Engineering audit data compliance checking method and system based on Transform model

Abstract

The invention discloses an engineering audit data compliance checking method and system based on a Transform model, and relates to the technical field of engineering audit data checking, and the method comprises the steps: S1, collecting financial data and material consumption data in the whole process of engineering audit; according to the method, structural analysis and deep processing are carried out on project auditing full-process data by relying on the context semantic understanding capability of a Transform model, and accurate semantic matching of auditing data and compliance terms is realized by combining dynamic construction of a compliance rule knowledge base and utilizing a self-attention mechanism, so that the auditing accuracy is improved. The problems that in the prior art, manual checking is low in efficiency and insufficient in accuracy, and a traditional rule matching method is insufficient in semantic understanding and poor in adaptability are solved, intelligence and high efficiency of engineering audit data compliance checking are achieved, the comprehensiveness of checking coverage and the accuracy of anomaly recognition are improved, the scene adaptability of the checking method is enhanced, and the checking efficiency is improved. And reliable technical support is provided for engineering audit.

Description

Technical Field

[0001] This invention relates to the field of engineering audit data verification technology, specifically to a method and system for engineering audit data compliance verification based on the Transformer model. Background Technology

[0002] Engineering auditing is an activity that comprehensively reviews the financial income and expenditure, construction process, and resource usage of an engineering project throughout its entire life cycle. It covers all stages, including project initiation, design, construction, and completion. Its core purpose is to ensure the rationality of the use of project funds and the standardization of the construction process, prevent risks such as waste of funds and illegal operations, provide important support for the quality control and cost control of engineering projects, and play a key role in maintaining the order of the engineering construction market and protecting the legitimate rights and interests of investors.

[0003] Compliance verification of engineering audit data is a core component of engineering audits. It specifically refers to the process of verifying whether the financial data and material usage data involved in the audit comply with relevant regulations, industry standards, project contract stipulations, and other compliance requirements. The accuracy and efficiency of compliance verification directly determine the reliability of engineering audit results. It is a key means to promptly detect data anomalies and identify violations, and it is of great significance to enhancing the authority and effectiveness of engineering audits.

[0004] However, in existing technologies, compliance verification of engineering audit data largely relies on manual verification or traditional rule matching methods. Manual verification is inefficient, struggling to achieve comprehensive coverage of the massive amounts of structured and unstructured data generated throughout the engineering audit process, and is easily influenced by subjective human factors, resulting in insufficient verification accuracy. Traditional rule matching methods rely on fixed rule bases, lacking a deep understanding of contextual semantics and failing to effectively handle the relationships between data. They are also difficult to identify hidden violations in complex scenarios. Furthermore, rule bases have high maintenance costs and poor adaptability, making it difficult to meet the intelligent and efficient requirements of engineering audit data compliance verification. Therefore, developing a compliance verification method and system for engineering audit data based on the Transformer model is of great significance. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method and system for compliance verification of engineering audit data based on the Transformer model. It can perform structured parsing and in-depth processing of data throughout the entire engineering audit process by relying on the contextual semantic understanding capability of the Transformer model. Combined with the dynamic construction of a compliance rule knowledge base, it uses a self-attention mechanism to achieve accurate semantic matching between audit data and compliance clauses, thereby realizing intelligent and efficient compliance verification of engineering audit data, improving the comprehensiveness of verification coverage and the accuracy of anomaly identification, reducing rule maintenance costs, and enhancing the scenario adaptability of the verification method.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a compliance verification method for engineering audit data based on the Transformer model. This method includes the following steps: S1, collecting financial data and material usage data throughout the entire engineering audit process; S2, cleaning and standardizing the collected data, converting unstructured data into structured data to form a unified format audit data set; S3, sorting out compliance requirements, breaking them down into structured compliance clauses, and constructing a compliance rule knowledge base; S4, inputting the audit data set into a pre-trained Transformer model, using the model's self-attention mechanism to mine the contextual semantic information and inter-data relationships, and simultaneously performing semantic depth matching between the audit data and the compliance clauses in the compliance rule knowledge base; S5, automatically identifying abnormal data points that do not comply with compliance clauses, classifying and labeling these abnormal data points, generating a compliance verification report, and feeding back new compliance scenarios or rule defects discovered during the verification process to the compliance rule knowledge base, thus completing the dynamic updating of the knowledge base.

[0007] Furthermore, step S2, which involves cleaning and standardizing the collected data, includes the following steps: Redundant data is removed from the collected data by comparing the repetition and invalid identifiers of data fields to filter and delete duplicate records and invalid data without actual audit significance; missing data is processed using a data completion strategy, selecting interpolation or mean-based filling based on the data type and logical relationship of related fields to complete missing information; data from different data sources is formatted uniformly, standardizing field names, data units, and data representation to ensure consistency across different data formats; key audit information is extracted from unstructured data using optical character recognition technology and converted into structured data, ultimately forming a unified format audit data set.

