Intelligent risk control routing method for enterprise payment to public, electronic device and storage medium

Abstract

The application discloses an intelligent risk control routing method for enterprise payment to public, an electronic device and a storage medium, and relates to the technical field of intelligent enterprise finance. The method comprises the following steps: collecting original invoice data of a heterogeneous business system and analyzing the original invoice data to generate a payment request; extracting features such as payment purpose, payee identifier, currency, time constraint and tax attribute through a pre-trained intention recognition model to generate an intention vector; performing pre-interception based on the intention features and a dynamic risk control rule set to screen out compliant payment requests; inputting the compliant payment requests into a multi-objective optimization routing module, selecting an optimal payment channel according to the real-time state of each channel, generating instructions meeting the agreement requirements and completing encrypted transmission and result feedback. The application significantly improves the degree of payment automation, the accuracy of risk control and the efficiency of fund use.

Description

Technical Field

[0001] This application relates to the field of intelligent enterprise finance technology, and in particular to an intelligent risk control routing method, electronic device and storage medium for corporate payments. Background Technology

[0002] In high-frequency corporate payment scenarios of large group enterprises, the payment intentions conveyed by business documents (such as contract terms "payment within 3 working days" and OA expense reports marked "urgent") lack structured expression. Existing systems cannot automatically identify multi-dimensional semantic features such as urgency, payment purpose, and tax attributes from unstructured text, leading to repeated manual verification. The risk control process relies on static field comparison (such as account / account name matching) and preset blacklists and whitelists, making it difficult to detect hidden risks such as logical conflicts between the payee and the intention (such as expense reimbursement intention corresponding to a corporate account) and splitting orders to circumvent limits. Payment channel selection generally adopts hard-coded configuration without dynamic adjustment based on real-time balance, fee rate, and system status, which can easily lead to high transaction fees or idle funds. Summary of the Invention

[0003] The purpose of this application is to provide an intelligent risk control routing method, electronic device and storage medium for corporate payments, so as to alleviate the above-mentioned technical problems existing in the prior art.

[0004] In a first aspect, the present invention provides an intelligent risk control routing method for corporate payments, comprising: Collect raw document data from multiple heterogeneous business systems within the enterprise, perform document parsing and recognition on the raw document data, and generate payment requests to be processed. By using a pre-trained payment intent recognition model, target intent features corresponding to the payment intent are extracted from the payment request to be processed, and a payment intent vector is generated. The target intent features include payment purpose category, unique identifier of the payee, settlement currency type, payment time constraints, and tax processing attributes. Based on the target intent characteristics and the pre-configured dynamic risk control rule set, the payment request is intercepted in advance to determine the payment request that passes the risk control verification. The payment request that has passed the risk control verification and its corresponding payment intent vector are input into the multi-objective optimization routing module so that the target payment channel can be selected according to the current availability status of each available payment channel, and a payment instruction that meets the protocol requirements of the selected channel can be generated for encrypted transmission and automatic feedback of execution results.

[0005] In an optional implementation, raw document data from multiple heterogeneous business systems within the enterprise is collected, and document parsing and recognition are performed on the raw document data to generate payment requests to be processed, including: By connecting to the purchase order system, contract management system, expense reimbursement system, and payment request system through middleware, structured document data and unstructured attachments can be obtained. Optical character recognition and layout analysis are performed on the unstructured attachments to extract text content and perform semantic cleaning, misspelling normalization and word segmentation. The cleaned text content is merged with the structured fields corresponding to the structured document data to generate a payment request to be processed.

[0006] In an optional implementation, a pre-trained payment intent recognition model is used to extract target intent features corresponding to the payment intent from the payment request to be processed, generating a payment intent vector, including: The payment request is processed by named entity recognition and intent classification using the BERT-BiLSTM-CRF joint model, which extracts the payee's name, bank account information, keywords for payment purpose, urgency indicators and tax-related statements. The extraction results are mapped to a three-level purpose classification system, a standardized payee master data identifier, a currency type identifier, a time constraint quantification window, and a tax attribute marker, respectively, to obtain the payment purpose category, payee unique identifier information, settlement currency type, payment time constraint conditions, and tax processing attributes. When some target intent features are missing, context completion is performed based on historical transaction behavior patterns or the status is marked as pending confirmation, and manual review is triggered.

[0007] In an optional implementation, payment requests are pre-emptively intercepted based on the target intent characteristics and a pre-configured set of dynamic risk control rules to determine payment requests that pass the risk control verification, including: Semantic fingerprint calculation is performed, and the cosine similarity of the current payment request summary and the historical payment request summary is compared. When the similarity exceeds a preset threshold and the amounts are similar, it is determined to be a potential duplicate payment. Within a set time window, multiple payment requests corresponding to the same payee identifier are detected. If there are characteristics of splitting the amount and the cumulative amount is close to the amount of a certain payment already made, a risk warning for split payment is triggered. The system calls upon real-time updated external sanctions lists and internal databases of untrustworthy suppliers to perform a thorough matching of the unique identifier information of the payee; if a match is found, the payment request is frozen and a compliance alert is generated. For payment requests classified as high-risk, the exemption threshold will be automatically lowered, and the payment request will be forcibly transferred to a multi-level review process.

[0008] In an optional implementation, the payment request that has passed the risk control verification and its corresponding payment intent vector are input into a multi-objective optimized routing module to select a target payment channel based on the current availability of each available payment channel, and generate a payment instruction that conforms to the protocol requirements of the selected channel for encrypted transmission and automatic feedback of execution results, including: A dynamic channel resource pool is pre-configured, including direct bank connection channels, bill payment channels, virtual account channels, and third-party aggregation channels, and the balance status, transaction fee rate, system availability, and single transaction limit information of each channel are collected in real time. Construct a multi-objective optimization routing module to perform weighted scoring and sorting of each channel based on the amount, payment time constraints, and payment purpose category in the payment intent vector; When the preferred channel is excluded due to insufficient balance or system unavailability, the degradation strategy is automatically activated, and the second-best channel is tried in turn; if all channels are unavailable, the fund aggregation logic is activated, and the fund is allocated through a preset path among the specified accounts before the routing decision is initiated again.

