Streamline processing method and system based on deep learning model

Abstract

The invention discloses a pipeline processing method and system based on a deep learning model, relates to the technical field of data processing, can be applied to the medical field and the financial field, and mainly aims to solve the problem that the pipeline processing efficiency is low in an existing complex and batch business scene. The method mainly comprises the following steps: determining a target business system according to flow information of a to-be-processed flow through a flow sorting engine, and issuing the to-be-processed flow to the target business system; identifying an associated contract matched with the to-be-processed flow and an associated fund in the associated contract according to the flow information through the target business system, and performing multi-dimensional cross check on the associated fund; under the condition that the multi-dimensional cross check is passed, performing multi-dimensional feature extraction on the pipeline information through a pre-trained deep learning model, and determining a cancel-after-verification rule combination according to an extracted multi-dimensional feature vector; and according to the cancel-after-verification rule combination, the amount of money of the to-be-processed flow is allocated to the associated money to complete cancel-after-verification. The method is mainly used for processing transaction flows.

Description

Technical Field

[0001] This invention relates to the field of communication data processing technology, which can be applied to digital healthcare and finance, and in particular to a pipeline processing method and system based on a deep learning model. Background Technology

[0002] With the deepening of digital transformation, the information systems of enterprises and institutions (such as medical institutions and financial institutions) are gradually forming an architecture where multiple business systems coexist. Taking the medical field as an example, hospitals need to manage hospital information systems, financial systems, and medical insurance settlement systems simultaneously; the financial field involves core transaction systems, risk control systems, and general ledger systems. Different systems achieve business collaboration through transaction logs (such as payment records and settlement documents), but the processing of these logs often relies on manual sorting or preset rule engines, resulting in low efficiency, poor flexibility, and difficulty in ensuring cross-system data consistency. Especially in scenarios such as medical billing and reconciliation, transaction logs need to be accurately matched with business elements such as contracts and payments, and traditional methods are difficult to cope with complex business rules and dynamically changing requirements.

[0003] Existing transaction processing systems typically rely on fixed rules to identify target business systems. However, business rules in fields such as healthcare and finance are complex and dynamically changing (e.g., adjustments to medical insurance policies, updates to financial products). Fixed rules cannot cover all scenarios, leading to incorrect transaction issuance or the need for manual reassignment, increasing processing delays. Transaction reconciliation requires precise matching of contracts and payments, but traditional methods, based on simple field matching, cannot handle complex situations such as contract changes and payment splits, easily causing reconciliation errors or omissions. Therefore, this results in low transaction processing efficiency. Summary of the Invention

[0004] In view of this, the present invention provides a pipeline processing method and system based on a deep learning model, the main purpose of which is to solve the problem of low pipeline processing efficiency in existing complex, batch business scenarios.

[0005] According to one aspect of the present invention, a pipeline processing method based on a deep learning model is provided, comprising: The process flow sorting engine determines the target business system based on the flow information of the flow to be processed, and sends the flow to be processed to the target business system. The target business system identifies the associated contracts and related payments in the contracts that match the transaction details based on the transaction information, and performs multi-dimensional cross-validation on the related payments. If the multidimensional cross-validation passes, a pre-trained deep learning model is used to extract multidimensional features from the flow information, and the combination of reimbursement rules is determined based on the extracted multidimensional feature vectors. According to the aforementioned reconciliation rules, the amount of the pending transaction is allocated to the related funds to complete the reconciliation.

[0006] Furthermore, before determining the target business system based on the flow information of the flow to be processed through the flow sorting engine, the method further includes: The system acquires the flow rate, flow characteristic distribution, and system performance data of the transaction to be processed, wherein the flow characteristic distribution includes the median amount and time density, and the system performance data includes resource utilization, processing latency, and processing error rate. The expected slice data is obtained by predicting the flow rate, the flow characteristic distribution, and the system performance data through a flow slice model. The expected slice data includes the number of slices and the degree of parallelism. The pipeline to be processed is sliced ​​according to the slice quantity, and the sliced ​​pipeline to be processed is sent to the pipeline clearing engine according to the parallelism, so that in the steps of determining the target business system based on the pipeline information of the pipeline to be processed and subsequent steps, each slice subset of the pipeline to be processed is processed respectively.

[0007] Furthermore, the step of determining the target business system based on the flow information of the flow to be processed includes: The clearing strategy is retrieved, and the clearing strategy includes a strategy generated by a speech recognition model based on real-time voice commands; Based on at least one of the transaction type judgment statement, account matching judgment statement, and tag type judgment statement in the clearing strategy, the corresponding content in the transaction information is judged to obtain the business type of the transaction to be processed. Identify the target business system that matches the business type from the business system mapping relationship set.

[0008] Furthermore, the step of identifying the associated contracts matching the transaction details to be processed and the associated payments within those contracts based on the transaction information includes: The structured data of the flow information is numerically standardized to obtain the standardized structured data of the flow to be processed; The transaction to be processed that matches the standardized structure data with a preset blacklist or special business tag is designated as a special transaction, and the rest of the transaction to be processed is designated as a normal transaction. For the special transaction flow, a transaction flow processing strategy matching the special transaction flow is retrieved to process the special transaction flow, wherein the transaction flow processing strategy includes a related contract identification strategy based on binding relationship, a blacklist account restriction strategy, and an overdue restriction strategy; For the ordinary transaction records, the structured numerical information of the ordinary transaction records is matched with the contract attribute information using a single-dimensional attribute, and the contracts that match the ordinary transaction records using a single-dimensional attribute are designated as associated contracts, and the associated contracts are bound to their respective ordinary transaction records. For ordinary transaction records that do not match associated contracts during single-dimensional attribute matching, keywords are extracted from the text information of the ordinary transaction records, and multi-dimensional information similarity matching is performed based on the keywords and structured numerical information. Contracts with similarity greater than a preset similarity threshold are identified as associated contracts, and the associated contracts are bound to their respective corresponding ordinary transaction records. Extract at least one payment from the associated contract as the associated payment for each pending transaction. For each transaction to be processed, the structured numerical information and keywords are matched and verified with their respective associated funds from the dimensions of amount, period, and pending verification status, respectively, to obtain the multi-dimensional cross-verification results of the transaction to be processed.

