Water conservancy industry electronic dark bidding document enterprise internal examination method and system based on artificial intelligence

Abstract

The invention provides a water conservancy industry electronic dark bidding document enterprise internal examination method and system based on artificial intelligence, and belongs to the field of water conservancy industry bidding. Performing automatic pre-auditing on the bidding file: performing compliance inspection and integrity inspection, identifying problems existing in the bidding file, and providing improvement suggestions to ensure that the problems conform to the basic requirements of electronic dark label review; carrying out data standardization and cleaning on the bidding file: carrying out semantic extension and ambiguity elimination, identifying images, tables and handwritten contents in the bidding file, identifying a main body structure of the water conservancy design drawing, extracting engineering quantity list data, and carrying out compliance verification; identifying an abnormal behavior in the bidding file through a machine learning model; and an internal examination decision tree is constructed according to preset key indexes and score weights of the water conservancy project, an interpretable artificial intelligence algorithm is utilized to carry out internal examination decision making on the bidding document, an internal examination score result is fed back, and specific score deduction reasons are explained. And the standardization and the accuracy of internal examination of the bidding document are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of bidding and tendering in the water conservancy industry, and in particular to an enterprise internal audit method and system for electronic blind bidding documents in the water conservancy industry based on artificial intelligence. Background Art

[0002] Blind bid review is a common bid evaluation method in the bidding and tendering sector, widely used for its fairness and confidentiality. However, traditional blind bid review methods are inefficient and time-consuming, lack standardized review processes, and pose risks of information leakage. To address these issues, the "Technical Specifications for Electronic Bidding Systems" have been developed, and several electronic bidding and tendering platforms, such as the mainstream China Bidding and Tendering Public Service Platform, have been established to support electronic blind bid review.

[0003] However, companies currently find that the following problems still exist when participating in electronic blind bidding:

[0004] 1. Difficulties in data standardization: Electronic blind bidding requires a highly standardized data format for bid documents, but currently, companies offer a wide variety of bid document formats, making it difficult to meet the system's standardization requirements. Companies lack data cleansing and standardization tools, making it difficult to convert unstructured data into the format required by the system. This difficulty in data standardization can lead to the following problems:

[0005] The bidding document is rejected by the system: due to the format not meeting the requirements, the bidding document may not pass the system's initial review, resulting in the enterprise losing the bidding qualification.

[0006] Data parsing errors: During the review process, the system may not be able to correctly parse non-standardized data, resulting in deviations in the review results, affecting the company's score and chances of winning the bid.

[0007] 2. Low quality of bidding documents: Many companies lack systematic planning and professional support when preparing bidding documents, resulting in low quality bidding documents that fail to fully demonstrate the company's advantages. Specific issues include:

[0008] The technical proposal is not detailed enough: The technical proposal is the core part of the bidding document, but many companies' technical proposals are too brief and cannot meet the requirements of the bidding documents, resulting in the evaluation experts being unable to fully understand the company's technical capabilities.

[0009] Unreasonable quotations: Some companies have obvious low-price bidding or high-price bidding phenomena. Unreasonable quotations may cause the company to score low in the evaluation or even be identified as malicious bidding.

[0010] Repetitive content and unclear logic: The repetitive content and unclear logic in the bidding documents affect the review experts' reading and understanding, resulting in the company's low score in the review and loss of the chance to win the bid.

[0011] 3. Insufficient understanding of electronic blind bidding rules: There are significant differences between electronic blind bidding and traditional open bidding. Many companies do not fully understand the rules and procedures of electronic blind bidding, leading to mistakes during the bidding process. Specific issues include:

[0012] Lack of understanding of anonymization requirements: Electronic bidding requires that bidding documents be completely anonymous, but many companies are unaware of this requirement, resulting in the disclosure of corporate information in bidding documents, violating anonymization regulations and potentially being disqualified from bidding.

[0013] Unfamiliarity with evaluation standards and processes: Many companies are unfamiliar with the evaluation standards and processes for electronic blind bidding and are unable to prepare bidding documents in a targeted manner, putting them at a disadvantage in the evaluation.

[0014] 4. Insufficient understanding of evaluation criteria: The evaluation criteria for electronic blind bidding are usually complex. Many companies do not fully understand the evaluation criteria, resulting in bid documents failing to fully meet the evaluation requirements. Specific issues include:

[0015] Inaccurate weight distribution of technical proposals and quotations: Enterprises failed to accurately grasp the weight distribution of technical proposals and quotations, causing the bid documents to deviate from the evaluation focus and affecting the evaluation score.

[0016] Insufficient understanding of detailed requirements: Detailed requirements such as construction measures and construction schedules in the evaluation standards are often ignored by enterprises, resulting in bidding documents not meeting the requirements and affecting the evaluation results. Summary of the Invention

[0017] In order to solve the problems of the existing technology, the present invention provides an artificial intelligence-based enterprise internal audit method and system for electronic blind bidding documents in the water conservancy industry. Based on artificial intelligence technology, combined with water conservancy industry project standards and blind bidding review processes, it aims to further improve the standardization and accuracy of the internal audit of electronic blind bidding documents for water conservancy industry projects.

