Purchase return risk early warning and management method and system based on multi-mode intelligent auditing, electronic equipment and computer readable storage medium

Abstract

The invention belongs to the crossing field of computer technology and business management, and discloses a purchase return risk early warning and management method and system based on multi-mode intelligent auditing, electronic equipment and a computer readable storage medium. The method comprises the following steps: segmenting a PDF invoice image through an improved Apache PDFBox analyzer; a CLIP model and an OCR technology are combined, a cross-modal attention layer is used for correcting an OCR text, a multi-modal consistency difference value is generated, and invoice authenticity is judged; when a false invoice is detected, generating a standardized risk event, and extracting supplier, purchaser and transaction edge data from the graph database to construct a time sequence sub-graph; and identifying an abnormal mode based on the graph neural network, verifying the risk by using a time sequence risk scoring model, and executing a treatment strategy. According to the method, through multi-modal fusion, dynamic graph analysis and time sequence modeling, the problems of low false invoice detection precision, long abnormal fund tracing time consumption, risk prediction lag and the like in traditional auditing are solved.

Description

Technical Field

[0001] The present invention belongs to the cross - field of computer technology and business management, and particularly relates to a procurement kickback risk early warning and management method and system, an electronic device, and a computer - readable storage medium based on multimodal intelligent auditing. Background Art

[0002] With the expansion of enterprise procurement scale, hidden benefit transfer behaviors such as procurement kickbacks show a trend of being technicalized and concealed, and traditional auditing means are difficult to cope with new fraud patterns. The existing technologies mainly have the following limitations: 1. Deficiencies of manual auditing and rule engines Traditional methods rely on manual review of paper contracts and invoices, supplemented by automated tools based on fixed rules (such as amount thresholds, supplier blacklists). For example, an enterprise resource planning (ERP) system screens abnormal transactions (such as contract amount over - limit) through preset rules, but there are significant defects: (1) High missed - detection rate: The missed - detection rate of manual review for tampered contracts and false invoices exceeds 46%; (2) Rigid rules: Static rules cannot adapt to complex tampering means (such as adversarial sample attacks), and the update cycle is as long as several weeks.

[0003] 2. Limitations of single - modality OCR technology Existing OCR technologies (such as Tesseract, Adobe PDF Extract) only extract text information and lack the ability to verify image semantics; 3. Defects of static transaction network analysis Transaction analysis based on graph databases (such as Neo4j) mostly uses static rules, such as marking purchasers with more than 10 monthly transactions or suppliers with annual transactions exceeding the budget.

[0004] In view of the above problems, there is an urgent need for a technical solution that integrates multimodal data analysis, dynamic graph calculation, and intelligent decision - making linkage to achieve accurate early warning and closed - loop management of procurement kickback risks. The present invention breaks through the limitations of traditional methods through innovative designs of CLIP - OCR multimodal correction, time - series sub - graph construction, and dynamic scoring models, significantly improving the auditing efficiency and accuracy. Summary of the Invention

[0005] The purpose of the present invention is to propose a procurement kickback risk early warning and management method and system, an electronic device, and a computer - readable storage medium based on multimodal intelligent auditing to solve problems such as low detection accuracy of false invoices, long time consumption for tracing abnormal funds, and lagging risk prediction in traditional auditing.

[0006] The present invention provides a procurement kickback risk early warning and management method based on multimodal intelligent auditing, and the method includes the following steps: Step 1: Input the PDF invoice image to be detected, segment it through an improved Apache PDFBox parser, and separate the directly editable text blocks and image blocks; apply OCR technology to the image blocks to generate an image-embedded text set {T} and the corresponding original image block set {I}; Step 2: Input the image-embedded text set {T} into the text encoder in the CLIP model to generate a text feature vector CLIP text (T); input the original image block set {I} into the image encoder in the CLIP model to generate an image feature vector CLIP image (I); use the cross-modal attention layer in the CLIP model to calculate the correlation weights between the text and image features , dynamically correct the OCR extraction results; based on the correlation weights , calculate the corrected text set ; Step 3: Calculate the difference value between the corrected text set and the image block set {I}, and detect the authenticity of the invoice by comparing it with a preset threshold ; ; Step 4: When it is determined that the invoice is a false invoice, generate a standardized risk event; according to the supplier_id and purchaser_id fields in the risk event, extract the supplier node, purchaser node, and transaction edge data information from the graph database; use the above data information to construct a temporal subgraph; Step 5: Based on the constructed temporal subgraph above, use the graph neural network GNN for preliminary detection to generate a list of high-risk nodes and anomaly pattern labels; then verify the high-risk nodes through a temporal risk scoring model and perform corresponding disposal operations based on the verification results.