[0009] Furthermore, the step of extracting multi-dimensional features from the flow information using a pre-trained deep learning model, and determining the combination of reimbursement rules based on the extracted multi-dimensional feature vectors, includes: Numerical and textual features are extracted from the transaction information through a feature extraction layer. The numerical features include monetary features, time features, and statistical features, while the textual features include semantic embeddings, keywords, and named entities. The numerical features and the text features are concatenated and fused to obtain a multidimensional feature vector; The attention mechanism fusion layer captures the correlation between the multidimensional feature vector and different reimbursement rules, and generates a combination of reimbursement rules based on the amount feature and the time feature.

[0010] Furthermore, the combination of reversal rules includes the priority of each reversal rule, and the method further includes: Obtain the transaction characteristics, account status, and contract status of the transaction to be processed, as well as the reversal success rate and processing time of different reversal rules within historical time periods; If the number of pending transactions with the same flow characteristics exceeds a preset threshold, the priority of each reimbursement rule is updated and calculated based on the flow characteristics using a reinforcement learning algorithm; or, For each reconciliation rule, a reconciliation efficiency parameter is calculated based on the reconciliation success rate and the processing time, and the priority of each reconciliation rule is updated based on the reconciliation efficiency parameter.

[0011] Furthermore, after allocating the amount of the pending transaction to the related funds according to the reconciliation rules to complete the reconciliation, the method further includes: After successful reconciliation, the reconciliation data is sent to the financial system to generate financial vouchers based on the financial system. The financial documents and the reconciliation data are sent to the reconciliation engine to complete the reconciliation process based on the reconciliation engine.

[0012] According to another aspect of the present invention, a pipeline processing system based on a deep learning model is provided, including a pipeline sorting engine and a target business system; The process flow sorting engine is used to determine the target business system based on the process flow information of the process to be processed, and to send the process flow to be processed to the target business system. The target business system is used to identify the associated contracts and related payments in the contracts that match the transaction information based on the transaction information to be processed, and to perform multi-dimensional cross-validation on the related payments; if the multi-dimensional cross-validation passes, a pre-trained deep learning model is used to extract multi-dimensional features from the transaction information, and a combination of reimbursement rules is determined based on the extracted multi-dimensional feature vectors; according to the combination of reimbursement rules, the amount of the transaction to be processed is allocated to the related payments to complete the reimbursement.

[0013] According to another aspect of the present invention, a storage medium is provided, wherein at least one executable instruction is stored therein, the executable instruction causing a processor to perform operations corresponding to the pipelined processing method based on the deep learning model described above.

[0014] According to another aspect of the present invention, a computer device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the pipelined processing method based on the deep learning model described above.

[0015] By employing the above-described technical solutions, the technical solutions provided by the embodiments of the present invention have at least the following advantages: This invention provides a transaction processing method and system based on a deep learning model. First, a transaction clearing engine determines the target business system based on the transaction information of the transaction to be processed and sends the transaction to the target business system. The target business system then identifies the associated contracts and related payments within the associated contracts that match the transaction information, and performs multi-dimensional cross-validation on the related payments. If the multi-dimensional cross-validation passes, a pre-trained deep learning model extracts multi-dimensional features from the transaction information and determines a combination of reimbursement rules based on the extracted multi-dimensional feature vectors. According to the combination of reimbursement rules, the amount of the transaction to be processed is allocated to the related payments to complete the reimbursement. This invention accurately locates the target business system through a transaction clearing engine, achieving efficient transaction distribution. The target business system can accurately identify associated contracts and payments and perform multi-dimensional validation. After successful validation, the pre-trained deep learning model extracts multi-dimensional features of the transaction to determine the combination of reimbursement rules, thereby completing the reimbursement. Compared to existing technologies, this method is more accurate and efficient, reduces manual intervention, and significantly improves the quality and efficiency of transaction processing.

[0016] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart of a pipeline processing method based on a deep learning model provided by an embodiment of the present invention is shown; Figure 2 A flowchart of another pipeline processing method based on a deep learning model provided by an embodiment of the present invention is shown; Figure 3 This invention provides an internal fund allocation matching and reconciliation flowchart according to an embodiment of the present invention. Figure 4 This invention provides a flowchart of a bank-lease contract transaction matching and reconciliation process. Figure 5 This diagram illustrates a block diagram of a deep learning-based water treatment system according to an embodiment of the present invention. Figure 6 A schematic diagram of the structure of a computer device provided in an embodiment of the present invention is shown. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0019] This invention provides a pipeline processing method based on a deep learning model, implemented through a pipeline processing system based on a deep learning model. The system includes a pipeline clearing engine, multiple business systems, at least one financial system, and a reconciliation engine. Each business system is configured with an automatic reconciliation engine for reconciling pipeline transactions. The system can be built on a distributed processing architecture, using microservice clustering, containerized deployment, etc., enabling rapid scaling up and down based on changes in business volume. Decoupling between services improves system stability and reduces system iteration and maintenance costs.

[0020] like Figure 1 As shown, the method includes: 101. Using the flow sorting engine, determine the target business system based on the flow information of the flow to be processed, and send the flow to be processed to the target business system.

[0021] In this embodiment of the invention, the transaction clearing engine analyzes key information of the transaction to be processed (such as transaction type, source channel, institution code, etc.) and combines it with preset business system mapping rules (such as "credit card transactions are sent to the core transaction system" and "medical insurance settlements are sent to the medical insurance clearing system" in the financial field) to dynamically determine the target business system to which the transaction should be distributed and completes automated distribution. If the transaction to be processed fails to match the corresponding template business system or has special tags (such as a transaction needing to be manually split into multiple business parties), it will be automatically marked into the manual clearing task pool for further manual clearing. The clearing results after manual clearing will be used as the learning corpus for the transaction clearing engine to achieve continuous optimization of the transaction clearing engine.

[0022] It's important to note that the pending transaction data is in batches and can be stored as a list. This list contains detailed information for each transaction, such as transaction number, amount, time, direction (in or out), and remarks. Examples include a list of fund transfer transactions for fund transfer reconciliation scenarios and a list of wire transfer transactions for bank lease contract reconciliation scenarios. After obtaining the pending transactions, in addition to sorting them into the corresponding business systems for reconciliation, it's also necessary to store them. This means generating a unique hash value and timestamp for each pending transaction, serving as its digital fingerprint, and storing it in an irreversible hash chain structure to achieve fast and accurate backtracking and location throughout the entire transaction lifecycle.