[0018] The present invention provides an artificial intelligence-based internal audit method for electronic blind bidding documents in the water conservancy industry, comprising the following steps:

[0019] Step S1: Automatically pre-examine the bidding documents: perform compliance checks and completeness checks, identify problems in the bidding documents, and provide improvement suggestions to ensure that they meet the basic requirements of electronic blind bidding review;

[0020] Step S2: Standardize and clean the data of the bidding documents obtained in step S1 that meet the basic requirements of the electronic blind bidding review. This includes semantic expansion and ambiguity elimination, recognition of images, tables, and handwritten content in the bidding documents, and identification of the main structure of the water conservancy design drawings. The data of the bill of quantities is extracted and compliance verification is performed.

[0021] Step S3, identifying abnormal behaviors in the bidding documents through a machine learning model;

[0022] Step S4: construct an internal audit decision tree based on the preset key indicators and scoring weights of the water conservancy project, use an explainable artificial intelligence algorithm to make internal audit decisions on the bidding documents, feedback the internal audit scoring results, and explain the specific reasons for the deductions.

[0023] Preferably, the problems existing in the bidding documents in step S1 include format errors, missing content, and inconsistent data.

[0024] Preferably, in step S1, a system interface is designed for compliance checking. The system interface includes a file upload interface and a water conservancy specification data interface. The file upload interface supports XML or JSON format that complies with the "Technical Specifications for Electronic Bidding and Tendering Systems" to ensure that the bidding documents and review results are compatible with mainstream platforms; the water conservancy specification data interface is connected to the network through OAuth.

[0025] 2.0 protocol accesses the national water industry standards database and obtains specification entries in related fields in real time for compliance checks.

[0026] Preferably, the integrity check in step S1 specifically involves identification of key chapters of interest and content integrity verification, specifically including:

[0027] Use the NLP model to segment the bidding documents into paragraphs, and use the BERT model to identify the titles of key sections of interest. If a section is missing, a prompt will be given.

[0028] Through keyword extraction and semantic analysis, the BERT model is used to verify whether each chapter includes the necessary content of interest. If the necessary content of interest is missing, a prompt will be given.

[0029] Preferably, step S2 specifically includes:

[0030] Input data, process it using NLP models, build a dynamic terminology database covering the field of water conservancy engineering, and achieve semantic expansion of terminology;

[0031] In the case of polysemous words or abbreviations in the bidding documents, a context-aware model is used to eliminate ambiguity;

[0032] Computer vision is used to process images, tables, and handwritten content in bidding documents, identify key structures of water conservancy design drawings in bidding documents, extract bill of quantities data from bidding documents, and simultaneously link to the water conservancy industry standard library for compliance verification.

[0033] Preferably, in step S3, through the reinforcement learning Q-learning model, in the dynamic score allocation, the state is defined as the feature set of the bidding document, the action is defined as the adjustment of the weights of the technical solution, quotation, enterprise qualification, and historical performance, and the reward is defined as: a positive reward if the scoring result is consistent with the expert review result, a negative reward if the scoring result is inconsistent with the expert review result, and a zero reward if the scoring result is not verified.

[0034] Preferably, in step S3, an SBERT model is used to perform anomaly detection on the data to be detected to generate anomaly detection results.

[0035] Preferably, in step S3, the SBERT model combines two twin networks or triple network structures with BERT. In the twin network structure, two identical BERT encoders share weights, each receiving a sentence input and generating a semantic vector for calculating sentence similarity; in the triple network structure, three identical BERT encoders are used to process anchor sentences, positive sample sentences, and negative sample sentences respectively, so that semantically similar sentences are closer in the embedding space, and semantically different sentences are farther away in the embedding space.

[0036] Preferably, using the SBERT model to perform anomaly detection on the data to be tested specifically includes the following steps:

[0037] Data collection: Collect annotated sentence pairs or sentence groups. This data includes similarity labels for the sentence pairs, where 0 indicates dissimilarity and 1 indicates similarity, or anchor-positive-negative sentence groups for triplet networks.

[0038] Data preprocessing: Segment the collected text data, add tags [CLS] and [SEP], and convert sentence pairs or sentence groups into a format suitable for input into the SBERT model;

[0039] Data anomaly detection: The bidding documents are input into the SBERT model, which outputs semantic vectors. The semantic similarity of the semantic vectors of different bidding documents is calculated. Based on the semantic similarity, it is determined whether the technical solutions are identical or whether the quotations fluctuate abnormally.

[0040] Preferably, data preprocessing includes concatenating two sentences, separating them with [SEP] and adding a [CLS] tag at the beginning.

