AI large model-based fund dynamic routing decision system

Abstract

The application relates to the technical field of financial technology, and particularly discloses a fund dynamic routing decision system based on an AI large model. The system comprises: a historical transaction knowledge base for storing historical transaction data and evaluation labels; a business rule knowledge graph construction module for extracting and verifying fund business core rules and converting the rules into a structured constraint feature vector; a dynamic feature perception module for collecting channel parameters in real time to generate a dynamic environment feature vector; a rule and state fusion module for fusing the two through an attention mechanism to form a composite decision state; a coupling effect prediction module for outputting multi-step expected cumulative return values of each selectable action based on a deep Q network; a strategy optimization and routing decision module for generating optimal fund routing instructions by using a deep deterministic policy gradient algorithm; and a routing instruction execution and feedback module for executing the instructions and forming a closed-loop feedback to dynamically optimize rule verification standards and evaluation labels. The application realizes intelligent, dynamic and accurate decision of fund routing.

Description

Technical Field

[0001] This invention relates to the field of financial technology, and in particular to a dynamic fund routing decision system based on AI big data models. Background Technology

[0002] In the field of modern financial payment and settlement, dynamic fund routing decision-making is a core element for improving fund utilization efficiency, reducing transaction costs, and ensuring payment success rates. With the diversification of payment channels and the increasing complexity of business rules, traditional fund routing systems often rely on static rule configuration, manual experience-based decision-making, or models based on simple historical statistics. This makes it difficult to cope with multi-dimensional real-time dynamic factors such as market rate fluctuations, changes in channel status, and adjustments in regulatory policies. Existing technical solutions typically handle business rule constraints and real-time environmental data in a fragmented manner, lacking a mechanism for deeply integrating domain-specific prior knowledge with dynamic scenario awareness. This makes it difficult to achieve an optimal balance between compliance, timeliness, and cost-effectiveness in routing decisions, preventing true adaptive dynamic optimization.

[0003] The main shortcomings of existing technologies lie in their failure to construct an intelligent decision-making system capable of integrating rigid business rules with flexible environmental characteristics in real time and continuously evolving through closed-loop feedback. Specifically: First, business rules are typically treated as hard constraints and cannot effectively interact and collaboratively optimize with real-time data-driven machine learning models; second, they lack the ability to quantitatively predict the dynamic coupling effects between funding channels (such as load transfer and risk transmission), limiting the decision-making perspective to single-step optimization; third, they lack an automated feedback learning mechanism from decision results to rules and model parameters, resulting in stagnant system intelligence. Therefore, a systematic solution is urgently needed that can connect the entire chain of "rules, perception, prediction, decision-making, and feedback" to achieve dynamic and precise control of funding routing decisions.

[0004] Therefore, this invention proposes a dynamic funding routing decision system based on an AI large model. Summary of the Invention

[0005] This invention provides a dynamic fund routing decision-making system based on an AI large model. By constructing a closed-loop intelligent system of "rules-perception-fusion-prediction-decision-feedback", it overcomes the technical defects of traditional fund routing systems, such as the separation of rules and data, short-sighted decision-making, and system rigidity, and achieves the technical effect of dynamic and accurate optimization and autonomous continuous evolution of fund routing.

[0006] This invention provides a dynamic funding routing decision system based on an AI large-scale model, comprising: Historical transaction knowledge base, used to store historical transaction data of the entire funding route and corresponding route evaluation tags; The business rule knowledge graph construction module is used to extract core rules of capital business from the capital business system, verify the effectiveness of the core rules of capital business based on historical transaction data in the historical transaction knowledge base, and transform the verified core rules of capital business into structured constraint feature vectors. The dynamic feature perception module is used to collect dynamic parameters of the funding channel in real time and generate dynamic environmental feature vectors. The rule and state fusion module is used to treat structured constraints as a prior knowledge layer and dynamic environmental feature vectors as a real-time perception layer. It fuses the two layers of information through an attention mechanism to form a composite decision state that integrates business rule constraints and real-time environmental features. The coupling effect prediction module is used to build a coupling effect prediction model based on a deep Q-network, and combined with the composite decision state, outputs the multi-step expected cumulative return value of each optional fund routing action in the optional fund routing action space. The strategy optimization and routing decision module is used to optimize the strategy network parameters in real time and generate the optimal funding routing instructions based on the deep deterministic strategy gradient algorithm with the goal of maximizing the multi-step expected cumulative return. The routing instruction execution and feedback module is used to execute the optimal funding routing instruction and feed the execution result back to the historical transaction knowledge base to update the corresponding routing evaluation label; the execution result is also fed back to the business rule knowledge graph construction module to dynamically adjust the validity verification standards of the core rules of funding business.

[0007] Preferably, the business rule knowledge graph construction module includes: The rule parsing unit is used to extract the original fee rate policy, channel permission rules, and settlement time limit rules from the fund business system, and to use natural language processing technology to identify rule entities and relationships to obtain the core rules and rules relationships of fund business. The rule validity verification unit is used to calculate the compliance rate and effectiveness contribution of each core fund business rule in actual transactions based on historical transaction data in the historical transaction knowledge base, and to screen and verify valid core fund business rules based on the compliance rate and effectiveness contribution. The graph construction unit is used to construct a knowledge graph of fund business rules by taking the rule entities corresponding to the core rules of the verified fund business as nodes and the rule relationships as edges; the node attributes include fee priority, maximum channel limit, and settlement time window. The feature vector generation unit is used to generate structured constraint feature vectors based on the structured constraints of the knowledge graph of fund business rules.

[0008] Preferably, the business rule knowledge graph construction module also includes a rule vitality value calculation unit, used to perform the following steps: Real-time monitoring of the frequency with which each core rule of a financial transaction is triggered in historical decisions, serving as the rule activation frequency; The rule failure index is calculated based on the deviation between the actual routing result and the expected effect of the rule after the core rules of capital business are triggered. The rule activation frequency is converted into an activation intensity coefficient, and the rule failure index is converted into a failure rate coefficient. Through the nonlinear combination of the activation intensity coefficient and the failure rate coefficient, the real-time vitality value of each core rule of the capital business is calculated. When the structured constraint feature vector generation unit generates structured constraint feature vectors based on the knowledge graph of capital business rules, the feature dimensions of the corresponding capital business core rules are dynamically weighted according to the real-time vitality value of the capital business core rules, so that rules with high vitality values ​​occupy higher weights in the structured constraint feature vectors. When the real-time vitality value of a certain core rule of capital business continues to be lower than the evolution threshold, the rule reconstruction process is automatically triggered to generate new candidate rules based on recent successful transaction patterns. The candidate rule verification unit is used to verify the effectiveness of new candidate rules, including: conducting backtesting in similar scenarios based on the historical transaction knowledge base, calculating the simulation compliance rate and simulation effect contribution of candidate rules; and monitoring the actual execution effect of candidate rules within a preset period in a low-volume live trading environment; when the simulation compliance rate and simulation effect contribution both reach the preset verification standard, and the actual execution effect meets expectations, the candidate rule is automatically adopted and updated to the capital business rule knowledge graph.

[0009] Preferably, the dynamic feature perception module includes: The multi-source data access unit is used to access bank channel status, foreign exchange rate fluctuations, and system load monitoring data in real time through data interfaces to obtain real-time data streams. The message transmission strength analysis unit is used to calculate the state correlation strength between two adjacent channels based on real-time data streams, forming an influence transmission strength matrix between channels. The feature synthesis unit is used to fuse the channel's own dynamic indicators with the influence transmission strength matrix to generate a dynamic environmental feature vector containing the coupling relationship between channels.

[0010] Preferably, the step of the message transmission strength analysis unit calculating the state correlation strength between pairs of adjacent channels based on real-time data streams and establishing an influence transmission network between channels includes: For each channel's real-time load data, calculate the fluctuation amplitude and frequency within the sliding time window; Based on the fluctuation characteristics and state differences of adjacent channels, the message transmission tendency value between channels is calculated. All message transmission tendency values ​​are normalized to form an influence transmission strength matrix between channels, which is used for dynamic feature fusion. Based on the influence transmission strength matrix, the indirect influence weights between non-adjacent channels are calculated by matrix iterative multiplication, and the indirect influence weights are merged with the direct influence weights to update the influence transmission strength matrix.

[0011] Preferably, the rule and state fusion module includes: The prior knowledge encoding unit is used to encode the structured constraints and generate prior knowledge feature vectors. The real-time sensing coding unit is used to encode the feature vectors of the dynamic environment and generate real-time sensing feature vectors. The attention fusion unit is used to calculate the mutual influence weight between the prior knowledge feature vector and the real-time perceived feature vector based on the multi-head attention mechanism, and to perform weighted fusion of the two types of feature vectors according to the mutual influence weight to obtain the fused feature vector. The state generation unit is used to map the fused feature vectors into composite decision states that the policy network can handle.

[0012] Preferably, the deep Q-network in the coupling effect prediction module employs a hierarchical attention mechanism, specifically including: The state coding layer is used to receive the composite decision state and perform deep feature encoding to obtain the encoded features. The attention fusion layer is used to calculate the mutual influence weights between different feature dimensions in a composite decision state using an attention mechanism. The feature coupling layer is used to perform weighted fusion of encoded features based on mutual influence weights to form a deeply coupled state representation. The value output layer is used to output the multi-step expected cumulative return value of each optional funding routing action in the optional funding routing action space based on the deeply coupled state representation.