[0023] 102. Through the target business system, identify the associated contracts and related payments in the associated contracts that match the transaction information, and perform multi-dimensional cross-validation on the related payments.

[0024] In this embodiment of the invention, the target business system retrieves matching related contracts from the contract database based on transaction information (such as contract number, patient ID, transaction time, etc.), further locates specific payments under the contract (such as installment periods for financial loans, fee details for medical projects), and ensures consistency between the transaction details and contract payments (such as amount, time, participating parties, etc.) through cross-validation. For example, for a "loan early repayment" transaction, after matching the loan contract, the system verifies whether the repayment amount equals the remaining principal + handling fee (as stipulated in the contract) and verifies whether the repayment time is within the early repayment period allowed by the contract. As another example, when a patient pays for a combined "CT scan + medication" fee, the system matches the corresponding medical contract based on the outpatient number, verifies whether the transaction amount equals the CT scan fee (500 yuan) + medication fee (200 yuan) in the contract, and verifies whether the payment time is on the day of the visit to prevent duplicate charges or cross-period reimbursement. Multi-dimensional verification (amount, time, business rules) covers hidden errors (such as incorrect calculation of handling fees in financial transactions, omissions or duplicate charges in medical projects), improving the reconciliation accuracy rate to over 99%.

[0025] 103. If the multidimensional cross-validation passes, the pre-trained deep learning model is used to extract multidimensional features from the flow information, and the combination of reimbursement rules is determined based on the extracted multidimensional feature vectors.

[0026] In this embodiment of the invention, after multidimensional cross-validation passes, the system calls a pre-trained multidimensional matching model to perform end-to-end feature learning on the contextual data of transaction information (amount, time, account, remarks, etc.) and related contract payments. For example, the model can numerically encode structured data (such as amount and timestamp) and extract semantic features from unstructured data (such as text like "emergency" and "installment" in remarks, and account transaction history) using NLP and graph embedding techniques to form a unified multidimensional feature vector. A self-attention mechanism is used to capture the dynamic relationships between features (such as the matching degree between "amount" and "remaining contract amount," and the constraint relationship between "time" and "contract validity period"), replacing manually preset fixed rules. For example, in financial transactions, "early repayment amount = remaining principal × 0.95" (contractual discount). When the model learns that "time → contract validity period is less than 30 days," it automatically adjusts the discount coefficient to 0.98, optimizing the allocation of the write-off amount.

[0027] It's worth noting that the model automatically uncovers complex relationships hidden within multidimensional data, matching multidimensional rules to identify and combine reimbursement rules for orders. This replaces hundreds of fixed rules summarized manually, reducing rule maintenance costs by over 80%. After determining the reimbursement rule combinations, the model automatically assigns reimbursements based on combination priority, automatically completing the reimbursement of both borrower and lender transactions, significantly improving the accuracy and efficiency of reimbursement. Furthermore, in scenarios involving adjustments to financial policies (such as changes in the LPR interest rate) or updates to the medical insurance catalog, the model quickly adapts to new rules through online learning, eliminating the need for system downtime for upgrades.

[0028] 104. According to the aforementioned reconciliation rules, allocate the amount of the pending transaction to the related funds to complete the reconciliation.

[0029] In this embodiment of the invention, the reconciliation rule combination includes the priority of different reconciliation rules. During the reconciliation process based on the reconciliation rule combination, the amount of the transaction to be processed is split in real time according to the rule priority, and the remaining reconcilable amount of related funds is automatically matched to ensure that the allocation result complies with business constraints. If the reconciliation fails due to rule constraints, such as financial transactions triggering the "minimum repayment amount" protection, the system automatically triggers a rollback process, restores the state before allocation, and pushes alarm information to business personnel. After the reconciliation is completed, the processing status of the corresponding related funds in the corresponding contract and the relevant data in the financial system (such as the general ledger system and the clearing system) are updated synchronously, and standardized reconciliation vouchers are generated. Through dynamic rule combination and amount splitting algorithm, it is ensured that the allocation result is 100% compliant with business constraints. In addition, the reconciliation process time is greatly shortened, and the reconciliation efficiency is improved while ensuring the accuracy of the reconciliation.

[0030] In one embodiment of the present invention, for further illustration and limitation, such as Figure 2As shown, before determining the target business system based on the flow information of the flow to be processed using the flow sorting engine, the method further includes: 201. Obtain the flow rate, flow characteristic distribution, and system performance data of the water to be processed.

[0031] 202. The expected slice data is obtained by predicting the flow rate, the flow characteristic distribution, and the system performance data through the flow slice model.

[0032] 203. The pipeline to be processed is sliced ​​according to the slice quantity, and the sliced ​​pipeline to be processed is sent to the pipeline clearing engine according to the parallelism, so that in the steps of determining the target business system based on the pipeline information of the pipeline to be processed and subsequent steps, each pipeline slice subset to be processed is processed respectively.

[0033] In this embodiment of the invention, to improve system throughput and processing efficiency, the transaction flow to be processed is sliced ​​before the target business system is determined based on the transaction flow information using the transaction flow sorting engine. Specifically, various data of the transaction flow to be processed are collected, including transaction volume, representing the total number of transactions; transaction flow characteristic distribution, including the median amount reflecting the median level of transaction amount and the time density reflecting the concentration of transactions in the time dimension, i.e., the transaction flow characteristic distribution includes the median amount and time density. System performance data includes resource utilization, such as the usage of resources like CPU and memory, processing latency, and processing error rate, i.e., system performance data includes resource utilization, processing latency, and processing error rate. The collected data is then predictively processed using a transaction flow slicing model to obtain expected slice data, which includes the number of slices and the degree of parallelism. The number of slices represents the number of subsets into which the transaction flow is divided, and the degree of parallelism represents the extent to which these subsets are processed simultaneously. Finally, the pipeline to be processed is divided into multiple subsets according to the number of slices, and these subsets are sent to the pipeline clearing engine based on the degree of parallelism. This allows for subsequent determination of the target business system for each subset and for reconciliation operations on each subset. By automatically adjusting the slicing strategy based on real-time business data and system performance, dynamic selection of slicing methods and quantities can be achieved, thereby improving system throughput and processing efficiency.

