reports, constructing a multi-type initial feature set encompassing numerical, computational, word / vector, and knowledge graph categories. By combining four criterion-level indicators—comprehensiveness, standardization, logic, and accuracy—along with their subordinate secondary evaluation indicators, it achieves full-dimensional coverage of the report's content completeness, format standardization, scheme logic, and data accuracy, ensuring thorough verification.

[0126] (2) Enhance the accuracy of inspection and reduce subjective error:

[0127] By leveraging artificial intelligence technologies such as natural language processing and deep learning image processing, we can accurately extract entity names, entity relationships, and image features, generating standardized feature values ​​and knowledge graphs to provide objective data support for verification. The AHP evaluation method scientifically assigns weights to indicators at each level, avoiding subjective judgments from manual evaluation. Differentiated verification rules are developed for different feature types, and standardized verification is performed using computational program models, significantly reducing data errors and judgment biases, and improving the reliability of verification results.

[0128] (3) Achieve automated inspection, improve efficiency and reduce costs:

[0129] The fully automated inspection system constructed by this invention eliminates the need for deep human intervention in feature extraction, model construction, evaluation system establishment, inspection execution, and result output. This significantly improves inspection efficiency, reduces labor and time costs, and can meet the rapid inspection needs of design reports for the reinforcement of large-scale dilapidated reservoirs.

[0130] (4) Optimize inspection standards to ensure the quality of the renovation project:

[0131] By setting pre-defined quality standards, constructing standardized calculation program models and evaluation systems, a unified inspection benchmark and operating procedure are established to avoid inconsistencies in inspection results caused by differences in knowledge structure and experience among different reviewers. The inspection results provide design units with clear optimization directions, ensuring that design reports meet industry standards and guaranteeing the effectiveness of the reinforcement and renovation project for dilapidated reservoirs from the outset. Attached Figure Description

[0132] Figure 1 This is a flowchart illustrating the quality inspection method for the preliminary design report of a dangerous reservoir reinforcement project in an embodiment of the present invention.

[0133] Figure 2 This is a framework diagram of the quality inspection system for the preliminary design report of the reinforcement project for a dilapidated reservoir in this embodiment of the invention. Detailed Implementation

[0134] This invention aims to provide a method and system for quality inspection of preliminary design reports for the reinforcement and renovation of dilapidated reservoirs, addressing the problems of traditional methods that rely on subjective expert evaluation, are difficult to quantify in terms of quality, lack comprehensive indicators, are inefficient, and are easily influenced by subjective factors. Its core idea is to deeply integrate artificial intelligence technology with water conservancy project quality control to construct an automated, standardized, and precise quality inspection system for preliminary design reports of reinforcement and renovation projects for dilapidated reservoirs, replacing the traditional inspection model that relies on subjective expert evaluation.

[0135] Specifically, the present invention achieves the above core idea through the following means:

[0136] (1) Multi-source data integration and feature extraction: By collecting various types of data such as design reports, industry standards, and basic census data, a corpus and an image library are constructed. Text entities, relationships and knowledge graphs are extracted through natural language processing (Bert+BiLSTM+CRF). Image vector features are extracted through deep learning image processing (YOLO+ResNet50) to form a multi-dimensional initial feature set covering text and images.

[0137] (2) Construction of standardized model and evaluation system: In view of the core calculation needs of water conservancy projects, a program model containing 17 key calculations is established to provide a unified calculation benchmark; the AHP method is used to construct a hierarchical evaluation system of "target layer - criterion layer - indicator layer - sub-indicator layer", scientifically assigning weights to indicators to realize the quantification of evaluation standards.

[0138] (3) Categorized Precise Verification and Closed-Loop Validation: Based on feature types (numerical, computational, word / vector, knowledge graph), differentiated verification rules are formulated, and precise verification is achieved through methods such as difference ratio, threshold comparison, cosine similarity, and comprehensive comparison. The effectiveness of the feature set and evaluation system is first verified through the reviewed reports, and then the full-process automated verification is performed on the reports to be reviewed.

[0139] (4) Full-process automation and result visualization output: Based on the automated inspection method that does not require deep human intervention in the entire process of feature extraction, model construction, inspection execution and result output, the final output includes a labeled report, a modification suggestion comparison table and a quality inspection report, providing a clear basis for design optimization and quality control.

[0140] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0141] This embodiment first provides a method for quality inspection of preliminary design reports for the reinforcement and renovation of dilapidated reservoirs. (See [link to relevant documentation]). Figure 1 It includes the following implementation process:

[0142] S1. Collect data and build corpora and image libraries:

[0143] The core logic of this step is to provide comprehensive and authoritative basic data support for subsequent feature extraction and verification analysis, ensuring that the data covers the design report itself, industry standards, and basic background information.

[0144] In one exemplary implementation, the relevant data collected includes: the preliminary design report for the reinforcement and renovation of dilapidated reservoirs that has been reviewed and approved, the preliminary design report for the reinforcement and renovation of dilapidated reservoirs before the review of the corresponding approved report, the design specifications for the reinforcement and renovation of dilapidated reservoirs, dam registration data, national water conservancy census data, and statistical yearbooks of the regions where the dilapidated reservoirs are located.

