A method, system and medium for checking and reviewing inspection reports based on a large model

Abstract

The present invention discloses a method, system and medium for checking and auditing inspection reports based on a large model, and belongs to the field of special equipment inspection technology. The method includes: S1: model selection and adjustment stage, selecting a pre-trained model that has been pre-trained, and adjusting the pre-trained model according to the inspection report data set to obtain an inspection model; S2: inspection report data preprocessing stage, preprocessing the target inspection report data to obtain pre-processed inspection report data; S3: verification and audit stage, using the inspection model to verify or audit the pre-processed inspection report data to obtain the first inspection report data; S4: verification and audit result output stage, converting the first inspection report data into an operable verification and audit result and outputting it. It can improve the accuracy and reliability of inspection report verification and audit; improve the efficiency of inspection report verification and audit, and save labor costs.

Description

Technical Field

[0001] The present invention relates to the technical field of special equipment inspection, and in particular to a large model-based inspection report verification and auditing method, system and medium. Background Art

[0002] With the increase in the number of special equipment, the number of inspection and testing reports has continued to grow. In order to ensure the quality of the inspection and testing reports, relevant specifications and quality management systems clearly require the verification and review of the inspection and testing reports. However, there are many problems with manual verification or review: 1. Visual fatigue and misjudgment: Prolonged concentration may lead to visual fatigue, which affects the accurate capture of details and increases the risk of misjudgment.

[0003] 2. Susceptible to external interference: External factors such as noise, light, and temperature in the working environment may affect the auditor's concentration and judgment. Personal emotions, health status, and other factors may also affect the objectivity and accuracy of the audit. 3. Subjectivity and consistency issues: Different auditors may have different understandings and judgments of the report, resulting in different audit standards. Personal experience, preferences, or biases may subtly affect the audit results, and there is a lack of unified quantitative standards. 4. Inefficiency: The speed of manual audit is limited by personal ability, and the efficiency is low when processing a large number of reports, which is difficult to meet high-throughput requirements. Repetitive labor accounts for a large proportion, which is easy to cause resource waste and rising labor costs. 5. Cost issues: The training and maintenance costs of high-quality auditors are high, especially in professional fields such as special equipment inspection. Manual audits have poor scalability, and costs increase linearly with business growth. 6. Reliability and traceability: Manual audit records may not be systematic enough, and once problems arise, tracing and auditing become difficult. Relying on personal memory and paper document storage methods increases the risk of information loss or damage. 7. Safety considerations: The safety risks involved in special equipment inspection require extremely high accuracy and reliability. Any omissions in manual review may have serious consequences. Summary of the invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and to provide a method, system and medium for checking and reviewing inspection reports based on a large model.

[0005] The object of the present invention is achieved through the following technical solutions: The first aspect of the present invention provides: a large model-based inspection report verification and auditing method, comprising the following steps:

[0006] S1: Model selection and adjustment stage, select a pre-trained model that has been pre-trained, and adjust the pre-trained model according to the test report data set to obtain the test model;

[0007] S2: Inspection report data preprocessing stage, preprocessing the target inspection report data to obtain preprocessed inspection report data, wherein the preprocessing includes data cleaning, text standardization, de-identification, structured data extraction and data integration mapping;

[0008] S3: Verification and review phase, using the inspection model to verify or review the pre-processed inspection report data to obtain the first inspection report data, including: semantic analysis, anomaly detection, pattern matching and rule engine configuration, model readjustment, and model iteration;

[0009] S4: Verification and audit result output stage, converting the first inspection report data into actionable verification and audit results and outputting them, including: error highlighting and marking, error type classification, suggestions and correction plans, and report generation.

[0010] Preferably, the S1: model selection and adjustment stage further includes the following steps:

[0011] Obtaining a test report data set, performing data cleaning on the test report data set, removing noise, errors, and duplicates, and standardizing the text format to obtain a first data set; then annotating the first data through supervised learning to obtain a second data set;

[0012] Use a tokenizer to tokenize text data, truncate or pad all sequences to a preset length to obtain tokenization results; generate the input ID and attention mask of the pre-trained model based on the tokenization results;

[0013] Add the corresponding output layer to the pre-trained model according to the target task; set hyperparameters, including learning rate, batch size, and number of training rounds;

[0014] Load the training weights of the pre-trained model; use the test report training set to train the pre-trained model to obtain the first model; use the test report validation set to evaluate the performance of the first model, adjust the hyperparameters according to the evaluation results to obtain the second model, thereby obtaining the optimal model weights and saving them;

[0015] The application feedback of the second model is collected, and the model is tuned according to the application feedback to obtain a test model.