[0008] Furthermore, step S3, which involves sorting out compliance requirements and breaking them down into structured compliance clauses to build a compliance rule knowledge base, includes the following steps: comprehensively collecting industry standards, project contractual agreements, and other relevant compliance requirements; screening effective compliance information using a combination of manual sorting and keyword extraction; breaking down the screened compliance requirements according to applicable scenarios, violation situations, and constraints, and converting them into structured compliance clauses, with each compliance clause clearly specifying the applicable audit process, data type, and specific constraint standards; and calculating the importance weight of each compliance clause using the following formula: ,in, For the first The importance and weight of each compliance clause For the first The coverage of applicable scenarios for a compliance clause, i.e., the ratio of the number of engineering audit stages that the clause can cover to the total number of audit stages. For the first The historical violation correlation of a compliance clause, i.e., the ratio of the number of violation data items corresponding to that clause in historical audits to the total number of violation data items. and The weighting coefficients were obtained by collecting data from audit experts on the importance of over 1,000 historical compliance clauses and optimizing them using the gradient descent method. A classification index was established for the structured compliance clauses according to their importance weights and dimensions of financial compliance and material usage compliance, forming an initial compliance rule knowledge base. A version management mechanism for compliance clauses was set up to record the creation time, scope of application, and update records of each compliance clause, providing support for subsequent maintenance and traceability.

[0009] Furthermore, step S4, through the model's self-attention mechanism, mines the contextual semantic information and inter-data relationships of the data. Simultaneously, when performing semantic depth matching with compliance clauses in the compliance rule knowledge base, it includes the following steps: converting the preprocessed audit data set into a vector form recognizable by the Transformer model; encoding the data through the model's embedding layer; mapping it to a high-dimensional semantic space; and obtaining the audit data vector. , The audit data sequence number is used; the model's self-attention mechanism is activated to calculate the attention weights between data fields. , For data field serial numbers, we mine the contextual semantic relationships between different data and the correspondence between data and audit stages; we convert each compliance clause in the compliance rule knowledge base into vector form. , For each compliance clause number, calculate the semantic matching degree between the audit data vector and the compliance clause vector. The matching degree calculation formula is as follows: ,in, For the first The audit data and the first The semantic matching degree of the compliance clauses The dot product of two vectors. The attention weighting of audit data and corresponding fields in compliance clauses is coordinated. The total number of data fields. For the attention weight of compliance clauses, and They are respectively and The feature dimension value, and It is automatically learned by the Transformer model during pre-training and fine-tuning in the field of engineering auditing. and The fixed feature dimension value is preset during the vector encoding process; a similarity threshold is set as a compliance judgment standard, and data with similarity below the threshold is marked. The data is further verified for violations by combining the constraints of compliance clauses and importance weights, and finally abnormal data points are identified and the corresponding compliance clause association information is recorded.

[0010] Furthermore, in step S1, when collecting financial data and material usage data throughout the entire engineering audit process, the data includes structured and unstructured data. The structured data comes from database tables and Excel spreadsheets in the engineering project management system, financial accounting system, and material management system. The unstructured data comes from scanned documents, Word documents, PDF files, and images of handwritten documents. During the data collection process, a data verification mechanism is established to verify the integrity of the collected data. Data with missing key fields is marked and fed back to the data source system for supplementary collection. The collected data covers the information required for key audit nodes throughout the entire engineering audit process.

[0011] Furthermore, the compliance rule knowledge base supports dynamic updates. During the compliance review process, when new compliance requirements are discovered or existing compliance clauses are found to be missing or incomplete, the new compliance requirements are entered through a preset rule editing interface. After being processed according to the structured decomposition standard of compliance clauses, they are added to the knowledge base. At the same time, existing compliance clauses with defects are modified or deleted. During the update process, the compliance and accuracy of the newly added or modified compliance clauses are verified through an audit mechanism. After verification, they are included in the compliance rule knowledge base and applied to subsequent compliance review processes.

[0012] Furthermore, the pre-trained Transformer model supports fine-tuning and optimization based on professional data from the engineering auditing field. Historical data labeled with compliance in the engineering auditing field is selected as the fine-tuning dataset. This dataset is divided into training and validation sets according to a preset ratio, and a loss function for model fine-tuning is constructed. The loss function calculation formula is as follows: ,in, This is the total loss value. The loss for classifying outlier data is calculated using the cross-entropy loss function. The semantic matching loss is calculated using the mean squared error loss function. and The loss weight coefficients are obtained by iteratively optimizing the balance between semantic matching accuracy and anomaly detection recall on the validation set. The optimization process aims to maximize the harmonic mean of these two values ​​by adjusting the model's learning rate and iteration count hyperparameters, ultimately resulting in a total loss value. The model is trained with the goal of minimization, and semantic matching accuracy is used as the evaluation metric during training. The model parameters are continuously optimized to adapt the model to the professional terminology and data association characteristics of the engineering audit field.

[0013] Furthermore, the compliance verification report generated in step S5 includes basic information about abnormal data points, the basis for violations, classification statistics, and related data indexes. The basic information includes the data source, data field name, specific data value, and the audit stage in which the violation occurred. The basis for violations clearly specifies the corresponding compliance clause number and specific content. The classification statistics summarize the data according to the type of violation and the audit stage in which the violation occurred. The related data index provides query paths for other audit data related to the abnormal data. The report supports export in multiple formats and displays the distribution of abnormal data and violation trends through a visual interface.