[0009] In an optional implementation, a payment instruction conforming to the selected channel protocol is generated for encrypted transmission and automatic feedback of execution results, including: Based on the message standard of the target payment channel, load the pre-configured field mapping template and convert the payment intent vector and risk control verification result into the standardized message format required by the channel; Before transmission, the message is digitally signed and encrypted using the national cryptographic algorithm, and the instruction signing is completed through the bank-enterprise direct connection dedicated line or hardware U-Key server. After receiving the processing status code returned by the bank, identify false failures and initiate a status polling mechanism; if it is confirmed to be a substantive failure, perform retry, channel switching or appointment delay processing according to the reason for failure. After successful payment, the electronic receipt is automatically downloaded, key fields are parsed and uniquely associated with the original payment request, and then synchronously written back to the business system and automatically generated accounting vouchers.

[0010] In an optional implementation, the method further includes: After the payment instruction is executed, the payment intent vector, risk control interception log, routing decision path and execution result data are structured, labeled and persistently stored. Based on continuously accumulated labeled data, the training corpus of the payment intent recognition model is periodically updated, and the dynamic risk control rule set and multi-objective optimization routing module are iteratively optimized online.

[0011] In an optional implementation, the method further includes: When the payment purpose category is identified as tax payment or fiscal payment, the system automatically verifies whether the payee's unique identifier belongs to the tax target account; if they do not match, the system blocks the generation of payment instructions and prompts an abnormality in the payee account.

[0012] In a second aspect, the present invention provides an electronic device including a processor and a memory, wherein the memory stores computer-executable instructions that can be executed by the processor, and the processor executes the computer-executable instructions to implement the intelligent risk control routing method for corporate payments as described in any of the foregoing embodiments.

[0013] Thirdly, the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the intelligent risk control routing method for corporate payments as described in any of the foregoing embodiments.

[0014] This application provides an intelligent risk control routing method, electronic device, and storage medium for corporate bank payments. By collecting and parsing multi-source heterogeneous documents to generate payment requests, it overcomes the problem of inaccurate intent recognition caused by the disconnect between business documents and payment instructions, ensuring the integrity and authenticity of original semantic information. It extracts a payment intent vector containing purpose, recipient identifier, currency, time constraints, and tax attributes, enabling the system to accurately capture implicit business intent and support subsequent fine-grained risk control and routing decisions. Pre-interception based on intent features and dynamic risk control rule sets overcomes the limitations of traditional field-level comparisons, identifying conflicts between intent and behavioral logic, such as matching reimbursement intent to corporate accounts and semantic duplication risks. Inputting into a multi-objective optimized routing module and selecting routes based on real-time channel status solves the problem of high fees or idle funds caused by static configuration. Overall, it significantly improves the degree of payment automation, risk control accuracy, and fund utilization efficiency. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating an intelligent risk control routing method for corporate payments provided in this application embodiment; Figure 2 A schematic diagram of an intelligent payment routing and risk control system architecture based on intent recognition is provided for embodiments of this application; Figure 3This application provides a schematic diagram of an intent recognition and dynamic routing decision logic. Figure 4 A risk control interception flowchart is provided as an embodiment of this application; Figure 5 A structural diagram of an intelligent risk control routing device for corporate payments provided in this application embodiment; Figure 6 This is a structural diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0018] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0019] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0020] This application provides an intelligent risk control routing method for corporate bank payments, see [link to relevant documentation]. Figure 1 As shown, the method mainly includes the following steps: Step S110: Collect original document data from multiple heterogeneous business systems within the enterprise, perform document parsing and recognition on the original document data, and generate payment requests to be processed.

[0021] like Figure 2The intelligent payment routing and risk control system architecture based on intent recognition is shown. The system docks with the procurement order system (SRM / ERP), contract management system (CLM), expense reimbursement system (OA / expense control system), and payment request form system through middleware to obtain structured document data and unstructured attachments. These attachments usually include scanned contracts, photos of paper invoices, payment request forms in PDF format, etc. The system performs optical character recognition (OCR) and layout analysis on the unstructured attachments, not only identifying the text but also distinguishing the positions of key fields such as "Party A / Party B", "bank account opening", "account number", etc. in the contract. For the manually filled remarks field (such as "Pay the goods payment to Company A, urgent"), the system performs text cleaning, removing meaningless characters, correcting typos (such as normalizing "account number" to "account number"), and performing word segmentation. Subsequently, the cleaned text content is merged with the corresponding structured fields of the structured document data (such as amount, applying department, budget item) to generate a unified JSON format object as the payment request to be processed. This object contains fields such as source system identifier, original text, amount, original information of the payee, etc., and is pushed into the message queue for subsequent processing.

[0022] In practical applications, there are a variety of heterogeneous systems within an enterprise, and the data formats vary greatly. To build a unified data collection layer, the system not only actively pulls data through API interfaces but also supports integrating those old systems that do not provide open interfaces through RPA. For the collected contract documents, the system focuses on extracting clause information directly related to payments, such as "payment node", "settlement method", "liquidated damages clause", etc. When the OCR engine processes images or PDFs, it first performs image preprocessing (such as denoising, binarization, skew correction), and then uses a deep learning-based text detection and recognition model to output the recognition results in a structured manner. For the recognized text, further semantic cleaning is performed: unifying "RMB" and "Renminbi" to "CNY", normalizing "account" and "账" to "account", and using regular expressions to remove irrelevant symbols. Finally, all source documents are converted into a unified payment request object, which contains fields such as source system ID, original text summary, amount, original description of the payee, etc., providing a standardized input for subsequent intent recognition.

[0023] Step S120: Extract the target intent features corresponding to the payment intent from the payment request to be processed through a pre-trained payment intent recognition model, and generate a payment intent vector; the target intent features include payment purpose category, unique identification information of the payee, settlement currency type, payment time constraint conditions, and tax treatment attributes.