[0041] Preferably, step S4 specifically includes the following steps:

[0042] During the internal review of bidding documents, the decision-making process of the artificial intelligence algorithm is visualized, and an internal audit decision tree is constructed based on the key indicators and scoring weights preset for the water conservancy project. Internal audit experts make decisions based on the decision tree and generate internal audit reports.

[0043] Build a database of water conservancy industry standards and regulations, automatically associate internal audit reports with the standards and regulations in the database, provide feedback on internal audit scores, and explain the specific reasons for deductions.

[0044] Preferably, an internal audit decision tree is constructed based on the preset key indicators and scoring weights of the water conservancy project. The internal audit experts make decisions based on the decision tree and generate an internal audit report that specifically includes:

[0045] Based on the internal review requirements for electronic blind bidding documents in the water conservancy industry, we first determined the scoring factors and weights, and collected and pre-processed historical bidding document data; scoring factors included technical solutions and price rationality;

[0046] Use ID3 algorithm to train decision tree model;

[0047] Combining expert experience with the Q-learning algorithm to dynamically optimize weights, the Q-learning algorithm is used to achieve weight optimization, text parsing, and anomaly detection, ultimately generating a visual display decision tree.

[0048] Internal audit experts make decisions based on the decision tree;

[0049] Build a database of water conservancy industry standards and regulations, automatically associate internal audit reports with the standards and regulations in the database, provide feedback on internal audit scores, and explain the specific reasons for deductions.

[0050] The present invention provides an enterprise internal audit system for electronic blind bidding documents in the water conservancy industry based on artificial intelligence, including a processor capable of executing a computer program, and the computer program capable of implementing the steps of the above-mentioned enterprise internal audit method for electronic blind bidding documents in the water conservancy industry based on artificial intelligence.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] (1) Improve industry compliance

[0053] The system is embedded with water conservancy industry standards and specifications such as the "Specifications for Construction Organization Design of Water Conservancy and Hydropower Projects". During the system internal audit process, the technical proposals in the bidding documents are compared item by item with the system's embedded standards and specifications, thereby improving the system's ability to detect missing technical proposals and further improving the compliance of the bidding technical proposals.

[0054] (2) Improved internal audit efficiency

[0055] Traditionally, manual internal review takes 3-5 man-days (2-3 days for technical review + 1-2 days for commercial review), and can easily extend to 1-2 weeks due to repeated document revisions. This application's fully automated process (data cleaning + AI review) takes ≤ 2 hours, increasing efficiency by over 20 times.

[0056] (3) Strengthening industry adaptability

[0057] The analysis of water conservancy professional data has been strengthened. Among them, through computer vision (CV) technology, the structural layout drawings in the water conservancy design drawings in the bidding documents are identified, and the processing capabilities of complex water conservancy structure drawings such as tunnel lining layers and gate opening and closing machine installation drawings are enhanced. The accuracy rate of drawing parameter extraction is ≥98%.

[0058] At the same time, combined with the water conservancy professional terminology database and NLP parsing optimization technology, the semantic expansion of water conservancy terminology in the system is realized, thereby improving the parsing effect of water conservancy terminology. For example, the semantic parsing accuracy of non-standard terms such as grass gabion foot guards and inverted siphon anchors is increased to 95%, avoiding scoring deviations due to differences in expression.

[0059] (4) Rating adaptability optimization

[0060] Dynamic weight allocation makes the technical solution scoring more in line with the characteristics of water conservancy projects. For example, the weight of "channel anti-seepage technology" in irrigation projects is increased by 15%, thereby optimizing the technical solution from the perspective of scoring adaptability, thereby increasing the winning rate of the bidding solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying any creative work.

[0062] Figure 1 This is a flow chart of an enterprise internal audit method for electronic blind bidding documents in the water conservancy industry based on artificial intelligence, according to an embodiment of the present invention. DETAILED DESCRIPTION

[0063] The specific embodiments of the present invention are described in detail below.

[0064] The present invention provides an artificial intelligence-based internal audit method for electronic blind bidding documents in the water conservancy industry, comprising the following steps:

[0065] Step S1: Automatically pre-examine the bidding documents: perform compliance checks and completeness checks, identify problems in the bidding documents, and provide improvement suggestions to ensure that they meet the basic requirements of electronic blind bidding review;

[0066] Step S2: Standardize and clean the data of the bidding documents obtained in step S1 that meet the basic requirements of the electronic blind bidding review. This includes semantic expansion and ambiguity elimination, recognition of images, tables, and handwritten content in the bidding documents, and identification of the main structure of the water conservancy design drawings. The data of the bill of quantities is extracted and compliance verification is performed.

[0067] Step S3, identifying abnormal behaviors in the bidding documents through a machine learning model;

[0068] Step S4: construct an internal audit decision tree based on the preset key indicators and scoring weights of the water conservancy project, use an explainable artificial intelligence algorithm to make internal audit decisions on the bidding documents, feedback the internal audit scoring results, and explain the specific reasons for the deductions.