[0014] This engineering audit data compliance verification system based on the Transformer model is applicable to the aforementioned engineering audit data compliance verification method based on the Transformer model. The system includes: a data acquisition module, a data preprocessing module, a compliance rule knowledge base construction module, a semantic matching and anomaly identification module, and a result output and feedback module. The data acquisition module collects financial data and material usage data throughout the entire engineering audit process. The data preprocessing module, connected to the data acquisition module, cleans and standardizes the collected data, converting unstructured data into structured data to form a unified format audit data set. The compliance rule knowledge base construction module organizes compliance requirements and breaks them down into structured compliance clauses to build a compliance rule knowledge base. The semantic matching and anomaly identification module, connected to both the data preprocessing module and the compliance rule knowledge base construction module, incorporates a pre-trained Transformer model to mine the contextual semantic information and inter-data relationships of the audit data, performs deep semantic matching with compliance clauses, and identifies anomalous data points. The result output and feedback module, connected to the semantic matching and anomaly identification module, classifies and labels anomalous data points, generates a compliance verification report, and transmits relevant feedback information to the compliance rule knowledge base construction module.

[0015] Furthermore, the data acquisition module establishes data connections with the project management system, financial accounting system, and material management system through API interfaces, supporting both batch data acquisition and real-time data synchronization modes. In batch acquisition mode, audit data within a specified time period is extracted from each data source at preset time intervals. In real-time synchronization mode, updated data is captured and collected immediately when the data source changes. The data preprocessing module has built-in algorithms for redundant data removal, missing data completion, format standardization, and unstructured data conversion. Each algorithm can be called according to the data type and processing requirements. The semantic matching and anomaly identification module provides model parameter adjustment interfaces and fine-tuning training interfaces. The result output and feedback module sets up a feedback information collection interface. Auditors submit new compliance scenario information and rule defect feedback through this interface. After the feedback information is reviewed and approved, it is synchronized to the compliance rule knowledge base construction module.

[0016] Compared with existing technologies, this engineering audit data compliance verification method and system based on the Transformer model has the following beneficial effects: By relying on the contextual semantic understanding capability of the Transformer model, this invention performs structured analysis and in-depth processing of the entire engineering audit process data. Combined with the dynamic construction of a compliance rule knowledge base, it utilizes a self-attention mechanism to achieve accurate semantic matching between audit data and compliance clauses. This solves the problems of low efficiency and insufficient accuracy of manual verification in existing technologies, as well as the lack of semantic understanding and poor adaptability of traditional rule matching methods. It realizes intelligent and efficient compliance verification of engineering audit data, improves the comprehensiveness of verification coverage and the accuracy of anomaly identification, reduces rule maintenance costs, enhances the scenario adaptability of the verification method, provides reliable technical support for engineering audits, and further ensures the authority and effectiveness of engineering audit results.

[0017] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0019] Figure 1 A flowchart illustrating a compliance verification method for engineering audit data based on the Transformer model;

[0020] Figure 2 A flowchart illustrating a compliance verification method for engineering audit data based on the Transformer model;

[0021] Figure 3 This is a schematic diagram of the structure of an engineering audit data compliance verification system based on the Transformer model. Detailed Implementation

[0022] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0023] This invention addresses the problems of low efficiency and insufficient accuracy of manual verification in existing engineering audit data compliance checks, as well as the lack of semantic understanding and poor adaptability of traditional rule matching methods. It provides an engineering audit data compliance verification method and system based on the Transformer model, thereby achieving high efficiency and accuracy in audit data compliance verification.

[0024] See Figure 1 and Figure 2 The core process of the verification method is as follows: First, financial data and material usage data from the entire project audit process are collected, covering structured data (database tables and Excel spreadsheets from project management, financial accounting, and other systems) and unstructured data (scanned documents, files, images of handwritten documents, etc.). During collection, a data verification mechanism ensures data integrity; data lacking key fields is marked and collected. Second, the collected data undergoes cleaning and standardization, eliminating redundant and invalid data, using adaptation methods to complete missing data, unifying the format specifications of different data sources, and converting unstructured data into structured data using optical character recognition technology, forming a unified format audit data set.

[0025] Subsequently, compliance requirements from industry standards and project contracts were reviewed. Through manual review and keyword extraction, valid information was filtered and broken down into structured compliance clauses with clearly defined applicable scenarios and binding standards. These clauses were then categorized and indexed according to importance and compliance dimensions, and a version management mechanism was implemented to build an initial compliance rule knowledge base. This knowledge base supports dynamic updates, and any added or modified clauses must be reviewed and verified before being incorporated. Next, the audit data set was input into a pre-trained Transformer model. The model uses a self-attention mechanism to mine the semantic context and relationships within the data, while simultaneously performing deep semantic matching with the clauses in the compliance rule knowledge base to identify anomalous data points. This model can be fine-tuned and optimized based on professional data from the engineering audit field to adapt to domain terminology and data characteristics. Finally, anomalous data was categorized and labeled, generating a compliance verification report containing basic data information, violation justification, classification statistics, and related data indexes. This report supports export in multiple formats and visualization, and new compliance scenarios or rule deficiencies are fed back to the knowledge base for dynamic updates.

