Artificial intelligence-based engineering cost intelligent auditing method and system

Abstract

The application discloses an engineering cost intelligent auditing method and system based on artificial intelligence, and relates to the technical field of engineering cost management. The method structures cost files by OCR, document analysis and lightweight BIM technology, and constructs a standardized data pool; adopts a hybrid architecture AI model to perform compliance checking, reasonableness research and correlation cross-examination in parallel, and establishes a drawing-list-attestation semantic-space correlation by combining NLP and CV technology; generates a traceable structured auditing report, and realizes closed-loop self-adaptive iteration through incremental learning. The system includes data standardization, a knowledge base, an AI auditing engine, an interactive report and a self-learning optimization module, can significantly improve auditing efficiency and accuracy, unify auditing standards, and has dynamic self-evolution capability.

Description

Technical Field

[0001] This invention relates to the field of engineering cost management technology, specifically to an intelligent engineering cost auditing method and system that integrates multimodal artificial intelligence, lightweight BIM, and incremental learning, applicable to scenarios such as construction project settlement auditing, cost control, and cost compliance review. Background Technology

[0002] Construction cost auditing is a core component of investment control and cost management in construction projects, aiming to ensure the compliance, rationality, and accuracy of documents such as bills of quantities, unit prices, fee standards, and design changes. Traditional auditing methods heavily rely on the personal experience of cost engineers, manually comparing massive amounts of heterogeneous documents (such as PDF drawings, Excel lists, and change orders), which inherently has the following drawbacks: Low review efficiency: Large projects involve thousands of list items and supporting drawings, and the manual review cycle can take several weeks or even months, which is difficult to meet the "fast-paced and high-intensity" settlement and auditing needs of modern engineering.

[0003] Insufficient accuracy and consistency: Audit results are easily affected by staff fatigue and differences in professional level. The verification of complex logical connections across documents (such as drawings and lists) is prone to omissions and misjudgments. Differences in the understanding of quota rules among different auditors lead to inconsistent audit standards.

[0004] Lack of dynamic adaptability: Dynamic changes such as market price fluctuations, regional quota updates, and the emergence of new technologies require continuous iteration of audit rules. In the traditional model, new knowledge needs to be integrated into the audit process through training or manual review, which cannot achieve self-evolution of audit capabilities.

[0005] In recent years, OCR, Natural Language Processing (NLP), and Computer Vision (CV) technologies have been applied in the field of document processing. However, existing technical solutions have obvious limitations: they mostly focus on a single link (such as automatic calculation of engineering quantities or quota application checks), lacking a comprehensive review framework covering all dimensions of "compliance-reasonableness-relevance"; they have failed to effectively establish deep semantic and spatial relationships between "text description-graphic components-price data"; and they generally lack closed-loop adaptive learning mechanisms, requiring redevelopment or extensive manual intervention after rule updates, thus failing to form technological innovation and systematic integration.

[0006] Therefore, this invention proposes an intelligent auditing method and system for engineering costs based on artificial intelligence. By innovating a hybrid architecture AI model, optimizing multimodal correlation algorithms, and constructing a closed-loop incremental learning mechanism, it solves the core pain points of traditional auditing and existing technologies, and achieves intelligent auditing that is comprehensive, highly accurate, and self-evolving. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide an intelligent auditing method and system for engineering costs based on artificial intelligence.

[0008] To achieve the aforementioned objectives, the technical solution of this invention is implemented as follows: An intelligent engineering cost review method and system based on a hybrid AI architecture and multimodal correlation. The core innovation is: constructing a three-in-one hybrid AI model of "compliance verification - rationality assessment - cross-relationship review," integrating OCR, NLP, CV, and lightweight BIM technologies to achieve multimodal correlation of diagrams, data, and text, and establishing a closed-loop incremental learning mechanism of "review - feedback - learning - update" to solve the technical challenges of cross-document logical conflict identification, unified review standards, and system adaptive iteration.

[0009] The process includes the following steps: S1: Data Acquisition and Structured Construction. Acquire a set of engineering cost documents to be reviewed, including at least the bill of quantities, comprehensive unit price analysis table, fee schedule, design drawings, and change orders. Utilize a self-developed OCR + document structure parsing algorithm and a lightweight BIM engine to perform structured processing on the heterogeneous documents. OCR converts scanned drawings and PDF / image files into structured text and pixel data, with a character recognition accuracy of ≥98%. WBS coding is applied to the inventory items according to the ISO12006-2 standard to achieve unique identification of each item; The design drawings are deconstructed at the component level to extract geometric attributes, material labels, and topological relationships, generating a lightweight BIM model for calculations with a compression rate of ≥80% without losing core calculation parameters. This ultimately forms a standardized audit data pool containing unified coding, component parameters, and semantic information.