[0024] Such as Figure 3The system employs a BERT-BiLSTM-CRF joint model architecture to perform named entity recognition and intent classification on payment requests, based on the logic of schematic diagram recognition and dynamic routing decisions. Specifically, the named entity recognition (NER) task extracts key entities from the text, such as extracting the payee name "XXX Computing Co., Ltd.", the keyword "service fee," and the implied amount information (combined with structured fields) from "XXX Computing Co., Ltd. 2025 Service Fee." The intent classification task determines transaction attributes, such as urgency level (P0 - real-time, P1 - same day, P2 - normal) and tax attributes. The system maps the extracted results to a three-level usage classification system (e.g., operating expenses → raw materials → chemical raw materials), standardized payee master data identifiers (aligned with the enterprise master data system through fuzzy matching to obtain the unified social credit code and bank account provincial and municipal codes), currency type identifiers, time constraint quantification windows (e.g., "express" is mapped to T+0, "normal" is mapped to T+3), and tax attribute markers (e.g., whether it includes tax, whether it needs to be withheld and paid on behalf of others, and whether the invoice has been certified).

[0025] When some target intent features are missing, such as a payment request that only states "pay rent" without specifying the month, the system performs context completion based on historical transaction behavior patterns (e.g., automatically inferring the payment month based on the supplier's contract period), or marks it as "pending confirmation" and triggers manual review. Finally, the system assembles the extracted information into a payment intent vector, which has a five-tuple form: [purpose category, payee identifier, currency, time constraint, tax attribute].

[0026] In terms of model building and training, this step uses the pre-trained language model BERT as the basic encoder to capture deep semantic information in the text. A bidirectional long short-term memory network (BiLSTM) is superimposed on BERT to model contextual sequence features. Finally, a conditional random field (CRF) layer is used for label decoding to achieve accurate entity boundary annotation. The corpus used for model training comes from historical payment documents accumulated by enterprises and manually annotated intent tags, with annotation content including usage classification, urgency level, tax attributes, etc.

[0027] To improve the model's generalization ability, this application also employs methods such as randomly replacing synonyms and inserting noisy characters. For identifying the payee's name, fuzzy matching algorithms (such as edit distance and pinyin similarity) are used to align it with the standard supplier database in the Enterprise Master Data (MDM) to obtain the unified social credit code and bank account information, thereby avoiding payment errors due to non-standard name spelling. For automatic urgency classification, not only are explicit keywords in the text analyzed ("urgent," "late payment fee," "supply disruption risk"), but also business context information, such as the contractual payment deadline and supplier credit rating, are combined to comprehensively calculate a quantified time window requirement. If there are missing or contradictory information in the extracted data (e.g., the purpose is "tax payment" but the payee is not a treasury account), the system will mark it as "pending confirmation" and return it to the initiator for supplementation, ensuring the accuracy of the intent features.

[0028] Step S130: Based on the target intent characteristics and the pre-configured dynamic risk control rule set, perform pre-risk interception on the payment request and determine the payment request that passes the risk control verification.

[0029] like Figure 4 The risk control interception process shown in the diagram extracts deep intent features from step S120 before funds flow out of the enterprise, rather than simply using surface-level fields, for risk control interception. Multi-dimensional risk detection is employed during pre-emptive risk interception: First, intelligent duplicate payment interception is performed. Traditional systems only compare document numbers; this system calculates based on semantic fingerprints: it calculates the cosine similarity between the current payment request summary text and the historical payment request summary. When the similarity exceeds a preset threshold (e.g., 90%) and the amounts are similar (relative error less than 5%), it is determined to be a potential duplicate payment, preventing duplicate reimbursements for "different document numbers but the same payment". Simultaneously, within a set time window (e.g., 1 hour), multiple payment requests corresponding to the same payee are detected. If there are characteristics of splitting the amount (e.g., two consecutive payments of 9999 yuan and 10000 yuan, totaling close to a previously paid 19999 yuan) and the cumulative amount approaches a previously paid amount, a split payment risk warning is triggered. Secondly, the system uses real-time updated external sanctions lists (such as OFAC and UN sanctions lists) and business and judicial blacklists (dishonest judgment debtors) to match the unique identifier of the payee with its ultimate controlling shareholder. If a match is found, the payment request is immediately frozen and a compliance alert is generated. Thirdly, for payments classified as high-risk (such as "entertainment expenses," "consulting fees," or "marketing expenses"), the system automatically lowers the exemption threshold (e.g., from 100,000 yuan to 10,000 yuan) and forces the payment request to undergo a multi-level review process.

[0030] In addition, further business logic consistency checks are performed: for requests with the purpose of "purchase payment", the procurement system interface is automatically called to verify the consistency of the "contract-order-invoice" three documents; it is also checked whether the budget balance of the department corresponding to this intent is sufficient; for cases where the purpose is "employee reimbursement" but the receiving account is a corporate account, it is determined that the intent-behavior is inconsistent and is blocked. Only payment requests that pass all the above risk control checks will be marked as "compliant and pending payment" and proceed to the next step of the routing decision-making process.

[0031] In its implementation, the risk control engine combines a rule engine with a machine learning model. The rule engine handles explicit, enumerable rules (such as blacklist matching and limit control), while the machine learning model handles ambiguous risk scenarios that require learning historical patterns (such as semantic similarity judgment for duplicate payments). The duplicate payment detection module not only compares precise amounts, account numbers, and dates, but also uses NLP technology to calculate summary semantic similarity and employs a time-sliding window algorithm to detect "split payment" patterns. For blacklist screening, the system not only matches the payee's name but also uses enterprise graph technology to trace back to its legal representative, shareholders, and other actual controllers to identify indirect associated risks. For strict budget control, the system connects to the budget management system in real time to obtain the used and remaining quotas across three dimensions: department, project, and purpose. When the requested amount exceeds the remaining quota, it automatically intercepts and prompts a budget shortage warning. For high-risk usage categories, the system has a built-in dynamically adjusted exemption limit table, which automatically lowers the threshold based on historical violation records and external risk trends. Requests that pass risk control verification are assigned a "risk control passed" label, along with their risk score and interception record, for subsequent auditing use.

[0032] Step S140: Input the payment request and its corresponding payment intent vector through risk control verification into the multi-target optimization routing module so as to select the target payment channel according to the current availability status of each available payment channel, and generate a payment instruction that meets the protocol requirements of the selected channel for encrypted transmission and automatic feedback of execution results.