[0026] See Figure 3 The corresponding verification system includes five modules: a data acquisition module that connects to various data sources via API interfaces, supporting batch acquisition and real-time synchronization; a data preprocessing module with built-in algorithms to achieve standardized data cleaning; a compliance rule knowledge base construction module responsible for sorting out and decomposing compliance clauses and building and maintaining the knowledge base; a semantic matching and anomaly identification module with built-in Transformer model to complete semantic matching and anomaly data identification; and a results output and feedback module that generates reports and achieves feedback loop.

[0027] This solution leverages the semantic understanding capabilities of the Transformer model, combined with a dynamically updated compliance rule knowledge base, to significantly improve the comprehensiveness of audit coverage and the accuracy of anomaly identification, while reducing rule maintenance costs and providing reliable technical support for engineering audits.

[0028] Example 1: This example is applied to the compliance verification of engineering audit data for a large-scale municipal road construction project. The project covers multiple stages, including project approval, design bidding, construction, and final acceptance, involving massive amounts of financial and material usage data. Due to the large number of project participants and the complexity of data flow, traditional manual verification methods are insufficient to achieve full-process data coverage and are prone to omissions of non-compliant data due to human error. Traditional rule-matching methods cannot accurately identify compliance risks arising from implicit relationships between data. Therefore, the engineering audit data compliance verification method and system based on the Transformer model of this invention are adopted to achieve efficient and accurate compliance verification of the project's audit data.

[0029] See Figure 1 , Figure 2 and Figure 3 The specific implementation process of this embodiment is as follows: Data collection for the entire engineering audit process is carried out. Stable data connections are established with the municipal road project's project management system, financial accounting system, and material management system via the API interface of the data collection module. Simultaneously, both batch data collection and real-time data synchronization modes are enabled. In batch collection mode, payment records, material procurement details, and engineering quantity calculation data for the most recent audit period are extracted at fixed daily intervals. In real-time synchronization mode, updated data is captured and collected immediately when changes occur in the data source system, such as design changes leading to cost adjustments or updates to the quantity of materials arriving on site.

[0030] The collected data encompasses both financial and material usage data. Financial data includes detailed expenditures such as labor costs, machinery usage fees, and management fees. Material usage data includes the arrival, consumption, and inventory data of major building materials such as steel bars, cement, and asphalt. A data integrity verification mechanism was established during the data collection process. Data lacking key fields such as payment amounts and material specifications was marked and fed back to the corresponding data source system. The relevant personnel then supplemented and improved the data before re-collection, ensuring that the collected data comprehensively covers the information required for key audit nodes throughout the entire project audit process.

[0031] Data cleaning and standardization were performed. First, redundant data was removed by comparing the duplication and invalid identifiers of each data field, and filtering and deleting redundant and invalid data such as duplicated project payment records and test data without actual audit significance. For missing data such as arrival time and supplier information in some material arrival records, interpolation was used to fill in missing time-related data based on data type and logical relationship of related fields, and the average value of historical material procurement data of similar municipal road projects was used to fill in discrete missing data such as supplier information.

[0032] Subsequently, the data from different data sources were formatted uniformly, field names were standardized, and different expressions such as "project payment amount" and "total payment amount" were unified into "project payment amount". The data units were standardized, and the "tons" and "kilograms" of some material usage were converted to "tons". The data representation format was standardized, and the date format was unified into "year-month-day". Finally, for unstructured data such as scanned copies of construction drawings and images of handwritten visa documents, optical character recognition technology was used to extract key audit information, such as the engineering quantity markings in the construction drawings and the visa reasons and amounts in the visa documents, and these were converted into structured data, ultimately forming a unified set of audit data.

[0033] Build a compliance rules knowledge base. Comprehensively collect relevant compliance requirements from the municipal road construction industry, including engineering audit standards, project bidding documents, material procurement contracts, and project payment agreements. Using a combination of manual review and keyword extraction, select effective compliance information such as "project payments must comply with the payment milestones stipulated in the contract" and "the amount of major building materials used must not exceed the loss rate range approved in the design drawings." Break down the selected compliance requirements according to applicable scenarios, violation situations, and constraints, transforming them into structured compliance clauses. Clearly define the audit stage, data type, and specific constraint standards for each clause. For example, if a clause specifically applies to the project payment audit stage, the constraint standard for the payment amount field in the financial data is "the payment amount must not exceed 10% of the amount approved for the corresponding project progress milestone."

[0034] In the specific implementation of this embodiment, the importance weight of each compliance clause is calculated using the following formula: ,in For the first The importance and weight of each compliance clause For the first The coverage of applicable scenarios for a compliance clause, i.e., the ratio of the number of audit stages in municipal road projects that the clause can cover to the total number of audit stages. For the first The historical non-compliance correlation of a compliance clause is the ratio of the number of non-compliance data items corresponding to that clause to the total number of non-compliance data items in the historical audit and verification of municipal road projects. and The weighting coefficients were obtained by collecting importance-based data from auditing experts on over 1000 historical compliance clauses for municipal road projects and optimizing them using the gradient descent method. Subsequently, a categorized index was created for the structured compliance clauses based on their importance weights and dimensions of financial compliance and material usage compliance, forming an initial compliance rule knowledge base. A version management mechanism for the compliance clauses was also established to record the creation time, scope of application, and update history of each clause, providing support for subsequent maintenance and traceability.