[0013] Preferably, the hierarchical attention mechanism employs an attention meta-learning controller, and the steps performed by the attention meta-learning controller as the core component of the attention fusion layer include: Based on successful transaction records in the historical transaction knowledge base, the corresponding attention weight distribution is extracted to establish an attention pattern library containing various typical decision-making logics; Encode the current complex decision state to generate a current context feature vector; The similarity between the current context feature vector and the context feature vectors of each historical attention pattern in the attention pattern library is calculated, and the historical attention pattern with the highest similarity is retrieved. The attention weight distribution of the retrieved historical attention patterns is used as the initial weight configuration for the attention fusion layer; An online learning approach is adopted, using the current complex decision-making state as the only training sample. The initial weight configuration is optimized and adjusted step by step through the gradient descent algorithm to generate an attention weight allocation scheme that is adapted to the current decision-making context. When the routing instruction execution and feedback module confirms that the current routing decision is successful, it stores the valid attention weight allocation scheme in the current decision as a new candidate pattern into the attention pattern library, and evaluates the activity of historical patterns in the attention pattern library and eliminates expired patterns to achieve continuous evolution of attention patterns.

[0014] Preferably, the strategy optimization and routing decision module adopts a dual-objective optimization architecture, specifically including: The strategy network is divided into a main network and an auxiliary network. The main network is responsible for generating routing instructions that maximize the expected cumulative return value of multiple steps, while the auxiliary network assesses the potential risk level of the current routing instructions based on risk cases in the historical trading knowledge base. The instruction filtering unit comprises a risk review subunit and an instruction refactoring subunit, wherein: The risk review subunit is used to receive routing instructions generated by the main network and potential risk levels assessed by the auxiliary network. When the potential risk level does not exceed the preset threshold, the routing instruction generated by the main network that has passed the risk review is selected as the optimal funding routing instruction; When the potential risk level exceeds a preset threshold, a risk avoidance instruction is sent to the instruction reconfiguration subunit. The instruction reconstruction subunit is used to retrieve successful routing instructions with potential risk levels below a preset threshold in similar situations from the historical transaction knowledge base based on risk avoidance instructions, and generate alternative routing instructions based on the current composite decision state as the optimal funding routing instructions. The auxiliary network assessment of potential risk levels employs a multi-dimensional risk factor fusion method, including: Based on the knowledge graph of business rules, the degree of compliance between inspection instructions and regulatory rules is checked, and the compliance risk factor is calculated. Based on the channel's historical fraud records and real-time abnormal transaction pattern detection results, calculate the security risk factor; Based on the characteristics of channel fund settlement costs and settlement cycles, a liquidity risk factor is calculated. The compliance risk factor, security risk factor, and liquidity risk factor are weighted and integrated into a comprehensive risk score, which serves as the potential risk level of the current routing instruction.

[0015] Preferably, the routing instruction execution and feedback module includes: The instruction issuing unit is used to send the optimal fund routing instruction to the corresponding payment gateway or fund clearing system; The execution monitoring unit is used to track the entire execution process of the optimal fund routing instruction, and record the actual transaction success rate, actual fund cost, and actual arrival time as the execution result; The feedback distribution unit is used to simultaneously distribute the execution results to the historical transaction knowledge base and the business rule knowledge graph construction module. The knowledge base update unit is used to update the route evaluation tags in the historical transaction knowledge base based on the execution results. The rule verification and adjustment unit is used to dynamically adjust the validity verification standards of core rules for fund business based on the statistical data of execution results within a preset time window.

[0016] The beneficial effects of this invention compared to existing technologies are as follows: By innovatively constructing a closed-loop intelligent system that is "rule-knowledge-based, perception-real-time, decision-making-intelligent, and evolution-automated," it fundamentally solves the three core defects of traditional fund routing technology: 1) the separation of rules and data, i.e., effectively integrating business rules with dynamic environmental data; 2) short-sighted decision-making vision, i.e., achieving accurate prediction of channel coupling effects and long-term chain effects of decisions; 3) system rigidity, i.e., establishing an automated feedback closed loop from execution results to rules and model parameters. This technological breakthrough has achieved the following comprehensive effects: First, by transforming business rules into evolvable structured knowledge and deeply integrating them with real-time perceived features, a composite decision-making state that comprehensively reflects the decision-making context is formed. Second, based on the coupling effect prediction and strategy optimization mechanism of deep reinforcement learning, accurate assessment of the long-term value of fund routing and generation of globally optimal instructions are achieved. Finally, through a closed-loop feedback-driven continuous evolution mechanism, the system can autonomously adapt to market changes and dynamically optimize decision-making logic. Thus, under multiple objectives such as compliance, timeliness, cost-effectiveness, and security, dynamic and accurate optimization and autonomous continuous evolution of fund routing decisions are achieved, significantly improving fund efficiency and risk management levels.

[0017] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall architecture diagram of the AI-based large-scale model-based dynamic funding routing decision system in an embodiment of the present invention; Figure 2 This is a diagram illustrating the core decision-making process and technical interaction in an embodiment of the present invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] like Figure 1 and Figure 2 As shown, this invention provides an embodiment of a dynamic funding routing decision system based on an AI large-scale model, comprising: Historical transaction knowledge base, used to store historical transaction data of the entire funding route and corresponding route evaluation tags; The business rule knowledge graph construction module is used to extract core rules of capital business from the capital business system, verify the effectiveness of the core rules of capital business based on historical transaction data in the historical transaction knowledge base, and transform the verified core rules of capital business into structured constraint feature vectors. The dynamic feature perception module is used to collect dynamic parameters of the funding channel in real time and generate dynamic environmental feature vectors. The rule and state fusion module is used to treat structured constraints as a prior knowledge layer and dynamic environmental feature vectors as a real-time perception layer. It fuses the two layers of information through an attention mechanism to form a composite decision state that integrates business rule constraints and real-time environmental features. The coupling effect prediction module is used to build a coupling effect prediction model based on a deep Q-network, and combined with the composite decision state, outputs the multi-step expected cumulative return value of each optional fund routing action in the optional fund routing action space. The strategy optimization and routing decision module is used to optimize the strategy network parameters in real time and generate the optimal funding routing instructions based on the deep deterministic strategy gradient algorithm with the goal of maximizing the multi-step expected cumulative return. The routing instruction execution and feedback module is used to execute the optimal funding routing instruction and feed the execution result back to the historical transaction knowledge base to update the corresponding routing evaluation label; the execution result is also fed back to the business rule knowledge graph construction module to dynamically adjust the validity verification standards of the core rules of funding business.

[0022] In this embodiment, end-to-end fund routing refers to the complete path decision-making and execution process that a fund undergoes from initiation to final receipt, involving all aspects such as the initiator, routing decision system, multiple intermediate fund channels, and the recipient.

[0023] In this embodiment, historical transaction data records details of each successful or failed fund routing transaction in the past, including transaction time, amount, selected fund channel, actual cost, arrival time, and final status.

[0024] In this embodiment, the routing evaluation label is a quantitative scoring label that the system assigns to each funding channel or historical transaction record to identify its comprehensive performance in terms of success rate, cost efficiency, timeliness, etc. This label is dynamically calculated by analyzing the historical transaction performance of the channel.

[0025] In this embodiment, the funds business system refers to the software platform in the financial system of a bank, payment institution, or enterprise used to manage and execute funds receipt and payment business, which defines the business logic, rules, and strategies for funds routing.

[0026] In this embodiment, the core rules of the funds business are the rigid policies and constraints that must be followed in the funds business system for making fund routing decisions. Their textual forms include, but are not limited to, fee schedules, channel blacklists and whitelists, single or daily cumulative limits, settlement deadlines, and compliance requirements.

[0027] In this embodiment, extracting the core rules of the capital business from the capital business system refers to using natural language processing technologies, such as named entity recognition and relation extraction, to automatically parse the rule text in the configuration document or database of the capital business system and identify the structured information of "rule entities" and "relationships between entities".

[0028] In this embodiment, verifying the effectiveness of core rules for capital operations based on historical transaction data in the historical transaction knowledge base refers to calculating the frequency with which each rule is followed in historical transactions and the degree to which following the rule contributes to positive results such as transaction success and cost savings through statistical analysis, thereby selecting rules with high compliance rates and high positive contributions in practice.

[0029] In this embodiment, the dynamic parameters of the funding channel refer to the external and internal state variables that are acquired in real time at the time of routing decision and affect the availability and efficiency of the channel, including but not limited to the current connectivity status of each bank channel, the real-time foreign exchange rate, the current queuing load rate of the channel, or the response delay.

[0030] In this embodiment, the composite decision state is a fixed-dimensional numerical vector output by the rule-state fusion module. It is formed by weighted fusion of structured constraint feature vectors and dynamic environment feature vectors through an attention mechanism, comprehensively representing all rule constraints and real-time environmental information that need to be considered in the current decision.

[0031] In this embodiment, a coupling effect prediction model is constructed based on a deep Q-network. Combining the composite decision state, it outputs the multi-step expected cumulative return value for each optional funding route action in the action space. This involves constructing a deep neural network model with the composite decision state as input and the evaluation action as its core. This model is trained using a deep Q-learning algorithm. Its input is the composite decision state, and its output is a vector. Each element in the vector represents the expected discounted cumulative reward within multiple consecutive decision steps after selecting the corresponding optional funding route action. In this embodiment, the deep Q-network adopts a three-layer fully connected structure with 256 and 128 neurons in the hidden layers, using ReLU activation function, and a linear activation function in the output layer. The policy network (main network) adopts two fully connected layers with 128 neurons in the hidden layers, using Tanh activation function, and the output layer is mapped according to the action space dimension.