[0033] Continue to refer to Figure 3 The second half of the routing decision logic and Figure 2At the instruction execution layer, the system pre-configures a dynamic channel resource pool including direct bank connections (bank-enterprise direct connections), bill payment channels (bank acceptance bills / commercial acceptance bills), virtual account channels, and third-party aggregation channels. It also collects real-time information on the balance status, fee rates, system availability, and single-transaction limits for each channel. Specifically, the system maintains a dynamic "channel capability matrix," including the attributes of each bank's interface: supported currencies and limits (e.g., a single-transaction limit of 5 million for direct connection with Bank A), timeliness (e.g., the large-value payment system closes at 17:00), cost model (e.g., a fee of 5 yuan per transaction for Bank B, and 0.1% of the amount for Bank C), and current status (balance adequacy, interface health). A multi-objective optimization routing module constructs a routing function that comprehensively considers three objectives: cost, speed, and compliance. For each payment request that passes risk control, the system weights and ranks the available channels based on the amount in its intent vector, payment time constraints (urgency), and payment purpose category. For example, for non-urgent large-amount purchases where the company holds unmatured bank acceptance bills, the system prioritizes payment via bill endorsement transfer, which is cost-free and optimizes cash flow. For urgent tax payments or compensation, the system routes payments to the online banking real-time payment channel. For scenarios requiring balancing account balances, if the preferred account has insufficient funds, the system automatically activates fund aggregation logic, first aggregating funds from sub-accounts before payment, or routing to other bank accounts with sufficient funds. When the preferred channel is excluded due to insufficient funds or system unavailability, a fallback strategy is automatically activated, trying the next best channel in turn. If all channels are unavailable, the fund aggregation logic is activated, performing a preset path fund allocation among designated accounts before initiating a routing decision again.

[0034] After selecting the target payment channel, the system automatically converts the original payment request into the standard message format required by that channel. Different banks have vastly different direct connection message formats (XML, JSON, fixed-length messages). The system has a built-in "message conversion factory" that automatically converts messages based on configured mapping templates, supporting standards such as CNAPS, CIPS, and SWIFT MT / MX. At the transport layer, national cryptographic algorithms (SM2 / SM3 / SM4) are used to digitally sign and encrypt the messages. A dedicated bank-enterprise direct connection line is established, or a bank-provided U-Key server is used for hardware signing to ensure the instructions are tamper-proof. The system sends payment instructions through an encrypted dedicated line and monitors the bank's processing status code in real time. If the bank returns "processing" or a timeout, the system automatically initiates a "status query" poll instead of directly reporting an error to prevent duplicate payments. For false failures (such as timeouts caused by network jitter), the system automatically retryes up to three times, with increasing intervals. If a substantial failure is confirmed (such as a mismatch between account name and account holder), retrying, channel switching, or scheduled delay processing are performed based on the cause of the failure. For example, if the payment is rejected due to "insufficient balance," an automatic transfer process is triggered for a retry; if it is rejected due to "outside of working hours," it is automatically added to the "scheduled payment queue" for automatic execution the following day. Upon successful payment, the system automatically downloads the electronic receipt, parses key fields (such as bank transaction number, actual deduction time, and handling fee amount), and uniquely associates them with the original payment request. This information is then synchronously written back to the business system (ERP), automatically generating accounting vouchers (Debit: Accounts Payable, Credit: Bank Deposit), thus automating financial accounting. Finally, the payment result is fed back to the initiator, completing the closed loop.

[0035] In the specific implementation of routing decisions, the system constructs a multi-objective optimization model. The objective function is to minimize the weighted comprehensive cost, where the weights can be dynamically adjusted according to the enterprise's strategy (e.g., financially sound enterprises can increase the compliance weight, while enterprises with tight funds can decrease the transaction fee weight). For each channel i, the comprehensive score Score_i = w1 (cost normalized value) + w2 (1 - time normalized value) + w3 (compliance score) is calculated. Costs include explicit transaction fees and implicit capital occupation costs (e.g., using bill payments can postpone cash outflow, thereby obtaining interest income). Timeliness is matched according to the channel's expected arrival time and the urgency of the request: if the request requires T+0 real-time arrival, only channels supporting real-time payments are retained; if it is a T+3 ordinary payment, it can be included in the batch payment channel. The compliance score is based on the channel's historical success rate, anti-money laundering compliance record, etc. The system updates the channel status (balance, fee rate, interface availability) every 5 minutes to ensure that routing decisions are based on the latest data. For cross-border payment scenarios, the routing module also needs to consider factors such as exchange rate costs and SWIFT intermediary bank fees. When the system detects a change in the transaction fee of a certain channel or a system maintenance announcement from a bank, it automatically adjusts the scoring weight of that channel to avoid routing transactions to unavailable or excessively costly channels. In addition, the routing module also supports special channels such as "virtual accounts" and "agent payments": for scenarios where direct payment is not possible (such as cross-border small payments), transactions are routed to virtual credit cards (VCC) or third-party payment aggregation channels to expand payment coverage.

[0036] In one implementation, a bank return code knowledge base is maintained, categorizing bank return status codes into "Success," "Processing," "Soft Rejection" (retryable), and "Hard Rejection" (non-retryable). For "Soft Rejection" types (such as insufficient balance or system busy), the system automatically executes a retry strategy, with retry intervals using an exponential backoff algorithm (e.g., 1 second, 2 seconds, 4 seconds). For "Hard Rejection" types (such as invalid account or blacklist), the system is directly marked as failed, triggering a manual review process, and the reason for the failure is simultaneously notified to finance personnel via the company's instant messaging tools.

[0037] Furthermore, if the preferred channel fails twice consecutively, the same request is automatically switched to the secondary channel to re-initiate the payment, ensuring payment continuity. For large payments, a preset cooling-off period (e.g., 30 minutes) is allowed before the instruction is sent, giving finance managers sufficient time for post-transaction checks. All payment execution logs (including routing decision paths, retries, and final status) are structured and stored for subsequent auditing and performance analysis.