[0007] A procurement kickback risk warning and management method based on multimodal intelligent auditing as described above, The image-embedded text set {T} in Step 1 refers to the set composed of the text recognized from the image blocks through OCR technology.

[0008] A procurement kickback risk warning and management method based on multimodal intelligent auditing as described above, The correlation weight in Step 2 has the following calculation formula: , where d is the feature dimension; is the transpose operation; The corrected text set is calculated based on the correlation weight ​ , and the calculation formula is: , where T is the set of text embedded in the image; is the text decoder; is the image feature vector.

[0009] A procurement kickback risk warning and management method based on multimodal intelligent auditing as described above The difference value between the corrected text set and the image block set in step 3 The specific calculation formula is: , where, is the L2 norm; The threshold is default set to 0.12, and it can be adaptively adjusted according to the credit score of the supplier. The specific adjustment formula is: .

[0010] A procurement kickback risk warning and management method based on multimodal intelligent auditing as described above Step 4 is specifically: Step 4.1: When , it is determined as a false invoice, and the generation of a standardized event is triggered; Step 4.2: Execute the generation of standardized risk events. The standardized risk events include the following core fields: event_id, supplier_id, purchaser_id, invoice_id, timestamp, risk_level; Step 4.3: Push it to the graph neural network GNN in real time through the message queue; Step 4.4: According to the supplier_id and purchaser_id in the risk event, use the nGQL query language to extract the supplier node, purchaser node, and transaction edge data information from the graph database, and construct a time series subgraph based on the above data information.

[0011] A procurement kickback risk warning and management method based on multimodal intelligent auditing as described above The preliminary detection using the graph neural network GNN in step 5 includes using the time series graph attention network T-GAT for high-frequency trading detection and procurement kickback identification; among them, high-frequency trading detection means that if the trading edge weight between a certain supplier and purchaser is significantly higher than the historical average in a short time, it is marked as "high-frequency trading risk"; and when detecting the closed-loop path, calculate the total path weight, and if it exceeds the threshold, it is marked as "procurement kickback".

[0012] A procurement kickback risk early warning and management method based on multimodal intelligent auditing as described above, When verifying high-risk nodes through the time-series risk scoring model in step 5, the specific formula for time-series risk scoring is:

[0013] where n is the number of transaction records; The default value is 0.1, representing the risk scoring time decay coefficient; is the transaction time interval, with the unit of days.

[0014] The present invention also discloses a procurement kickback risk early warning and management system based on multimodal intelligent auditing, and the system includes: Graphic and text segmentation and extraction module: used for the PDF invoice image to be detected, segmented by an improved Apache PDFBox parser, separating directly editable text blocks and image blocks; applying OCR technology to the image blocks to generate an image embedded text set {T} and the corresponding original image block set {I}; CLIP module: used to encode the image embedded text set {T} by the text encoder in the CLIP module to generate a text feature vector CLIP text (T); encoding the original image block set {I} by the image encoder in the CLIP module to generate an image feature vector CLIP image (I); using the cross-modal attention layer in the CLIP model to calculate the correlation weight between text and image features to dynamically correct the OCR extraction result; based on the correlation weight to calculate the corrected text set ; Graphic and text consistency determination module: used to calculate the difference value between the corrected text set and the image block set {I}, and detect the authenticity of the invoice by comparing with a preset threshold ; Time-series subgraph construction module: used to generate a standardized risk event when determining that the invoice is a false invoice; extract supplier nodes, purchaser nodes, and transaction edge data information from the graph database according to the supplier_id and purchaser_id fields in the risk event; construct a time-series subgraph using the above data information; Risk grading and disposal module: used to perform preliminary detection using a graph neural network GNN based on the constructed time-series subgraph above to generate a list of high-risk nodes and abnormal pattern labels; then verify the high-risk nodes through a time-series risk scoring model, and perform corresponding disposal operations based on the verification results.