[0032] In this embodiment, the optional fund routing action space is a discrete, predefined set that includes all technically executable fund routing operations at the current decision moment, such as "remitting funds through channel A", "remitting funds through channel B", or "splitting the amount into channels A and C".

[0033] In this embodiment, the optional fund routing action refers to a specific operation selection in the optional fund routing action space.

[0034] In this embodiment, the multi-step expected cumulative return is a scalar value calculated by the coupling effect prediction model. It predicts the expected total return obtained by executing a specific funding routing action starting from the current state and following the current optimal strategy in subsequent steps. The total return is the sum of the immediate rewards of each step after being decayed by a future discount factor. In this embodiment, the multi-step expected cumulative return is calculated using a standard discounted return format. Specifically, the system sets a future return discount factor less than 1, such as 0.99, during prediction to reflect the value reduction of future returns relative to current returns. During calculation, starting from the current decision moment, the immediate return of the next step estimated by the coupling effect prediction model is added to the immediate return of the next step after one discount, then to the immediate return of the third step after two discounts, and so on, with the future multi-step returns being discounted step by step and then summed up. The sum obtained is the multi-step expected cumulative return. Among them, the more discounts are applied to returns further in the future, the smaller their contribution weight in the sum.

[0035] In this embodiment, based on the deep deterministic policy gradient algorithm, the strategy network parameters are optimized in real time to maximize the multi-step expected cumulative return value and generate the optimal fund routing instruction. This means that a deep neural network is used as the strategy network. The network takes the composite decision state as input and directly outputs a specific fund routing instruction. Through the deep deterministic policy gradient reinforcement learning algorithm, the internal connection weights of the strategy network are continuously adjusted according to the gradient information provided by the coupling effect prediction model, so that the multi-step expected cumulative return value corresponding to the output instruction becomes larger and larger. Finally, the optimized strategy network generates the instruction in real time during the decision-making process.

[0036] In this embodiment, the policy network parameters refer to all adjustable weights and biases in the policy neural network. The set of these values ​​determines the mapping relationship between the network from state to action.

[0037] In this embodiment, the execution result is fed back to the historical transaction knowledge base to update the corresponding route evaluation label. This means that after each route instruction is executed, the actual result data is stored in the knowledge base as a new historical transaction record, and the route evaluation label of the involved fund channel is recalculated based on the result.

[0038] In this embodiment, dynamically adjusting the validity verification standard of the core rules of the capital business refers to re-evaluating the calculation method or threshold of the rule compliance rate and effect contribution in historical transaction data based on the new feedback execution results, so as to determine the updated screening standard to be used when extracting and verifying rules in the future.

[0039] To ensure the accurate embedding and quantitative expression of business rules in intelligent decision-making through rule validity verification and feature vector generation, a business rule knowledge graph construction module is proposed, including: The rule parsing unit is used to extract the original fee rate policy, channel permission rules, and settlement time limit rules from the fund business system, and to use natural language processing technology to identify rule entities and relationships to obtain the core rules and rules relationships of fund business. The rule validity verification unit is used to calculate the compliance rate and effectiveness contribution of each core fund business rule in actual transactions based on historical transaction data in the historical transaction knowledge base, and to screen and verify valid core fund business rules based on the compliance rate and effectiveness contribution. The graph construction unit is used to construct a knowledge graph of fund business rules by taking the rule entities corresponding to the core rules of the verified fund business as nodes and the rule relationships as edges; the node attributes include fee priority, maximum channel limit, and settlement time window. The feature vector generation unit is used to generate structured constraint feature vectors based on the structured constraints of the knowledge graph of fund business rules.

[0040] In this embodiment, the original fee rate policy, channel permission rules, and settlement time limit rules are extracted from the funds business system. Natural language processing technology is used to identify rule entities and relationships to obtain the core rules and relationship of the funds business. This means obtaining the original text describing the rules from the database or configuration file of the business system, using named entity recognition technology to find key entities such as "channel A", "fee rate 0.1%" and "single transaction limit of 1 million" from the text, and using dependency parsing or relation extraction technology to determine the relationship between entities, such as "channel A has a fee rate of 0.1%" and "channel B prohibits cross-border transactions", thereby transforming unstructured text into structured rule and relationship data.

[0041] In this embodiment, based on historical transaction data in the historical transaction knowledge base, the compliance rate and effectiveness contribution of each core rule of capital business in actual transactions are calculated. For each extracted rule, the total number of transaction instances that should have triggered the rule in the historical transaction data, and the number of transaction instances in which the actual decision conforms to the rule are counted. The latter is divided by the former to obtain the compliance rate. At the same time, the degree of improvement or enhancement in key indicators such as success rate, cost, and timeliness of transactions that follow the rule compared to transactions that do not follow it (or compared to the historical average performance) is analyzed, and this degree of improvement is quantified as the effectiveness contribution.

[0042] In this embodiment, core rules for fund operations that are verified to be valid are selected based on compliance rate and effect contribution. This means setting a compliance rate threshold and an effect contribution threshold, for example, requiring a compliance rate higher than 95% and an effect contribution of a positive value. Rules that meet both of these thresholds are judged as "verified valid" rules. Rules with low compliance rates or negative effect contributions are considered invalid or rules that need to be optimized and are not included in the subsequent knowledge graph construction.

[0043] In this embodiment, the rule entities corresponding to the core rules of the verified and valid fund business are used as nodes, and the rule relationships are used as edges to construct a fund business rule knowledge graph. This means that the entities identified in the filtered and valid rules, such as specific fund channels, fee rates, limit values, time windows, etc., are used as nodes in the graph; and the relationships between entities, such as "applicable", "less than", "earlier than", etc., are used as directed edges connecting the nodes. In this way, all valid rules are organized into an interconnected network graph structure, which intuitively displays the rule system.

[0044] In this embodiment, node attributes include rate priority, maximum channel limit, and settlement time window. This means that in the knowledge graph, in addition to the node type, each node is also assigned specific attribute values. For example, for a node representing "channel X", its "maximum channel limit" attribute is set to "5 million yuan"; for a node representing "settlement timeliness", its "settlement time window" attribute is set to "T+1"; the "rate priority" attribute is used to identify which rule should be given priority when multiple rate rules conflict.

[0045] In this embodiment, the structured constraint feature vector is generated based on the structured constraints of the knowledge graph of fund business rules. This means traversing all nodes and edges in the knowledge graph, converting node attribute values, node types, and the relationship types represented by edges into numerical values ​​through feature encoding methods (such as one-hot encoding and embedding encoding); then concatenating all these numerical values ​​into a fixed-length, one-dimensional numerical array in a predetermined order. This array is the structured constraint feature vector, which is a machine-readable digital representation of the entire current effective rule set.

[0046] To achieve adaptive evolution and intelligent maintenance of business rules through dynamic vitality value calculation and weighting mechanisms, a business rule knowledge graph construction module is proposed, which also includes a rule vitality value calculation unit to perform the following steps: Real-time monitoring of the frequency with which each core rule of a financial transaction is triggered in historical decisions, serving as the rule activation frequency; The rule failure index is calculated based on the deviation between the actual routing result and the expected effect of the rule after the core rules of capital business are triggered. The rule activation frequency is converted into an activation intensity coefficient, and the rule failure index is converted into a failure rate coefficient. Through the nonlinear combination of the activation intensity coefficient and the failure rate coefficient, the real-time vitality value of each core rule of the capital business is calculated. When the structured constraint feature vector generation unit generates structured constraint feature vectors based on the knowledge graph of capital business rules, the feature dimensions of the corresponding capital business core rules are dynamically weighted according to the real-time vitality value of the capital business core rules, so that rules with high vitality values ​​occupy higher weights in the structured constraint feature vectors. When the real-time vitality value of a certain core rule of capital business continues to be lower than the evolution threshold, the rule reconstruction process is automatically triggered to generate new candidate rules based on recent successful transaction patterns. The candidate rule verification unit is used to verify the effectiveness of new candidate rules. This includes: backtesting in similar scenarios based on the historical transaction knowledge base to calculate the simulated compliance rate and simulated effect contribution of the candidate rule; and monitoring the actual execution effect of the candidate rule within a preset period in a low-volume live trading environment. When both the simulated compliance rate and simulated effect contribution meet the preset verification standards, and the actual execution effect meets expectations, the candidate rule is automatically adopted and updated to the capital business rule knowledge graph. Specifically, after generating new candidate rules in the rule reconstruction process, the system initiates a two-stage verification process to ensure its safety and effectiveness. The first stage is offline backtesting verification: the system selects past transaction records from the historical transaction knowledge base that are similar to the transaction scenario characteristics targeted by the candidate rule, simulates the application of the candidate rule for decision-making, and calculates the rate at which the rule is followed in these simulated transactions (simulated compliance rate), as well as the degree to which compliance improves key indicators such as transaction success rate and cost (simulated effect contribution). The second phase is limited live trading validation: The system applies candidate rules to real, ongoing fund routing decisions at a small percentage (e.g., 1% of the traffic) and continuously monitors them for a preset time period (e.g., 24 hours), collecting data on the success rate and costs of actual execution. Only when the simulated compliance rate and contribution of the candidate rule in the backtesting phase both exceed preset thresholds, and the actual execution effect in the live trading validation phase also meets the expected goals, will the system automatically adopt the candidate rule and update it as a valid structured constraint in the fund business rule knowledge graph, replacing or supplementing the original low-activity rules.