[0010] S2: Dynamically invokes a multi-dimensional audit knowledge base. This pre-built multi-dimensional audit rule knowledge base uses a hybrid storage of graph database and relational database, integrating national / regional quota standards, real-time market price interfaces, enterprise historical cost databases, and vectorized historical audit case features; it supports real-time online updates with an update response time of ≤10s, adapting to external rules and market dynamics.

[0011] S3: Intelligent multi-round review based on a hybrid model inputs a standardized review data pool into a pre-trained and optimized hybrid AI review model, performing three-dimensional parallel analysis. The model's innovations are as follows: Overall architecture: a generative rule expert system + a numerical anomaly detection model + a multimodal correlation neural network, which work in parallel and collaboratively to output unified audit results; Training method: Transfer learning + semi-supervised learning is adopted, with 100,000+ historical cost data as the basic training set, and fine-tuned for the engineering field to avoid the problem of insufficient adaptability of conventional AI models in cost scenarios.

[0012] S31: Automatic compliance verification is based on a production rule expert system (innovative optimization: embedding a quota rule logic reasoning engine to support automatic verification of complex formulas). It performs logical matching and formal verification on engineering quantity calculation rules, quota application, and fee collection procedures, automatically marking suspected violations and directly outputting precise violation evidence (specific clauses and page numbers). Mathematical logic: Let the compliance verification function be F(C,R), where C is the list data and R is the knowledge base rule. When F(C,R)=0, compliance is determined; when F(C,R)=1, violation is determined, and the violation rule index R is output. x .

[0013] S32: The rationality assessment utilizes an improved isolated forest algorithm, an optimized Local Loop (LOF) anomaly detection algorithm, and a Gradient Boosting Decision Tree (GBDT) price prediction model (innovation: introducing a time decay factor and project category weights to improve price prediction accuracy). It performs horizontal (same-period, similar projects) and vertical (historical enterprise data) comparisons of main material unit prices, measure item costs, and comprehensive unit price composition to identify anomalous data points. Mathematical expression: Abnormal scores: (α+β=1, To score for LOF (For horizontal comparison of deviation scores) Price prediction: , Based on characteristics such as material type, region, and time, the prediction error is ≤3%. Risk levels are classified according to abnormal scores, and data points that deviate from the normal range are located.

[0014] S33: Cross-review of relevance utilizes optimized NLP semantic parsing and improved CV image semantic segmentation technology to establish a semantic-spatial correlation graph of invoices, unit prices, drawings, and visas (core innovation: breaking through the bottleneck of cross-modal matching of graphs, data, and text), specifically including: The Mask R-CNN model is optimized to locate components in the drawings, extract geometric and material features, and perform cosine similarity matching with the features in the bill of quantities. The similarity threshold is set to 0.85. If the similarity is lower than the threshold, it is judged as "no drawing quantity" or "feature description mismatch". The domain-adjusted BERT model is used to parse change order texts, extracting engineering quantities, processes, and material features. The theoretical impact value is automatically calculated, and the deviation rate is checked against the reported change amount. A deviation rate > 10% is used to determine "duplicate measurement" or "abnormal change." Mathematical logic: similarity. Deviation rate .

[0015] S4: Structured audit report and interactive integration of S31-S33 analysis results to generate a hierarchical structured audit report, including a risk level-ranked list of issues, correction suggestions, audit basis tracing, and evidence chain; supports one-click export of audit working paper format, which can be embedded into mainstream cost software; through an interactive visual interface, it can achieve precise positioning of issue items and source files (drawings, lists, visas), with a positioning response time of ≤1s.

[0016] S5: Adaptive iterative optimization feeds manually confirmed correction rules, new case features, and review decisions back to the knowledge base and AI model through an incremental learning algorithm (innovation: small sample incremental training, no need for full retraining), completing rule updates and model parameter optimization, forming a closed-loop adaptive evolution of review-feedback-learning-update. The system's review accuracy continues to improve with the accumulation of cases, and the accuracy rate is ≥95% after accumulating 1000+ projects.