[0038] After completing the basic description of the above steps, the core technologies involved in each step will be further explained in detail below to demonstrate the inventiveness of this invention.

[0039] Regarding the construction and training of the intent recognition model, this invention adopts a deep learning scheme customized for enterprise payment scenarios. The base model uses a pre-trained Chinese BERT (e.g., BERT-base-chinese), but for special entities such as numbers, dates, amounts, and company names present in payment documents, domain adaptation is performed on top of the pre-training. Specifically, approximately 100,000 anonymized historical payment documents from within the enterprise are used for continued pre-training of the BERT model. Training tasks include masked language modeling (MLM) and next-sentence prediction (NSP), enabling the model to fully learn the terminology distribution and contextual patterns in the payment text. Based on this, a joint architecture of BERT-BiLSTM-CRF is constructed: the BERT layer outputs a vector representation of each character; the BiLSTM layer captures bidirectional sequence dependencies, alleviating the ambiguity of long-distance entity boundaries in BERT; the CRF layer uses the transition probability matrix to constrain the legality of the label sequence (e.g., "B-recipient" must be followed by "I-recipient"). The model training employs multi-task learning, simultaneously optimizing both the NER and intent classification loss functions while sharing BERT encoding layer parameters. The NER tagging system includes 15 entity types (e.g., payee, payer, amount, currency, date, contract number, invoice number, purpose keywords, urgency keywords, tax keywords, bank account number, bank name, summary, remarks, and comments). The intent classification task outputs three levels of purpose categories (approximately 50 subcategories), urgency levels (3 levels), and tax attributes (tax included / tax excluded / withholding required). Early stopping and a learning rate decay strategy are used during training, achieving an F1 score of over 95% on the validation set.

[0040] Furthermore, through an incremental learning mechanism, newly generated and manually reviewed payment request data are added to the training set weekly to perform lightweight fine-tuning of the model, enabling it to adapt to business changes (such as new payment purpose types or new supplier name patterns). This continuous learning mechanism ensures the long-term stability of the intent recognition model in a real-world production environment.

[0041] In the data processing flow for intent feature extraction, this application adopts a "structured-unstructured fusion" strategy. For unstructured text imported from OCR or Excel, a fast pre-parsing based on regular expressions is first used to extract obvious patterns such as numbers and account numbers; then the remaining text is input into the BERT model. The model output is cross-validated with structured fields (such as application departments and budget items in ERP): for example, if the "application department" in the structured field is "marketing department," but the model identifies the purpose as "raw material procurement" from the remarks, the system will trigger a consistency check; if there is a discrepancy, the confidence level is reduced and it is marked as pending confirmation. For the identified payee name, the system calls the Enterprise Master Data Service (MDM) for fuzzy matching. The matching algorithm uses weighted edit distance, where the difference between "Limited Company" and "Limited Liability Company" is given a lower weight, and the core trade name is given a higher weight. After successful matching, the payee's unified social credit code, bank account number, and bank code are automatically completed to form complete payee identification information. For missing time constraints, the system analyzes keywords such as "urgent," "late payment fee," "supply interruption," and "contract expiration date" in the text, combined with the current time and the bank's payment system's deadline (e.g., the large-value payment system closes at 17:00), to automatically calculate the optimal expected payment time window. If "late payment fee" is detected and the current time is close to the contractually agreed payment deadline, the urgency level is set to the highest level, P0 (T+0 real-time), and routing is forced to the real-time payment channel. This processing method, which integrates structured and unstructured information, significantly improves the accuracy and practicality of intent recognition.

[0042] Regarding the pre-emptive interception of risk control rules, this application adopts the "intent-behavior consistency verification" paradigm. Traditional risk control only checks whether the payee is on a blacklist and whether the amount exceeds the limit, while this invention further verifies whether the "payment intent identified by the system" matches the "actual payee type / amount / time pattern". To this end, the system constructs an intent-behavior mapping knowledge base containing tens of thousands of rules. For example, when the intent is "paying taxes", the payee's account must conform to the national treasury account number range; when the intent is "employee reimbursement", the payee's name should be an individual's name or the name of the reimbursement account designated by the company, and the amount should not exceed the employee's monthly reimbursement limit; when the intent is "purchasing raw materials", the payee should be marked as a "raw material supplier" in the supplier master data, and the amount should deviate from the average unit price of historical purchase orders by no more than 30%. When the verification fails, the system not only intercepts the payment but also generates a detailed difference explanation report, such as "the intent is to pay taxes, but the payee's account is not a national treasury account (actually a corporate account of XX company), suspected of being a false tax application". This paradigm effectively prevents fraudulent activities that circumvent fund supervision by altering the intended use of documents.

[0043] Furthermore, in the duplicate payment detection module, the system not only compares the semantic similarity of the summaries but also introduces the concept of "transaction behavior fingerprints": The system inputs the payment amount sequence and time interval sequence of the same payee over the past 30 days into a temporal anomaly detection model (based on an LSTM autoencoder). When the amount and time interval of a new request deviate from the historical pattern by more than 3 standard deviations, a suspected duplicate or anomaly warning is triggered. For example, if a supplier historically receives a monthly payment of approximately 500,000, and suddenly two payments of 250,000 each appear on the same day, the system determines this as "order splitting to circumvent approval" and blocks them.