[0047] In this embodiment, the frequency at which each core rule of the financial business is triggered in historical decisions is monitored in real time. This frequency is used as the rule activation frequency, which refers to which core rules of the financial business are considered or evaluated as constraints in the current decision-making process after each routing decision. The activation frequency of a rule is obtained by counting the number of times each rule is triggered within a unit time window (e.g., the past 24 hours). This frequency reflects the activity level of the rule in the current decision-making environment.

[0048] In this embodiment, a rule failure index is calculated based on the deviation between the actual routing result and the expected effect of the core rules for fund operations after they are triggered. This index compares the actual result (e.g., actual cost, actual arrival time) of each transaction after the rule is triggered with the ideal effect (e.g., expected minimum cost, expected latest arrival time) of the rule, calculating the degree of deviation. The average deviation of the rule across all triggered transactions is normalized to obtain a value between 0 and 1 as the rule failure index. A higher index indicates a worse actual guiding effect of the rule. In this embodiment, the rule failure index is calculated as follows: For each transaction where a rule is triggered, the absolute value of the relative deviation between the actual effect value (e.g., actual cost) and the expected effect value (e.g., expected minimum cost) is calculated. The average of the relative deviations of the rule across all triggered transactions is then used as its rule failure index. Specifically, it can be expressed as: Failure Index = Average(|Actual Value - Expected Value| / Expected Value).

[0049] In this embodiment, the rule activation frequency is converted into an activation intensity coefficient, and the rule failure index is converted into a failure rate coefficient. This means that the rule activation frequency is converted by applying a nonlinear function (such as a logarithmic function or a sigmoid function) to map it to a standardized numerical range to obtain the activation intensity coefficient. This coefficient increases with the frequency but the rate of increase gradually slows down. The rule failure index is converted by applying another nonlinear function (such as an exponential function) to map it into a failure rate coefficient. This coefficient can amplify the impact of a high failure index.

[0050] In this embodiment, the real-time vitality value of each core rule of the capital business is calculated by a non-linear combination of the activation intensity coefficient and the failure rate coefficient. This means that the activation intensity coefficient is used as a positive factor, and the failure rate coefficient as a negative factor, calculated using a preset formula (e.g., vitality value = activation intensity coefficient). (1-failure rate coefficient) The two are combined to calculate a comprehensive score as the real-time vitality value; this vitality value reflects both the activity and effectiveness of the rule, and the higher the value, the more active and reliable the rule is.

[0051] In this embodiment, the feature dimensions of the corresponding core rules of the capital business are dynamically weighted according to the real-time vitality value of the core rules of the capital business, so that rules with high vitality values ​​have higher weights in the structured constraint feature vector. This means that when generating the structured constraint feature vector, the numerical part (feature dimension) of each rule in the vector is multiplied by a weight coefficient that is positively correlated with the real-time vitality value of the rule. For example, the higher the vitality value of a rule, the more its corresponding feature dimension value is amplified in the vector, thus having a greater influence in subsequent model calculations.

[0052] In this embodiment, the evolution threshold is a pre-set numerical constant or a dynamically calculated baseline used to determine whether a rule has "aged" or "failed" to the point that it needs to be re-examined and replaced. In this embodiment, the evolution threshold for the rule vitality value is set to 0.3. When the vitality value of a rule is below 0.3 for five consecutive evaluation periods, the rule reconstruction process is triggered.

[0053] In this embodiment, when the real-time vitality value of a certain core rule of capital business is continuously lower than the evolution threshold, the rule reconstruction process is automatically triggered. New candidate rules are generated based on recent successful transaction patterns. The system continuously monitors the vitality value of each rule. Once it is found that the vitality value of a certain rule is lower than the evolution threshold for multiple consecutive evaluation periods, a rule reconstruction subroutine is automatically started. The subroutine analyzes the recent successful transaction data that did not follow or violate the old rule, and summarizes new and potential regular constraints from them.

[0054] In this embodiment, the recent successful transaction pattern refers to transaction instances that occurred in the past period (e.g., the most recent week), were marked as successful, and whose routing decision logic may differ from the existing rule set, as well as common characteristics or patterns extracted from these instances regarding channel selection, timing, amount splitting, etc.

[0055] In this embodiment, a new candidate rule refers to one or more new rule statements generated by the rule reconstruction process based on recent successful transaction patterns, which may replace the original low-activity rules. These candidate rules are also described in a structured form and submitted after preliminary verification, awaiting subsequent evaluation and possible formal adoption.

[0056] To accurately capture the dynamic coupling relationship of the funding network through the analysis of the influence transmission strength between channels, a dynamic feature perception module is proposed, including: The multi-source data access unit is used to access bank channel status, foreign exchange rate fluctuations, and system load monitoring data in real time through data interfaces to obtain real-time data streams. The anomaly processing unit is used to detect and process anomalies in the real-time data stream. This includes: when missing data is detected, filling it with statistical values ​​from recent historical data and adding a missing data marker; when an abnormal jump in data value is detected that exceeds the normal fluctuation range, replacing the original abnormal value with a smoothed and corrected data value to ensure the reliability of data input to subsequent modules. Specifically, the dynamic feature perception module has a built-in dedicated data quality control component. This component continuously monitors the raw real-time data stream received from external interfaces. Its anomaly detection is based on a historical normal fluctuation range model established for each data indicator (such as the response latency of a channel), which is typically defined by the average and standard deviation of the indicator over a past period. When a data point does not arrive within the expected time window, the unit determines it as missing data and fills it with the moving average or last valid value of the indicator over a recent period (e.g., the past 5 minutes), adding a "missing data" marker signal to the data point so that subsequent modules are aware of the situation. When an arriving data point's value exceeds the normal fluctuation boundary calculated based on historical standard deviations (e.g., exceeding three times the standard deviation above and below the average), the unit determines it as an abnormal jump. For such outliers, the unit does not use them directly. Instead, it employs data smoothing algorithms such as exponentially weighted moving averages, combining the previous valid value of the indicator to calculate a corrected, smoother value to replace the original outlier. Through these two processes, the unit ensures that the real-time data stream output to the downstream message transmission strength analysis unit and feature synthesis unit is continuous, complete, and free from significant noise interference, thereby guaranteeing the reliability of the real-time environmental information upon which the system's core decisions depend.

[0057] The message transmission strength analysis unit is used to calculate the state correlation strength between two adjacent channels based on real-time data streams, forming an influence transmission strength matrix between channels. The feature synthesis unit is used to fuse the channel's own dynamic indicators with the influence transmission strength matrix to generate a dynamic environmental feature vector containing the coupling relationship between channels.

[0058] In this embodiment, real-time data streams are obtained by accessing bank channel status, foreign exchange rate fluctuations, and system load monitoring data in real time through data interfaces. This means that the system continuously acquires the latest data reflecting the external environment and internal operating status at a frequency of seconds or milliseconds by calling application programming interfaces provided by banks or payment service providers, subscribing to foreign exchange market data sources, and connecting to internal system monitoring components. These data are organized into a continuous sequence with timestamps, forming a real-time data stream.

[0059] In this embodiment, the strength of the state correlation between adjacent channels is calculated based on real-time data streams to form an influence transmission strength matrix between channels. This means that the system extracts the state indicators (such as response delay and success rate) of each funding channel at the same point in time from the real-time data stream, and then analyzes the synchronicity or causal correlation of the state changes of all possible channel pairs within a past time window. By calculating correlation coefficients, Granger causality tests, or prediction error reduction ratios, the influence of the state change of one channel on the state of another channel is quantified, resulting in an N×N numerical matrix (N is the total number of channels). The element value in the i-th row and j-th column of the matrix represents the strength of the state correlation or influence transmission of channel i to channel j.

[0060] In this embodiment, the channel's own dynamic indicators are fused with the influence propagation strength matrix to generate a dynamic environment feature vector containing the coupling relationship between channels. First, the real-time dynamic indicators of each channel (such as the current load rate and the latest exchange rate quote) are encoded into a sub-vector. Then, for each channel, the influence strength values ​​of all other channels on that channel are extracted from the influence propagation strength matrix, which also constitute a sub-vector. Finally, the channel's own indicator sub-vector and its received influence strength sub-vectors are concatenated or weighted to form a comprehensive feature vector. This vector not only describes the channel's own state but also contains potential influence information from other channels in the network. These comprehensive feature vectors of all channels together constitute the dynamic environment feature vector.

[0061] To achieve quantitative modeling of inter-channel correlation strength through fluctuation analysis and message passing tendency calculation, a message passing strength analysis unit is proposed to calculate the state correlation strength between pairwise adjacent channels based on real-time data streams, and to establish an influence transmission network between channels. The steps include: For each channel's real-time load data, calculate the fluctuation amplitude and frequency within the sliding time window; Based on the fluctuation characteristics and state differences of adjacent channels, the message transmission tendency value between channels is calculated. All message transmission tendency values ​​are normalized to form an influence transmission strength matrix between channels, which is used for dynamic feature fusion. Based on the influence propagation strength matrix, the indirect influence weights between non-adjacent channels are calculated through matrix iterative multiplication. These indirect influence weights are then merged with the direct influence weights to update the influence propagation strength matrix. Specifically, after calculating the direct influence weights between adjacent channels and forming an initial influence propagation strength matrix, the message propagation strength analysis unit further calculates the indirect influences generated through one or more intermediate channels in the channel network. The specific process involves performing multiple multiplications on the initial direct influence matrix (e.g., two matrix multiplications to calculate the second-level indirect influence through one intermediate channel, and three matrix multiplications to calculate the third-level indirect influence through two intermediate channels). The result of each matrix multiplication quantifies the strength of the indirect influence of the corresponding row channel on the corresponding column channel at a specific level. Then, the unit weights the calculated indirect influence weight matrices at different levels according to a preset attenuation coefficient (the deeper the level, the smaller the attenuation coefficient), resulting in a comprehensive indirect influence weight matrix. Finally, the unit linearly adds or weights and fuses this comprehensive indirect influence weight matrix with the original direct influence matrix to generate a new, more comprehensive influence transmission strength matrix that includes both direct and indirect correlations. This updated matrix replaces the original matrix and is input into the feature synthesis unit to generate more accurate dynamic environment feature vectors.