[0145] The above data was organized into two categories: text and images, to construct a corpus and an image library. The corpus contains plain text content such as standard text, report text, and statistical data, while the image library covers image data such as engineering drawings, BIM model diagrams, and hydrogeological profiles from the report. All data is stored in a database, supporting high-speed reading and retrieval.

[0146] S2. After preprocessing, a feature set is generated through deep learning:

[0147] The core logic of this step is to remove invalid information through standardized preprocessing, and then use artificial intelligence technology to deeply extract the core features of text and images to form a structured feature set that can be used for verification.

[0148] In one exemplary implementation, preprocessing the corpus includes:

[0149] Tables and images were removed from the corpus, leaving only plain text to form a text dataset. The plain text dataset was then cleaned to remove unnecessary characters, spaces, and punctuation marks. Word segmentation technology was used to divide the continuous text into semantically meaningful lexical units, which were then tagged with parts of speech to obtain labeled text.

[0150] The preprocessing of the image library includes: converting the initial images in the image library to JPG format; classifying and archiving all JPG images by recognizing the title bar to create a first-level image directory; based on the first-level image directory, performing Label annotation and YOLO instance segmentation operations on the images in sequence, and further refining the classification by combining the segmented instance names to create a second-level image directory; using the segmented instances in the second-level image directory as a basis, generating a reservoir holographic library through BIM modeling using Revit; slicing the holographic library to obtain more refined instance images; and creating a third-level image directory based on these sliced ​​images; and finally, combining all images from the first-level, second-level, and third-level image directories to form an instance image set.

[0151] Based on the labeled text, this embodiment generates a text feature set in the following way. :

[0152] Training parameters (learning rate, number of iterations, batch size, etc.) are set, and the labeled text is imported into the BERT pre-trained layer for semantic encoding. Then, it is input into the BiLSTM network layer to generate a bidirectional hidden layer text vector sequence. The label transition probability is constrained using a CRF model to select the optimal label sequence that conforms to the logic of the water conservancy engineering field. Based on the consecutive "B-entity type" and "I-entity type" labels in the optimal label sequence, entity names e and inter-entity relationships r are located and extracted. Furthermore, the performance of the deep learning model can be verified using precision, recall, and F1 score. If the requirements are met, the extracted entity names e and inter-entity relationships r are considered valid.

[0153] Then, in Neo4j, the entity name e and the relationship r between entities are created as nodes and relationships using Cypher statements to generate a knowledge graph, which is then stored in the storage module.

[0154] Based on the example image set, this embodiment generates the image feature set in the following manner. :

[0155] The example image set is divided into training, validation, and test sets. A graph feature library is built using YOLO learning, and the resulting numerical vector set is input into the ResNet50 model and stored. The hierarchical image resources of the example image set are integrated with the numerical vector set output by the ResNet50 model to form an image feature set. .

[0156] In the above text feature set and image feature set Based on this, the two are merged to generate an initial feature set. Its expression is:

[0157] ;

[0158] in, This represents the total number of feature terms in the initial feature set. Indicates the first One feature term;

[0159] ;

[0160] Where e is the entity name, extracted using named entity recognition technology, and is in the form of Chinese characters or English words;

[0161] r represents the relationships between entities, which are extracted using relation extraction techniques.

[0162] V is the corresponding standard feature value extracted from the entity name e using regularization, and its form includes words, numbers, numerical ranges, knowledge graphs, or vector sets;

[0163] t represents the feature type, which can be numerical, computational, word / vector, or knowledge graph type.

[0164] Address is the location information corresponding to this feature in the relevant data;

[0165] RuleV is the test rule, which uses the difference ratio, numerical range comparison, calculation result threshold comparison, cosine similarity comparison or comprehensive comparison.

[0166] ScoreM is the scoring standard, which calculates the score based on the feature type t and the test rule RuleV, and the score range is limited to [0,1].

[0167] It should be noted that when the standard feature value V is presented in the form of a knowledge graph, its expression is: ;

[0168] in, , is a set of entities; A set of relationships, used to define the types of associations between entities; , is a set of fact triples used to quantify the specific association between the head entity, relation, and tail entity; , is a set of entity attributes, where, For entities The tag parameters, For the label field name, This refers to the specific value of the label field; For entities Attribute parameters, For attribute field names, This refers to the specific value of the attribute field.

[0169] When the standard eigenvalues ​​V are represented as a vector set, its expression is: ,in, This represents the dimension of the vector set. This indicates that the vector is in the first position. The value of the dimension.

[0170] S3. Constructing a computational program model:

[0171] The core logic of this step is to establish standardized calculation benchmarks for the core calculation processes in water conservancy engineering design.

[0172] In one exemplary implementation, the calculation program model constructed based on the relevant normative clauses and approved report calculation examples of water conservancy engineering calculations in the corpus is represented as follows:

[0173] ;in, The name of the calculation program. For the calculation formula, For calculating parameters, For the calculation results, The number of input parameters. The number of calculation results, This is the computation function for the computation program model.

[0174] In this embodiment, the calculation program model covers 17 key calculation modules: design storm calculation, design peak flow calculation, design flood total volume calculation, design flood hydrograph derivation, design flood rationality analysis, phased flood calculation, sediment calculation, dead water level verification, reservoir flood regulation calculation, comprehensive safety assessment, current dam crest elevation verification, dam seepage analysis, dam anti-sliding stability analysis, spillway hydraulic verification, earth-rock cofferdam design, investment estimate calculation, and economic evaluation calculation. That is, for each of these 17 key calculation modules, the core elements need to be clearly defined: name, calculation formula, calculation parameters, and calculation results.