[0044] Regarding multi-objective optimization of intelligent routing, this invention creatively incorporates non-traditional payment methods such as bill payment and virtual accounts into the routing pool and designs a dedicated cost-timeliness-compliance joint optimization function. Specifically, for the bill payment channel, the system needs to query the bill management system (such as the Shanghai Commercial Paper Exchange interface) in real time to obtain information such as the face value, maturity date, and accepting bank of the available bills (bank acceptance bills or commercial acceptance bills) currently held by the enterprise. If the urgency level in the payment intention is "normal" (T+3 and above), and the payee is marked as "accepting bill endorsement" in the supplier master data, the system will prioritize the use of bill payment. At this time, the cost calculation includes not only the bill endorsement fee (usually 0) but also the implicit benefit: the interest income generated by delaying cash outflow (i.e., the time value of money). The interest income calculation formula is: bill face value × (enterprise weighted average cost of capital) × (remaining days of bill / 360). If the overall cost of bill payment (negative value) is lower than the bank transfer fee, then the system will route to the bill channel. For scenarios requiring balancing multiple account balances, the system employs a prediction-based balance scheduling algorithm: based on the projected income and expenditure of each account for the next 3 days (from the ERP payment and collection plan), it selects the account that will not lead to account overdraft after payment and maintains a minimum safety reserve as the payment account. If all account balances are insufficient, the system automatically triggers an atomic operation of "funds collection + payment": first, the required amount is collected from the funds pool account or sub-account (real-time collection), and then the payment is initiated, ensuring completion in one step. For cross-border payments, the routing module also dynamically compares the exchange rate quotes of different banks and SWIFT intermediary fees, selecting the path with the lowest overall cost. This routing engine obtains the latest exchange rate and fee tables from various banks every 5 minutes and trains a gradient boosting tree (GBDT) model using historical success rate and arrival time data to predict the "actual arrival probability" and "expected arrival time" of each channel at the current moment, using these as dynamic weight inputs to the scoring function. This machine learning-based dynamic routing significantly outperforms static rules.

[0045] In the routing decision-making section, this invention specifically designs a joint optimization function for cost, timeliness, and compliance to achieve optimal channel selection under multiple constraints. Specifically, the system constructs a multi-objective optimization model for each payment request, with the objective of minimizing the overall cost, as shown in the following formula: minCosttotal= wc Costexplicit+ wt Timepenalty+ wr Riskscore Cost explicit For explicit payment costs (such as bank fees, bill acceptance fees, and cross-border intermediary bank fees), Time penalty Risk is a time-sensitive penalty (calculated based on the time difference between the expected arrival time and the urgency requirement of the request). score A compliance risk score is assigned to the channel (based on the channel's historical failure rate, anti-money laundering violation records, etc.). Weighting coefficient w c ,w t ,w r It can be dynamically configured according to the company's financial strategy; for example, when pursuing low costs, w can be set. c =0.7,w t =0.2,w r =0.1, w can be set to 0.1 when high timeliness is desired. t =0.6.

[0046] For non-traditional channels such as bill payment, this invention introduces a hidden benefit item. implicit This allows the overall cost to be negative. The formula for calculating implicit benefits is: Benefit implicit =Amount × WACC × 360 Days float WACC is the company's weighted average cost of capital (provided in real time by the cash management system), Days float The actual number of days the cash outflow is delayed for payment by bill of exchange compared to bank transfer (i.e., the difference between the remaining maturity of the bill of exchange and the bank transfer settlement cycle). When Cost explicit Benefit implicit When the value is less than 0, the system determines that the bill payment has "negative costs" and should be selected first.

[0047] The timeliness penalty term uses a piecewise function:

[0048] In other words: there is no penalty when the expected arrival time meets the urgency requirement; when it exceeds the requirement, the penalty increases according to a quadratic function; for the highest urgency level (P0), a slight penalty is also incurred if the funds arrive early (to avoid wasting costs due to excessive pursuit of speed).

[0049] The compliance risk score is predicted in real time by the Gradient Boosting Tree (GBDT) model. Input features include: the recent (1 hour / 24 hours / 7 days) success rate of the channel, the status of the bank's system maintenance announcement, the number of times the channel has been hit in the blacklist history, and the sanction risk level corresponding to the region where the recipient of the current request is located.

[0050] The system collects the latest data from each channel every 5 minutes and uses the aforementioned optimization function to score and rank all available channels. Before ranking, hard constraint filtering is applied: channels with single-transaction limits less than the requested amount, unsupported currencies, or currently under system maintenance are directly removed. When the score difference between the best and second-best channels is less than a set threshold (e.g., 5%), the system automatically employs a "round-robin" or "minimum usage" strategy to achieve load balancing. Compared to static routing or simple priority rules, the aforementioned joint optimization function and its dynamic weight adjustment mechanism significantly improve the overall efficiency of fund disbursement.

[0051] Furthermore, this invention includes a closed-loop model self-optimization mechanism. After each payment instruction is executed, the system stores the complete processing log (including intent vector, risk control interception result, routing decision path, channel return status code, final success or failure, and manual review record) in a structured manner. The system runs an offline analysis task daily to calculate the actual success rate, average handling fee, and average arrival time of each channel, and compares them with the predicted values ​​during routing decision-making to calculate the deviation. If the actual performance of a channel is consistently lower than the prediction (e.g., frequent timeouts at a bank's interface), the system automatically reduces the timeliness weight of that channel in the scoring function and issues an alarm on the operation and maintenance monitoring panel. For the intent recognition model, the system collects manually corrected samples (i.e., cases where the model output is inconsistent with the manual review results) weekly, adds them to the training set as hard examples, and incrementally trains the model to improve its ability to identify boundary cases. The risk control rule engine also supports automatic adjustment of rule weights based on feedback: if a rule is frequently triggered within a period of time but is found to be a false alarm after manual review, the system automatically reduces the priority of that rule or suggests modifying the threshold. This continuous learning and self-optimizing architecture enables the system to adapt to business development and changes in the external environment, maintaining high accuracy over the long term.

[0052] Based on the above method embodiments, this application also provides an intelligent risk control routing device for corporate payments, see [link to relevant documentation]. Figure 5 As shown, the device includes the following parts: The document collection and parsing recognition module 510 is used to collect original document data from multiple heterogeneous business systems within the enterprise, and to perform document parsing and recognition on the original document data to generate payment requests to be processed. The payment intent feature extraction module 520 is used to extract the target intent features corresponding to the payment intent from the payment request to be processed through a pre-trained payment intent recognition model, and generate a payment intent vector; wherein, the target intent features include payment purpose category, unique identifier information of the payee, settlement currency type, payment time constraint conditions, and tax processing attributes; The pre-risk interception module 530 is used to intercept payment requests based on the target intent features and a pre-configured set of dynamic risk control rules, and to determine payment requests that pass the risk control verification. The multi-target optimization routing module 540 is used to input the payment request that has passed the risk control verification and its corresponding payment intention vector, so as to select the target payment channel according to the current availability status of each available payment channel, generate a payment instruction that conforms to the protocol requirements of the selected channel, encrypt and transmit the payment instruction, and receive and process the execution result to achieve automatic feedback.