[0034] In one embodiment of the present invention, for further explanation and limitation, the step of determining the target business system based on the flow information of the flow to be processed includes: Retrieve clearing strategy; Based on at least one of the transaction type judgment statement, account matching judgment statement, and tag type judgment statement in the clearing strategy, the corresponding content in the transaction information is judged to obtain the business type of the transaction to be processed. Identify the target business system that matches the business type from the business system mapping relationship set.

[0035] In this embodiment of the invention, the clearing strategy includes transaction type judgment statements, account matching judgment statements, and tag type judgment statements. These judgment statements are used to evaluate the relevant content in the transaction log information to determine the business type of the transaction to be processed. For example, the transaction type judgment statement can determine whether the transaction is a wire transfer, mail transfer, or other type; the account matching judgment statement can check whether the remittance account and the receiving account meet specific rules; and the tag type judgment statement can check whether the transaction has a specific tag. Then, from the business system mapping relationship set, based on the business type determined above, the target business system that matches it is found. The business system mapping relationship set is a pre-established set of correspondences between different business types and business systems. For example, the transaction type judgment statement determines that the transaction method is a wire transfer; the account matching judgment statement finds that the remittance account is not the bank's own account, and the receiving account is a corporate loan customer's account; since there are no manual tags, the tag type judgment statement also meets certain conditions. Combining these judgments, the business type of the transaction is determined to be a corporate loan repayment transaction. Then, based on the corporate loan repayment transaction, the target system is matched from the business system mapping relationship set to the bank's corporate loan business processing system. By comprehensively judging transaction information through multiple conditional statements, the business type can be more accurately determined, avoiding manual clearing and ensuring that the transaction data can be processed correctly, thereby improving the clearing efficiency in complex business scenarios.

[0036] It should be noted that the clearing strategy can be retrieved from preset strategies or strategies generated based on real-time voice configuration commands. For example, in specific business scenarios, staff can issue commands via voice, and the voice recognition model can convert these commands into executable strategy parameters. Dynamically adjusting the clearing strategy through real-time voice input increases the convenience and flexibility of strategy configuration.

[0037] In one embodiment of the present invention, for further explanation and limitation, the step of identifying the associated contract matching the transaction to be processed and the associated payment in the associated contract based on the transaction information includes: The structured data of the flow information is numerically standardized to obtain the standardized structured data of the flow to be processed; The transaction to be processed that matches the standardized structure data with a preset blacklist or special business tag is designated as a special transaction, and the rest of the transaction to be processed is designated as a normal transaction. For the special flow, a flow processing strategy that matches the special flow is retrieved and processed. For the ordinary transaction records, the structured numerical information of the ordinary transaction records is matched with the contract attribute information using a single-dimensional attribute, and the contracts that match the ordinary transaction records using a single-dimensional attribute are designated as associated contracts, and the associated contracts are bound to their respective ordinary transaction records. For ordinary transaction records that do not match associated contracts during single-dimensional attribute matching, keywords are extracted from the text information of the ordinary transaction records, and multi-dimensional information similarity matching is performed based on the keywords and structured numerical information. Contracts with similarity greater than a preset similarity threshold are identified as associated contracts, and the associated contracts are bound to their respective corresponding ordinary transaction records. Extract at least one payment from the associated contract as the associated payment for each pending transaction. For each transaction to be processed, the structured numerical information and keywords are matched and verified with their respective associated funds from the dimensions of amount, period, and pending verification status, respectively, to obtain the multi-dimensional cross-verification results of the transaction to be processed.

[0038] In this embodiment of the invention, blacklist and whitelist management is typically used to handle special cases, such as high-risk customers or special customers. Special business tags are identifiers carried by certain special transaction records. By matching a preset blacklist or special business tags with structured numerical information, the transactions to be processed are divided into special transactions and ordinary transactions. There can be multiple ordinary and special transactions, and the transaction information is stored in list form. In specific scenarios, the list of transactions to be processed can also be filtered into a special transaction list and a normal transaction list.

[0039] For special transaction flows, processing strategies are implemented based on several approaches, including a linked contract identification strategy, a blacklist account restriction strategy, and an overdue restriction strategy. The linked contract identification strategy directly identifies related contracts based on the binding relationship between specific business tags and contracts. Taking bank-leasing cooperation projects as an example, each project typically has a specific loan receipt number linked to a contract. When a transaction flow related to this cooperation project occurs, the system can automatically identify the loan receipt number and match it with the corresponding bank-leasing cooperation contract, achieving accurate transaction flow processing in special business scenarios and ensuring a clear correspondence between business and contract. The blacklist account restriction strategy imposes corresponding restrictions on customers on a blacklist. For example, for transfer flows from blacklisted customers, the system automatically identifies and performs special processing, such as restricting transfers or marking them as risky. The overdue restriction strategy processes transactions related to contracts at different overdue stages according to preset rules. For example, when a contract is severely overdue, the system restricts the write-off of certain types of transactions or adopts special processing procedures to adapt to different business risk situations.

[0040] For ordinary transaction records, the structured numerical information is matched with contract attribute information in a single dimension (such as contract number, customer ID, etc.). Contracts with matching information are identified as associated contracts and bound to the system. This is the initial method for screening associated contracts. When the initial screening method fails, keywords are further extracted from the text information and combined with the structured numerical information to perform multi-dimensional information similarity calculations. When the calculated similarity is greater than a preset similarity threshold, the associated contract is identified and bound to the system. The preset similarity threshold can be 95%, or it can be customized according to requirements. This embodiment of the invention does not impose specific limitations.

[0041] In one embodiment of the present invention, for further explanation and limitation, the step of extracting multi-dimensional features from the flow information using a pre-trained deep learning model and determining the combination of reimbursement rules based on the extracted multi-dimensional feature vectors includes: Numerical and textual features are extracted from the flow information through a feature extraction layer. The numerical features and the text features are concatenated and fused to obtain a multidimensional feature vector; The attention mechanism fusion layer captures the correlation between the multidimensional feature vector and different reimbursement rules, and generates a combination of reimbursement rules based on the amount feature and the time feature.