[0175] S4. Constructing an initial evaluation system:

[0176] The core logic of this step is to establish hierarchical and quantitative evaluation standards, and to scientifically allocate indicator weights through the AHP method to ensure the scientific and rational nature of the evaluation system.

[0177] In one exemplary implementation, this step involves determining a hierarchical structure model of quality evaluation indicators, using the Analytic Hierarchy Process (AHP) to determine the weights of indicators at each level, and constructing an initial evaluation system. .

[0178] The hierarchical structure model includes target layer A, criterion layer B, indicator layer C, and sub-indicator layer D.

[0179] The target layer A is the quality of the preliminary design report for the reinforcement and renovation project of the dilapidated reservoir.

[0180] Criterion layer B includes primary evaluation indicators encompassing comprehensiveness, standardization, logic, and accuracy. Each primary evaluation indicator is designated as follows: , Values ​​range from 1 to 4; comprehensive evaluation indicators include secondary evaluation indicators such as engineering characteristics and basic information of the project area; standardized evaluation indicators include secondary evaluation indicators such as qualification management, report format, and cost information; logical evaluation indicators include secondary evaluation indicators such as scheme design, reinforcement measures, construction plan, demolition design, environmental protection design, soil and water conservation design, and construction drawing design; and accurate evaluation indicators include secondary evaluation indicators such as design storm rainfall calculation, design peak flow calculation, design flood total volume calculation, design flood process line derivation, design flood rationality analysis, phased flood calculation, sediment calculation, dead water level verification, reservoir flood regulation calculation, comprehensive safety assessment, current dam crest elevation verification, dam seepage analysis, dam anti-sliding stability analysis, spillway hydraulic verification, earth-rock cofferdam design, investment estimate calculation, and economic evaluation calculation.

[0181] Indicator layer C has There are 1 primary evaluation indicator, and each primary evaluation indicator is defined as follows: , Value ;

[0182] Sub-index layer D has Sub-indicators, from the initial feature set The N2 sub-indicators are derived from the initial sub-indicator set; that is, the N2 sub-indicators are composed of entity names extracted from the initial sub-indicator set. The initial sub-index set By extracting the initial feature set All feature data items Construct the corresponding entity name.

[0183] Based on the hierarchical model described above, the weights of each evaluation indicator are calculated as follows:

[0184] 1. Calculate the indicator weights of target layer A. Target layer A has only one indicator, and its weight is 1.

[0185] 2. Calculate the criteria layer B. Indicator weights For criterion layer B The indicators were scored by q senior industry review experts using a nine-point scale, with pairwise comparisons made according to importance. The q experts evaluated each indicator. The indicator score is recorded as u takes values ​​from 1 to q; then the scores of each expert are merged using a matrix to obtain... The indicator score is calculated using the following formula:

[0186] ;

[0187] The criteria layer B is derived. After the final score is assigned to the indicator, the eigenvalues ​​are calculated using the product method based on the final score. After vector normalization, solving for the maximum eigenvalue, and random consistency detection, the first-level evaluation indicator is obtained. weight .

[0188] 3. Calculate the index layer C. Indicator weights For each primary evaluation indicator in indicator layer C Industry experts were invited to use the nine-scale method to conduct pairwise comparisons and scores of the indicators, using a formula. calculate The final score, of which, For the u-th expert on the indicator The score is given by q, where q is the number of experts who participated in the scoring.

[0189] Based on the final score, the eigenvalues ​​are calculated using the product method. After vector normalization, solving for the maximum eigenvalue, and random consistency detection, the primary evaluation index is obtained. weight .

[0190] 4. Calculate the sub-index layer D. The weights of the indicators are all 1.

[0191] The final result is an initial quality evaluation system for the preliminary design report of the reinforcement project for dilapidated reservoirs, consisting of the indicator names and weights of the target layer (A), criterion layer (B), indicator layer (C), and sub-indicator layer (D). .

[0192] S5. Feature set and initial evaluation system verification:

[0193] The core logic of this step is to verify the initial feature set using the reviewed report data. With the initial evaluation system The effectiveness of the calibration feature matching accuracy and the rationality of the weights are determined, forming a resource that can be directly used for actual testing.

[0194] In one exemplary implementation, based on an initial feature set The feature set of the preliminary design report for the reinforcement and renovation project of dangerous reservoirs was extracted from the corpus and image database. The inspection process is executed, wherein the computational features are calculated using the computational program model constructed in step S3. Accuracy verification was performed; based on the initial evaluation system. Conduct scoring to validate the initial feature set. To improve the quality and obtain a verified feature set. and verified evaluation system .

[0195] More specifically, the detailed implementation process for this step is as follows:

[0196] S51. Extract the feature data items and sub-indicators to be tested:

[0197] Initial feature set based on the preliminary design report of the reinforcement project for dilapidated reservoirs From the approved and pre-approval preliminary design reports of the dangerous reservoir reinforcement projects, extract the characteristic data items to be tested and record them as follows: Then, extract the entity names to form the sub-indicators to be tested. .