[0062] In this embodiment, for the real-time load data of each channel, the fluctuation amplitude and frequency within the sliding time window are calculated. The system maintains a data window that slides over time for each funding channel, for example, containing a load rate sequence of the most recent 100 time points. Within this window, the difference between the maximum and minimum values ​​of the load data is calculated as the fluctuation amplitude, and the number of times the load value crosses its average value within the window is counted as the fluctuation frequency, thereby quantifying the severity and speed of change of the channel load.

[0063] In this embodiment, based on the fluctuation characteristics and state differences of adjacent channels, a message transmission tendency value between channels is calculated. This means that for any two funding channels considered adjacent or potentially related, the calculated fluctuation amplitude and frequency characteristics of one channel are compared and combined with the corresponding characteristics of the other channel. Simultaneously, the absolute difference in the average state (e.g., average load rate) of the two channels within the current window is calculated. This is then processed using a preset synthesis function (e.g., tendency value = (fluctuation amplitude similarity factor / fluctuation frequency ratio)). The system uses a state difference function to process these comparison results and the difference, outputting a numerical value that represents the likelihood that a state fluctuation from one channel will "transfer" or "influence" another channel.

[0064] In this embodiment, all message passing tendency values ​​are normalized to form an influence transmission strength matrix between channels. This influence transmission strength matrix is ​​used for dynamic feature fusion. After calculating the message passing tendency values ​​between all channel pairs, a normalization method (such as Min-Max normalization or division by the sum of all tendency values) is applied to these original tendency values ​​to scale them to a uniform and comparable numerical range. Then, these normalized values ​​are filled into a matrix according to the channel order, where the rows and columns of the matrix represent the source channel and the target channel, respectively. This matrix is ​​the influence transmission strength matrix. During dynamic feature fusion, this matrix serves as a weighting coefficient to adjust and aggregate the feature information of different channels, so that the features of channels that are greatly influenced by other channels are more significantly reflected in the final environmental feature vector.

[0065] To integrate rules and real-time features through an attention mechanism to form a composite state that comprehensively reflects the decision-making context, a rule and state fusion module is proposed, including: The prior knowledge encoding unit is used to encode the structured constraints and generate prior knowledge feature vectors. The real-time sensing coding unit is used to encode the feature vectors of the dynamic environment and generate real-time sensing feature vectors. The attention fusion unit is used to calculate the mutual influence weight between the prior knowledge feature vector and the real-time perceived feature vector based on the multi-head attention mechanism, and to perform weighted fusion of the two types of feature vectors according to the mutual influence weight to obtain the fused feature vector. The state generation unit is used to map the fused feature vectors into composite decision states that the policy network can handle.

[0066] In this embodiment, the structured constraints are feature-encoded to generate a prior knowledge feature vector. This means that the structured constraints (such as rule nodes, attributes, and relationships) generated from the knowledge graph are input into an embedding layer or feature transformation layer. This layer maps discrete rule elements into continuous, low-dimensional numerical vector representations and concatenates or aggregates the vector representations of all elements in a fixed order to finally form a fixed-length numerical vector that can represent the comprehensive information of all current valid business rules, i.e., the prior knowledge feature vector.

[0067] In this embodiment, feature encoding is performed on the dynamic environment feature vector to generate a real-time perception feature vector. This means that the original dynamic environment feature vector output by the dynamic feature perception module is input into a fully connected neural network layer or a self-attention encoding layer. This layer performs nonlinear transformation and interactive learning on the features of each dimension in the original vector to extract a high-level representation that better reflects the essential features of the current real-time environment and outputs a numerical vector with optimized dimensions and stronger feature expression capabilities, namely the real-time perception feature vector.

[0068] In this embodiment, a multi-head attention mechanism is used to calculate the mutual influence weights between the prior knowledge feature vector and the real-time perceived feature vector. Based on these mutual influence weights, the two types of feature vectors are weighted and fused to obtain a fused feature vector. Specifically, the prior knowledge feature vector and the real-time perceived feature vector are used as the query, key, and value, respectively, and input into a multi-head attention calculation module. This module uses multiple parallel attention heads to calculate the correlation strength between elements in the two vectors from different feature subspaces, obtaining a set of attention weight matrices. Then, based on these weight matrices, the information in the real-time perceived feature vector is weighted and summed, and a similar operation may be performed on the prior knowledge feature vector in reverse. Finally, the outputs of each attention head are concatenated and linearly projected to generate a new feature vector that integrates rule constraints and environmental perception information. In this embodiment, the number of heads in the multi-head attention mechanism is set to 8. The online learning single-step optimization step size (i.e., learning rate) of the attention meta-learning controller is set to 0.001, and the Adam optimizer is used.

[0069] In this embodiment, mapping the fused feature vector to a composite decision state that the policy network can handle means inputting the feature vector obtained after attention fusion into a state projection layer. This projection layer is usually composed of one or more fully connected neural network layers. Its function is to transform (or map) the fused feature vector into a standardized state representation that meets the input requirements of the downstream policy network (such as a deep deterministic policy gradient network) in terms of dimension, numerical range, and feature distribution. The final output is the composite decision state.

[0070] To achieve accurate return prediction under a multi-step decision-making perspective through hierarchical attention and long-term value assessment, a deep Q-network in the coupling effect prediction module employs a hierarchical attention mechanism, specifically including: The state coding layer is used to receive the composite decision state and perform deep feature encoding to obtain the encoded features. The attention fusion layer is used to calculate the mutual influence weights between different feature dimensions in a composite decision state using an attention mechanism. The feature coupling layer is used to perform weighted fusion of encoded features based on mutual influence weights to form a deeply coupled state representation. The value output layer is used to output the multi-step expected cumulative return value of each optional funding routing action in the optional funding routing action space based on the deeply coupled state representation.

[0071] In this embodiment, the state encoding layer is used to receive the composite decision state and perform deep feature encoding to obtain the encoded features. This means that the composite decision state output by the rule-state fusion module is input into a deep neural network encoder, which is composed of multiple fully connected layers or transformer layers stacked together. Through nonlinear transformation and hierarchical feature extraction, the input composite decision state is transformed into a feature representation expressed in a higher-dimensional or more abstract feature space, that is, the encoded features. These features can capture the inherent patterns of state information more deeply.

[0072] In this embodiment, the attention fusion layer is used to calculate the mutual influence weights between different feature dimensions in the composite decision state using an attention mechanism. This means that the encoded features are input into a self-attention calculation module. This module first linearly projects the features into three sets of vectors: query, key, and value. Then, it calculates the dot product similarity between the query vector and all key vectors. After scaling and softmax normalization, a set of attention weights is obtained. These weights reflect the degree of mutual influence or dependence of each dimension (or position) in the encoded features on all other dimensions.

[0073] In this embodiment, the feature coupling layer is used to perform weighted fusion of the encoded features based on the mutual influence weights to form a deep coupled state representation. Specifically, the weight matrix calculated by the attention fusion layer is used to perform weighted summation on the encoded features (as value vectors). The operation is to multiply the weight matrix with the value vector to obtain a weighted feature matrix. This operation enables the new representation of each feature dimension to incorporate information from all other dimensions that is weighted according to the correlation, thereby forming a deep coupled state representation that can reflect the complex dependencies between features.

[0074] In this embodiment, the value output layer is used to output the multi-step expected cumulative return value of each optional funding routing action in the optional funding routing action space based on the deeply coupled state representation. This means that the deeply coupled state representation is input into a value evaluation network. The network is usually composed of one or more fully connected layers, and the number of neurons in its output layer is equal to the total number of actions in the optional funding routing action space. The network calculates a scalar value for each neuron (corresponding to a specific action). This value is the discounted cumulative expected value of the future multi-step return that can be obtained by selecting the action and assuming that the optimal strategy is followed under the current deeply coupled state representation. The outputs of all neurons together constitute a value vector.

[0075] To achieve rapid adaptation and continuous optimization of decision logic through attention meta-learning and pattern library evolution, a hierarchical attention mechanism is proposed, employing an attention meta-learning controller. The attention meta-learning controller, as the core component of the attention fusion layer, executes the following steps: Based on successful transaction records in the historical transaction knowledge base, the corresponding attention weight distribution is extracted to establish an attention pattern library containing various typical decision-making logics; Encode the current complex decision state to generate a current context feature vector; The similarity between the current context feature vector and the context feature vectors of each historical attention pattern in the attention pattern library is calculated, and the historical attention pattern with the highest similarity is retrieved. The attention weight distribution of the retrieved historical attention patterns is used as the initial weight configuration for the attention fusion layer; An online learning approach is adopted, using the current complex decision-making state as the only training sample. The initial weight configuration is optimized and adjusted step by step through the gradient descent algorithm to generate an attention weight allocation scheme that is adapted to the current decision-making context. When the routing instruction execution and feedback module confirms that the current routing decision is successful, it stores the valid attention weight allocation scheme in the current decision as a new candidate pattern into the attention pattern library, and evaluates the activity of historical patterns in the attention pattern library and eliminates expired patterns to achieve continuous evolution of attention patterns.