[0069] According to a specific embodiment of the present invention, the problems existing in the bidding documents in step S1 include format errors, missing content, and inconsistent data.

[0070] According to a specific implementation scheme of the present invention, in step S1, a system interface is designed to perform compliance checks. The system interface includes a file upload interface and a water conservancy specification data interface. The file upload interface supports XML or JSON format that complies with the "Technical Specifications for Electronic Tendering and Bidding Systems" to ensure that the bidding documents and review results are compatible with mainstream platforms; the water conservancy specification data interface accesses the national water conservancy industry standard database through the OAuth 2.0 protocol to obtain specification entries in related fields in real time for compliance checks.

[0071] According to a specific embodiment of the present invention, the integrity check in step S1 specifically involves identification of key chapters of interest and content integrity verification, specifically including:

[0072] Use the NLP model to segment the bidding documents into paragraphs, and use the BERT model to identify the titles of key sections of interest. If a section is missing, a prompt will be given.

[0073] Through keyword extraction and semantic analysis, the BERT model is used to verify whether each chapter includes the necessary content of interest. If the necessary content of interest is missing, a prompt will be given.

[0074] According to a specific embodiment of the present invention, step S2 specifically includes:

[0075] Input data, process it using NLP models, build a dynamic terminology database covering the field of water conservancy engineering, and achieve semantic expansion of terminology;

[0076] In the case of polysemous words or abbreviations in the bidding documents, a context-aware model is used to eliminate ambiguity;

[0077] Computer vision is used to process images, tables, and handwritten content in bidding documents, identify key structures of water conservancy design drawings in bidding documents, extract bill of quantities data from bidding documents, and simultaneously link to the water conservancy industry standard library for compliance verification.

[0078] According to a specific embodiment of the present invention, in step S3, through the reinforcement learning Q-learning model, in the dynamic score allocation, the state is defined as the feature set of the bidding document, the action is defined as the adjustment of the weights of the technical solution, quotation, enterprise qualification, and historical performance, and the reward is defined as: a positive reward if the scoring result is consistent with the expert review result, a negative reward if the scoring result is inconsistent with the expert review result, and a zero reward if the scoring result is not verified.

[0079] According to a specific embodiment of the present invention, in step S3, anomaly detection is performed on the data to be detected using the SBERT model to generate anomaly detection results.

[0080] According to a specific embodiment of the present invention, in step S3, the SBERT model combines two twin networks or triple network structures with BERT. In the twin network structure, two identical BERT encoders share weights, each receiving a sentence input and generating a semantic vector for calculating sentence similarity; in the triple network structure, three identical BERT encoders are used to process anchor sentences, positive sample sentences and negative sample sentences respectively, so that semantically similar sentences are closer in the embedding space, and semantically different sentences are farther away in the embedding space.

[0081] According to a specific embodiment of the present invention, using the SBERT model to perform anomaly detection on the data to be tested specifically includes the following steps:

[0082] Data collection: Collect annotated sentence pairs or sentence groups. This data includes similarity labels for the sentence pairs, where 0 indicates dissimilarity and 1 indicates similarity, or anchor-positive-negative sentence groups for triplet networks.

[0083] Data preprocessing: Segment the collected text data, add tags [CLS] and [SEP], and convert sentence pairs or sentence groups into a format suitable for input into the SBERT model;

[0084] Data anomaly detection: The bidding documents are input into the SBERT model, which outputs semantic vectors. The semantic similarity of the semantic vectors of different bidding documents is calculated. Based on the semantic similarity, it is determined whether the technical solutions are identical or whether the quotations fluctuate abnormally.

[0085] According to a specific embodiment of the present invention, data preprocessing includes concatenating two sentences, separating them with [SEP] and adding a [CLS] tag at the beginning.

[0086] According to a specific embodiment of the present invention, step S4 specifically includes the following steps:

[0087] During the internal review of bidding documents, the decision-making process of the artificial intelligence algorithm is visualized, and an internal audit decision tree is constructed based on the key indicators and scoring weights preset for the water conservancy project. Internal audit experts make decisions based on the decision tree and generate internal audit reports.

[0088] Build a database of water conservancy industry standards and regulations, automatically associate internal audit reports with the standards and regulations in the database, provide feedback on internal audit scores, and explain the specific reasons for deductions.

[0089] According to a specific embodiment of the present invention, an internal audit decision tree is constructed based on the preset key indicators and scoring weights of the water conservancy project. The internal audit experts make decisions based on the decision tree, and the generation of the internal audit report specifically includes:

[0090] Based on the internal review requirements for electronic blind bidding documents in the water conservancy industry, we first determined the scoring factors and weights, and collected and pre-processed historical bidding document data; scoring factors included technical solutions and price rationality;

[0091] Use ID3 algorithm to train decision tree model;

[0092] Combining expert experience with the Q-learning algorithm to dynamically optimize weights, the Q-learning algorithm is used to achieve weight optimization, text parsing, and anomaly detection, ultimately generating a visual display decision tree.