[0198] S52. Clarify the calculation formula for the actual score of sub-indicators:

[0199] The formula for calculating the actual score of a sub-indicator is as follows: ;

[0200] in, for The test result value, ; This is the function for calculating the actual score of the sub-indicator. For the feature data item being tested ; for eigenvalues; For feature type; Initial feature set for the preliminary design report of the reinforcement project for dilapidated reservoirs Feature data items in ; for eigenvalues; To verify the rules; This serves as the scoring standard.

[0201] S53. Match the corresponding test rules according to the feature type and calculate the actual score of the sub-indicator. :

[0202] Traverse the tested sub-indicator set For each sub-indicator in the set, the name of the currently tested sub-indicator will be compared with the initial sub-indicator set. All sub-indicator names are compared one by one in the initial sub-indicator set; if the name of the currently tested sub-indicator is not in the initial sub-indicator set... If a match is found, the actual score of the tested sub-indicator is determined. ;

[0203] If the names match, then... and Comparison, and when comparing, according to Feature type t-selection test rules:

[0204] ①If If the feature type t is numerical, then the difference ratio is used in RuleV1.1 mode to determine whether it meets the numerical requirements of the design specifications for the reinforcement and renovation of dilapidated reservoirs. Within, or through the RuleV1.2 mode, a comparison is made to determine whether the feature type t is within the numerical range;

[0205] The calculation method for RuleV1.1 mode is as follows:

[0206] ;

[0207] in, , for The eigenvalues ​​of are denoted as . ; for The eigenvalues ​​of are denoted as . ; The values ​​are those required by the design specifications for the reinforcement and renovation of dilapidated reservoirs.

[0208] The calculation method for RuleV1.2 mode is as follows:

[0209] ;

[0210] ②If If the feature type t is a calculation class, then the threshold comparison of the calculation results is performed using the RuleV3 mode.

[0211] The calculation method for the RuleV3 mode is as follows:

[0212] From the feature data items being examined eigenvalues Extract the set of calculation results to be tested and the set of calculated parameters to be tested ;Will Recorded as ;

[0213] Will Substitute into the calculation program model In the middle, the recalculation result is obtained. , recorded as ;

[0214] From the initial feature set Feature data items eigenvalues Extract the standard calculation result set , recorded as ;

[0215] Calculate relative error and : , ;

[0216] Determine relative error , Does it exceed the set threshold? ;

[0217] ;

[0218] ③If If the feature type t is a word / vector class, then cosine similarity comparison is used through RuleV4 mode;

[0219] The calculation method for the RuleV4 mode is as follows:

[0220] The set of features to be examined in the preliminary design report for the reinforcement and renovation project of dilapidated reservoirs Initial feature set of the preliminary design report for the reinforcement and renovation project of dilapidated reservoirs The eigenvalues ​​in the data are compared and tested, specifically:

[0221] calculate eigenvalues and eigenvalues Cosine similarity between them:

[0222] ;

[0223] in: This is the dot product operation for vectors; , They are respectively and The modulus length;

[0224] ;

[0225] ④If If the feature type t is a knowledge graph class, then a comprehensive comparison method is adopted through the RuleV5 mode;

[0226] The calculation method for the RuleV5 mode is as follows:

[0227] Perform entity alignment checks on the knowledge graph, including checking the completeness of tag similarity, attribute similarity, and relation similarity; the calculation formula is:

[0228] ;

[0229] in, , , These are the weighting coefficients; For tag similarity, For attribute similarity, For relational similarity;

[0230] Among them, label similarity The cosine similarity is calculated using word vectors, and the formula is as follows:

[0231] ;

[0232] in, for entity tags The corresponding word vectors; for entity tags The corresponding word vectors;

[0233] Attribute similarity The calculation uses normalized Euclidean distance for numerical attributes and word vector similarity weighted aggregation for text attributes. The formula is as follows:

[0234] ;

[0235] in, The number of common attributes, For the first Similarity of attributes;

[0236] Relationship similarity The calculation formula is:

[0237] ;

[0238] in, Eigenvalues A collection of relation types; Eigenvalues A set of relation types;

[0239] for Adjacent nodes connected along relation r; for Adjacent nodes connected along relation r; This is the similarity function.

[0240] S54. Calculate the actual scores of the indicator layer C, criterion layer B, and target layer A in sequence:

[0241] C in the indicator layer C j Actual score The calculation method is as follows: ;

[0242] In criterion layer B Actual score The calculation method is as follows: ; This is a first-level evaluation indicator in indicator layer C. The weights;

[0243] Calculate the actual score of target layer A The calculation method is as follows: ; This is the first-level evaluation indicator in criterion layer B. The weights;

[0244] The actual score of target layer A As the quality score result of the inspection report.

[0245] S55. Verify the correctness of the initial evaluation system, define the quality standards, and output the verified feature set. and verified evaluation system :

[0246] For both the preliminary design reports for the reinforcement and renovation of dilapidated reservoirs before review and the preliminary design reports for the reinforcement and renovation of dilapidated reservoirs after review and approval, the following procedures shall be performed:

[0247] Based on the method in step S51, the tested feature data items of the two types of reports are extracted. and extract The entity names in the test constitute the feature set. ;

[0248] Based on the method in steps S52-S54, the tested feature set Each sub-indicator in the report is sequentially subjected to name matching, rule adaptation, and hierarchical score calculation to obtain the quality score results for the two types of reports.

