Search

CN-121981810-A - Method, system and device for detecting back washing money based on graph convolution and RPA automatic execution

CN121981810ACN 121981810 ACN121981810 ACN 121981810ACN-121981810-A

Abstract

The invention relates to the technical field of financial risk management and discloses a money back-washing detection method, a system and a device based on graph convolution and RPA automatic execution, wherein the technical scheme is characterized in that multidimensional data are collected by an RPA robot and processed into a heterogeneous graph structure, and node embedded vectors are generated through unified vectorization of graph embedding technology; inputting node embedded vectors into a trained graph convolutional network, obtaining a final node vector through multi-layer neighbor feature aggregation, mapping the final node vector through a full-connection layer to obtain a first AML risk score, comparing and detecting a detection target in real time through a rule engine to obtain a second AML risk score, carrying out weighted fusion on the two scores to obtain a final risk score, mapping the risk score through a preset strategy matrix, obtaining a risk grade and a processing strategy corresponding to the risk score according to a preset risk grade rule, and executing the processing strategy through an RPA.

Inventors

  • HUANG CHEN
  • GU LILI

Assignees

  • 江苏苏商银行股份有限公司

Dates

Publication Date
20260505
Application Date
20251205

Claims (9)

  1. 1. A back-flushing detection method based on graph convolution and RPA automatic execution is characterized by comprising the following steps: S1, acquiring multidimensional data through an RPA robot, abstracting the multidimensional data into a heterogeneous graph structure, uniformly vectorizing the characteristics of the multidimensional data through a graph embedding technology, and generating a node embedding vector; S2, inputting node embedding vectors into a trained graph convolutional network, obtaining final node vectors through multi-layer neighbor feature aggregation, and mapping the final node vectors through a full-connection layer to obtain a first AML risk score; meanwhile, comparing and detecting the detection targets in real time through a rule engine to obtain a second AML risk score; Weighting and fusing the first AML risk score and the second AML risk score to obtain a final risk score; S3, mapping the risk scores through a preset strategy matrix, and obtaining risk grades and processing strategies corresponding to the risk scores according to preset risk grade rules; s4, executing the processing strategy through the RPA.
  2. 2. The method for detecting back-washing money based on graph convolution and RPA automation execution of claim 1, wherein the multidimensional data comprises user information, transaction data and account characteristics.
  3. 3. The method for back-flushing money detection based on graph convolution and RPA automation execution of claim 2, wherein nodes of the heterogeneous graph structure comprise users, accounts and transactions, and edges represent relationships between the nodes.
  4. 4. The method for detecting money laundering based on graph convolution and RPA automatic execution according to claim 3, wherein the RPA robot executes the acquisition task according to a preset period and is provided with a fault-tolerant mechanism, when one acquisition fails, the retry is automatically triggered, the preset number of retries is the maximum, if the retry still fails after reaching the preset number of retries, the current acquisition task is automatically marked as 'failure', the current task is automatically skipped, and the next acquisition task is continuously executed.
  5. 5. The method for detecting backwash money based on graph convolution and RPA automation execution of claim 4, wherein after the data acquisition task fails, alarm information is sent to the operation and maintenance monitoring platform, and a work order to be handled is created to notify manual intervention.
  6. 6. The method for back-flushing detection based on graph convolution and RPA automation of claim 5, wherein the rule engine's detection targets include high risk behavior, abnormal transaction patterns, and associated account links.
  7. 7. The method for detecting backwash money based on graph convolution and RPA automation execution of claim 6, wherein weights of the rule engine and the graph convolution network are dynamically adjusted according to performance of the graph convolution network.
  8. 8. A back-flushing detection system based on graph convolution and RPA automatic execution is characterized by comprising the following components: The data characteristic module is used for collecting multidimensional data through the RPA robot, abstracting the multidimensional data into a heterogeneous graph structure, uniformly vectorizing the characteristics of the multidimensional data through a graph embedding technology, and generating a node embedding vector; The risk identification module is used for inputting the node embedded vector into a trained graph convolutional network, obtaining a final node vector through multi-layer neighbor feature aggregation, mapping the final node vector through a full-connection layer to obtain a first AML risk score; the decision-making arrangement module is used for mapping the risk scores through a preset strategy matrix and obtaining risk grades and processing strategies corresponding to the risk scores according to preset risk grade rules; and the policy execution module is used for executing the processing policy through the RPA.
  9. 9. A money laundering detection device based on graph convolution and RPA automatic execution is characterized by comprising a processor and a memory, wherein the memory stores a computer program executable by the processor, and the processor realizes the method of any one of claims 1-7 when executing the computer program.