[0093] Internal audit experts make decisions based on the decision tree;

[0094] Build a database of water conservancy industry standards and regulations, automatically associate internal audit reports with the standards and regulations in the database, provide feedback on internal audit scores, and explain the specific reasons for deductions.

[0095] The present invention provides an enterprise internal audit system for electronic blind bidding documents in the water conservancy industry based on artificial intelligence, including a processor capable of executing a computer program, and the computer program capable of implementing the steps of the above-mentioned enterprise internal audit method for electronic blind bidding documents in the water conservancy industry based on artificial intelligence.

[0096] Example 1

[0097] The present invention provides an artificial intelligence-based internal audit method for electronic blind bidding documents in the water conservancy industry, comprising the following steps:

[0098] Step S1: Automatically pre-examine the bidding documents: perform compliance checks and completeness checks, identify problems in the bidding documents, and provide improvement suggestions to ensure that they meet the basic requirements of electronic blind bidding review;

[0099] Step S2: Standardize and clean the data of the bidding documents obtained in step S1 that meet the basic requirements of the electronic blind bidding review. This includes semantic expansion and ambiguity elimination, recognition of images, tables, and handwritten content in the bidding documents, and identification of the main structure of the water conservancy design drawings. The data of the bill of quantities is extracted and compliance verification is performed.

[0100] Step S3, identifying abnormal behaviors in the bidding documents through a machine learning model;

[0101] Step S4: construct an internal audit decision tree based on the preset key indicators and scoring weights of the water conservancy project, use an explainable artificial intelligence algorithm to make internal audit decisions on the bidding documents, feedback the internal audit scoring results, and explain the specific reasons for the deductions.

[0102] Example 2

[0103] The present invention provides an artificial intelligence-based internal audit method for electronic blind bidding documents in the water conservancy industry, comprising:

[0104] 1. Automatically pre-examine bidding documents

[0105] (1) Design system interface

[0106] Step 1: Standardize knowledge graph construction

[0107] Data access: Access the national water conservancy standards database through the OAuth 2.0 protocol to extract standard entries (such as "flood wall thickness ≥ 1.2m" in the "Water Conservancy Project Design Code").

[0108] Graph generation: Use Neo4j to build entity-relationship networks (such as "norm-parameter-threshold" associations).

[0109] Step 2: Semantic compliance verification

[0110] Parameter extraction: Extract key parameters (such as "concrete compressive strength 30MPa") from bidding documents through NLP models.

[0111] Dynamic matching: compare the parameters with the thresholds in the knowledge graph and generate an alarm if they do not match (e.g. "compression strength needs to be

[0112] ≥35MPa”).

[0113] (2) Content integrity check

[0114] Step 1: Domain corpus construction

[0115] Collect industry standard data and build a training set.

[0116] Step 2: Multimodal integrity check

[0117] 1) In-depth verification of qualification compliance: The system has a built-in water conservancy industry qualification database (such as the qualification level of general contracting for water conservancy and hydropower construction, and the status of production safety licenses). It automatically verifies whether the enterprise qualifications and project manager certificate qualifications in the bidding documents meet the requirements of water conservancy projects (for example, the bidding requires that the project manager must meet the requirements of a registered construction engineer in the water conservancy and hydropower field). It also links the credit rating (AAA / AA) and generates early warnings for enterprises that do not meet the qualifications or have a bad credit record.

[0118] 2) Consortium Division Compliance Review: For consortium bidding scenarios common in water conservancy projects, the system automatically identifies the division of responsibilities and rights in the consortium agreement (generally between the design and construction parties). It also checks whether the qualification documents provided by the consortium leader meet the qualification requirements for the consortium leader in the bidding documents.

[0119] 3) Content integrity check: The system uses natural language processing (NLP) technology to check whether the core water conservancy chapters in the technical plan, such as flood prevention and flood relief plans (such as diversion projects and cofferdam designs) and ecological protection measures (such as soil and water conservation plans and vegetation restoration processes), are complete.

[0120] 4) Compliance Check: The system automatically checks bid documents for compliance with relevant water conservancy industry specifications and standards. For example, based on the water conservancy industry standard SL631, "Regulations for Inspection and Assessment of Construction Quality of Water Conservancy Projects," it verifies whether technical parameters such as concrete impermeability and frost resistance, and foundation treatment processes (such as the depth of the anti-seepage wall) meet the specifications.

[0121] 5) Cross-modal consistency check: Compare the "concrete quantity" described in the text with the value marked on the drawing, and trigger an inconsistency alarm.

[0122] 2. Standardize and clean the data of bidding documents that meet the basic requirements of electronic blind bidding review

[0123] (1) Use natural language processing (NLP) technology to analyze text

[0124] Step 1: Text Parsing

[0125] After receiving the bid documents, the system first uses NLP technology to parse the text. Through techniques such as word segmentation, part-of-speech tagging, and named entity recognition, it extracts key information from the technical proposal, such as the construction period, construction process, and material specifications.