[0017] To implement the above method, the following modules are included. Each module achieves innovative functions through technology integration and collaborative algorithms, rather than simply piecing together existing technologies: The data acquisition and standardization module integrates an OCR recognition unit, a document parsing unit, a feature encoding unit, and a lightweight BIM engine sub-module (innovation: supports automatic conversion of CAD / PDF drawings into lightweight BIM models and extraction of component calculation parameters), completes the structuring processing of heterogeneous data, and outputs a standardized audit data pool.

[0018] The multi-dimensional audit rules knowledge base module uses a hybrid storage of graph database and relational database, including quota rule sub-library, market data sub-library, cost indicator sub-library, and case feature sub-library. It supports real-time dynamic updates and provides unified rule support for auditing.

[0019] AI-powered review engine module Compliance verification unit: Equipped with a quota logic reasoning expert system to achieve automatic compliance verification and basis traceability; Reasonableness assessment unit: Built-in improved isolated forest, LOF anomaly detection model and weighted GBDT price evaluation model to improve anomaly identification accuracy; The relevance review unit integrates and optimizes the BERT+MaskR-CNN multimodal relevance model to achieve deep relevance verification of graphs, data, and documents. These three components work together to complete comprehensive intelligent analysis, resolving the challenge of identifying logical conflicts across documents.

[0020] The audit report generation and interaction module generates hierarchical structured reports, provides an interactive visual interface, supports problem tracing, document location, collaborative annotation, and is compatible with mainstream cost estimation software plugins.

[0021] The model self-learning optimization module includes a data cleaning unit, a feature extraction unit, and an incremental training unit (innovation: small sample incremental learning, quickly adapting to new rules and processes), which collects feedback data to update the knowledge base and optimizes model parameters.

[0022] The collaborative review workflow module automatically assigns manual review tasks according to professional fields, tracks the review progress, and sends the review conclusions back to the self-learning module as a supervision signal, thereby improving the human-machine collaborative review mechanism.

[0023] It stores a computer program that, when executed by a processor, implements any of the above-described intelligent cost auditing methods.

[0024] The beneficial effects of this invention are reflected in: Full-dimensional high-precision review: The three-in-one hybrid model breaks through the bottleneck of identifying logical conflicts across images, data, and text documents, with a compliance verification accuracy of ≥96% and an abnormal data identification recall rate of ≥94%, far exceeding single-technology review solutions; True self-evolution capability: The "review-feedback-learning-update" closed loop enables the system to self-adapt and iterate without the need for manual secondary development. It can automatically adapt to quota updates, market fluctuations and new processes, completely getting rid of the "outdated as soon as it goes online" dilemma. Ultimate review efficiency: The entire process of unstructured data is automated, reducing the review cycle from weeks to hours, improving efficiency by over 90%; the interactive positioning function reduces the burden of manual review and enables efficient human-machine collaboration; Unified standards and traceability: Audits are conducted based on a unified knowledge base and quantitative models to eliminate subjective bias; each issue is accompanied by a complete chain of evidence, ensuring that audit results are authoritative and traceable, and meeting audit compliance requirements. Attached Figure Description

[0025] In the attached diagram: Figure 1 Flowchart of an AI-based intelligent auditing method for engineering costs; Figure 2 : Module architecture diagram of an AI-based intelligent engineering cost auditing system; Figure 3 : Internal structure diagram of the AI ​​review engine module. Detailed Implementation

[0026] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only a part of the embodiments of the invention, and not all of them. Unless otherwise specified, the embodiments and features described in this application can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0027] Please refer to the instruction manual appendix. Figures 1-3 Example 1: An AI-based intelligent method for reviewing engineering costs (application in residential projects) Taking the cost audit of a residential construction project as an example, the implementation process is as follows: Step S1: Data acquisition and structured processing. Acquire the project's PDF list, comprehensive unit price analysis table, fee schedule, scanned design drawings, and Word change orders. Convert the unstructured data using a self-developed OCR algorithm (98.5% accuracy). Complete the WBS coding of the list according to the ISO 12006-2 standard. Extract the geometric dimensions and material information of beams, slabs, and columns using a lightweight BIM engine to generate a lightweight BIM model (85% compression rate) and form a standardized audit data pool.

[0028] Step S2: Call the multi-dimensional audit knowledge base. Call the Jiangsu Province 2024 version of construction engineering quota, the real-time price of building materials in the second quarter of 2025, the enterprise's historical residential project cost data, and vectorized audit case features. The knowledge base update response time is 8 seconds.