[0042] In this embodiment of the invention, the pre-trained deep learning model includes a feature extraction layer and an attention mechanism fusion layer. The feature extraction layer extracts features from the numerical components (such as amount, time, statistical data, etc.) and textual components (such as semantics, keywords, named entities, etc.) of the transaction information. For numerical features, appropriate numerical processing methods are used to extract key numerical features, including amount features, time features, and statistical features. For textual features, natural language processing techniques, such as word embedding and keyword extraction algorithms, are used to obtain textual features, including semantic embedding, keywords, and named entities. The extracted numerical and textual features are then concatenated and fused to form a multi-dimensional feature vector. Subsequently, the attention mechanism fusion layer is used to learn the correlation between the multi-dimensional feature vector and different reimbursement rules.

[0043] The feature extraction layer can include a numerical feature extraction sublayer and a text feature extraction sublayer. The numerical feature extraction sublayer can be built based on a Multilayer Perceptron (MLP). An MLP consists of multiple fully connected layers. The first fully connected layer maps monetary data to a higher-dimensional space, and subsequent layers further extract more abstract features, thereby learning the non-linear relationships in the monetary data. Transaction timestamps can be converted into various time features, such as year, month, day, hour, and minute. Similarly, an MLP is used to extract and transform these time features to capture time-related patterns, such as certain transactions occurring more frequently within specific time periods. Statistical features can be calculated using statistical quantities as features, such as the transaction frequency, average transaction amount, and maximum transaction amount of an account within a certain period.

[0044] The text feature extraction sublayer can include pre-trained word embedding models, such as BERT and ELMo; keyword extraction algorithms, such as TF-IDF and TextRank; and named entity recognition models, such as those based on Conditional Random Fields (CRF). Word embedding models capture semantic information in text, mapping similar words to similar vector spaces to convert text data into semantic vectors. Keyword extraction algorithms extract important keywords from text such as notes and use word embedding techniques to convert these keywords into vector representations. Named entity recognition models can extract entity names such as person names, place names, and organization names. Named entities can provide important information about counterparties and business background, which are also converted into vector representations.

[0045] The attention mechanism fusion layer is used to capture the correlation between multi-dimensional feature vectors and different reimbursement rules. The main idea of ​​this layer is to dynamically adjust the contribution of different parts of the feature vector to the reimbursement rules based on their importance. Specifically, this involves: calculating attention weights: using an attention network (usually a multilayer perceptron), the attention weight of each feature in the multi-dimensional feature vector is calculated for different reimbursement rules. The attention weight reflects the correlation between the feature and the reimbursement rule; a larger weight indicates a stronger correlation. Weighted fusion: based on the calculated attention weights, the multi-dimensional feature vector is weighted and fused to obtain a weighted feature vector for each reimbursement rule. Generating a combination of reimbursement rules: the weighted feature vector is input into a classifier (such as a Softmax classifier) ​​to generate a probability distribution for each reimbursement rule. Based on the probability distribution, the reimbursement rules with higher probabilities are selected and combined into the final combination of reimbursement rules.

[0046] In one embodiment of the present invention, for further explanation and limitation, the method further includes: Obtain the transaction characteristics, account status, and contract status of the transaction to be processed, as well as the reversal success rate and processing time of different reversal rules within historical time periods; If the number of pending transactions with the same flow characteristics exceeds a preset threshold, the priority of each reimbursement rule is updated and calculated based on the flow characteristics using a reinforcement learning algorithm; or, For each reconciliation rule, a reconciliation efficiency parameter is calculated based on the reconciliation success rate and the processing time, and the priority of each reconciliation rule is updated based on the reconciliation efficiency parameter.

[0047] In this embodiment of the invention, the reconciliation rule combination includes the priority of each reconciliation rule. To ensure that the execution order of the rules can adapt to the actual reconciliation processing status, reinforcement learning or reconciliation rule scoring is used to update the priority of the reconciliation rules. The reinforcement learning method is triggered by the number of pending transactions with the same flow characteristics, while the reconciliation rule scoring can be triggered based on a preset time interval. When the number of pending transactions with the same flow characteristics exceeds a preset threshold (meaning there are many transactions with similar characteristics that need to be processed), the priority of each reconciliation rule is updated and calculated using a reinforcement learning algorithm based on these flow characteristics. The reinforcement learning algorithm can continuously adjust its strategy based on environmental feedback (which may be the result feedback after processing the transactions), thereby determining which reconciliation rule should be used first for transactions with these specific flow characteristics. For each reconciliation rule, a reconciliation efficiency parameter is calculated based on the previously obtained reconciliation success rate and processing time. For example, the success rate of reconciliation can be divided by the processing time to obtain an efficiency value, which can be used as a reconciliation efficiency parameter. The larger the value, the higher the efficiency, and the higher the priority of the corresponding reconciliation rule, so that it can be used first in subsequent processing of the transaction log.

[0048] In one embodiment of the present invention, for further explanation and limitation, after allocating the amount of the pending transaction to the related funds according to the reconciliation rules to complete the reconciliation, the method further includes: After successful reconciliation, the reconciliation data is sent to the financial system to generate financial vouchers based on the financial system. The financial documents and the reconciliation data are sent to the reconciliation engine to complete the reconciliation process based on the reconciliation engine.

[0049] In the embodiments of the present invention, when the amount of the to-be-processed transaction is successfully allocated to the associated funds according to the write-off rules, that is, after the write-off operation is successfully completed, the system sends the relevant data generated during the write-off process, such as the written-off transaction amount, associated contract information, write-off time, etc., to the financial system. After receiving these write-off data, the financial system will generate corresponding financial vouchers according to its own financial rules and logics. The financial vouchers include financial information such as accounting subjects, amounts, debit and credit directions, etc. After the financial system generates the financial vouchers, the system will further send the generated financial vouchers and the write-off data sent to the financial system before to the reconciliation engine. After receiving these data, the reconciliation engine will check the financial vouchers and write-off data according to the preset reconciliation rules and logics. Real-time reconciliation can promptly detect possible differences in financial processing and data recording of the write-off business, such as inconsistent amounts, mismatched information, etc. Once a difference is found, the system can issue a warning in a timely manner so that relevant personnel can conduct investigations and handling in a timely manner to ensure the accuracy and consistency of financial data and safeguard the safety and stability of enterprise fund management.