[0016] Preferably, the S2: test report data preprocessing stage further includes the following steps:

[0017] Clean the target inspection report data, remove or replace special symbols, meaningless blank characters, and line breaks in the target inspection report data, use statistical methods to identify outliers in the target inspection report data and correct or delete them, use interpolation methods to process missing data points in the target inspection report data, remove duplicate data, non-numerical noise, and data with incorrect formats in the target inspection report data, and finally obtain the first target inspection report data;

[0018] De-identifying the first target inspection report data using a hash function or a data masking method to obtain second target inspection report data;

[0019] Perform text standardization on the second target inspection report data, use industry terminology library or ontology to standardize the professional terms and abbreviations in the second target inspection report data, and then convert various measurement units in the second target inspection report data into the international standard unit system, and finally obtain the third target inspection report data;

[0020] Perform structured data extraction on the third-target inspection report data, use natural language processing methods and regular expressions to extract device ID, model, inspection date, inspection results, and location coordinates from the free text of the third-target inspection report data, interpret the inspection code and standard references and convert them into computable data, judge the safety status of the target device and record it to obtain preprocessed inspection report data.

[0021] Preferably, the S3: verification and review stage further includes the following steps:

[0022] Use the inspection model to perform semantic analysis on the pre-processed inspection report data. Through multi-task learning, the inspection model learns the language features and professional terms in the technical field contained in the pre-processed inspection report data, and adds an attention mechanism to the inspection model.

[0023] The test model then performs anomaly detection on the preprocessed test report data, establishes a statistical baseline based on historical data, and uses hypothesis testing methods to identify numerical anomalies; then the preprocessed test report data is checked for consistency;

[0024] Perform pattern matching and rule engine configuration on the inspection model, build or integrate the knowledge graph of the technical fields contained in the preprocessed inspection report data in the inspection model and set the rule engine to be configurable;

[0025] Re-adjust the test model, and perform transfer learning and domain adaptation on the test model;

[0026] Finally, the verification model is continuously iterated through online learning and model fusion.

[0027] Preferably, the S4: verification and audit result output stage further includes the following steps:

[0028] Error highlighting and marking of the first inspection report data, using color coding or different font styles to highlight detected errors or potential problems, and providing a location index to locate each error or potential problem;

[0029] Classifying the error types of the first inspection report data by fine-grained classification and assigning a unique error code to each type of error;

[0030] Provide suggestions and correction plans for detected errors. When a first-type error is detected, the verification model provides automatic correction options through an automatic correction method; when a second-type error is detected, the verification model provides correction suggestions based on expert system rules; when a third-type error is detected, the verification model automatically triggers a manual review request and attaches error analysis details;

[0031] Generate a verification and audit report, template the verification and audit report, provide data analysis in the verification and audit report, indicate the audit status, obtain the verification and audit results and output them.

[0032] Preferably, the pre-trained model is a BERT, RoBERTa or GPT series model.

[0033] The second aspect of the present invention provides: a large model-based inspection report verification and auditing system, used to implement any of the above-mentioned large model-based inspection report verification and auditing methods, including:

[0034] The large model module connects the user interaction interface, data preprocessing module, back-end service and database, and verification and audit result output module; the user interaction interface is used for file upload, progress tracking, result display, and interactive correction; the back-end service and database are used to provide API services, database management system, and model services.

[0035] The third aspect of the present invention provides: a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are loaded and executed by a processor, any of the above-mentioned inspection report verification and auditing methods based on large models is implemented.

[0036] The beneficial effects of the present invention are:

[0037] 1) Improve the accuracy and reliability of inspection report verification and audit; improve the efficiency of inspection report verification and audit, and save labor costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1It is a flow chart of the inspection report verification and audit method based on the big model;

[0039] Figure 2 This is the structural diagram of the inspection report verification and audit system based on the big model;

[0040] Figure 3 This is a functional diagram of the data preprocessing module;

[0041] Figure 4 It is a functional diagram of the large model module;

[0042] Figure 5 It is a functional diagram of the module for checking and reviewing the results output;

[0043] Figure 6 It is a functional diagram of the user interaction interface;

[0044] Figure 7 Schematic diagram of the database function serving the backend. DETAILED DESCRIPTION