[0029] Step S3: Intelligent Review using Hybrid AI Model S31: Compliance Verification: The quota logic reasoning expert system compares the "site leveling" calculation rules and determines that calculating based on "land area" is illegal, based on Article 1.0.1 of the "Jiangsu Province Construction and Decoration Engineering Pricing Quota (2024 Edition)"; it verifies the safety and civilized construction fee rate and determines that it is not implemented according to the Class I project standard, marking the violation item. Mathematical verification: F(C₍site leveling₎, R₍calculation rules₎)=1, indicating a violation.

[0030] S32: Reasonableness Assessment: A horizontal comparison of the unit price of C30 concrete using the improved isolated forest model shows that the declared price is 15% higher than the industry average, exceeding the ±5% reasonable range. The weighted GBDT model predicts a reasonable price, classifying it as a medium-risk anomaly. Mathematical calculation: S(x) = 0.6 × 0.82 + 0.4 × 1.2 = 0.972 > anomaly threshold of 0.7, thus classified as an anomaly.

[0031] S33: Cross-relationship review: Optimized Mask R-CNN extracted the cross-sectional dimensions of frame beam KL1 as 250mm×500mm, which has a cosine similarity of 0.78 < 0.85 with the listed dimensions of 250mm×450mm, thus determining "feature description mismatch"; Domain BERT analysis of the roof insulation layer visa shows a theoretical engineering volume of 800㎡ and an amount of 48,000 yuan, while the reported amount is 62,000 yuan, resulting in a deviation rate of 29.17% > 10%, thus determining a high-risk abnormal change. Mathematical calculation: η = |6.2 - 4.8| / 4.8 × 100% = 29.17%.

[0032] Step S4: Generate a structured audit report. Generate a report containing 5 issues, risk levels, correction suggestions, and source tracing. It supports clicking on issues to directly locate the corresponding positions of visas and checklists, and allows one-click export of audit working papers.

[0033] Step S5: Adaptive iterative optimization and manual review to confirm all issues. The new feature "insulation layer change deviation rate > 20% is judged as high risk" and the review conclusion are updated to the knowledge base and model through incremental learning algorithm. The training time is ≤ 30 minutes, and there is no need for full retraining, which improves the accuracy of subsequent change visa review.

[0034] Example 2: Composition of an AI-based Intelligent Engineering Cost Review System The system's modules are integrated through innovative technologies to achieve collaborative operation, rather than simply piecing together existing technologies. Data acquisition and standardization module: Employs self-developed OCR algorithm, document parsing unit, WBS feature encoding unit, and lightweight BIM engine (self-optimized) to complete data structuring; Multi-dimensional audit rules knowledge base module: Neo4j graph database + MySQL hybrid storage, real-time synchronization of quotas, market prices, costs, and case data; Artificial intelligence audit engine module: developed in Python, integrating a quota reasoning expert system, an improved anomaly detection model, a multimodal correlation neural network, three-unit parallel analysis, and self-optimized core algorithm; Audit report generation and interaction module: A visual interface developed with Vue.js, supporting file location, annotation, and export; Model self-learning optimization module: data cleaning, feature extraction, small sample incremental training unit, and rapid update of rules and model; Collaborative review workflow module: The Activiti engine assigns professional review tasks, tracks progress, and provides feedback on conclusions to the self-learning module.

[0035] System interaction process: Data collection and standardization module → Artificial intelligence review engine module (calling knowledge base) → Report interaction module → Collaborative workflow module → Self-learning optimization module → Update knowledge base / engine model, forming a complete closed loop.