[0015] The present invention also discloses an electronic device, including one or more processors; a storage device for storing one or more computer programs, which, when executed by the one or more processors, enable the electronic device to implement the above-mentioned procurement kickback risk warning and management method based on multimodal intelligent auditing.

[0016] The present invention also discloses a computer-readable storage medium, characterized in that a computer program is stored thereon, which, when executed by a processor of an electronic device, enables the electronic device to execute the above-mentioned procurement kickback risk warning and management method based on multimodal intelligent auditing. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings without creative efforts based on these drawings.

[0018] Figure 1 It is a flowchart of the steps of the procurement kickback risk warning and management method based on multimodal intelligent auditing. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] The present invention will be described in detail below through embodiments.

[0020] Please refer to Figure 1 , which shows a flowchart of the steps of the procurement kickback risk warning and management method based on multimodal intelligent auditing provided by an embodiment of the present invention. As Figure 1 shown, the procurement kickback risk warning and management method based on multimodal intelligent auditing provided by the embodiments of the present application includes the following steps 1 to 5: Step 1: Input the PDF invoice image to be detected, and perform segmentation through an improved Apache PDFBox parser to separate the directly editable text block and the image block; apply OCR technology to the image block to generate an image embedded text set {T} and the corresponding original image block set {I}; It should be noted that the improved design of the Apache PDFBox parser lies in: (1) Splitting the PDF file into independent tasks by page and improving the parsing speed to at least 300 pages per minute through multi-threaded concurrent processing; (2) Preloading the common font library into memory when the parser starts, reducing the rendering time by more than 20%; (3) Adopting an image-text alignment algorithm, based on the PDFTextStripperByArea class of Apache PDFBox, improving the region segmentation logic, and accurately extracting directly editable text blocks T and their corresponding image blocks I through adaptive grid division (the grid density is dynamically adjusted, and the error tolerance is ≤ 5px). Among them, (1) Directly editable text blocks: Directly extracted from the PDF text layer by PDFBox, not participating in multi-modal verification, and directly used for business logic (such as contract amount comparison); (2) Image-embedded text set {T}: Only refers to the text recognized from the image block through OCR; (3) Image block set {I}: Retains the original image data, which refers to the set of image blocks obtained by splitting the PDF invoice image through the improved Apache PDFBox parser for the CLIP image encoder to extract visual features.

[0021] Step 2: Input the image-embedded text set {T} into the text encoder in the CLIP model to generate the text feature vector CLIP text (T); Input the original image block set {I} into the image encoder in the CLIP model to generate the image feature vector CLIP image (I); Use the cross-modal attention layer in the CLIP model to calculate the correlation weight between the text and image features , dynamically correct the OCR extraction result; Based on the correlation weight , calculate the corrected text set ; It should be noted that: Use the cross-modal attention layer in the CLIP model to calculate the correlation weight between the text and image features , and its calculation formula is: , where d is the feature dimension; is the transpose operation; In the above formula, the core is to assign weights by calculating the similarity between features; In particular, the dot product operation of the transpose of the text feature vector and the image feature vector is used to measure the semantic correlation between the text feature and the image feature. The larger the dot product value, the higher the consistency between the two; The scaling factor is to avoid the dot product result value being too large when the dimension d is large, by dividing by Control the gradient stability; the transpose operation is to achieve dimensional matching between vectors; and the Softmax normalization is to convert the dot product result into a probability distribution to ensure the weights , thereby reasonably allocating the contribution ratio of text and image features. The physical meaning of this weight lies in the credibility weight of text features. For example, if the image features are highly consistent with the text features (such as clear invoice printing and accurate OCR extraction), then , mainly based on the OCR result; if the image features are significantly different from the text features (such as the text being tampered with but the image background remaining intact), then , mainly based on the image decoding result.