[0035] Semantic matching and anomaly identification are performed. First, the preprocessed audit data set is converted into a vector form recognizable by the Transformer model. The data is then encoded through the model's embedding layer, mapping it to a high-dimensional semantic space to obtain the audit data vector. , This is the audit data sequence number. The model's self-attention mechanism is activated, calculating the attention weights between data fields. , For data field serial numbers, we can mine the contextual semantic relationships between different data and the correspondence between data and auditing stages, such as identifying the semantic relationship between project payment amount and project progress nodes and contract payment agreements.

[0036] In the specific implementation of this embodiment, each compliance clause in the compliance rule knowledge base is converted into vector form. , Using the compliance clause number as the reference, the semantic matching degree between the audit data vector and the compliance clause vector is calculated using the following formula: ,in For the first The audit data and the first The semantic matching degree of the compliance clauses The dot product of two vectors. The attention weighting of audit data and corresponding fields in compliance clauses is coordinated. The total number of data fields. and They are respectively and The feature dimension value, and It was automatically learned by the Transformer model during pre-training and fine-tuning in the field of municipal road engineering auditing. and These are fixed feature dimension values ​​preset during the vector encoding process. A similarity threshold is set as a compliance judgment standard. Data with similarity below the threshold is marked, and further verification is performed in conjunction with the constraints and importance weights of compliance clauses. For example, if the matching degree of a certain project payment data with the "payment node compliance" clause is lower than the threshold, and the clause has a high importance weight, after verification, it is confirmed that the payment has not reached the payment node stipulated in the contract, and it is finally determined as an abnormal data point and the corresponding compliance clause association information is recorded.

[0037] Meanwhile, the pre-trained Transformer model is fine-tuned and optimized for specific domains. Historical data with compliance already marked in the field of municipal road engineering audits are selected as the fine-tuning dataset and divided into training and validation sets according to a preset ratio.

[0038] In the specific implementation of this embodiment, the loss function for model fine-tuning is constructed, and the calculation formula is as follows: ,in This is the total loss value. The loss for classifying outlier data is calculated using the cross-entropy loss function. The semantic matching loss is calculated using the mean squared error loss function. and The loss weight coefficients are obtained by iteratively optimizing the balance between semantic matching accuracy and anomaly detection recall on the validation set. The optimization process aims to maximize the harmonic mean of these two values ​​by adjusting the model's learning rate and iteration count hyperparameters, ultimately resulting in a total loss value. The model is trained with the goal of minimization, and semantic matching accuracy is used as the evaluation metric during training. The model parameters are continuously optimized to adapt the model to the professional terminology and data association characteristics in the field of municipal road engineering auditing.

[0039] Generate compliance verification reports and dynamically update the knowledge base. Identified anomalous data points are categorized by violation type, such as payment node violations, material usage overruns, and non-compliant expense listings, and a compliance verification report is generated. The report includes basic information on the anomalous data points, clearly specifying the data source, data field name, specific data value, and the audit stage in which each anomalous data point occurred. The violation basis section indicates the corresponding compliance clause number and specific content. The classification statistics section summarizes data by violation type and audit stage, such as the total number of material usage overrun violations and the amount involved during the construction stage. The associated data index provides query paths for other audit data related to the anomalous data, such as project progress ledgers and material procurement contracts. The report supports export to multiple formats such as PDF and Excel, and a visual interface displays the distribution of anomalous data and violation trends at each audit stage.

[0040] Auditors submit new compliance scenarios discovered during the verification process, such as "compliance requirements for the procurement of new environmentally friendly materials," and imperfections in existing compliance clauses regarding "accounting for design change costs," through the feedback information collection interface of the results output and feedback module. After the feedback information is reviewed and approved by audit experts, it is synchronized to the compliance rule knowledge base construction module. The new compliance requirements are processed according to the structured decomposition standard and added to the knowledge base. Existing clauses with defects are modified to complete the dynamic update of the knowledge base.

[0041] In summary, this embodiment effectively solves the problems of low efficiency, incomplete coverage, and insufficient identification accuracy of traditional verification methods by applying the engineering audit data compliance verification method and system based on the Transformer model to municipal road construction projects. The dual-mode acquisition mechanism of the data acquisition module enables complete acquisition of audit data throughout the entire process; standardized data cleaning transforms disorganized structured and unstructured data into a unified format audit data set, laying a solid foundation for subsequent verification; the structured construction and dynamic updating mechanism of the compliance rule knowledge base ensures the comprehensiveness and timeliness of compliance clauses; the self-attention mechanism and domain fine-tuning optimization of the Transformer model significantly improve the data semantic association mining capability and compliance matching accuracy, successfully identifying hidden violations that are difficult to detect using traditional methods; and the detailed and standardized compliance verification report provides auditors with clear decision-making basis.