[0126] Step 2: Data normalization

[0127] Convert the extracted key information into a unified format. For example, convert the construction period into days and material specifications into national or industry standard units.

[0128] Step 3: Data Cleansing

[0129] The extracted data is cleaned to remove noisy data such as duplicate content and fill in missing values. For example, if a bidding document does not clearly indicate the construction period, the system will automatically fill in a reasonable construction period range based on historical data.

[0130] (2) Using computer vision (CV) technology to identify images and table content

[0131] Step 1: Image Recognition

[0132] The system uses CV technology to recognize images and table content in bidding documents. For example, it can identify key nodes in construction drawings and table data in bills of quantities.

[0133] Step 2: Data extraction and transformation

[0134] Convert recognized images and table content into structured data. For example, convert quantities in construction drawings into a processable digital format, or extract quotation data from tables into a unified format.

[0135] Step 3: Data Verification

[0136] Validate the extracted data to ensure its accuracy and completeness. For example, check that the total price in the bill of quantities equals the unit price multiplied by the quantity.

[0137] 3. Identify abnormal behavior in bidding documents through machine learning models

[0138] (1) Implementation of reinforcement learning algorithms

[0139] Q-learning is a value-based reinforcement learning algorithm that enables an agent to take actions in the environment to maximize the cumulative reward by learning a policy. The core of Q-learning is the Q-table, which stores the expected reward value (Q value) under different state and action combinations. In this embodiment, the agent is the internal audit system, the environment is the internal audit process of the bidding documents, and the reward is the accuracy and rationality of the internal audit results. In this embodiment, the reinforcement learning algorithm dynamically adjusts the scoring weights by learning historical internal audit data to adapt to the special needs of different water conservancy projects. The following are the specific implementation steps:

[0140] Step 1: Study historical data

[0141] First, we collected historical bid documents for water conservancy projects over the past three years, along with their corresponding scoring results and internal audit feedback, to establish a preliminary database. Second, based on this database, we extracted key features from the bid documents, such as the level of detail in the technical proposal, the rationality of the quote, and the compliance of the company's qualifications. Finally, we used weight initialization to assign initial weights to each feature. These weights will be continuously adjusted as the system continues to run and learn.

[0142] Step 2: Design the reward function

[0143] The reward function here refers to the accuracy and rationality of the internal audit results. During the initial design phase, the same bid document undergoes both expert review and system-based internal audit. The system-based internal audit results are compared with the expert review results. If the internal audit results agree with the expert review results, a positive reward is awarded; if they disagree, a negative reward is given. Therefore, by defining a reward function, we can measure the accuracy and rationality of the internal audit results.

[0144] Step 3: Algorithm training

[0145] Initialization state: Select a bidding document from historical data as the initial state.

[0146] Choose an action:

[0147] Randomly choose an action with probability ε (exploration).

[0148] Select the action (utilization) with the highest Q value in the current state with probability 1-ε.

[0149] Execute action: The system executes the selected action (such as increasing the weight of the technical solution) and observes the environmental feedback (reward).

[0150] Update the Q table:

[0151] Calculate the new Q value:

[0152] Q(s,a)←Q(s,a)+α[r+γmax(Q(s′,a′))-Q(s,a)]

[0153] Where s is the current state, a is the current action, r is the reward, s′ is the next state, and a′ is the possible action for the next state.

[0154] State transfer: Update the current state to the next state.

[0155] Cyclic training: Repeat the above steps until the algorithm converges (the Q value change is less than a certain threshold).

[0156] Step 4: Dynamic Update

[0157] Adjusting and updating the weighting of review content within the system primarily relies on selecting an appropriate reinforcement learning strategy. The state space and action space are regularly updated to adapt to new review criteria and project requirements. Furthermore, new historical data is continuously collected and the model is regularly retrained to ensure adaptability and accuracy. For example, if a feature (such as the level of detail in the technical proposal) has a significant impact on the scoring results, its weight is increased; otherwise, its weight is decreased.

[0158] (2) Abnormal behavior identification

[0159] Step 1: Data preparation and preprocessing

[0160] Collect technical proposals from historical water conservancy project bidding documents, which cover key technical details of water conservancy projects, such as dam construction technology, sluice design parameters, and irrigation system layout plans;

[0161] Annotate technical proposal texts to identify similar texts (e.g., documents showing the same water conservancy company using similar dam reinforcement technical solutions in different reservoir projects) and dissimilar texts (e.g., documents showing different companies using different diversion technical solutions for river management projects).

[0162] The annotated text was segmented using a specialized word segmentation tool for the water conservancy industry, accurately segmenting it into water conservancy professional vocabulary units such as "concrete pouring process," "anti-seepage wall depth," and "flood discharge calculation," thus fully preparing for subsequent model processing.