[0022] Based on the correlation weight , the corrected text set is calculated, and its calculation formula is: , where T is the text set embedded in the image; is the text decoder; is the image feature vector.

[0023] In this formula, Z-score normalization can be used to make its mean 0 and variance 1, eliminating the dimension difference; and using α and (1−α) as complementary weights to ensure that the weighted features are still within the range of the original feature distribution.

[0024] And is not the inverse model of CLIP natively, but a fully connected network independently trained with the following structure: Input: CLIP image features (dimension d = 512); Hidden layer: 2-layer MLP (dimension 512→256→512), with the activation function being GELU; Output: Decoded text features, with the same dimension as the OCR features.

[0025] It can be seen that in this step 2, by using the multi-modal fusion of the CLIP model and OCR technology and through two-way feature interaction, the isolated application mode of traditional OCR and CLIP is broken through, achieving the following technical effects: using the semantic understanding ability of CLIP to correct the OCR output and resist adversarial attacks; achieving in-depth fusion of text and image features through cross-modal attention to accurately identify hidden fraud such as "text tampering but image remaining intact".

[0026] Step 3: Calculate the difference value between the corrected text set and the image block set {I}, and detect the authenticity of the invoice by comparing it with the preset threshold ; It should be noted that: the calculation formula for the difference value between the corrected text set and the image block set is: , where, is the L2 norm (Euclidean distance), and the smaller its value, the higher the semantic consistency between the text and the image; while the threshold Δ is default preset to 0.12, which can be adaptively adjusted dynamically according to the supplier's credit score. The specific adjustment formula is:

[0027] By dynamically adjusting the value of Δ, suppliers with high credit can relax the threshold to reduce false alarms, while for suppliers with low credit, tighten the threshold to improve the detection rate. For example: when the credit score of supplier A is 70 points, the threshold is adjusted to Δ = 0.12 1.2 = 0.144, to avoid misjudging high-credit suppliers due to some minor formatting issues; when the credit score of supplier B is 30 points, the threshold is adjusted to Δ = 0.12 0.8 = 0.096, the threshold is tightened to improve the detection rate; It can be seen that in this step 3, By quantifying the semantic differences between text and image features, it provides a multimodal fusion detection method for procurement kickback audits, achieving the following technical effects: making up for the single-modal defects of traditional OCR and realizing a more comprehensive anti-tampering ability; the dynamic threshold is linked with the credit score to balance risk control and business efficiency.

[0028] Step 4: When it is determined that the invoice is a false invoice, generate a standardized risk event; according to the supplier_id and purchaser_id fields in the risk event, extract the supplier node, purchaser node, and transaction edge data information from the graph database; use the above data information to construct a temporal subgraph; It should be noted that: the graph database, as a persistent storage system, is responsible for storing the original data of all static nodes (suppliers, purchasers) and edges (transaction records), and provides basic query functions; the temporal subgraph is dynamically generated, a temporary data structure for specific analysis objectives (such as high-risk events), constructed based on part of the data in the graph database for real-time or near-real-time analysis.

[0029] Standardized risk event generation and data extraction: False invoice determination and event generation: 1. Trigger condition: When the difference value > Δ, it is determined to be a false invoice.

[0030] 2. Event structure: Generate a standardized risk event, including the following core fields: { "event_id": "RISK_001", "supplier_id": "S_123", "purchaser_id": "P_456", "invoice_id": "INV_789", "timestamp": "2023-10-05T14:30:00Z", "risk_level": "high" } 3. Event Transmission: Push to the Graph Neural Network (GNN) module in real time through the message queue.

[0031] Graph Database Data Extraction: 1. Query Target: According to the supplier_id and purchaser_id in the event, extract the following data: Supplier Node: Include attributes such as supplier ID, credit score, number of historical risk events, registration information, etc.; Purchaser Node: Include attributes such as purchaser ID, approval authority, job rank, number of associated contracts, etc.; Bank Account Node (Optional): If it is necessary to analyze the fund flow path, it can include attributes such as account ID, bank of deposit, balance change, etc.; Transaction Edge: All transaction records within a limited time window (such as the last 3 months), including amount, timestamp, contract number, etc.; Fund Flow Edge (Optional): If the bank account node is included, record the fund flow direction (such as purchaser → account → supplier).