[0050] In an application example of internal fund transfer matching write-off, the process of transaction processing is as Figure 3 shown. Figure 3 It describes the complete process of fund transfer transactions from acquisition to processing, write-off, and record update, specifically including: 1. Obtain the transaction list: First, obtain the fund transfer transaction list, and the transaction remarks start with "ZJDB".

[0051] 2. Process transactions in a loop: Process the fund transfer transaction list in a loop, and each transaction is operated according to the following steps: 3. Query transfer records: Query transfer records according to the remarks.

[0052] 4. Write-off processing under different conditions: If it is "transfer of wrongly remitted funds by the customer", no write-off is performed and the transfer transaction processing is exited.

[0053] If it is "transfer of new lease disposal funds" and the write-off transaction arrival entity is Tianjin, principal write-off is called.

[0054] If it is "early settlement transfer of new lease" and the write-off transaction arrival entity is Tianjin, early settlement write-off is performed.

[0055] If it is "transfer of new lease purchase price" and the write-off transaction arrival entity is Tianjin, lease write-off is performed.

[0056] 5. Update the record table and exit: Update the transfer record table in batches according to the transfer ID, and then exit the transfer transaction processing.

[0057] The entire process involves using different criteria to determine and specifically verify fund transfer transactions, and updating relevant records accordingly.

[0058] In an application example of bank lease contract transaction matching and reconciliation, the transaction processing procedure is as follows: Figure 4 As shown. Figure 4 This demonstrates the entire process from obtaining the list of wire transfer records to completing automatic claiming and verification. Specifically, it includes: 1. Querying wire transfer records that meet the criteria and returning the filtered records. The system first queries the list of wire transfer records, filters out the records that meet the requirements based on preset criteria, and returns the filtered results, providing basic data for subsequent processing.

[0059] 2. Extract the 16-digit IOU number from the transaction notes. For the filtered transaction records, the system extracts the crucial 16-digit IOU number from the notes of each transaction. This IOU number serves as important identification information and will be used for subsequent matching with the contract.

[0060] 3. Search for contracts from Rongyi Tianfen or Beijing Bank Shangfen based on the loan receipt number. Using the extracted 16-digit loan receipt number, the system searches the contract database of Rongyi Tianfen or Beijing Bank Shangfen to try to find the contract information corresponding to that loan receipt number.

[0061] 4. Return successfully matched contracts. After the query operation, the system will return the information of the successfully matched contracts. This contract information will be used for subsequent operations such as contract status verification.

[0062] 5. Verify whether the rent has been cancelled in the contract and determine if it is overdue. The system verifies the returned contract to check whether the rent cancellation operation has been completed and determines whether the contract is overdue. This step is crucial for determining the subsequent processing procedure.

[0063] 6. If no contract associated with the IOU number is found in the previous steps, the system will extract the lessee's keywords from the relevant data; if an associated contract is found, the subsequent operations will continue according to the normal process.

[0064] 7. When there is no remarks in the contract-related transaction records, the system matches the corresponding lessee based on the payer information. In this way, the lessee entity related to the transaction records can be identified as accurately as possible.

[0065] 8. If there are remarks in the contract-related transaction records, the system will extract the name information from the remarks and use it as a basis for determining the lessee and other relevant information.

[0066] 9. The system assesses the matched or extracted information to determine if only one related contract exists. This assessment will affect subsequent operations such as amount verification.

[0067] 10. Verify that the total principal amount of the bank-lease agreement matches the transaction amount. The system verifies that the total principal amount of the bank-lease agreement matches the corresponding transaction amount. If they match: Perform the bank-lease agreement claim operation, which confirms the association between the transaction and the corresponding bank-lease agreement, completes the relevant write-off operations, and finally ends the automatic claim and write-off process. The entire process, through the sequential execution of multiple steps and conditional judgments, achieves automatic write-off processing of bank-lease agreement transactions, improving the efficiency and accuracy of business processing.

[0068] This invention provides a transaction processing method based on a deep learning model. First, a transaction clearing engine determines the target business system based on the transaction information of the transaction to be processed and sends the transaction to the target business system. The target business system then identifies the associated contracts and related payments within those contracts that match the transaction information, and performs multi-dimensional cross-validation on the related payments. If the multi-dimensional cross-validation passes, a pre-trained deep learning model extracts multi-dimensional features from the transaction information and determines a combination of reimbursement rules based on the extracted feature vectors. According to the reimbursement rule combination, the amount of the transaction to be processed is allocated to the related payments to complete the reimbursement. This invention accurately locates the target business system through a transaction clearing engine, achieving efficient transaction distribution. The target business system can accurately identify associated contracts and payments and perform multi-dimensional validation. After successful validation, the pre-trained deep learning model extracts multi-dimensional features from the transaction to determine the combination of reimbursement rules, thereby completing the reimbursement. Compared to existing technologies, this method is more accurate and efficient, reduces manual intervention, and significantly improves the quality and efficiency of transaction processing.

[0069] Furthermore, as a response to the above Figure 1 The implementation of the method shown in this invention provides a pipeline processing system based on a deep learning model, such as... Figure 5 As shown, the system includes a flow sorting engine 31 and a target business system 32; The flow sorting engine 31 is used to determine the target business system based on the flow information of the flow to be processed, and to send the flow to be processed to the target business system. The target business system 32 is used to identify the associated contracts and related payments in the contracts that match the transaction information based on the transaction information, and to perform multi-dimensional cross-validation on the related payments; if the multi-dimensional cross-validation passes, the system extracts multi-dimensional features from the transaction information using a pre-trained deep learning model, and determines the reconciliation rule combination based on the extracted multi-dimensional feature vectors; and allocates the amount of the transaction to be processed to the related payments to complete the reconciliation according to the reconciliation rule combination.

[0070] Furthermore, the system also includes a slicing engine, which is used to acquire the flow volume, flow characteristic distribution, and system performance data of the transaction to be processed. The flow characteristic distribution includes the median amount and time density, and the system performance data includes resource utilization, processing latency, and processing error rate. The system uses a flow slicing model to predict the flow volume, flow characteristic distribution, and system performance data to obtain expected slice data, which includes the number of slices and the degree of parallelism. The system slices the transaction to be processed according to the number of slices and sends the sliced ​​transaction to the flow sorting engine according to the degree of parallelism, so that in the steps of determining the target business system based on the flow information of the transaction to be processed and subsequent steps, each subset of the transaction slices to be processed is processed separately.