[0045] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0046] See also Figure 1-Figure 7 The first aspect of the present invention provides: a method for checking and reviewing an inspection report based on a large model, comprising the following steps:

[0047] S1: Model selection and adjustment stage, select a pre-trained model that has been pre-trained, and adjust the pre-trained model according to the test report data set to obtain the test model;

[0048] S2: Inspection report data preprocessing stage, preprocessing the target inspection report data to obtain preprocessed inspection report data, wherein the preprocessing includes data cleaning, text standardization, de-identification, structured data extraction and data integration mapping;

[0049] S3: Verification and review phase, using the inspection model to verify or review the pre-processed inspection report data to obtain the first inspection report data, including: semantic analysis, anomaly detection, pattern matching and rule engine configuration, model readjustment, and model iteration;

[0050] S4: Verification and audit result output stage, converting the first inspection report data into actionable verification and audit results and outputting them, including: error highlighting and marking, error type classification, suggestions and correction plans, and report generation.

[0051] In this embodiment, a model that has been pre-trained on large-scale text data is selected in S1, such as BERT, RoBERTa or GPT series. Based on the pre-trained deep learning model, the pre-trained model is fine-tuned on the data set of the inspection report to adapt it to the specific language and structure of the inspection report to obtain the inspection model, and the fine-tuned inspection model is specifically used to understand and check or review the inspection report. In S2, the original inspection report received is subjected to data cleaning, text standardization, de-identification, structured data extraction and data integration mapping, etc., for subsequent model processing. In S3, the inspection model model after fine-tuning is used to check or review the inspection report after preprocessing, including: understanding and semantic analysis of the inspection report text, anomaly detection, pattern matching and rule engine, model fine-tuning, continuous learning and updating, etc. In S4, the analysis results of the large model module are converted into actionable information for use by users or downstream systems. Including: error highlighting and marking in the inspection report, error type classification, suggestions and correction solutions, and report generation.

[0052] In some embodiments, the S1: model selection and adjustment stage further includes the following steps:

[0053] Obtaining a test report data set, performing data cleaning on the test report data set, removing noise, errors, and duplicates, and standardizing the text format to obtain a first data set; then annotating the first data through supervised learning to obtain a second data set;

[0054] Use a tokenizer to tokenize text data, truncate or pad all sequences to a preset length to obtain tokenization results; generate the input ID and attention mask of the pre-trained model based on the tokenization results;

[0055] Add the corresponding output layer to the pre-trained model according to the target task; set hyperparameters, including learning rate, batch size, and number of training rounds;

[0056] Load the training weights of the pre-trained model; use the test report training set to train the pre-trained model to obtain the first model; use the test report validation set to evaluate the performance of the first model, adjust the hyperparameters according to the evaluation results to obtain the second model, thereby obtaining the optimal model weights and saving them;

[0057] The application feedback of the second model is collected, and the model is tuned according to the application feedback to obtain a test model.

[0058] In this embodiment, a large number of relevant inspection reports are collected as inspection report data sets, and these reports cover all types and formats that the model is expected to understand. Through supervised learning (such as classification or named entity recognition), the data is annotated, including: marking key information points or classification labels in the report. The word segmenter is a word segmenter that can be used by the pre-trained model. For example, BERT uses the WordPiece word segmenter. All sequences are truncated or padded to ensure that the input sequence length meets the model requirements. The target tasks include classification, regression, sequence labeling, etc. When training the pre-trained model, indicators such as loss and accuracy are monitored during the training process. The performance of the first model is evaluated using the inspection report validation set to prevent overfitting. At the same time, the generalization ability of the inspection model can be evaluated on unseen test data, the performance of the inspection model in various situations can be analyzed, possible problem areas can be identified, and feedback can be collected from users or actual applications to understand the shortcomings of the inspection model.

[0059] In some embodiments, the S2: inspection report data preprocessing stage further includes the following steps:

[0060] Clean the target inspection report data, remove or replace special symbols, meaningless blank characters, and line breaks in the target inspection report data, use statistical methods to identify outliers in the target inspection report data and correct or delete them, use interpolation methods to process missing data points in the target inspection report data, remove duplicate data, non-numerical noise, and data with incorrect formats in the target inspection report data, and finally obtain the first target inspection report data;

[0061] De-identifying the first target inspection report data using a hash function or a data masking method to obtain second target inspection report data;

[0062] Perform text standardization on the second target inspection report data, use industry terminology library or ontology to standardize the professional terms and abbreviations in the second target inspection report data, and then convert various measurement units in the second target inspection report data into the international standard unit system, and finally obtain the third target inspection report data;

[0063] Perform structured data extraction on the third-target inspection report data, use natural language processing methods and regular expressions to extract device ID, model, inspection date, inspection results, and location coordinates from the free text of the third-target inspection report data, interpret the inspection code and standard references and convert them into computable data, judge the safety status of the target device and record it to obtain preprocessed inspection report data.