[0032] 2. Query Language: Use the graph database query language (such as nGQL of Nebula Graph); 3. Output Result: Form a data set containing supplier nodes, purchaser nodes, and transaction edges.

[0033] Temporal Subgraph Construction and Feature Enhancement: 1. Node and Edge Screening: Only retain the supplier, purchaser nodes, and transaction edges related to the risk event; 2. Time Window Constraint: The timestamp of the transaction edge needs to be within the preset window (such as the last 3 months); 3. Dynamic Weight Calculation: Attach a time decay weight to each transaction edge:

[0034] where t is the transaction time interval, that is, the time difference between the current time and the transaction occurrence time, in days; is the graph network decay coefficient (default 0.1), which controls the weight decay speed of historical transactions, and the value range is [0.05, 0.2]; i represents the transaction initiator node, such as the supplier; j represents the transaction recipient node, such as the purchaser; Specifically, for time-decaying weights, as t increases, the weights decrease exponentially, making it conform to the business logic that "recent transactions are more important than distant transactions"; and by adjusting , the rate of weight decay can be flexibly controlled. For example: when =0.01 , the decay is slower, retaining the influence of longer-term historical transactions; when =0.2 , the decay is faster, only focusing on recent transactions. Of course, to eliminate the problem of weights being too small or too large, weight normalization or setting a lower limit for weights in combination with business rules can be adopted to ensure that distant transactions still retain a certain influence, and these can be achieved through conventional settings.

[0035] Feature Fusion and Enhancement 1. Node feature enhancement: Supplier node: Add the proportion of recent false invoices (e.g., 2 high-risk events detected in the last month); Buyer node: Add the total approved amount and the number of abnormal transaction marks.

[0036] 2. Edge feature extension: Mark labels such as contract type (normal purchase / supplementary agreement), transaction status (completed / awaiting payment), etc.

[0037] It can be seen that in this step 4, the core of constructing the time-series subgraph depends on the supplier nodes, buyer nodes, and transaction edges extracted from the graph database. However, through dynamic weight calculation, time window screening, and multi-feature fusion, a fine-grained modeling of transaction behavior is achieved. This method not only ensures the transparency of data sources but also significantly improves the accuracy and efficiency of risk detection by quantifying the time decay effect and abnormal patterns.

[0038] Step 5: Based on the time-series subgraph constructed above, use the graph neural network GNN for preliminary detection to generate a list of high-risk nodes and abnormal pattern labels; then verify the high-risk nodes through a time-series risk scoring model and perform corresponding disposal operations based on the verified risk levels.

[0039] It should be noted that: First, when designing the GNN model, the time-series graph attention network (T-GAT) is adopted, and the input features include: Node features: Credit score, number of risk events, approval authority level; Edge features: Transaction amount, time decay weight, contract type label.

[0040] Abnormal pattern detection includes: High-frequency transaction detection and procurement kickback identification, specifically: If the weight of the transaction edge between a certain supplier and a buyer is significantly higher than the historical average in a short period of time, it is marked as "high-frequency transaction risk"; Detect the closed-loop path (such as Supplier → Purchaser → Associated Account → Supplier), calculate the total path weight, and mark it as "Procurement Rebate" if it exceeds the threshold. The specific process is as follows: Since the weight of each edge is jointly determined by the transaction amount and the time decay coefficient , an example of calculating the total path weight is as follows: Assume that the current time is October 8, 2023, and the information of the three edges of a closed-loop path is as follows: 1. Transaction edge: amount 50,000 yuan, time difference 7 days, corresponding w1 ≈ 24850; 2. Transfer edge (Purchaser → Account): amount 50,000 yuan, time difference 7 days, corresponding w2 ≈ 24850; 3. Transfer edge (Account → Supplier): amount 50,000 yuan, time difference 7 days, corresponding w3 ≈ 24850; Then the total path weight: Total Weight = w1 + w2 + w3 = 74,550; Finally, determine whether there is a suspected procurement rebate by comparing the total path weight with the preset threshold.