Description

Method, system and device for detecting back washing money based on graph convolution and RPA automatic execution Technical Field The invention relates to the technical field of financial risk management, in particular to a money laundering detection method, a money laundering detection system and a money laundering detection device based on graph convolution and RPA automatic execution. Background Financial back money laundering (AML) is a key line of defense in maintaining national financial security and in fighting criminal activities. With the expansion of global financial transaction scale and the increasing complexity of cross-border payment, money laundering means are continuously updated, the concealment is stronger, the technology is higher, and the traditional detection means are difficult to deal with. Therefore, the high-efficiency and intelligent money-back system is constructed, which is not only a hard requirement for supervision compliance, but also an important task for financial institutions to prevent risks and fulfill social responsibility. The current bank money laundering monitoring system mainly relies on manual rule configuration, and has the following problems: 1. The rule engine is stiff, the adaptability is poor, the novel money laundering method is difficult to identify depending on a static threshold (such as large transaction amount and transaction frequency), such as complex modes of split transaction, multi-layer funds nesting and the like, and the identification rate of the complex money laundering mode is less than 40% in the traditional rule. 2. The data integration capability is insufficient, customer information, transaction records and cross-border data are scattered in a plurality of systems, a unified data management framework is lacking, and cross-system association analysis is difficult. About 60% of suspicious transaction analysis in financial institutions cannot complete deep tracing of data islanding problems, taking an average of over 30 minutes across system queries. 3. The manual verification is inefficient in that compliance personnel need to manually log in 5-8 systems to verify suspicious transactions, the average single audit takes about 45 minutes, report generation is completely dependent on manual operation, and the error rate is up to 15%. This not only results in limited daily throughput, but also fails to meet the real-time reporting requirements of regulatory authorities on Suspicious Transaction Reports (STRs). In contrast, existing mainstream anti-money laundering systems (e.g., oracle Mantas, SAS AML) have certain machine learning capabilities, but still use rule engines as cores, and face the following limitations: 1. cross-system synergy is weak, most schemes lack deep integration with Robot Process Automation (RPA) and real-time data pipelines, and manual intervention data extraction and integration are still needed. 2. The cost effectiveness is unbalanced, the system deployment and maintenance cost is high, but the efficiency improvement is limited. It is therefore necessary to design a new detection scheme to solve the current problems. Disclosure of Invention The invention provides a money laundering detection method, a money laundering detection system and a money laundering detection device based on graph convolution and RPA automatic execution. The invention is realized by the following technical scheme that the money laundering detection method based on graph convolution and RPA automatic execution comprises the following steps: S1, acquiring multidimensional data through an RPA robot, abstracting the multidimensional data into a heterogeneous graph structure, uniformly vectorizing the characteristics of the multidimensional data through a graph embedding technology, and generating a node embedding vector; S2, inputting node embedding vectors into a trained graph convolutional network, obtaining final node vectors through multi-layer neighbor feature aggregation, and mapping the final node vectors through a full-connection layer to obtain a first AML risk score; meanwhile, comparing and detecting the detection targets in real time through a rule engine to obtain a second AML risk score; Weighting and fusing the first AML risk score and the second AML risk score to obtain a final risk score; S3, mapping the risk scores through a preset strategy matrix, and obtaining risk grades and processing strategies corresponding to the risk scores according to preset risk grade rules; s4, executing the processing strategy through the RPA. As a preferred technical scheme of the invention, the multidimensional data comprises user information, transaction data and account characteristics. As a preferred technical scheme of the invention, the nodes of the heterogeneous graph structure comprise users, accounts and transactions, and edges represent the relationship among the nodes. As a preferable technical scheme of the invention, the RPA robot executes