[0064] In this embodiment, the statistical methods include Z-Score, IQR, etc., and after the abnormal value is identified, it is corrected or deleted according to industry standards or historical data. The interpolation method includes an interpolation method based on neighboring equipment data, an average interpolation method based on equipment type, etc. De-identification is to protect personal or corporate information such as equipment owners and inspectors, ensure the anonymization of customer information, and comply with relevant data privacy regulations. Industry terminology libraries or ontologies include ASME, ISO standard terms, etc. The natural speech processing method includes named entity recognition methods (NER) and other NLP methods to extract key information from free text, such as equipment type, inspection date, inspector, inspection item, result value, etc., and convert it into structured data. Natural language processing technology (such as named entity recognition NER) and domain-specific regular expressions are used to identify and extract key information from free text, such as equipment ID, model, inspection date, inspection result, defect type, location coordinates, etc. Interpret inspection codes and standard references and convert them into computable forms, such as converting "ASME BPVC Section VIII" into a corresponding inspection item list. Based on the inspection results and standard specifications, the safety status of the equipment is automatically determined, such as qualified, unqualified, maintenance required, etc.

[0065] In some embodiments, the S3: verification and review phase further includes the following steps:

[0066] Use the inspection model to perform semantic analysis on the pre-processed inspection report data. Through multi-task learning, the inspection model learns the language features and professional terms in the technical field contained in the pre-processed inspection report data, and adds an attention mechanism to the inspection model.

[0067] The test model then performs anomaly detection on the preprocessed test report data, establishes a statistical baseline based on historical data, and uses hypothesis testing methods to identify numerical anomalies; then the preprocessed test report data is checked for consistency;

[0068] Perform pattern matching and rule engine configuration on the inspection model, build or integrate the knowledge graph of the technical fields contained in the preprocessed inspection report data in the inspection model and set the rule engine to be configurable;

[0069] Re-adjust the test model, and perform transfer learning and domain adaptation on the test model;

[0070] Finally, the verification model is continuously iterated through online learning and model fusion.

[0071] In this embodiment, the inspection model first trains a large amount of text data such as special equipment inspection reports to learn the language features and professional terms in the field of special equipment. For example, a bidirectional Transformer model similar to BERT is used, which can understand the semantic relationship in the special equipment inspection report, including basic information, equipment category, inspection personnel, inspection items, inspection results, diagnostic conclusions, etc. The self-attention mechanism is added to the Transformer architecture so that the model can focus on key parts of the report, such as outliers, key terms, etc., to improve the efficiency of interpretation. The hypothesis testing method includes a t-test method and a chi-square test method to identify numerical anomalies that deviate from the normal range. Logical reasoning and rule engines are used to verify the consistency of information in the report, such as the matching degree between equipment type and inspection items. By building or integrating a knowledge graph in the field of special equipment, the model can query information such as equipment type, standard, inspection process, etc. to assist in report review. Based on the special equipment knowledge base and safety technical specification standards, check whether the information in the report complies with the safety technical specification standards related to special equipment. A configurable rule engine is implemented to allow dynamic adjustment of inspection rules and thresholds according to the latest industry standards and regulations. The large model can integrate the knowledge graph or rule base in the field of special equipment to ensure that its decisions and recommendations meet the latest specifications and standards for special equipment inspection and testing. In this way, even in the face of rare or complex situations, the model can give judgments based on professional knowledge. Based on domain-specific training data, the model is further optimized to improve its accuracy and robustness in inspection report review. Through transfer learning strategies, the knowledge of the inspection model is transferred to the field of special equipment to accelerate the learning of the model on new tasks. Use domain adaptation techniques (such as adversarial training and joint training) to reduce the differences between the source domain (pre-training data) and the target domain (inspection report data) to improve the performance of the model on the target task. By continuously learning new special equipment inspection knowledge and updated data, the large model can continuously improve its accuracy and applicability and gradually adapt to the development and changes in the field of special equipment. The inspection model has the ability to learn online, and can learn from new inspection reports in real time or regularly, and automatically update model parameters to cope with the latest developments and changes in the field of special equipment. Combine the predictions of multiple models and use model fusion techniques (such as weighted average and stacked generalization) to improve the accuracy and stability of the final decision.