[0076] In this embodiment, based on successful transaction records in the historical transaction knowledge base, the corresponding attention weight distribution is extracted, and an attention pattern library containing multiple typical decision-making logics is established. The system analyzes the transaction records marked as successful in the historical transaction knowledge base. For each record, it backtracks and extracts the attention weight matrix generated by the attention fusion layer at the time of the transaction decision. These weight matrices are associated with their corresponding transaction context (i.e., the composite decision state at that time) and stored as a "historical attention pattern". All the stored patterns together constitute a pattern library. Each pattern represents a decision attention point distribution method that has been verified to be effective in a specific context.

[0077] In this embodiment, the current composite decision state is feature-encoded to generate a current context feature vector. This means that the composite decision state formed when a decision needs to be made is input into a context encoding network. This network is usually composed of a lightweight fully connected neural network. Its function is to compress or map the high-dimensional composite decision state into a vector representation with a lower dimension that can better summarize the core features of the current decision scenario, i.e., the current context feature vector.

[0078] In this embodiment, the similarity between the current context feature vector and the context feature vectors of each historical attention pattern in the attention pattern library is calculated, and the historical attention pattern with the highest similarity is retrieved. This means calculating the similarity between the current context feature vector and the context feature vector associated with each historical attention pattern in the pattern library (such as calculating the cosine similarity or the reciprocal of the Euclidean distance). By comparing all the calculated similarity values, the one with the largest value is found, and the corresponding historical attention pattern is retrieved as the pattern most similar to the current context.

[0079] In this embodiment, the attention weight distribution of the retrieved historical attention patterns is used as the initial weight configuration of the attention fusion layer. This means that the attention weight matrix stored in the retrieved historical attention patterns is directly loaded into the weight parameters of the current attention fusion layer, replacing its original random initialization or the weight state after the last decision, thus serving as a high-starting point initialization configuration for the current attention calculation.

[0080] In this embodiment, an online learning method is employed, using the current composite decision state as the sole training sample. The initial weight configuration is optimized step-by-step using the gradient descent algorithm to generate an attention weight allocation scheme adapted to the current decision context. Specifically, based on the initial weight configuration, the current composite decision state is used as input data. According to a preset loss function related to the current decision task (e.g., a loss function designed to assign higher weights to key features), the gradient of the loss function with respect to the attention fusion layer weights is calculated. Then, the gradient descent algorithm (or its variants such as Adam) is used to perform a small step-by-step update of the weights along the reverse gradient direction. The updated weights are the final attention weight allocation scheme that is more adapted to the current specific decision context.

[0081] In this embodiment, when the routing instruction execution and feedback module confirms the success of the current routing decision, it stores the valid attention weight allocation scheme in the current decision as a new candidate pattern into the attention pattern library. It also evaluates the activity of historical patterns in the attention pattern library and eliminates expired patterns, thus achieving continuous evolution of attention patterns. Specifically, after the routing instruction is executed and confirmed as successful, the system binds the attention weight allocation scheme ultimately used in this decision with the current context feature vector, as a new candidate attention pattern. Simultaneously, the system periodically scans the pattern library, maintaining an "activity" score for each historical pattern. This score may be calculated based on the frequency of its retrieval and the most recent retrieval time. Patterns with activity scores below a preset elimination threshold or that have not been retrieved for a long time are removed from the library. Through this mechanism of "adding effective patterns and eliminating outdated patterns," the pattern library can dynamically reflect the latest and most effective decision-making logic.

[0082] To effectively control decision-making risk while maximizing returns through dual-objective optimization and intelligent risk filtering, a dual-objective optimization architecture is proposed for the strategy optimization and routing decision modules, specifically including: The strategy network is divided into a main network and an auxiliary network. The main network is responsible for generating routing instructions that maximize the expected cumulative return value of multiple steps, while the auxiliary network assesses the potential risk level of the current routing instructions based on risk cases in the historical trading knowledge base. The dynamic balancing decision-making unit dynamically adjusts the relative weights of the return maximization objective and the risk minimization objective based on the potential risk level assessed by the auxiliary network. When the potential risk level is low, the return maximization objective is given a higher weight; when the potential risk level is medium to high, the weight of the risk minimization objective is gradually increased, and the adjusted weights comprehensively guide the generation of the optimal funding routing instruction. Specifically, it refers to a coordinating decision-making component set in the strategy optimization and routing decision-making module. This unit receives a quantified potential risk level score output from the auxiliary network. It internally pre-defines a risk level range division rule (e.g., a score below 0.3 is low risk, 0.3 to 0.7 is medium risk, and above 0.7 is high risk) and a corresponding objective weight mapping table. When determined to be low risk, the unit sets a higher decision weight (e.g., 0.8) for the return maximization objective (i.e., pursuing the highest multi-step expected cumulative return value) and a lower weight (e.g., 0.2) for the risk minimization objective. When determined to be medium risk, the unit balances the weights of the two objectives (e.g., 0.5 for each). When a scenario is deemed high-risk, the unit prioritizes the risk minimization objective (e.g., 0.7) while simultaneously reducing the weight of the reward objective (e.g., 0.3). These dynamically determined weights guide the final instruction generation logic: in low-risk scenarios, the system prioritizes instructions generated by the main network with the highest expected reward; in high-risk scenarios, the system tends to adopt lower-risk alternative instructions generated by the instruction reconstruction subunit; and in medium-risk scenarios, the system may fine-tune the main network instructions or introduce stronger risk constraints. In this way, the unit ensures that the system instruction generation process achieves an adaptive, quantitative, and dynamic balance between reward and safety. The instruction filtering unit comprises a risk review subunit and an instruction refactoring subunit, wherein: The risk review subunit is used to receive routing instructions generated by the main network and potential risk levels assessed by the auxiliary network. When the potential risk level does not exceed the preset threshold, the routing instruction generated by the main network that has passed the risk review is selected as the optimal funding routing instruction; When the potential risk level exceeds a preset threshold, a risk avoidance instruction is sent to the instruction reconfiguration subunit. The instruction reconstruction subunit is used to retrieve successful routing instructions with potential risk levels below a preset threshold in similar situations from the historical transaction knowledge base based on risk avoidance instructions, and generate alternative routing instructions based on the current composite decision state as the optimal funding routing instructions. The auxiliary network assessment of potential risk levels employs a multi-dimensional risk factor fusion method, including: Based on the knowledge graph of business rules, the degree of compliance between inspection instructions and regulatory rules is checked, and the compliance risk factor is calculated. Based on the channel's historical fraud records and real-time abnormal transaction pattern detection results, calculate the security risk factor; Based on the characteristics of channel fund settlement costs and settlement cycles, a liquidity risk factor is calculated. The compliance risk factor, security risk factor, and liquidity risk factor are weighted and integrated into a comprehensive risk score, which serves as the potential risk level of the current routing instruction.

[0083] In this embodiment, the strategy network is divided into a main network and an auxiliary network. The main network is responsible for generating routing instructions that maximize the multi-step expected cumulative return value. The auxiliary network evaluates the potential risk level of the current routing instruction based on risk cases in the historical transaction knowledge base. Specifically, within the strategy optimization and routing decision module, two functionally independent deep neural networks are deployed: one is the main network, whose structure and training objective are focused on generating specific routing instructions that can bring the highest expected long-term return (multi-step expected cumulative return value) based on the composite decision state; the other is the auxiliary network, whose structure and training objective are focused on analyzing risk events recorded in the historical transaction knowledge base (such as transaction failure, high costs, compliance violations, etc.) and, based on the patterns learned from these historical cases, assessing the risk of any candidate routing instruction generated by the main network and outputting a quantified potential risk level.

[0084] In this embodiment, the preset threshold is a numerical limit set by the system administrator based on risk tolerance, or dynamically adjusted by the system based on historical risk data. It is used to determine whether the potential risk level assessed by the auxiliary network is high enough to trigger a risk avoidance mechanism. In this embodiment, the preset threshold for the potential risk level is 0.7 (risk score normalized to 0-1). Exceeding this threshold is considered high risk, triggering the risk avoidance mechanism.

[0085] In this embodiment, the main network is responsible for generating routing instructions that maximize the multi-step expected cumulative return value. This means that the main network takes the composite decision state as input, calculates through its internal parameters, and directly outputs a specific fund routing operation instruction. The training objective of this network is to make the multi-step expected cumulative return value corresponding to its output instruction as high as possible. Its training process is achieved through the interaction of deep reinforcement learning algorithm and coupling effect prediction model.

[0086] In this embodiment, the risk avoidance instruction refers to an internal trigger signal generated by the risk review subunit when it determines that the potential risk level of the routing instruction generated by the main network exceeds a preset threshold. This signal does not contain specific routing information, but is only a clear command to notify the instruction reconstructing subunit to start the process of generating alternative instructions.