[0163] Step 2: Model loading

[0164] Select a Sentence-BERT (SBERT) pre-trained model that has been fine-tuned on a large number of water conservancy documents, water conservancy project reports, and other corpora, and is suitable for the Chinese water conservancy industry context. Ensure that the model can better adapt to the semantic characteristics of water conservancy technical texts and is familiar with water conservancy professional terminology and expressions.

[0165] Step 3: Text Encoding

[0166] The text of the water conservancy project technical plan is input into the loaded SBERT model. With the help of the model's powerful semantic understanding ability, a semantically rich vector representation is generated. This vector can accurately capture the core semantic information of water conservancy technology in the text, such as "pump station head design standards" and "channel lining material selection", and convert complex water conservancy technical plan text into a numerical form that can be processed by computers.

[0167] Step 4: Similarity calculation

[0168] From the vectors output by the SBERT model, we extract the vector representations corresponding to the technical proposals of the water conservancy projects to be tested. These vectors condense the key points of the water conservancy technical proposals.

[0169] The cosine similarity formula is used to calculate the similarity between the text vectors of the technical proposals of two water conservancy projects. The formula is: Similarity = (Vector 1·Vector 2) / (||Vector 1||*||Vector 2||). The similarity value is used to judge the degree of similarity between the technical proposal texts of the water conservancy project bidding documents, thereby realizing the detection of abnormal behavior (similarity behavior) of the water conservancy project bidding documents. For example, it is detected whether the technical proposal texts submitted in two different bidding documents have abnormal similarities in key parts such as "Seismic Design Method of Hydraulic Structures" and "Layout Planning of Water Conservancy Hubs".

[0170] 4. Construct an internal audit decision tree based on the preset key indicators and scoring weights of the water conservancy project, use an explainable artificial intelligence algorithm to make internal audit decisions on bidding documents, provide feedback on internal audit scoring results, and explain the specific reasons for deductions

[0171] Implementation of Visualization Technology for Internal Audit Decision-Making Path

[0172] Step 1: Decision tree generation

[0173] The system generates a decision tree for the internal audit results, displaying the specific scores and weights for each scoring item. For example, the decision tree will display specific information such as the technical solution score and the quotation score.

[0174] Step 2: Visualization

[0175] The decision tree is presented to internal auditors in a visual format. For example, the system generates a chart showing the score and weight of each scoring item, making it easier for internal auditors to understand and verify.

[0176] Step 3: Generate internal audit report

[0177] The system generates detailed internal audit reports that demonstrate the decision-making process behind internal audit findings. For example, the report explains the specific basis and scoring process for each scoring item, ensuring transparency and traceability of internal audit results.

[0178] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. An artificial intelligence-based internal audit method for electronic blind bidding documents in the water conservancy industry, characterized by: The steps include: Step S1: Automatically pre-examine the bidding documents: perform compliance checks and completeness checks, identify problems in the bidding documents, and provide improvement suggestions to ensure that they meet the basic requirements of electronic blind bidding review; Step S2: Standardize and clean the data of the bidding documents obtained in step S1 that meet the basic requirements of the electronic blind bidding review. This includes semantic expansion and ambiguity elimination, recognition of images, tables, and handwritten content in the bidding documents, and identification of the main structure of the water conservancy design drawings. The data of the bill of quantities is extracted and compliance verification is performed. Step S3, identifying abnormal behaviors in the bidding documents through a machine learning model; Step S4: construct an internal audit decision tree based on the preset key indicators and scoring weights of the water conservancy project, use an explainable artificial intelligence algorithm to make internal audit decisions on the bidding documents, feedback the internal audit scoring results, and explain the specific reasons for the deductions.

2. The artificial intelligence-based internal audit method for electronic blind bidding documents in the water conservancy industry according to claim 1 is characterized in that: Problems in the bidding documents in step S1 include format errors, missing content, and inconsistent data.

3. The artificial intelligence-based internal audit method for electronic blind bidding documents in the water conservancy industry according to claim 2 is characterized in that: In step S1, a system interface is designed to perform compliance checks. The system interface includes a file upload interface and a water conservancy specification data interface. The file upload interface supports XML or JSON format that complies with the "Technical Specifications for Electronic Tendering and Bidding Systems" to ensure that the bidding documents and review results are compatible with mainstream platforms; the water conservancy specification data interface accesses the national water conservancy industry standard database through the OAuth 2.0 protocol to obtain specification entries in related fields in real time for compliance checks.

4. The artificial intelligence-based internal audit method for electronic blind bidding documents in the water conservancy industry according to claim 2 is characterized in that: The integrity check in step S1 specifically involves identifying the key chapters of interest and verifying their content integrity, including: Use the NLP model to segment the bidding documents into paragraphs, and use the BERT model to identify the titles of key sections of interest. If a section is missing, a prompt will be given. Through keyword extraction and semantic analysis, the BERT model is used to verify whether each chapter includes the necessary content of interest. If the necessary content of interest is missing, a prompt will be given.