[0072] In some embodiments, the S4: verification and audit result output stage further includes the following steps:

[0073] Error highlighting and marking of the first inspection report data, using color coding or different font styles to highlight detected errors or potential problems, and providing a location index to locate each error or potential problem;

[0074] Classifying the error types of the first inspection report data by fine-grained classification and assigning a unique error code to each type of error;

[0075] Provide suggestions and correction plans for detected errors. When a first-type error is detected, the verification model provides automatic correction options through an automatic correction method; when a second-type error is detected, the verification model provides correction suggestions based on expert system rules; when a third-type error is detected, the verification model automatically triggers a manual review request and attaches error analysis details;

[0076] Generate a verification and audit report, template the verification and audit report, provide data analysis in the verification and audit report, indicate the audit status, obtain the verification and audit results and output them.

[0077] In this embodiment, color coding or different font styles are used to highlight detected errors or potential problems, such as red for serious errors, yellow for warnings, and green for confirmation. Accurate line numbers and paragraph indexes are provided for each error or exception to facilitate quick positioning. Classification is performed according to the nature of the error, such as numerical errors, logical errors, format errors, etc. In addition to basic numerical errors, logical errors, and format errors, professional categories such as misuse of terminology, non-compliance with standards, and missing information can also be included. A unique error code is assigned to each type of error for easy tracking and management. For detected errors, possible correction suggestions or solutions are provided. Once a potential error or anomaly is found, the model can make automatic correction suggestions or remind the checker or auditor to conduct a manual review. For simple errors, such as spelling errors, inconsistent units, etc., automatic correction options are provided. For complex errors, such as the interpretation of abnormal test results, the model can make correction suggestions based on expert system rules. When the model cannot determine the correction solution, a manual review request is automatically triggered with model analysis details. An audit report is automatically generated to summarize the verification results, including an error summary, correction suggestions, and overall audit status. Use templates to automatically generate audit reports to ensure that the report format is uniform and easy to read and understand. The report contains error statistics, trend analysis, and risk assessment to help users understand the overall quality of the report and the direction of improvement. Clearly mark the audit status of the report, such as "audited", "pending review", or "audit failed".

[0078] In some embodiments, the pre-trained model is a BERT, RoBERTa or GPT series model.

[0079] The second aspect of the present invention provides: a large model-based inspection report verification and auditing system, used to implement any of the above-mentioned large model-based inspection report verification and auditing methods, including:

[0080] The large model module connects the user interaction interface, data preprocessing module, back-end service and database, and verification and audit result output module; the user interaction interface is used for file upload, progress tracking, result display, and interactive correction; the back-end service and database are used to provide API services, database management system, and model services.

[0081] In this embodiment, the user interface supports uploading inspection report files in multiple formats; it can display the audit progress, let users understand the current status and clearly display the audit results, including error lists and correction suggestions; it also allows users to modify the report directly on the interface and save the changes. The backend service and database provide RESTful APIs for the front-end interface or other systems to call; store user data, audit records, model versions and configuration information; deploy and manage large model instances to ensure efficient and secure execution of audit tasks.

[0082] The third aspect of the present invention provides: a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are loaded and executed by a processor, any of the above-mentioned inspection report verification and auditing methods based on large models is implemented.

[0083] The above is only a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art shall not deviate from the spirit and scope of the present invention, and shall be within the scope of protection of the claims attached to the present invention.