[0053] In one feasible implementation, the aforementioned document collection and parsing recognition module 510 is specifically used for: By connecting to the purchase order system, contract management system, expense reimbursement system, and payment request system through middleware, structured document data and unstructured attachments can be obtained. For unstructured attachments, perform optical character recognition and layout analysis to extract text content and perform semantic cleaning, misspelling normalization and word segmentation; The cleaned text content is merged with the structured fields corresponding to the structured document data to generate a payment request to be processed.

[0054] In one feasible implementation, the payment intent feature extraction module 520 is specifically used for: The BERT-BiLSTM-CRF joint model is used to perform named entity recognition and intent classification on payment requests, extracting the payee's name, bank account information, keywords for payment purpose, urgency indicators, and tax-related statements. The extraction results are mapped to a three-level purpose classification system, a standardized payee master data identifier, a currency type identifier, a time constraint quantification window, and a tax attribute marker, respectively, to obtain the payment purpose category, payee unique identifier information, settlement currency type, payment time constraint conditions, and tax processing attributes. When some target intent features are missing, context completion is performed based on historical transaction behavior patterns or the status is marked as pending confirmation, and manual review is triggered.

[0055] In one feasible implementation, the aforementioned pre-risk interception module 530 is specifically used for: Semantic fingerprint calculation is performed, and the cosine similarity of the current payment request summary and the historical payment request summary is compared. When the similarity exceeds a preset threshold and the amounts are similar, it is determined to be a potential duplicate payment. Within a set time window, multiple payment requests corresponding to the same payee identifier are detected. If there are characteristics of splitting the amount and the cumulative amount is close to the amount of a certain payment already made, a risk warning for split payment is triggered. The system calls upon real-time updated external sanctions lists and internal databases of untrustworthy suppliers to perform a thorough matching of the unique identifier information of the payee; if a match is found, the payment request is frozen and a compliance alert is generated. For payment requests classified as high-risk, the exemption threshold will be automatically lowered, and the payment request will be forcibly transferred to a multi-level review process.

[0056] In one feasible implementation, the multi-objective optimized routing module 540 described above is specifically used for: A dynamic channel resource pool is pre-configured, including direct bank connection channels, bill payment channels, virtual account channels, and third-party aggregation channels, and the balance status, transaction fee rate, system availability, and single transaction limit information of each channel are collected in real time. Construct a multi-objective optimization routing module to perform weighted scoring and sorting of each channel based on the amount, payment time constraints, and payment purpose category in the payment intent vector; When the preferred channel is excluded due to insufficient balance or system unavailability, the degradation strategy is automatically activated, and the second-best channel is tried in turn; if all channels are unavailable, the fund aggregation logic is activated, and the fund is allocated through a preset path among the specified accounts before the routing decision is initiated again.

[0057] In one feasible implementation, the multi-objective optimized routing module 540 described above is further used for: Based on the message standard of the target payment channel, load the pre-configured field mapping template and convert the payment intent vector and risk control verification results into the standardized message format required by the channel; Before transmission, the message is digitally signed and encrypted using the national cryptographic algorithm, and the instruction signing is completed through the bank-enterprise direct connection dedicated line or hardware U-Key server. After receiving the processing status code returned by the bank, identify false failures and initiate a status polling mechanism; if it is confirmed to be a substantive failure, perform retry, channel switching or appointment delay processing according to the reason for failure. After successful payment, the electronic receipt is automatically downloaded, key fields are parsed and uniquely associated with the original payment request, and then synchronously written back to the business system and automatically generated accounting vouchers.

[0058] In one feasible implementation, the apparatus further includes an optimization module for: After the payment instruction is executed, the payment intent vector, risk control interception log, routing decision path and execution result data are structured, labeled and persistently stored. Based on continuously accumulated labeled data, the training corpus of the payment intent recognition model is periodically updated, and the dynamic risk control rule set and multi-objective optimization routing module are iteratively optimized online.

[0059] In one feasible implementation, the device further includes an error notification module for: When the payment purpose category is identified as tax payment or fiscal payment, the system automatically verifies whether the payee's unique identifier belongs to the tax target account; if they do not match, the system blocks the generation of payment instructions and prompts an abnormality in the payee account.

[0060] The intelligent risk control routing device for corporate payments provided in this application has the same implementation principle and technical effects as the aforementioned method embodiments. For the sake of brevity, any parts of the intelligent risk control routing device for corporate payments not mentioned in the embodiments can be referred to the corresponding content in the aforementioned intelligent risk control routing method embodiments for corporate payments.

[0061] This application also provides an electronic device, such as... Figure 6 The diagram shows the structure of the electronic device 100, which includes a processor 61 and a memory 60. The memory 60 stores computer-executable instructions that can be executed by the processor 61. The processor 61 executes the computer-executable instructions to implement any of the above-mentioned intelligent risk control routing methods for corporate payments.

[0062] exist Figure 6 In the illustrated embodiment, the electronic device further includes a bus 62 and a communication interface 63, wherein the processor 61, the communication interface 63, and the memory 60 are connected via the bus 62.

[0063] The memory 60 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 63 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 62 may be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. The bus 62 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 6 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0064] Processor 61 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 61 or by instructions in software form. The processor 61 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in the memory. The processor 61 reads the information in the memory and, in conjunction with its hardware, completes the steps of the intelligent risk control routing method for corporate payments in the aforementioned embodiment.

[0065] This application also provides a computer-readable storage medium storing computer-executable instructions. When these computer-executable instructions are called and executed by a processor, they cause the processor to implement the aforementioned intelligent risk control routing method for corporate payments. For specific implementation details, please refer to the foregoing method embodiments, which will not be repeated here.

[0066] The computer program product of the intelligent risk control routing method, electronic device and storage medium for corporate payments provided in this application includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here.

[0067] Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application.