[0041] The preliminary detection output result is: generate a list of high-risk nodes and abnormal pattern labels (such as "High-frequency Transaction", "Procurement Rebate"); an example is: { "supplier_id": "S_123", "purchaser_id": "P_456", "risk_type": "high_frequency_transaction", "confidence": 0.92 } Then, the specific operation of using the time-series risk scoring model for verification is as follows: 1. Input data: high-risk nodes output by GNN and their associated transaction records; 2. Calculate the corresponding time-series risk score, and the specific formula is:

[0042] where n is the number of transaction records; The default value is 0.1, representing the risk score time decay coefficient, and the value range is [0.05, 0.3]; is the transaction time interval, in days.

[0043] For example: Supplier S_123 has 5 transactions in the past 3 months, and the ratios of transaction amount to contract amount are 1.2, 0.8, 1.5, 1.1, and 1.3 respectively, and the time intervals are 7, 15, 30, 60, and 90 days: E risk = 1.2×e -0.7 + 0.8×e -1.5 + 1.5×e -3 + 1.1×e -6 + 1.3×e -9 ≈ 8.6 3. Compare the obtained time series analysis score with the preset threshold, classify the risk into levels, and execute corresponding dispositions; Assume the preset threshold is 5.0. If E risk > ≥ 5.0, confirm that the GNN detection result is valid; (1) Hierarchical response: When E risk ≥ 8.0, it is a high risk: immediately freeze the contract and initiate an internal investigation; When 5.0 E risk < < 8.0, it is a medium risk: restrict the bidding authority and require the submission of explanatory materials; When E risk <5 < 5.0, it is a low risk: mark as an observation object and conduct regular reviews.

[0044] (2) Automatically execute the corresponding level of disposition measures, including but not limited to: contract freezing, authority downgrading, account monitoring, credit score update, etc.

[0045] Therefore, in step 5 of the present invention, preliminary detection is first performed through a graph neural network and then verified through a time series risk scoring model. The former captures implicit associations from topological relationships, and the latter quantifies time-sensitive risks. The two-dimensional cross reduces the false alarm rate.

[0046] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention without departing from the principles and purposes of the present invention.

Claims

1. A procurement kickback risk early warning and management method based on multimodal intelligent auditing, characterized in that, The method includes the following steps: Step 1: Input the PDF invoice image to be detected, and segment it through an improved Apache PDFBox parser to separate directly editable text blocks and image blocks; apply OCR technology to the image blocks to generate an image embedded text set {T} and the corresponding original image block set {I}; Step 2: Input the set of image embedded text {T} into the text encoder in the CLIP model to generate the text feature vector CLIP text (T); Input the set of original image blocks {I} into the image encoder in the CLIP model to generate the image feature vector CLIP image (I); Use the cross-modal attention layer in the CLIP model to calculate the correlation weights between the text and image features , dynamically correct the OCR extraction results; Based on the correlation weights , calculate the corrected text set ; Step 3: Calculate the corrected text set The difference value from the image block set {I} , and detect the authenticity of the invoice by comparing with a preset threshold ​ Step 4: When it is determined that the invoice is a false invoice, generate a standardized risk event; according to the supplier_id and purchaser_id fields in the risk event, extract the supplier node, purchaser node, and transaction edge data information from the graph database; use the above data information to construct a temporal subgraph; Step 5: Based on the constructed temporal subgraph above, use the graph neural network GNN for preliminary detection to generate a list of high-risk nodes and anomaly pattern labels; then verify the high-risk nodes through a temporal risk scoring model, and perform corresponding disposal operations based on the verification results.

2. The procurement kickback risk early warning and management method based on multimodal intelligent auditing according to claim 1, wherein, The image embedded text set {T} in Step 1 refers to the set composed of the text recognized from the image blocks through OCR technology.