[0071] Furthermore, the clearing engine 31 is specifically used to retrieve the clearing strategy, which includes a strategy generated by a speech recognition model based on real-time voice commands; and to judge the corresponding content in the transaction information based on at least one of the transaction type judgment statement, account matching judgment statement, and tag type judgment statement in the clearing strategy to obtain the business type of the transaction to be processed. Identify the target business system that matches the business type from the business system mapping relationship set.

[0072] Furthermore, the target business system 32 is specifically used to perform numerical standardization on the structured data of the flow information to obtain standardized structured data of the flow to be processed; The transaction to be processed that matches the standardized structure data with a preset blacklist or special business tag is designated as a special transaction, and the rest of the transaction to be processed is designated as a normal transaction. For the special transaction flow, a transaction flow processing strategy matching the special transaction flow is retrieved to process the special transaction flow, wherein the transaction flow processing strategy includes a related contract identification strategy based on binding relationship, a blacklist account restriction strategy, and an overdue restriction strategy; For the ordinary transaction records, the structured numerical information of the ordinary transaction records is matched with the contract attribute information using a single-dimensional attribute, and the contracts that match the ordinary transaction records using a single-dimensional attribute are designated as associated contracts, and the associated contracts are bound to their respective ordinary transaction records. For ordinary transaction records that do not match associated contracts during single-dimensional attribute matching, keywords are extracted from the text information of the ordinary transaction records, and multi-dimensional information similarity matching is performed based on the keywords and structured numerical information. Contracts with similarity greater than a preset similarity threshold are identified as associated contracts, and the associated contracts are bound to their respective corresponding ordinary transaction records. Extract at least one payment from the associated contract as the associated payment for each pending transaction. For each transaction to be processed, the structured numerical information and keywords are matched and verified with their respective associated funds from the dimensions of amount, period, and pending verification status, respectively, to obtain the multi-dimensional cross-verification results of the transaction to be processed.

[0073] Furthermore, the target business system 32 is specifically used to extract numerical features and textual features from the transaction information through a feature extraction layer, wherein the numerical features include monetary features, time features, and statistical features, and the textual features include semantic embeddings, keywords, and named entities; The numerical features and the text features are concatenated and fused to obtain a multidimensional feature vector; The attention mechanism fusion layer captures the correlation between the multidimensional feature vector and different reimbursement rules, and generates a combination of reimbursement rules based on the amount feature and the time feature.

[0074] Furthermore, the target business system 32 is also used to acquire the transaction characteristics, account status, and contract status of the transaction to be processed, as well as the reconciliation success rate and processing time of different reconciliation rules within a historical period; when the number of transaction to be processed with the same transaction characteristics is greater than a preset threshold, the priority of each reconciliation rule is updated and calculated based on the transaction characteristics using a reinforcement learning algorithm; or, for each reconciliation rule, a reconciliation efficiency parameter is calculated based on the reconciliation success rate and the processing time, and the priority of each reconciliation rule is updated based on the reconciliation efficiency parameter.

[0075] Furthermore, the system also includes a financial system and a reconciliation engine; The target business system 32 is also used to send the reconciliation data to the financial system and the reconciliation engine after the reconciliation is successful. The financial system is used to generate financial vouchers based on the reconciliation data and send the financial vouchers to the reconciliation engine; The reconciliation engine is used to perform reconciliation based on the financial documents and the write-off data.

[0076] This invention provides a transaction processing system based on a deep learning model. First, a transaction clearing engine determines the target business system based on the transaction information of the transaction to be processed and sends the transaction to the target business system. The target business system then identifies the associated contracts and related payments within those contracts that match the transaction information, and performs multi-dimensional cross-validation on the related payments. If the multi-dimensional cross-validation passes, a pre-trained deep learning model extracts multi-dimensional features from the transaction information and determines a combination of reimbursement rules based on the extracted feature vectors. According to the reimbursement rule combination, the amount of the transaction to be processed is allocated to the related payments to complete the reimbursement. This invention accurately locates the target business system through a transaction clearing engine, achieving efficient transaction distribution. The target business system can accurately identify associated contracts and payments and perform multi-dimensional validation. After successful validation, the pre-trained deep learning model extracts multi-dimensional features from the transaction to determine the combination of reimbursement rules, thereby completing the reimbursement. Compared to existing technologies, this method is more accurate and efficient, reduces manual intervention, and significantly improves the quality and efficiency of transaction processing.

[0077] According to one embodiment of the present invention, a storage medium is provided, the storage medium storing at least one executable instruction, the computer-executable instruction being able to execute the pipelined processing method based on a deep learning model in any of the above method embodiments.

[0078] Figure 6 The diagram illustrates a structural schematic of a computer device according to an embodiment of the present invention. The specific embodiments of the present invention do not limit the specific implementation of the computer device.

[0079] like Figure 6 As shown, the computer device may include: a processor 402, a communications interface 404, a memory 406, and a communications bus 408.

[0080] The processor 402, communication interface 404, and memory 406 communicate with each other via communication bus 408.

[0081] Communication interface 404 is used to communicate with other network elements such as clients or other servers.

[0082] The processor 402 is used to execute program 410, specifically to execute the relevant steps in the above-described embodiment of the pipelined processing method based on a deep learning model.

[0083] Specifically, program 410 may include program code that includes computer operation instructions.

[0084] Processor 402 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The computer device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0085] Memory 406 is used to store program 410. Memory 406 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0086] Specifically, program 410 can be used to cause processor 402 to perform the following operations: The process flow sorting engine determines the target business system based on the flow information of the flow to be processed, and sends the flow to be processed to the target business system. The target business system identifies the associated contracts and related payments in the contracts that match the transaction details based on the transaction information, and performs multi-dimensional cross-validation on the related payments. If the multidimensional cross-validation passes, a pre-trained deep learning model is used to extract multidimensional features from the flow information, and the combination of reimbursement rules is determined based on the extracted multidimensional feature vectors. According to the aforementioned reconciliation rules, the amount of the pending transaction is allocated to the related funds to complete the reconciliation.