Claims

1. A method for checking and reviewing inspection reports based on a large model, characterized in that: The following steps are involved: S1: Model selection and adjustment stage, select a pre-trained model that has been pre-trained, and adjust the pre-trained model according to the test report data set to obtain the test model; S2: Inspection report data preprocessing stage, preprocessing the target inspection report data to obtain preprocessed inspection report data, wherein the preprocessing includes data cleaning, text standardization, de-identification, structured data extraction and data integration mapping; S3: Verification and review phase, using the inspection model to verify or review the pre-processed inspection report data to obtain the first inspection report data, including: semantic analysis, anomaly detection, pattern matching and rule engine configuration, model readjustment, and model iteration; S4: Verification and audit result output stage, converting the first inspection report data into actionable verification and audit results and outputting them, including: error highlighting and marking, error type classification, suggestions and correction plans, and report generation; The S1: model selection and adjustment phase also includes the following steps: Obtaining a test report data set, performing data cleaning on the test report data set, removing noise, errors, and duplicates, and standardizing the text format to obtain a first data set; then annotating the first data through supervised learning to obtain a second data set; Use a tokenizer to tokenize text data, truncate or pad all sequences to a preset length to obtain tokenization results; generate the input ID and attention mask of the pre-trained model based on the tokenization results; Add the corresponding output layer to the pre-trained model according to the target task; set hyperparameters, including learning rate, batch size, and number of training rounds; Load the training weights of the pre-trained model; use the test report training set to train the pre-trained model to obtain the first model; use the test report validation set to evaluate the performance of the first model, adjust the hyperparameters according to the evaluation results to obtain the second model, thereby obtaining the optimal model weights and saving them; Collect application feedback of the second model, and optimize the model according to the application feedback to obtain a test model; The S3: verification and review phase also includes the following steps: Use the inspection model to perform semantic analysis on the pre-processed inspection report data. Through multi-task learning, the inspection model learns the language features and professional terms in the technical field contained in the pre-processed inspection report data, and adds an attention mechanism to the inspection model. The test model then performs anomaly detection on the preprocessed test report data, establishes a statistical baseline based on historical data, and uses hypothesis testing methods to identify numerical anomalies; then the preprocessed test report data is checked for consistency; Perform pattern matching and rule engine configuration on the inspection model, build or integrate the knowledge graph of the technical fields contained in the preprocessed inspection report data in the inspection model and set the rule engine to be configurable; Re-adjust the test model, and perform transfer learning and domain adaptation on the test model; Finally, the verification model is continuously iterated through online learning and model fusion.

2. The inspection report verification and auditing method based on a large model according to claim 1 is characterized in that: The S2: inspection report data preprocessing stage further includes the following steps: Clean the target inspection report data, remove or replace special symbols, meaningless blank characters, and line breaks in the target inspection report data, use statistical methods to identify outliers in the target inspection report data and correct or delete them, use interpolation methods to process missing data points in the target inspection report data, remove duplicate data, non-numerical noise, and data with incorrect formats in the target inspection report data, and finally obtain the first target inspection report data; De-identifying the first target inspection report data using a hash function or a data masking method to obtain second target inspection report data; Perform text standardization on the second target inspection report data, use industry terminology library or ontology to standardize the professional terms and abbreviations in the second target inspection report data, and then convert various measurement units in the second target inspection report data into the international standard unit system, and finally obtain the third target inspection report data; Perform structured data extraction on the third-target inspection report data, use natural language processing methods and regular expressions to extract device ID, model, inspection date, inspection results, and location coordinates from the free text of the third-target inspection report data, interpret the inspection code and standard references and convert them into computable data, judge the safety status of the target device and record it to obtain preprocessed inspection report data.

3. The inspection report verification and auditing method based on a large model according to claim 1 is characterized in that: The S4: verification and audit result output stage also includes the following steps: Error highlighting and marking of the first inspection report data, using color coding or different font styles to highlight detected errors or potential problems, and providing a location index to locate each error or potential problem; Classifying the error types of the first inspection report data by fine-grained classification and assigning a unique error code to each type of error; Provide suggestions and correction plans for detected errors. When a first-type error is detected, the verification model provides automatic correction options through an automatic correction method; when a second-type error is detected, the verification model provides correction suggestions based on expert system rules; when a third-type error is detected, the verification model automatically triggers a manual review request and attaches error analysis details; Generate a verification and audit report, template the verification and audit report, provide data analysis in the verification and audit report, indicate the audit status, obtain the verification and audit results and output them.

4. The inspection report verification and auditing method based on a large model according to any one of claims 1 to 3, characterized in that: The pre-trained model is a BERT, RoBERTa or GPT series model.

5. A large-model-based inspection report verification and auditing system, characterized by: The method for verifying and auditing an inspection report based on a large model according to any one of claims 1 to 4 comprises: The large model module connects the user interaction interface, data preprocessing module, back-end service and database, and verification and audit result output module; the user interaction interface is used for file upload, progress tracking, result display, and interactive correction; the back-end service and database are used to provide API services, database management system, and model services.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are loaded and executed by the processor, the large model-based inspection report verification and auditing method as described in any one of claims 1 to 4 is implemented.

Images