[0249] Integrate the quality scores from the reviewed and approved reports to form a benchmark score set; statistically analyze the score distribution range of this benchmark score set to verify the initial evaluation system. The rationality;

[0250] The quality score results of the pre-review report are compared with the score range of the above benchmark score set to determine the quality standard threshold of the preliminary design report for the reinforcement and renovation project of the dangerous reservoir.

[0251] The initial feature set is selected based on the quality standard threshold. The feature set that meets the standard is obtained from the preliminary design report of the reinforcement project of the dilapidated reservoir. The optimized preliminary design report for the reinforcement and renovation project of the dilapidated reservoir has been simultaneously output, and its quality has been verified and evaluated. .

[0252] S6. Quality Inspection and Result Output of Reports to be Reviewed:

[0253] The core logic of this step is to perform fully automated inspection of the report to be reviewed based on the validated feature set and evaluation system, and output clear and practical inspection results to meet the needs of actual applications.

[0254] In one exemplary implementation, based on a verified feature set Extract the feature set to be tested from the preliminary design report of the dangerous reservoir reinforcement project to be reviewed. The inspection process is executed, wherein the computational features are calculated using the computational program model constructed in step S3. Accuracy verification was performed, and the results were based on a validated evaluation system. The review report is evaluated and scored.

[0255] It should be noted that the verification process and scoring calculation method in this step are the same as in step S5, but the difference lies in the input feature set: in step S5, the input is the feature set to be verified. The input for step S6 is a set of features extracted from the preliminary design report of the current reservoir reinforcement project to be reviewed. Furthermore, the evaluation system in step S5 adopts the initial evaluation system. The evaluation system used in step S6 is a validated evaluation system. Therefore, the specific verification process and scoring calculation method will not be elaborated further in this step.

[0256] Finally, the feature comparison results, scores at each level, and details of nonconformities during the inspection process are stored in a distributed database and output in three formats through the output publishing module:

[0257] The suggested modification table lists the non-compliant sub-indicators, corresponding defects, rectification basis, and rectification suggestions.

[0258] Inspected and marked preliminary design report: Add markings to the original report to be reviewed, mark the locations of non-conforming items in red, and attach brief inspection comments.

[0259] Quality inspection report: Summary report scores, pass / fail results, etc.

[0260] The above describes an implementation example of a quality inspection method for preliminary design reports of dilapidated reservoir reinforcement projects. To implement this method, this embodiment also provides a quality inspection system for preliminary design reports of dilapidated reservoir reinforcement projects. (See attached image.) Figure 2It includes:

[0261] The system comprises a data collection module, a data preprocessing module, a data transmission module, a model generation module, an evaluation system module, a verification module, a storage module, and a publishing output module. The functions of each module are as follows:

[0262] The data collection module is used to acquire historical data related to the preliminary design report of the dangerous reservoir reinforcement project, and to build a corpus and an image library;

[0263] The data preprocessing module is used to receive the corpus and image library output by the data collection module, and perform classification preprocessing according to the two data formats of text and images respectively;

[0264] The data transmission module is used to establish a data interaction channel between modules using a transmission protocol;

[0265] The model generation module is used to receive preprocessed text data and image data, and generate text feature sets respectively through deep learning. and image feature set The initial feature set is obtained after merging. Based on the calculation content related to water conservancy projects in the corpus, a calculation program model was constructed. ;in, The name of the calculation program. For the calculation formula, For calculating parameters, For the calculation results, The number of input parameters. The number of calculation results, For the computation function of the computation program model;

[0266] The evaluation system module is used to determine the hierarchical structure model of quality evaluation indicators, and uses the Expert Hierarchy Process (AHP) method to determine the weights of indicators at each level, thus constructing the initial evaluation system. ;

[0267] The verification module is based on the initial feature set. Extract the tested feature set from the corpus and image database. Call the computational program model Accuracy verification of computational features; based on the initial evaluation system. Conduct scoring to validate the initial feature set. To improve the quality and obtain a verified feature set. and verified evaluation system It is also used based on verified feature sets. Extract the feature set to be tested from the preliminary design report of the dangerous reservoir reinforcement project to be reviewed. Call the calculation program model again Accuracy verification is performed on computational features, based on a validated evaluation system. The review reports are evaluated and scored to obtain the report quality inspection results;

[0268] The storage module is used to perform data storage and provide support for data retrieval by various modules;

[0269] The output module is used to output the report quality inspection results to the user via screen output or file output.

[0270] It is understood that since the corresponding functions of each functional module in the above-mentioned report quality inspection system are consistent with the steps in the above-mentioned report quality inspection method, and given that the specific implementation means of the steps in the report quality inspection method have been described in detail in this embodiment, the specific implementation means of each functional module in the report quality inspection system will not be repeated.

[0271] Finally, it should be noted that although embodiments of the present invention have been described above, those skilled in the art will understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and all such changes and variations do not depart from the protection scope of the present invention.