[0042] Example 2: This example is applied to the compliance verification of engineering audit data for a super-large commercial complex construction project. The project encompasses multiple business formats, including commercial podiums, office buildings, and serviced apartments, and involves multiple stages such as preliminary planning, design, construction, leasing, and preparation. The data types are more complex and diverse. Due to the complexity of the project's business formats and the multiple dimensions of compliance requirements, such as building safety, commercial operation, environmental protection, and energy conservation, traditional verification methods are insufficient to handle the cross-matching needs of multi-dimensional compliance rules. Furthermore, the large volume of cross-departmental data transfers is prone to compliance loopholes. Therefore, the verification method and system of this invention are adopted, optimizing the data collection and processing flow based on the aforementioned examples to improve the accuracy of compliance verification in multi-business scenarios.

[0043] See Figure 1 , Figure 2 and Figure 3The specific implementation process of this embodiment is as follows: Based on the data acquisition mechanism of the aforementioned embodiment, the data sources and acquisition methods are further expanded. In addition to establishing connections with various systems through API interfaces, the data acquisition module adds access to IoT devices at the construction site to collect real-time dynamic data such as the number of large electromechanical equipment on site, usage time, and energy consumption data; simultaneously, it connects to the compliance document database of a third-party auditing institution to obtain relevant data on specific compliance requirements in the commercial building sector. The batch acquisition mode extracts cross-stage audit data at fixed weekly intervals, while the real-time synchronization mode focuses on capturing key data changes such as adjustments to investment attraction policies and changes in rent calculation standards.

[0044] The collected data covers specific expenditures in financial data, such as investment promotion fees, advance rent payments, and equipment purchase payments. Material usage data includes procurement, consumption, and inventory data for various types of materials, including decoration materials, electromechanical equipment, and fire-fighting equipment. During the data collection process, a data integrity verification mechanism was used, focusing on verifying key information such as rent payment cycles in lease contracts and the applicable conditions for preferential policies in investment promotion documents. This ensures that the data covers the specific audit requirements of the entire commercial complex audit process.

[0045] Building upon the aforementioned standardized data cleaning process, data preprocessing enhances the in-depth processing of unstructured data. In the redundant data removal stage, in addition to filtering duplicate records, the focus is on deleting invalid data without practical audit significance, such as test data from investment promotion activities. For missing data, addressing the issue of missing rent calculation parameters specific to commercial formats, a mean-based filling method based on historical rent data from similar commercial complexes in the same area is used to fill in the missing parameters. For continuously missing data such as equipment technical parameters, interpolation is used. During format standardization, commercially specific field names are standardized, and expressions such as investment promotion discount rates and rent unit prices from different systems are unified and standardized, along with data representation formats such as currency units. For complex unstructured data such as lease contract design schemes, in addition to extracting key information using optical character recognition technology, natural language processing technology is introduced to analyze semantic relationships in the text, extracting deep compliance information such as payment constraints and breach of contract liabilities from lease contracts, converting it into structured data, and ultimately forming a unified format audit data set.

[0046] The compliance rules knowledge base is built upon the aforementioned construction process, expanding the scope of compliance requirements and optimizing the structured decomposition logic. It comprehensively collects compliance requirements from industry standards for commercial buildings, local commercial operation regulations, environmental protection and energy conservation requirements, and project-specific leasing and cooperation agreements. Valid information is screened using a combination of manual review and keyword extraction, with a focus on specific compliance information such as rent collection compliance and environmentally friendly material usage standards in commercial operations. The screened compliance requirements are then broken down into structured compliance clauses based on the specific business scenarios and non-compliance constraints of commercial buildings, clearly defining the applicable audit stages for each clause, including leasing, construction, and operational audits.

[0047] In the calculation of the weighting of compliance clauses, a formula is used. We completed weighting assignments and established a categorized index based on multiple dimensions, including importance, financial compliance, material usage compliance, business operations compliance, and environmental compliance. We improved the version management mechanism, focusing on recording the update trajectory of compliance clauses specific to commercial formats, thus forming a compliance rule knowledge base adapted to commercial complex projects. Simultaneously, leveraging a pre-defined rule editing interface, we established a rapid update channel for compliance requirement changes caused by frequent adjustments to investment attraction policies in commercial projects. Newly added compliance clauses are incorporated into the knowledge base after joint review by audit experts and business operations experts.

[0048] Semantic matching and anomaly detection, based on the aforementioned process, optimize model fine-tuning and matching logic. The preprocessed audit data set is converted into vector form, encoded through the model embedding layer to obtain audit data vectors, and a self-attention mechanism is activated to mine semantic relationships between data, focusing on identifying the correlation between rent payment data, investment promotion policy material procurement data, and environmental protection standards.

[0049] In the semantic matching degree calculation stage, the formula is used. Complete in-depth matching of audit data and compliance clauses, set differentiated similarity thresholds, and set lower thresholds for key clauses such as environmental compliance to improve identification sensitivity.

[0050] During model fine-tuning, labeled historical data from the field of commercial complex auditing were selected as the fine-tuning dataset. The training and validation sets were divided according to a preset ratio, and a formula was used. A loss function was constructed, and the optimization process focused on improving the model's ability to identify hidden violations related to commercial operations. By adjusting the learning rate and the number of iterations, the model was trained with the goal of minimizing the total loss value, making the model adaptable to the compliance rule expressions and data correlation characteristics of multiple business formats in commercial complex projects. The labeled data was validated using compliance clause constraints and importance weights, with a focus on identifying abnormal data points such as excessive leasing expenses, substandard materials and environmental protection standards, and non-compliant rent collection, and recording the associated compliance clause information.