[0087] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing systems. They can be centralized on a single computing system or distributed across a network of multiple computing systems. Optionally, they can be implemented using program code executable by a computing system, thereby storing them in a storage system for execution by the computing system. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0088] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A pipeline processing method based on a deep learning model, characterized in that, include: The process flow sorting engine determines the target business system based on the flow information of the flow to be processed, and sends the flow to be processed to the target business system. The target business system identifies the associated contracts and related payments in the contracts that match the transaction details based on the transaction information, and performs multi-dimensional cross-validation on the related payments. If the multidimensional cross-validation passes, a pre-trained deep learning model is used to extract multidimensional features from the flow information, and the combination of reimbursement rules is determined based on the extracted multidimensional feature vectors. According to the aforementioned reconciliation rules, the amount of the pending transaction is allocated to the related funds to complete the reconciliation.

2. The method according to claim 1, characterized in that, Before determining the target business system based on the flow information of the flow to be processed through the flow sorting engine, the method further includes: The system acquires the flow rate, flow characteristic distribution, and system performance data of the transaction to be processed, wherein the flow characteristic distribution includes the median amount and time density, and the system performance data includes resource utilization, processing latency, and processing error rate. The expected slice data is obtained by predicting the flow rate, the flow characteristic distribution, and the system performance data through a flow slice model. The expected slice data includes the number of slices and the degree of parallelism. The pipeline to be processed is sliced ​​according to the slice quantity, and the sliced ​​pipeline to be processed is sent to the pipeline clearing engine according to the parallelism, so that in the steps of determining the target business system based on the pipeline information of the pipeline to be processed and subsequent steps, each slice subset of the pipeline to be processed is processed respectively.

3. The method according to claim 2, characterized in that, The step of determining the target business system based on the flow information of the flow to be processed includes: The clearing strategy is retrieved, and the clearing strategy includes a strategy generated by a speech recognition model based on real-time voice commands; Based on at least one of the transaction type judgment statement, account matching judgment statement, and tag type judgment statement in the clearing strategy, the corresponding content in the transaction information is judged to obtain the business type of the transaction to be processed. Identify the target business system that matches the business type from the business system mapping relationship set.

4. The method according to claim 1, characterized in that, The step of identifying the associated contracts and related payments in the contracts that match the transaction details based on the transaction information includes: The structured data of the flow information is numerically standardized to obtain the standardized structured data of the flow to be processed; The transaction to be processed that matches the standardized structure data with a preset blacklist or special business tag is designated as a special transaction, and the rest of the transaction to be processed is designated as a normal transaction. For the special transaction flow, a transaction flow processing strategy matching the special transaction flow is retrieved to process the special transaction flow, wherein the transaction flow processing strategy includes a related contract identification strategy based on binding relationship, a blacklist account restriction strategy, and an overdue restriction strategy; For the ordinary transaction records, the structured numerical information of the ordinary transaction records is matched with the contract attribute information using a single-dimensional attribute, and the contracts that match the ordinary transaction records using a single-dimensional attribute are designated as associated contracts, and the associated contracts are bound to their respective ordinary transaction records. For ordinary transaction records that do not match associated contracts during single-dimensional attribute matching, keywords are extracted from the text information of the ordinary transaction records, and multi-dimensional information similarity matching is performed based on the keywords and structured numerical information. Contracts with similarity greater than a preset similarity threshold are identified as associated contracts, and the associated contracts are bound to their respective corresponding ordinary transaction records. Extract at least one payment from the associated contract as the associated payment for each pending transaction. For each transaction to be processed, the structured numerical information and keywords are matched and verified with their respective associated funds from the dimensions of amount, period, and pending verification status, respectively, to obtain the multi-dimensional cross-verification results of the transaction to be processed.

5. The method according to claim 1, characterized in that, The step of extracting multi-dimensional features from the flow information using a pre-trained deep learning model and determining the combination of reimbursement rules based on the extracted multi-dimensional feature vectors includes: Numerical and textual features are extracted from the transaction information through a feature extraction layer. The numerical features include monetary features, time features, and statistical features, while the textual features include semantic embeddings, keywords, and named entities. The numerical features and the text features are concatenated and fused to obtain a multidimensional feature vector; The attention mechanism fusion layer captures the correlation between the multidimensional feature vector and different reimbursement rules, and generates a combination of reimbursement rules based on the amount feature and the time feature.

6. The method according to claim 5, characterized in that, The combination of reversal rules includes the priority of each reversal rule, and the method further includes: Obtain the transaction characteristics, account status, and contract status of the transaction to be processed, as well as the reversal success rate and processing time of different reversal rules within historical time periods; If the number of pending transactions with the same flow characteristics exceeds a preset threshold, the priority of each reimbursement rule is updated and calculated based on the flow characteristics using a reinforcement learning algorithm; or, For each reconciliation rule, a reconciliation efficiency parameter is calculated based on the reconciliation success rate and the processing time, and the priority of each reconciliation rule is updated based on the reconciliation efficiency parameter.

7. The method according to claim 1, characterized in that, After allocating the amount of the pending transaction to the related funds according to the reconciliation rules to complete the reconciliation, the method further includes: After successful reconciliation, the reconciliation data is sent to the financial system to generate financial vouchers based on the financial system. The financial documents and the reconciliation data are sent to the reconciliation engine to complete the reconciliation process based on the reconciliation engine.

8. A pipeline processing system based on a deep learning model, characterized in that, This includes a flow sorting engine and multiple target business systems; The process flow sorting engine is used to determine the target business system based on the process flow information of the process to be processed, and to send the process flow to be processed to the target business system. The target business system is used to identify the associated contracts that match the transaction to be processed and the associated payments in the associated contracts based on the transaction information, and to perform multi-dimensional cross-validation on the associated payments. If the multidimensional cross-validation passes, a pre-trained deep learning model is used to extract multidimensional features from the transaction information, and a combination of reimbursement rules is determined based on the extracted multidimensional feature vectors. According to the combination of reimbursement rules, the amount of the transaction to be processed is allocated to the associated funds to complete the reimbursement.

9. A storage medium, characterized in that, The storage medium stores at least one executable instruction that causes the processor to perform the operation corresponding to the pipelined processing method based on the deep learning model as described in any one of claims 1-7.

10. A computer device, characterized in that, include: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the pipelined processing method based on the deep learning model as described in any one of claims 1-7.

Images