Claims

1. A method for quality inspection of preliminary design reports for the reinforcement and renovation of dilapidated reservoirs, characterized in that, Includes the following steps: S1. Obtain relevant historical data and construct a corpus and image library; S2. After preprocessing the corpus and image database, text feature sets are generated separately using deep learning. and image feature sets The initial feature set is formed by merging these features. ; Generating text feature sets through deep learning The methods include: Set training parameters, import the labeled text into the Bert pre-training layer and BiLSTM network layer in sequence to generate a bidirectional hidden layer text vector sequence, and then connect the bidirectional hidden layer text vector sequence into the CRF model to construct a feature function sequence. Combine the label transition probability constraint to select the optimal label sequence. Based on the consecutive "B-entity type" and "I-entity type" labels in the optimal label sequence, the entity name e and the relationship r between entities are located and extracted; Import the extracted entity name e and the relationship r between entities into the Neo4j graph database, create the corresponding nodes and relationships using Cypher statements, generate a knowledge graph and store it. By integrating the extracted entity names e, inter-entity relationships r, and the generated knowledge graph, a text feature set is formed. ; Image feature sets are generated using deep learning. The methods include: The example image set is divided into training, validation, and test sets. A graph feature library is built using YOLO learning, and the resulting numerical vector set is input into the ResNet50 model and stored. The hierarchical image resources of the example image set are integrated with the numerical vector set output by the ResNet50 model to form an image feature set. ; In step S2, the initial feature set is generated. Represented as: ; in, This represents the total number of feature terms in the initial feature set. Indicates the first One feature term; ； Where e is the entity name, extracted using named entity recognition technology, and is in the form of Chinese characters or English words; r represents the relationships between entities, which are extracted using relation extraction techniques. V is the corresponding standard feature value extracted from the entity name e using regularization, and its form includes words, numbers, numerical ranges, knowledge graphs, or vector sets; t represents the feature type, which can be numerical, computational, word / vector, or knowledge graph type. Address is the location information corresponding to this feature in the relevant data; RuleV is the test rule, which uses the difference ratio, numerical range comparison, calculation result threshold comparison, cosine similarity comparison or comprehensive comparison. ScoreM is the scoring standard, which calculates the score based on the feature type t and the test rule RuleV, and the score range is limited to [0,1]. S3. Construct a calculation program model based on the calculation content related to water conservancy projects in the corpus. ;in, The name of the calculation program. For the calculation formula, For calculating parameters, For the calculation results, The number of input parameters. The number of calculation results, The computational functions are for the computational program model; the constructed computational program model covers: Design storm calculation, design peak flow calculation, design flood total volume calculation, design flood hydrograph derivation, design flood rationality analysis, phased flood calculation, sediment calculation, dead water level verification, reservoir flood regulation calculation, comprehensive safety assessment, current dam crest elevation verification, dam seepage analysis, dam anti-sliding stability analysis, spillway hydraulic verification, earth-rock cofferdam design, investment estimate calculation and economic evaluation calculation; S4. Determine the hierarchical structure model of quality evaluation indicators, use the Analytic Hierarchy Process (AHP) to determine the weights of indicators at each level, and construct the initial evaluation system. The hierarchical model includes: target layer A, criterion layer B, indicator layer C, and sub-indicator layer D; target layer A is the quality of the preliminary design report for the reinforcement project of the dilapidated reservoir. The criteria layer B includes primary evaluation indicators such as comprehensiveness, standardization, logic, and accuracy. Comprehensiveness includes secondary evaluation indicators such as engineering characteristics and basic information about the project area. Standardization includes secondary evaluation indicators such as qualification management, report format, and cost information. Logic includes secondary evaluation indicators such as scheme design, reinforcement measures, construction schemes, demolition design, environmental protection design, soil and water conservation design, and construction drawing design. Accuracy includes secondary evaluation indicators such as design storm calculation, design peak flow calculation, design flood total volume calculation, design flood hydrograph derivation, design flood rationality analysis, phased flood calculation, sediment calculation, dead water level verification, reservoir flood regulation calculation, comprehensive safety assessment, current dam crest elevation verification, dam seepage analysis, dam anti-sliding stability analysis, spillway hydraulic verification, earth-rock cofferdam design, investment estimate calculation, and economic evaluation calculation. Indicator layer C has There are 1 primary evaluation indicator, and each primary evaluation indicator is defined as follows: ; Sub-index layer D has Sub-indicators Sub-indicators are derived from the initial sub-indicator set. The initial sub-index set By extracting the initial feature set All feature data items Construct the corresponding entity name; In step S4, the AHP method is used to determine the weights of indicators at each level, including: Target layer A: Target layer A has only one indicator, with a weight set to 1; Quasi-testing layer B: This refers to the primary evaluation indicators within quasi-testing layer B. Industry experts used a nine-scale method to compare and score the indicators pairwise, using a formula... calculate The final score, of which, For the first One expert on the indicators The score, The number of experts who participated in the scoring; Based on the final score, the eigenvalues ​​are calculated using the product method. After vector normalization, solving for the maximum eigenvalue, and random consistency detection, the primary evaluation index is obtained. weight ; Indicator Layer C: This refers to each primary evaluation indicator within Indicator Layer C. Industry experts used a nine-scale method to compare and score the indicators pairwise, using a formula... calculate The final score, of which, For the first One expert on the indicators The score, The number of experts who participated in the scoring; Based on the final score, the eigenvalues ​​are calculated using the product method. After vector normalization, solving for the maximum eigenvalue, and random consistency detection, the primary evaluation index is obtained. weight ; Sub-indicator layer D: The weight of each sub-indicator in sub-indicator layer D is set to 1; The initial quality evaluation system for the preliminary design report of the reinforcement and renovation project of a dilapidated reservoir is composed of the indicator names and corresponding indicator weights of the target layer A, criterion layer B, indicator layer C, and sub-indicator layer D. ; S5. Based on the initial feature set The feature set of the preliminary design report for the reinforcement and renovation project of dangerous reservoirs was extracted from the corpus and image database. The inspection process is executed, wherein the computational features