[0051] Building upon the aforementioned embodiments, the report generation and knowledge base updates optimize the content and feedback mechanism of compliance audit reports. The generated reports now include new statistics on compliance categories for commercial operations, categorizing violations by stage (land acquisition, construction, and operation / maintenance) and clarifying the distribution of violations across different business sectors. The associated data index adds query paths specific to commercial projects, such as lease contract numbers and equipment purchase contract numbers, and supports export to multiple formats including PDF, Excel, and HTML. The visualization interface now features comparison charts of violation trends across different compliance dimensions. Auditors submit new compliance scenarios discovered in commercial complex projects, such as compliance requirements for commercial advertising and rule deficiencies in existing clauses regarding cross-sector material allocation, through the feedback information collection interface. After joint review and approval, the feedback information is synchronized to the compliance rule knowledge base construction module, enabling dynamic updates to the knowledge base and ensuring that subsequent audits can adapt to the complex compliance needs of commercial projects.

[0052] In summary, this embodiment, based on the aforementioned embodiments, expands the data sources and processing methods to address the multi-format and multi-dimensional compliance verification needs of large commercial complexes. It optimizes the adaptability of the model and knowledge base, effectively solving problems such as the difficulty in identifying implicit violations through cross-matching of compliance rules across business formats in commercial projects. By adding IoT device collection and connecting to third-party data sources, comprehensive data collection is achieved; natural language processing technology improves the processing accuracy of complex unstructured data; multi-dimensional compliance rule indexing and a fast update channel ensure the timeliness and adaptability of the knowledge base; and targeted fine-tuning of the model further improves the accuracy of identifying abnormal data in commercial scenarios.

[0053] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for compliance verification of engineering audit data based on the Transformer model, characterized in that, The method includes the following steps: S1. Collect financial data and material usage data throughout the entire engineering audit process; S2. Clean and standardize the collected data to convert unstructured data into structured data, forming a unified format audit data set; S3. Organize compliance requirements, break them down into structured compliance clauses, and build a compliance rules knowledge base; S4. Input the audit data set into the pre-trained Transformer model, and use the model's self-attention mechanism to mine the contextual semantic information and relationships between data. At the same time, perform semantic deep matching between the audit data and the compliance clauses in the compliance rule knowledge base. S5. Automatically identify abnormal data points that do not comply with compliance clauses, classify and label abnormal data points, generate compliance verification reports, and feed back new compliance scenarios or rule defects found during the verification process to the compliance rule knowledge base to complete the dynamic update of the knowledge base.

2. The method for compliance verification of engineering audit data based on the Transformer model according to claim 1, characterized in that, Step S2, which involves cleaning and standardizing the collected data, includes the following steps: Redundant data is removed from the collected data by comparing the duplication and invalid identifiers of data fields to filter and delete duplicate records and invalid data that have no actual audit significance. Data completion strategies are used to process missing data. Based on the data type and the logical relationship between related fields, interpolation or the mean-based filling method based on historical data of similar projects is selected to complete the missing information. Standardize the format of data from different data sources, and standardize field names, data units and data representations to ensure that the formats of various data types are consistent. Key audit information is extracted from unstructured data using optical character recognition technology, converted into structured data, and finally formed a unified set of audit data.

3. The method for compliance verification of engineering audit data based on the Transformer model according to claim 1, characterized in that, Step S3, which involves sorting out compliance requirements and breaking them down into structured compliance clauses to build a compliance rule knowledge base, includes the following steps: We comprehensively collect industry standards, project contractual agreements, and other relevant compliance requirements, and use a combination of manual sorting and keyword extraction to screen effective compliance information. The selected compliance requirements are broken down according to applicable scenarios, violation situations, and constraints, and transformed into structured compliance clauses. Each compliance clause clearly defines the applicable audit process, data type, and specific constraint standards. The importance weight of each compliance clause is calculated using the following formula: ,in, For the first The importance and weight of each compliance clause For the first The scope of application scenarios covered by the compliance clauses For the first The correlation of historical violations of the compliance clauses and These are the weighting coefficients; A classification index was established for the structured compliance clauses according to their importance weight and the dimensions of financial compliance and material usage compliance, forming an initial compliance rule knowledge base; Establish a version management mechanism for compliance clauses, recording the creation time, scope of application, and update history of each compliance clause to support subsequent maintenance and traceability.

4. The method for compliance verification of engineering audit data based on the Transformer model according to claim 1, characterized in that, Step S4, which mines the contextual semantic information and inter-data relationships of the data through the model's self-attention mechanism, and performs semantic deep matching with compliance clauses in the compliance rule knowledge base, includes the following steps: The preprocessed audit data set is converted into a vector form recognizable by the Transformer model. The data is then encoded through the model's embedding layer and mapped to a high-dimensional semantic space to obtain audit data vectors. ; The model's self-attention mechanism is activated, and attention weights between data fields are calculated. To uncover the contextual semantic relationships between different data and the correspondence between data and auditing processes; Convert the compliance clauses in the compliance rules knowledge base into vector form. Calculate the semantic matching degree between the audit data vector and the compliance clause vector. The matching degree calculation formula is as follows: ,in, For the first The audit data and the first The semantic matching degree of the compliance clauses The dot product of two vectors. The attention weighting of audit data and corresponding fields in compliance clauses is coordinated. For the attention weight of compliance clauses, The total number of data fields. and They are respectively and Feature dimension values; A similarity threshold is set as a compliance judgment standard. Data with similarity below the threshold is marked. The data is further verified for violations by combining the constraints of compliance clauses and their importance weights. Finally, abnormal data points are identified and the corresponding compliance clause information is recorded.