3. The procurement kickback risk early warning and management method based on multimodal intelligent auditing according to claim 2, wherein The associated weight in the said step 2 The calculation formula is as follows: ， where d is the feature dimension; is the transpose operation; The above-mentioned based on the association weight , the corrected text set is calculated , and the calculation formula is: , where T is the set of text embedded in the image; is the text decoder; is the image feature vector.

4. The procurement kickback risk early warning and management method based on multimodal intelligent auditing according to claim 3, wherein The difference value between the corrected text set and the image block set in the said step 3 The specific calculation formula is as follows: Among them, is the L2 norm; The threshold value The default value is set to 0.12, which can be adaptively adjusted according to the supplier's credit score. The specific adjustment formula is as follows: 。 5. The procurement kickback risk early warning and management method based on multimodal intelligent auditing according to claim 4, wherein, The specific content of Step 4 is as follows: Step 4.1: When > Δ, it is determined as a false invoice, triggering the generation of a standardization event; Step 4.2: Execute the generation of a standardized risk event, and the standardized risk event includes the following core fields: event_id, supplier_id, purchaser_id, invoice_id, timestamp, risk_level; Step 4.3: Push it to the graph neural network GNN in real time through a message queue; Step 4.4: According to the supplier_id and purchaser_id in the risk event, use the nGQL query language to extract the supplier node, purchaser node, and transaction edge data information from the graph database, and construct a temporal subgraph based on the above data information.

6. The procurement kickback risk early warning and management method based on multimodal intelligent auditing according to claim 5, characterized in that, The preliminary detection using the graph neural network GNN in Step 5 includes using the temporal graph attention network T-GAT for high-frequency transaction detection and procurement kickback identification; among them, high-frequency transaction detection means that if the weight of the transaction edge between a certain supplier and purchaser is significantly higher than the historical average in a short period of time, it is marked as "high-frequency transaction risk"; when detecting a closed-loop path, calculate the total path weight, and if it exceeds the threshold, it is marked as "procurement kickback".

7. The procurement kickback risk early warning and management method based on multimodal intelligent auditing according to claim 6, characterized in that When verifying the high-risk nodes through a temporal risk scoring model in Step 5, the specific formula for temporal risk scoring is: Where n is the number of transaction records; The default value is 0.1, representing the risk score time decay coefficient; is the transaction time interval, in days.

8. A procurement kickback risk early warning and management system based on multimodal intelligent auditing, characterized in that, The system includes: A graphic and text segmentation and extraction module: used to segment the PDF invoice image to be detected through an improved Apache PDFBox parser to separate directly editable text blocks and image blocks; apply OCR technology to the image blocks to generate an image embedded text set {T} and the corresponding original image block set {I}; CLIP module: used to encode the image embedded text set {T} using the text encoder in this CLIP module to generate the text feature vector CLIP text (T); use the image encoder in this CLIP module to encode the original image block set {I} to generate the image feature vector CLIP image (I); use the cross-modal attention layer in the CLIP model to calculate the correlation weights between the text and image features , dynamically correct the OCR extraction results; based on the correlation weights , calculate the corrected text set ; Text-image consistency determination module: used to calculate the corrected text set and the difference value from the image block set {I} , and detect the authenticity of the invoice by comparing with a preset threshold ​ Timing sub-graph construction module: used to generate standardized risk events when it is determined that the invoice is a false invoice; extract supplier nodes, purchaser nodes, and transaction edge data information from the graph database according to the supplier_id and purchaser_id fields in the risk event; construct a timing sub-graph using the above data information. Risk grading and disposal module: used to perform preliminary detection using the graph neural network GNN based on the constructed timing sub-graph above, generate a list of high-risk nodes and abnormal pattern labels; then verify the high-risk nodes through a timing risk scoring model and perform corresponding disposal operations based on the verification results.

9. An electronic device, characterized in that, It includes one or more processors; a storage device for storing one or more computer programs, which when executed by the one or more processors, enable the electronic device to implement the procurement kickback risk early warning and management method based on multi-modal intelligent auditing as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, A computer program is stored thereon, which when executed by the processor of the electronic device, causes the electronic device to execute the procurement kickback risk early warning and management method based on multi-modal intelligent auditing as described in any one of claims 1-7.

Images