are calculated using the computational program model constructed in step S3. Accuracy verification was performed; based on the initial evaluation system. Conduct scoring to validate the initial feature set. To improve the quality and obtain a verified feature set. and verified evaluation system ; S6. Based on validated feature sets Extract the feature set to be tested from the preliminary design report of the dangerous reservoir reinforcement project to be reviewed. The inspection process is executed, wherein the computational features are calculated using the computational program model constructed in step S3. Accuracy verification was performed, and the results were based on a validated evaluation system. Evaluate and score the review report; Step S5 specifically includes: S51. Extract the feature data items and sub-indicators to be tested: Initial feature set based on the preliminary design report of the reinforcement project for dilapidated reservoirs From the approved and pre-approval preliminary design reports of the dangerous reservoir reinforcement projects, extract the characteristic data items to be tested and record them as follows: Then extract the entity names to form the sub-indicators to be tested. ; S52. Clarify the calculation formula for the actual score of sub-indicators: The formula for calculating the actual score of the sub-indicator is as follows: ; in, for The test result value, ; This is the function for calculating the actual score of the sub-indicator. For the feature data item being tested ; for eigenvalues; For feature type; Initial feature set for the preliminary design report of the reinforcement project for dilapidated reservoirs Feature data items in ; for eigenvalues; To verify the rules; As the scoring criteria; S53. Match the corresponding test rules according to the feature type and calculate the actual score of the sub-indicator. : Traverse the tested sub-indicator set For each sub-indicator in the set, the name of the currently tested sub-indicator will be compared with the initial sub-indicator set. All sub-indicator names are compared one by one in the initial sub-indicator set; if the name of the currently tested sub-indicator is not in the initial sub-indicator set... If a match is found, the actual score of the tested sub-indicator is determined. ; If the names match, then... and Comparison, and when comparing, according to Feature type selection test rules: like Feature types For numerical data, the difference ratio is used in RuleV1.1 mode to determine whether the numerical values ​​meet the requirements of the design specifications for the reinforcement and renovation of dilapidated reservoirs. Internally, or through the RuleV1.2 pattern for feature types. Determine whether it falls within the range of values ​​by comparison; like Feature types For calculation-related cases, the calculation results are compared and judged using the RuleV3 mode. like Feature types For word / vector classes, cosine similarity comparison is used in RuleV4 mode; like Feature types For knowledge graphs, a comprehensive comparison method is used through the RuleV5 pattern; S54. Calculate the actual scores of the indicator layer C, criterion layer B, and target layer A in sequence: C in the indicator layer C j Actual score The calculation method is as follows: ; In criterion layer B Actual score The calculation method is as follows: ; This is a first-level evaluation indicator in indicator layer C. The weights; Calculate the actual score of target layer A The calculation method is as follows: ; This is the first-level evaluation indicator in criterion layer B. The weights; The actual score of target layer A As the quality score result of the inspection report; S55. Verify the correctness of the initial evaluation system, define the quality standards, and output the verified feature set. and verified evaluation system : For both the preliminary design reports for the reinforcement and renovation of dilapidated reservoirs before review and the preliminary design reports for the reinforcement and renovation of dilapidated reservoirs after review and approval, the following procedures shall be performed: Based on the method in step S51, the tested feature data items of the two types of reports are extracted. and extract The entity names in the test constitute the feature set. ; Based on the method in steps S52-S54, the tested feature set Each sub-indicator in the report is sequentially subjected to name matching, rule adaptation, and hierarchical score calculation to obtain the quality score results for the two types of reports. Integrate the quality scores from the reviewed and approved reports to form a benchmark score set; statistically analyze the score distribution range of this benchmark score set to verify the initial evaluation system. The rationality; The quality score results of the pre-review report are compared with the score range of the above benchmark score set to determine the quality standard threshold of the preliminary design report for the reinforcement and renovation project of the dangerous reservoir. The initial feature set is selected based on the quality standard threshold. The feature set that meets the standard is obtained from the preliminary design report of the reinforcement project of the dilapidated reservoir. The optimized preliminary design report for the reinforcement and renovation project of the dilapidated reservoir has been simultaneously output, and its quality has been verified and evaluated. ; The calculation method for the RuleV1.1 mode is as follows: ； in , for The eigenvalues ​​of are denoted as . ; for The eigenvalues ​​of are denoted as . ; The numerical values ​​are those required by the design specifications for the reinforcement and renovation of dilapidated reservoirs. The calculation method for the RuleV1.2 mode is as follows: ； The calculation method for the RuleV3 mode is as follows: From the feature data items being examined eigenvalues Extract the set of calculation results to be tested and the set of calculated parameters to be tested ;Will Recorded as ; Will Substitute into the calculation program model In the middle, the recalculation result is obtained. , recorded as ; From the initial feature set Feature data items eigenvalues Extract the standard calculation result set , recorded as ; Calculate relative error and : , ; Determine the relative error , Does it exceed the set threshold? ; ； The calculation method for the RuleV4 mode is as follows: The set of features to be examined in the preliminary design report for the reinforcement and renovation project of dilapidated reservoirs Initial feature set of the preliminary design report for the reinforcement and renovation project of dilapidated reservoirs The eigenvalues ​​in the data are compared and tested, specifically: calculate eigenvalues and eigenvalues Cosine similarity between them: ； in: This is the dot product operation for vectors; , They are respectively and The modulus length; ； The calculation method for the RuleV5 mode is as follows: Perform entity alignment checks on the knowledge graph, including checking the completeness of tag similarity, attribute similarity, and relation similarity; the calculation formula is: ； in, , , These are the weighting coefficients; For tag similarity, For attribute similarity, For relational similarity; Among them, label similarity The cosine similarity is calculated using word vectors, and the formula is as follows: ； in, for entity tags The corresponding word vectors; for entity tags The corresponding word vectors; Attribute similarity The calculation uses normalized Euclidean distance for numerical attributes and word vector similarity weighted aggregation for text attributes. The formula is as follows: ； in, The number of common attributes, For the first Similarity of attributes; Relationship similarity The calculation formula is: ； in, Eigenvalues A collection of relation types; Eigenvalues A set of relation types; for Adjacent nodes connected along relation r; for Adjacent nodes connected along relation r; This is the similarity function.