5. The method for compliance verification of engineering audit data based on the Transformer model according to claim 1, characterized in that, When collecting financial and material usage data throughout the entire engineering audit process in step S1, the data includes structured and unstructured data. The structured data comes from database tables and Excel spreadsheets in the engineering project management system, financial accounting system, and material management system. The unstructured data comes from scanned documents, Word documents, PDF files, and images of handwritten documents. During the data collection process, a data verification mechanism is established to verify the integrity of the collected data. Data with missing key fields is marked and fed back to the data source system for supplementary collection. The collected data covers the information required for key audit nodes throughout the entire engineering audit process.

6. The method for compliance verification of engineering audit data based on the Transformer model according to claim 1, characterized in that, The compliance rules knowledge base supports dynamic updates. During the compliance review process, when new compliance requirements are found or existing compliance clauses are missing or incomplete, the new compliance requirements are entered through a preset rule editing interface. After being processed according to the structured decomposition standard of compliance clauses, they are added to the knowledge base. At the same time, existing compliance clauses with defects are modified or deleted. During the update process, the compliance and accuracy of the newly added or modified compliance clauses are verified through an audit mechanism. After verification, they are included in the compliance rules knowledge base and applied to subsequent compliance review processes.

7. The method for compliance verification of engineering audit data based on the Transformer model according to claim 1, characterized in that, The pre-trained Transformer model supports fine-tuning and optimization based on professional data from the engineering auditing field. Historical data with compliance annotations in the engineering auditing field is selected as the fine-tuning dataset. This dataset is divided into training and validation sets according to a preset ratio. A loss function for model fine-tuning is constructed, and the formula for calculating the loss function is as follows: ,in, This is the total loss value. For outlier data classification loss, For semantic matching loss, and The loss weight coefficients are determined by adjusting the model's learning rate and iteration count hyperparameters to obtain the total loss value. The model is trained with the goal of minimization, and semantic matching accuracy is used as the evaluation metric during training. The model parameters are continuously optimized to adapt the model to the professional terminology and data association characteristics of the engineering audit field.

8. The method for compliance verification of engineering audit data based on the Transformer model according to claim 1, characterized in that, The compliance verification report generated in step S5 includes basic information about abnormal data points, the basis for violations, classification statistics, and related data indexes. The basic information includes the data source, data field name, specific data value, and the audit stage in which the violation occurred. The basis for violations clearly states the corresponding compliance clause number and specific content. The classification statistics summarize the data according to the type of violation and the audit stage in which the violation occurred. The related data index provides query paths for other audit data related to the abnormal data. The report supports export in multiple formats and displays the distribution of abnormal data and violation trends through a visual interface.

9. A compliance verification system for engineering audit data based on the Transformer model, applicable to the compliance verification method for engineering audit data based on the Transformer model as described in any one of claims 1-8, characterized in that, The system includes: a data acquisition module, a data preprocessing module, a compliance rule knowledge base construction module, a semantic matching and anomaly identification module, and a result output and feedback module; The data acquisition module is used to collect financial data and material usage data throughout the entire engineering audit process; The data preprocessing module is connected to the data acquisition module and is used to clean and standardize the acquired data, converting unstructured data into structured data to form a unified format audit data set. The compliance rules knowledge base construction module is used to sort out compliance requirements and break them down into structured compliance clauses to build a compliance rules knowledge base; The semantic matching and anomaly identification module is connected to the data preprocessing module and the compliance rule knowledge base construction module, respectively. It has a built-in pre-trained Transformer model to mine the contextual semantic information and inter-data relationships of audit data, perform semantic deep matching with compliance clauses, and identify abnormal data points. The result output and feedback module is connected to the semantic matching and anomaly identification module, which is used to classify and label abnormal data points, generate compliance verification reports, and transmit relevant feedback information to the compliance rule knowledge base construction module.

10. The engineering audit data compliance verification system based on the Transformer model according to claim 9, characterized in that, The data acquisition module establishes data connections with the project management system, financial accounting system, and material management system through API interfaces, supporting both batch data acquisition and real-time data synchronization modes. In batch acquisition mode, audit data within a specified time period is extracted from each data source at preset time intervals. In real-time synchronization mode, updated data is captured and collected immediately when the data source changes. The data preprocessing module has built-in algorithms for redundant data removal, missing data completion, format standardization, and unstructured data conversion. Each algorithm can be called according to the data type and processing requirements. The semantic matching and anomaly identification module provides interfaces for adjusting model parameters and fine-tuning training. The result output and feedback module sets up a feedback information collection interface. Auditors can submit new compliance scenario information and rule defect feedback through this interface. After the feedback information is reviewed and approved, it is synchronized to the compliance rule knowledge base construction module.

Images