[0036] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0037] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0038] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An intelligent engineering cost auditing method based on artificial intelligence, characterized in that, Includes the following steps: S1: Obtain the project cost documents to be reviewed, including the bill of quantities, comprehensive unit price analysis table, fee schedule, design drawings and change orders; use OCR recognition + document parsing + lightweight BIM engine to perform structured processing on the diverse and heterogeneous documents, complete the WBS coding of the bill of quantities according to ISO12006-2 standard, deconstruct the drawings at the component level and generate a lightweight BIM model for calculation, forming a standardized review data pool; S2: Call the multi-dimensional audit rule knowledge base that uses a hybrid storage of graph database and relational database. The knowledge base includes quota specifications, real-time market prices, enterprise cost data and vectorized historical audit case features, and supports real-time online updates. S3: Input a standardized data pool into the hybrid architecture AI auditing model to perform three-dimensional intelligent auditing in parallel: S31: Compliance Verification: Based on a generative rule expert system with an embedded quota logic reasoning engine, the system performs logical matching verification on the engineering quantity calculation rules, quota application, and fee collection procedures, marks violations, and outputs the basis for traceability. S32: Reasonableness assessment: Using an improved isolated forest, an optimized LOF anomaly detection algorithm and a weighted GBDT price prediction model, the unit price of materials, the cost of measures, and the comprehensive unit price are compared with the same period in the horizontal and with the historical data in the vertical. The risk level is divided according to the anomaly score and the abnormal data points are located. S33: Cross-relationship review: Jointly optimize Mask R-CNN image semantic segmentation and domain fine-tuning BERT model, establish a semantic-spatial correlation map of list-drawings-approval, and identify undrawn quantities, duplicate measurements, inconsistent feature descriptions, and abnormal change conflicts through feature cosine similarity matching and change amount deviation rate verification. S4: Integrate audit results to generate a hierarchical structured audit report. The report includes a risk-ranked issue list, remedial suggestions, evidence traceability and chain of evidence, and supports interactive visualization and precise location of source files. S5: New rules, new case features, and review decisions confirmed by manual review are fed back to the knowledge base and AI model through a small-sample incremental learning algorithm to complete rule updates and model parameter optimization, achieving a closed-loop adaptive iteration of review-feedback-learning-update.

2. The method according to claim 1, characterized in that, In step S31, compliance verification uses mathematical logic: let the compliance verification function be F(C, R), where C is the manifest data and R is the knowledge base rule; when F(C, R) = 0, compliance is determined, and when F(C, R) = 1, violation is determined, and the corresponding violation rule index R is output. x .

3. The method according to claim 1, characterized in that, In step S32, the anomaly score and price prediction satisfy the following: Abnormal scores: (α+β=1, To score for LOF (For horizontal comparison of deviation scores) Price prediction: x n It incorporates weighted features based on material type, region, time, and project category, with a prediction error of ≤3%.

4. The method according to claim 1, characterized in that, In step S33, the association review matching rule is as follows: Component feature similarity: The threshold is set to 0.85; values ​​below this threshold will trigger a conflict. Change amount deviation rate: An abnormal change is determined if η > 10%.

5. The method according to claim 1, characterized in that, In step S1, the lightweight BIM engine can automatically parse CAD / PDF drawings, extract component geometric properties, material information and spatial topological relationships, with a model compression rate of ≥80% and no loss of core calculation parameters.

6. The method according to claim 1, characterized in that, In step S4, the structured audit report supports one-click export of standard audit working paper format and is embedded into mainstream engineering cost software as a plug-in.

7. An intelligent engineering cost auditing system based on artificial intelligence, used to implement the method described in any one of claims 1-6, characterized in that, include: Data acquisition and standardization module: integrates OCR recognition unit, document parsing unit, WBS feature coding unit and lightweight BIM drawing parsing submodule, completes the structured processing of diverse heterogeneous documents and outputs a standardized audit data pool; Multi-dimensional audit rules knowledge base module: It adopts a hybrid storage of graph database and relational database, and includes quota rules sub-library, market data sub-library, enterprise cost indicator sub-library, and historical case feature sub-library, which supports real-time dynamic updates; AI-powered audit engine module: This is the core processing unit, integrating a compliance verification unit, a rationality assessment unit, and a relevance review unit. These three units work in parallel and collaboratively to complete a full-dimensional intelligent audit. Audit report generation and interaction module: Used to generate hierarchical structured audit reports, providing an interactive visual interface, supporting issue tracing, file location, collaborative annotation, and format export; The model self-learning optimization module includes a data cleaning unit, a feature extraction unit, and a small sample incremental training unit, which completes knowledge base updates and model parameter optimization based on review feedback. Collaborative review workflow module: Automatically assigns manual review tasks according to professional fields, tracks review progress, and sends review conclusions back to the self-learning optimization module as a supervision signal.

8. The system according to claim 7, characterized in that... In the AI-powered review engine module: The rationality assessment unit incorporates an improved isolated forest model, an optimized LOF anomaly detection model, and a weighted GBDT price rationality assessment model. The relevance review unit integrates and optimizes the Mask R-CNN image recognition model with the domain-fine-tuned BERT natural language processing model to achieve multimodal semantic-spatial relevance matching of graphs, data, and text.

9. The system according to claim 7, characterized in that... The model self-learning optimization module adopts an incremental learning algorithm, which supports rapid iteration with small samples, eliminates the need for full retraining, and takes ≤30 minutes to update the model.

10. Computer-readable storage medium A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the AI-based intelligent auditing method for engineering costs as described in any one of claims 1-6.

Images