2. The method for quality inspection of preliminary design reports for the reinforcement and renovation of dilapidated reservoirs as described in claim 1, characterized in that, In step S1, the relevant historical data includes the preliminary design report of the reinforcement project for the dilapidated reservoir before review, the preliminary design report of the reinforcement project for the dilapidated reservoir after review and approval, the design specifications for the reinforcement project for the dilapidated reservoir, the dam registration data, the national water conservancy census data, and the statistical yearbook of the region where the dilapidated reservoir is located.

3. The quality inspection method for the preliminary design report of a dangerous reservoir reinforcement project as described in claim 1, characterized in that, In step S2, the preprocessing of the corpus and image database includes: Tables and images were removed from the corpus, leaving only plain text to form a text dataset. The plain text dataset was then cleaned to remove unnecessary characters, spaces, and punctuation marks. Word segmentation technology was used to divide the continuous text into semantically meaningful lexical units, and part-of-speech tagging was performed to obtain tagged text. The initial images in the image library are uniformly converted to JPG format. First, all JPG images are categorized and archived using title bar recognition to create a first-level image directory. Based on the first-level image directory, the images are sequentially annotated with Labels and segmented using YOLO. The categories are further refined based on the segmented instance names to create a second-level image directory. Based on the segmented instances in the second-level image directory, BIM modeling is performed using Revit to generate a reservoir holographic library. The holographic library is then sliced ​​to obtain more detailed instance images, and a third-level image directory is created based on these sliced ​​images. All images in the first-level, second-level, and third-level image directories are combined to form an instance image set.

4. A quality inspection system for preliminary design reports of dilapidated reservoir reinforcement projects, used to implement the quality inspection method for preliminary design reports of dilapidated reservoir reinforcement projects as described in any one of claims 1 to 3, characterized in that, The system includes: a data collection module, a data preprocessing module, a data transmission module, a model generation module, an evaluation system module, a verification module, a storage module, and a publishing output module; The data collection module is used to acquire historical data related to the preliminary design report of the dangerous reservoir reinforcement project, and to build a corpus and an image library; The data preprocessing module is used to receive the corpus and image library output by the data collection module, and perform classification preprocessing according to the two data formats of text and images respectively; The data transmission module is used to establish a data interaction channel between modules using a transmission protocol; The model generation module is used to receive preprocessed text data and image data, and generate text feature sets respectively through deep learning. and image feature sets The initial feature set is obtained after merging. Based on the calculation content related to water conservancy projects in the corpus, a calculation program model was constructed. ;in, The name of the calculation program. For the calculation formula, For calculating parameters, For the calculation results, The number of input parameters. The number of calculation results, For the computation function of the computation program model; The evaluation system module is used to determine the hierarchical structure model of quality evaluation indicators, and uses the Expert Hierarchy Process (AHP) method to determine the weights of indicators at each level, thus constructing the initial evaluation system. ; The verification module is based on the initial feature set. Extract the tested feature set from the corpus and image database. Call the computational program model Accuracy verification of computational features; based on the initial evaluation system. Conduct scoring to validate the initial feature set. To improve the quality and obtain a verified feature set. and verified evaluation system It is also used based on verified feature sets. Extract the feature set to be tested from the preliminary design report of the dangerous reservoir reinforcement project to be reviewed. Call the calculation program model again Accuracy verification is performed on computational features, based on a validated evaluation system. The review reports are evaluated and scored to obtain the report quality inspection results; The storage module is used to perform data storage and provide support for data retrieval by various modules; The output module is used to output the report quality inspection results to the user via screen output or file output.

Images