[0068] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A smart risk control routing method for corporate bank payments, characterized in that, include: Collect raw document data from multiple heterogeneous business systems within the enterprise, perform document parsing and recognition on the raw document data, and generate payment requests to be processed. By using a pre-trained payment intent recognition model, the target intent features corresponding to the payment intent are extracted from the payment request to be processed, and a payment intent vector is generated. The target intent features include payment purpose category, unique identifier of the payee, settlement currency type, payment time constraints, and tax processing attributes. Based on the target intent characteristics and the pre-configured dynamic risk control rule set, the payment request is intercepted in advance to determine the payment request that passes the risk control verification. The payment request that has passed the risk control verification and its corresponding payment intent vector are input into the multi-objective optimization routing module so that the target payment channel can be selected according to the current availability status of each available payment channel, and a payment instruction that meets the protocol requirements of the selected channel can be generated for encrypted transmission and automatic feedback of execution results.

2. The intelligent risk control routing method for corporate payments according to claim 1, characterized in that, The system collects raw document data from multiple heterogeneous business systems within the enterprise, performs document parsing and recognition on the raw document data, and generates payment requests to be processed, including: By connecting to the purchase order system, contract management system, expense reimbursement system, and payment request system through middleware, structured document data and unstructured attachments can be obtained. Optical character recognition and layout analysis are performed on the unstructured attachments to extract text content and perform semantic cleaning, misspelling normalization and word segmentation. The cleaned text content is merged with the structured fields corresponding to the structured document data to generate a payment request to be processed.

3. The intelligent risk control routing method for corporate payments according to claim 1, characterized in that, Using a pre-trained payment intent recognition model, target intent features corresponding to the payment intent are extracted from the payment request to be processed, generating a payment intent vector, including: The payment request is processed by named entity recognition and intent classification using the BERT-BiLSTM-CRF joint model, which extracts the payee's name, bank account information, keywords for payment purpose, urgency indicators and tax-related statements. The extraction results are mapped to a three-level purpose classification system, a standardized payee master data identifier, a currency type identifier, a time constraint quantification window, and a tax attribute marker, respectively, to obtain the payment purpose category, payee unique identifier information, settlement currency type, payment time constraint conditions, and tax processing attributes. When some target intent features are missing, context completion is performed based on historical transaction behavior patterns or the status is marked as pending confirmation, and manual review is triggered.

4. The intelligent risk control routing method for corporate payments according to claim 1, characterized in that, Based on the target intent characteristics and a pre-configured dynamic risk control rule set, pre-emptive risk interception is performed on payment requests to determine those that pass the risk control verification, including: Semantic fingerprint calculation is performed, and the cosine similarity of the current payment request summary and the historical payment request summary is compared. When the similarity exceeds a preset threshold and the amounts are similar, it is determined to be a potential duplicate payment. Within a set time window, multiple payment requests corresponding to the same payee identifier are detected. If there are characteristics of splitting the amount and the cumulative amount is close to the amount of a certain payment already made, a risk warning for split payment is triggered. The system calls upon real-time updated external sanctions lists and internal databases of untrustworthy suppliers to perform a thorough matching of the unique identifier information of the payee; if a match is found, the payment request is frozen and a compliance alert is generated. For payment requests classified as high-risk, the exemption threshold will be automatically lowered, and the payment request will be forcibly transferred to a multi-level review process.

5. The intelligent risk control routing method for corporate payments according to claim 1, characterized in that, The payment request that has passed the risk control verification and its corresponding payment intent vector are input into the multi-objective optimized routing module to select a target payment channel based on the current availability of each available payment channel, and generate a payment instruction that conforms to the protocol requirements of the selected channel for encrypted transmission and automatic feedback of execution results, including: A dynamic channel resource pool is pre-configured, including direct bank connection channels, bill payment channels, virtual account channels, and third-party aggregation channels, and the balance status, transaction fee rate, system availability, and single transaction limit information of each channel are collected in real time. Construct a multi-objective optimization routing module to perform weighted scoring and sorting of each channel based on the amount, payment time constraints, and payment purpose category in the payment intent vector; When the preferred channel is excluded due to insufficient balance or system unavailability, the degradation strategy is automatically activated, and the second-best channel is tried in turn; if all channels are unavailable, the fund aggregation logic is activated, and the fund is allocated through a preset path among the specified accounts before the routing decision is initiated again.

6. The intelligent risk control routing method for corporate payments according to claim 1, characterized in that, Generate payment instructions that comply with the selected channel protocol for encrypted transmission and automatic feedback of execution results, including: Based on the message standard of the target payment channel, load the pre-configured field mapping template and convert the payment intent vector and risk control verification result into the standardized message format required by the channel; Before transmission, the message is digitally signed and encrypted using the national cryptographic algorithm, and the instruction signing is completed through the bank-enterprise direct connection dedicated line or hardware U-Key server. After receiving the processing status code returned by the bank, identify false failures and initiate a status polling mechanism; if it is confirmed to be a substantive failure, perform retry, channel switching or appointment delay processing according to the reason for failure. After successful payment, the electronic receipt is automatically downloaded, key fields are parsed and uniquely associated with the original payment request, and then synchronously written back to the business system and automatically generated accounting vouchers.

7. The intelligent risk control routing method for corporate payments according to claim 1, characterized in that, The method further includes: After the payment instruction is executed, the payment intent vector, risk control interception log, routing decision path and execution result data are structured, labeled and persistently stored. Based on continuously accumulated labeled data, the training corpus of the payment intent recognition model is periodically updated, and the dynamic risk control rule set and multi-objective optimization routing module are iteratively optimized online.

8. The intelligent risk control routing method for corporate payments according to claim 1, characterized in that, The method further includes: When the payment purpose category is identified as tax payment or fiscal payment, the system automatically verifies whether the payee's unique identifier information belongs to the tax target account; if they do not match, the system blocks the generation of payment instructions and prompts an abnormality in the payee account.

9. An electronic device, characterized in that, The system includes a processor and a memory, the memory storing computer-executable instructions that can be executed by the processor, the processor executing the computer-executable instructions to implement the intelligent risk control routing method for corporate payments as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the intelligent risk control routing method for corporate payments as described in any one of claims 1 to 7.

Images