[0087] In this embodiment, successful routing instructions with potential risk levels below a preset threshold in similar scenarios are retrieved from the historical transaction knowledge base as templates. Alternative routing instructions are generated based on the current composite decision state and used as the optimal funding routing instruction. Specifically, upon receiving a risk aversion instruction, the instruction reconstruction subunit first searches the historical transaction knowledge base for historical transaction records where the decision-making situation was similar to the current one, was ultimately marked as a successful transaction, and whose historical risk level, when assessed using the current auxiliary network, would be below a preset threshold. The specific routing instructions executed at that time are extracted as reference templates. Then, based on the specific details of the current composite decision state (such as amount, target account, etc.), the template instruction is adaptively adjusted or replanned to generate a new alternative routing instruction suitable for the current transaction, which is then used as the final output optimal funding routing instruction. When no historical successful routing instructions with similarity reaching a preset similarity threshold are found, the instruction reconstruction subunit generates alternative routing instructions prioritizing risk minimization based on core rules in the business rule knowledge graph where the real-time vitality value is higher than the vitality threshold, combined with the current composite decision state.

[0088] In this embodiment, a similar scenario refers to a situation where the decision state in historical transaction records and the current composite decision state are highly comparable in terms of core features. The system defines "similarity" by calculating the similarity (such as cosine similarity) between two state vectors and setting a similarity threshold. In this embodiment, the threshold for judging whether the historical scenario and the current scenario are "similar" is set to a cosine similarity of 0.8.

[0089] In this embodiment, a successful routing instruction refers to a specific fund routing operation instruction generated by the system or manually specified at the time, corresponding to those transaction instances recorded in the historical transaction knowledge base whose final execution result is "success".

[0090] In this embodiment, a compliance risk factor is calculated based on the degree of conformity between the business rule knowledge graph and the inspection instructions and the regulatory rules. This means that the auxiliary network analyzes the routing instructions generated by the main network and compares them one by one with various regulatory rules (such as anti-money laundering rules, cross-border payment restrictions, customer identity restrictions, etc.) encoded in the business rule knowledge graph. Based on the number and severity of the rules violated by the instructions and the consequences of similar violations in the past, a quantitative score is calculated as the compliance risk factor. The higher the score, the greater the possibility or severity of the violation.

[0091] In this embodiment, a security risk factor is calculated based on the channel's historical fraud records and the detection results of real-time abnormal transaction patterns. This involves the auxiliary network querying the historical fraud transaction statistics of the selected funding channel and analyzing whether the characteristics of the current transaction (such as amount, time, and counterparty) match the real-time abnormal transaction patterns being monitored (such as frequent small-amount probes, large-amount transfers outside business hours, etc.). By combining the historical fraud probability and the real-time abnormal matching degree, a quantitative score is calculated as the security risk factor. The higher the score, the greater the risk that the transaction is suspected of fraud or abnormality.

[0092] In this embodiment, based on the characteristics of channel capital stagnation cost and settlement cycle, a liquidity risk factor is calculated. This refers to the auxiliary network analysis of the potential capital occupation situation that the selected capital channel may have after completing the current instruction. Considering the settlement cycle of the channel (such as T+1), the cost of funds in transit, and other factors that may affect the speed of capital return, the liquidity cost or opportunity cost incurred by executing the instruction is estimated and quantified into a score as a liquidity risk factor. The higher the score, the greater the risk of adverse impact on capital liquidity.

[0093] In this embodiment, compliance risk factors, security risk factors, and liquidity risk factors are weighted and fused into a comprehensive risk score, which serves as the potential risk level of the current routing instruction. This means the auxiliary network possesses a preset or learnable weight vector, corresponding to the compliance risk factor, security risk factor, and liquidity risk factor, respectively. This vector is multiplied by its corresponding weight, then summed (or subjected to other forms of aggregation calculation) to obtain a single comprehensive risk score. This comprehensive score serves as a quantitative assessment of the potential risk level of the current routing instruction. In this embodiment, the comprehensive risk score is calculated as: Comprehensive Risk Score = 0.5 Compliance risk factor +0.3 Safety risk factor +0.2 Liquidity risk factors. Each factor must be pre-normalized to the [0,1] range.

[0094] To achieve coordinated optimization and continuous improvement of system module parameters through closed-loop feedback distribution and dynamic adjustment, a routing instruction execution and feedback module is proposed, including: The instruction issuing unit is used to send the optimal fund routing instruction to the corresponding payment gateway or fund clearing system; The execution monitoring unit is used to track the entire execution process of the optimal fund routing instruction, and record the actual transaction success rate, actual fund cost, and actual arrival time as the execution result; The feedback distribution unit is used to simultaneously distribute the execution results to the historical transaction knowledge base and the business rule knowledge graph construction module. The knowledge base update unit is used to update the route evaluation tags in the historical transaction knowledge base based on the execution results. The rule verification and adjustment unit is used to dynamically adjust the validity verification standards of core rules for fund business based on the statistical data of execution results within a preset time window.

[0095] In this embodiment, the optimal fund routing instruction is sent to the corresponding payment gateway or fund clearing system. This means that the instruction issuing unit in the routing instruction execution and feedback module directly and automatically transmits the optimal fund routing instruction (containing clear channel, amount, account and other information) generated by the decision to the third-party gateway or the fund clearing system inside the institution responsible for the specific execution of the payment through a standardized application programming interface or proprietary communication protocol, so as to initiate the actual fund flow.

[0096] In this embodiment, the entire process of executing the optimal fund routing instruction is tracked, and the actual transaction success rate, actual fund cost, and actual arrival time are recorded as execution results. The execution monitoring unit continuously listens to the status updates returned by the payment gateway or clearing system from the moment the instruction is issued until the transaction is finally completed or fails. It records the actual execution outcome (success or failure) of this instruction. If successful, it accurately records the total cost incurred by the fund transfer and the total time taken from the issuance of the instruction to the receipt of funds by the other party (arrival time). These three core indicators together constitute the execution result data of this routing decision.

[0097] In this embodiment, the execution result data is simultaneously distributed to the historical transaction knowledge base and the business rule knowledge graph construction module. This means that after the feedback distribution unit collects the complete execution result data, it immediately copies it into two copies. One copy is sent to the historical transaction knowledge base for data storage, and the other copy is sent to the business rule knowledge graph construction module for adjusting the rule verification standards. This ensures that the feedback information can simultaneously drive system evolution in both the data accumulation and rule optimization directions.

[0098] In this embodiment, updating the routing evaluation labels in the historical transaction knowledge base based on the execution results means that after receiving the execution result data, the knowledge base update unit locates the funding channel involved in the transaction; based on the actual success rate, cost, duration, and other results of this transaction, it recalculates the corresponding rating labels of these channels according to a preset label update algorithm (such as average calculation based on a sliding window or Bayesian update), and uses the newly calculated label values ​​to cover or smoothly update the original label values ​​in the knowledge base.

[0099] In this embodiment, based on the statistical data of execution results within a preset time window, the validity verification standard of the core rules of the capital business is dynamically adjusted. This means that after receiving the execution result data, the rule verification and adjustment unit analyzes the impact of the result on historical transaction data statistics. It may recalculate the "compliance rate" and "effect contribution" of relevant rules based on new success or failure cases, and adjust the thresholds or calculation formulas used to screen "valid" rules accordingly, so that the rule validity verification standard can adapt to the latest transaction performance and business environment. Specifically, this includes: when the average effect contribution of a certain type of rule changes beyond the contribution change threshold within the time window, the compliance rate threshold of that type of rule is adjusted synchronously by a preset adjustment range; the adjustment range has an upper limit to avoid drastic fluctuations in the verification standard. In other words, the rule verification and adjustment unit executes a specific quantitative process of its core adjustment logic. This unit continuously monitors and statistically analyzes the average effect contribution of each type of core rules of the capital business (such as all rules related to fees) within the most recent preset time window (e.g., the past 30 decision cycles). It calculates the percentage change between the average effect contribution of this type of rule in the current time window and the value in the previous time window. The system presets a contribution change threshold (e.g., 20%). If the calculated percentage change is positive and exceeds this threshold (i.e., a significant improvement in effectiveness), the unit lowers the compliance verification threshold for this type of rule by a preset, fixed downward adjustment (e.g., 5%). This means that in the future, such rules only need to be followed by a lower percentage of transactions to be considered effective. Conversely, if the percentage change is negative and its absolute value exceeds the threshold (i.e., a significant decrease in effectiveness), the unit raises the compliance verification threshold for this type of rule by a preset, fixed upward adjustment (e.g., 8%). This means that in the future, such rules need to be followed by a higher percentage of transactions to be considered effective. Simultaneously, to prevent extreme changes in the verification standard due to short-term fluctuations, the entire adjustment mechanism sets an upper limit on the adjustment range (e.g., regardless of how many adjustments are triggered consecutively, the single adjustment range does not exceed 15%, and the cumulative adjustment range is limited to within ±30% of the initial threshold). This mechanism ensures that the business rule effectiveness verification standard can be stably and gradually optimized based on the actual performance of the rules, thereby supporting the continuous and reliable evolution of the system's knowledge base.

[0100] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A dynamic funding routing decision-making system based on an AI large-scale model, characterized in that: include: Historical transaction knowledge base, used to store historical transaction data of the entire funding route and corresponding route evaluation tags; The business rule knowledge graph construction module is used to extract core rules of capital business from the capital business system, verify the effectiveness of the core rules of capital business based on historical transaction data in the historical transaction knowledge base, and transform the verified core rules of capital business into structured constraint feature vectors. The dynamic feature perception module is used to collect dynamic parameters of the funding channel in real time and generate dynamic environmental feature vectors. The rule and state fusion module is used to treat structured constraints as a prior knowledge layer and dynamic environmental feature vectors as a real-time perception layer. It fuses the two layers of information through an attention mechanism to form a composite decision state that integrates business rule constraints and real-time environmental features. The coupling effect prediction module is used to build a coupling effect prediction model based on a deep Q-network, and combined with the composite decision state, outputs the multi-step expected cumulative return value of each optional fund routing action in the optional fund routing action space. The strategy optimization and routing decision module is used to optimize the strategy network parameters in real time and generate the optimal funding routing instructions based on the deep deterministic strategy gradient algorithm with the goal of maximizing the multi-step expected cumulative return. The routing instruction execution and feedback module is used to execute the optimal funding routing instruction and feed the execution result back to the historical transaction knowledge base to update the corresponding routing evaluation label; the execution result is also fed back to the business rule knowledge graph construction module to dynamically adjust the validity verification standards of the core rules of funding business.