5. The artificial intelligence-based internal audit method for electronic blind bidding documents in the water conservancy industry according to claim 1 is characterized in that: Step S2 specifically includes: Input data, process it using NLP models, build a dynamic terminology database covering the field of water conservancy engineering, and achieve semantic expansion of terminology; In the case of polysemous words or abbreviations in the bidding documents, a context-aware model is used to eliminate ambiguity; Computer vision is used to process images, tables, and handwritten content in bidding documents, identify key structures of water conservancy design drawings in bidding documents, extract bill of quantities data from bidding documents, and simultaneously link to the water conservancy industry standard library for compliance verification.

6. The artificial intelligence-based internal audit method for electronic blind bidding documents in the water conservancy industry according to claim 1 is characterized in that: In step S3, through the reinforcement learning Q-learning model, in the dynamic score allocation, the state is defined as the feature set of the bidding document, the action is defined as the adjustment of the weights of the technical solution, quotation, enterprise qualification, and historical performance, and the reward is defined as: if the scoring result is consistent with the expert review result, it is a positive reward; if the scoring result is inconsistent with the expert review result, it is a negative reward; if the scoring result is not verified, it is a zero reward.

7. The artificial intelligence-based internal audit method for electronic blind bidding documents in the water conservancy industry according to claim 1 is characterized in that: In step S3, the SBERT model is used to perform anomaly detection on the data to be detected and generate anomaly detection results.

8. The artificial intelligence-based internal audit method for electronic blind bidding documents in the water conservancy industry according to claim 7 is characterized in that: In step S3, the SBERT model combines two twin networks or triplet network structures with BERT. In the twin network structure, two identical BERT encoders share weights, each receiving a sentence input and generating a semantic vector for calculating sentence similarity; in the triplet network structure, three identical BERT encoders are used to process anchor sentences, positive sample sentences, and negative sample sentences respectively, so that semantically similar sentences are closer in the embedding space, and semantically different sentences are farther apart in the embedding space.

9. The artificial intelligence-based internal audit method for electronic blind bidding documents in the water conservancy industry according to claim 8 is characterized in that: Using the SBERT model to perform anomaly detection on the data to be tested specifically includes the following steps: Data collection: Collect annotated sentence pairs or sentence groups. This data includes similarity labels for the sentence pairs, where 0 indicates dissimilarity and 1 indicates similarity, or anchor-positive-negative sentence groups for triplet networks. Data preprocessing: Segment the collected text data, add tags [CLS] and [SEP], and convert sentence pairs or sentence groups into a format suitable for input into the SBERT model; Data anomaly detection: The bidding documents are input into the SBERT model, which outputs semantic vectors. The semantic similarity of the semantic vectors of different bidding documents is calculated. Based on the semantic similarity, it is determined whether the technical solutions are identical or whether the quotations fluctuate abnormally.

10. The artificial intelligence-based internal audit method for electronic blind bidding documents in the water conservancy industry according to claim 7 is characterized in that: Data preprocessing includes concatenating two sentences, separating them with [SEP] and adding a [CLS] tag at the beginning.

11. The artificial intelligence-based internal audit method for electronic blind bidding documents in the water conservancy industry according to claim 1 is characterized in that: Step S4 specifically includes the following steps: During the internal review of bidding documents, the decision-making process of the artificial intelligence algorithm is visualized, and an internal audit decision tree is constructed based on the key indicators and scoring weights preset for the water conservancy project. Internal audit experts make decisions based on the decision tree and generate internal audit reports. Build a database of water conservancy industry standards and regulations, automatically associate internal audit reports with the standards and regulations in the database, provide feedback on internal audit scores, and explain the specific reasons for deductions.

12. The artificial intelligence-based internal audit method for electronic blind bidding documents in the water conservancy industry according to claim 11 is characterized in that: An internal audit decision tree is constructed based on the preset key indicators and scoring weights of the water conservancy project. Internal audit experts make decisions based on the decision tree and generate an internal audit report that specifically includes: Based on the internal review requirements for electronic blind bidding documents in the water conservancy industry, we first determined the scoring factors and weights, and collected and pre-processed historical bidding document data; scoring factors included technical solutions and price rationality; Use ID3 algorithm to train decision tree model; Combining expert experience with the Q-learning algorithm to dynamically optimize weights, the Q-learning algorithm is used to achieve weight optimization, text parsing, and anomaly detection, ultimately generating a visual display decision tree. Internal audit experts make decisions based on the decision tree; Build a database of water conservancy industry standards and regulations, automatically associate internal audit reports with the standards and regulations in the database, provide feedback on internal audit scores, and explain the specific reasons for deductions.

13. An artificial intelligence-based internal audit system for electronic blind bidding documents in the water conservancy industry, characterized by: It includes a processor, which can execute a computer program, and the computer program can implement the steps of the enterprise internal review method of electronic blind bidding documents in the water conservancy industry based on artificial intelligence as described in any one of claims 1 to 12.

Images