2. The AI-based large-scale model-based dynamic funding routing decision system according to claim 1, characterized in that, The business rules knowledge graph construction module includes: The rule parsing unit is used to extract the original fee rate policy, channel permission rules, and settlement time limit rules from the fund business system, and to use natural language processing technology to identify rule entities and relationships to obtain the core rules and rules relationships of fund business. The rule validity verification unit is used to calculate the compliance rate and effectiveness contribution of each core fund business rule in actual transactions based on historical transaction data in the historical transaction knowledge base, and to screen and verify valid core fund business rules based on the compliance rate and effectiveness contribution. The graph construction unit is used to construct a knowledge graph of fund business rules by taking the rule entities corresponding to the core rules of the verified fund business as nodes and the rule relationships as edges; the node attributes include fee priority, maximum channel limit, and settlement time window. The feature vector generation unit is used to generate structured constraint feature vectors based on the structured constraints of the knowledge graph of fund business rules.

3. The AI-based large-scale model-based dynamic funding routing decision system according to claim 2, characterized in that, The business rule knowledge graph construction module also includes a rule vitality value calculation unit, used to perform the following steps: Real-time monitoring of the frequency with which each core rule of a financial transaction is triggered in historical decisions, serving as the rule activation frequency; The rule failure index is calculated based on the deviation between the actual routing result and the expected effect of the rule after the core rules of capital business are triggered. The rule activation frequency is converted into an activation intensity coefficient, and the rule failure index is converted into a failure rate coefficient. Through the nonlinear combination of the activation intensity coefficient and the failure rate coefficient, the real-time vitality value of each core rule of the capital business is calculated. When the structured constraint feature vector generation unit generates structured constraint feature vectors based on the knowledge graph of capital business rules, the feature dimensions of the corresponding capital business core rules are dynamically weighted according to the real-time vitality value of the capital business core rules, so that rules with high vitality values ​​occupy higher weights in the structured constraint feature vectors. When the real-time vitality value of a certain core rule of capital business continues to be lower than the evolution threshold, the rule reconstruction process is automatically triggered to generate new candidate rules based on recent successful transaction patterns. The candidate rule verification unit is used to verify the effectiveness of new candidate rules, including: conducting backtesting on similar scenarios based on the historical transaction knowledge base, and calculating the simulation compliance rate and simulation effect contribution of the candidate rules; It also monitors the actual execution effect of candidate rules within a preset period in a low-volume live trading environment; when the simulation compliance rate and the contribution of simulation effect both reach the preset verification standards, and the actual execution effect meets expectations, the candidate rules are automatically adopted and updated to the capital business rule knowledge graph.

4. The AI-based large-scale model-based dynamic funding routing decision system according to claim 1, characterized in that, The dynamic feature perception module includes: The multi-source data access unit is used to access bank channel status, foreign exchange rate fluctuations, and system load monitoring data in real time through data interfaces to obtain real-time data streams. The message transmission strength analysis unit is used to calculate the state correlation strength between two adjacent channels based on real-time data streams, forming an influence transmission strength matrix between channels. The feature synthesis unit is used to fuse the channel's own dynamic indicators with the influence transmission strength matrix to generate a dynamic environmental feature vector containing the coupling relationship between channels.

5. The AI-based large-scale model-based dynamic funding routing decision system according to claim 4, characterized in that, The message propagation strength analysis unit calculates the state correlation strength between pairs of adjacent channels based on real-time data streams and establishes the influence propagation network between channels, including the following steps: For each channel's real-time load data, calculate the fluctuation amplitude and frequency within the sliding time window; Based on the fluctuation characteristics and state differences of adjacent channels, the message transmission tendency value between channels is calculated. All message transmission tendency values ​​are normalized to form an influence transmission strength matrix between channels, which is used for dynamic feature fusion. Based on the influence transmission strength matrix, the indirect influence weights between non-adjacent channels are calculated by matrix iterative multiplication, and the indirect influence weights are merged with the direct influence weights to update the influence transmission strength matrix.

6. The AI-based large-scale model-based dynamic funding routing decision system according to claim 1, characterized in that, The rules and state fusion module includes: The prior knowledge encoding unit is used to encode the structured constraints and generate prior knowledge feature vectors. The real-time sensing coding unit is used to encode the feature vectors of the dynamic environment and generate real-time sensing feature vectors. The attention fusion unit is used to calculate the mutual influence weight between the prior knowledge feature vector and the real-time perceived feature vector based on the multi-head attention mechanism, and to perform weighted fusion of the two types of feature vectors according to the mutual influence weight to obtain the fused feature vector. The state generation unit is used to map the fused feature vectors into composite decision states that the policy network can handle.

7. The AI-based large-scale model-based dynamic funding routing decision system according to claim 1, characterized in that, The deep Q-network in the coupling effect prediction module employs a hierarchical attention mechanism, specifically including: The state coding layer is used to receive the composite decision state and perform deep feature encoding to obtain the encoded features. The attention fusion layer is used to calculate the mutual influence weights between different feature dimensions in a composite decision state using an attention mechanism. The feature coupling layer is used to perform weighted fusion of encoded features based on mutual influence weights to form a deeply coupled state representation. The value output layer is used to output the multi-step expected cumulative return value of each optional funding routing action in the optional funding routing action space based on the deeply coupled state representation.

8. The AI-based large-scale model-based dynamic funding routing decision system according to claim 7, characterized in that, The hierarchical attention mechanism employs an attention meta-learning controller. The steps executed by the attention meta-learning controller, as the core component of the attention fusion layer, include: Based on successful transaction records in the historical transaction knowledge base, the corresponding attention weight distribution is extracted to establish an attention pattern library containing various typical decision-making logics; Encode the current complex decision state to generate a current context feature vector; The similarity between the current context feature vector and the context feature vectors of each historical attention pattern in the attention pattern library is calculated, and the historical attention pattern with the highest similarity is retrieved. The attention weight distribution of the retrieved historical attention patterns is used as the initial weight configuration for the attention fusion layer; An online learning approach is adopted, using the current complex decision-making state as the only training sample. The initial weight configuration is optimized and adjusted step by step through the gradient descent algorithm to generate an attention weight allocation scheme that is adapted to the current decision-making context. When the routing instruction execution and feedback module confirms that the current routing decision is successful, it stores the valid attention weight allocation scheme in the current decision as a new candidate pattern into the attention pattern library, and evaluates the activity of historical patterns in the attention pattern library and eliminates expired patterns to achieve continuous evolution of attention patterns.

9. The AI-based large-scale model-based dynamic funding routing decision system according to claim 1, characterized in that, The strategy optimization and routing decision module adopts a dual-objective optimization architecture, which specifically includes: The strategy network is divided into a main network and an auxiliary network. The main network is responsible for generating routing instructions that maximize the expected cumulative return value of multiple steps, while the auxiliary network assesses the potential risk level of the current routing instructions based on risk cases in the historical trading knowledge base. The instruction filtering unit comprises a risk review subunit and an instruction refactoring subunit, wherein: The risk review subunit is used to receive routing instructions generated by the main network and potential risk levels assessed by the auxiliary network. When the potential risk level does not exceed the preset threshold, the routing instruction generated by the main network that has passed the risk review is selected as the optimal funding routing instruction; When the potential risk level exceeds a preset threshold, a risk avoidance instruction is sent to the instruction reconfiguration subunit. The instruction reconstruction subunit is used to retrieve successful routing instructions with potential risk levels below a preset threshold in similar situations from the historical transaction knowledge base based on risk avoidance instructions, and generate alternative routing instructions based on the current composite decision state as the optimal funding routing instructions. The auxiliary network assessment of potential risk levels employs a multi-dimensional risk factor fusion method, including: Based on the knowledge graph of business rules, the degree of compliance between inspection instructions and regulatory rules is checked, and the compliance risk factor is calculated. Based on the channel's historical fraud records and real-time abnormal transaction pattern detection results, calculate the security risk factor; Based on the characteristics of channel fund settlement costs and settlement cycles, a liquidity risk factor is calculated. The compliance risk factor, security risk factor, and liquidity risk factor are weighted and integrated into a comprehensive risk score, which serves as the potential risk level of the current routing instruction.

10. The AI-based large-scale model-based dynamic funding routing decision system according to claim 1, characterized in that, The routing instruction execution and feedback module includes: The instruction issuing unit is used to send the optimal fund routing instruction to the corresponding payment gateway or fund clearing system; The execution monitoring unit is used to track the entire execution process of the optimal fund routing instruction, and record the actual transaction success rate, actual fund cost, and actual arrival time as the execution result; The feedback distribution unit is used to simultaneously distribute the execution results to the historical transaction knowledge base and the business rule knowledge graph construction module. The knowledge base update unit is used to update the route evaluation tags in the historical transaction knowledge base based on the execution results. The rule verification and adjustment unit is used to dynamically adjust the validity verification standards of core rules for fund business based on the statistical data of execution results within a preset time window.

Images