A supply chain inventory invoice intelligent verification management system

Abstract

The present application relates to the technical field of supply chain management, in particular to a supply chain inventory invoice intelligent verification management system to solve the technical problem that the existing invoice verification method cannot dynamically determine the verification priority and allocate verification resources according to business impact, leading to high-risk invoice processing delay and low-risk invoice overuse of resources, the technical scheme is that: the invoice-inventory coupling twin body directed graph is established through the twin construction module, the business urgency score of the invoice to be verified is calculated through the model calculation module, the target processing level of the three-level verification microservice is determined according to the score through the decision scheduling module, and the verification task is executed through the execution module calling the corresponding microservice; the present application can realize real-time coupling of invoice and inventory data, improve verification efficiency and business adaptability, and reduce supply chain business and tax compliance risk.

Description

Technical Field

[0001] This invention relates to the field of supply chain management technology, and more specifically, to a supply chain inventory invoice intelligent verification management system. Background Technology

[0002] In enterprise supply chain management, the verification of purchase invoices is a key link in financial settlement and inventory confirmation. Invoice verification not only involves the accuracy of the information on the invoice, but also requires cross-referencing with the actual inventory received to ensure that the accounts match the actual situation. As the scale of enterprise business expands and the complexity of the supply chain increases, invoice verification faces challenges such as large processing volume, high timeliness requirements, and differences in verification depth.

[0003] In existing technologies, a common approach is to use rule-based automatic invoice verification systems. These systems typically perform the same verification process for each invoice, including extracting invoice information via Optical Character Recognition (OCR), matching it with purchase orders and receiving slips, and verifying tax compliance. The processing hierarchy is fixed; regardless of whether the material corresponding to the invoice is urgently needed for production or a non-critical auxiliary material, and regardless of whether the supplier has a good reputation or has a history of invoicing errors, the system allocates the same verification resources and performs the same level of verification. The problem with this approach is that the allocation of verification resources lacks correlation with business impact. When the material corresponding to an invoice is about to run out and the production line faces the risk of shutdown, that invoice may still be queued, waiting for the same processing flow as a large number of low-risk invoices. This can amplify the verification delays of high-risk invoices into production accidents.

[0004] Another existing technology involves introducing invoice verification and inventory receipt status comparison functions into the inventory management system. These systems query the current inventory quantity during invoice verification and perform a simple comparison with the invoice quantity to determine if there are any anomalies where the invoice quantity exceeds the inventory. While this approach establishes a link between invoices and inventory, this link is static and superficial. The system only knows the current inventory quantity and cannot understand the dynamic impact of inventory level changes and material consumption rates on the urgency of verification. It also cannot comprehensively assess risks through multi-dimensional risk quantification based on supplier historical behavior patterns and inventory health status. This results in a single criterion for determining verification priority, making it impossible to accurately identify business-critical invoices requiring priority processing among multiple invoices awaiting verification.

[0005] Therefore, a smart verification and management system for supply chain inventory invoices is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a supply chain inventory invoice intelligent verification management system to solve the technical problem that existing invoice verification methods cannot dynamically determine verification priorities and allocate verification resources based on business impact, resulting in delays in processing high-risk invoices and excessive resource consumption by low-risk invoices.

[0007] To address the aforementioned technical problems, the present invention aims to provide an intelligent verification and management system for supply chain inventory invoices, comprising: The twin construction module is used to obtain the structured invoice data and corresponding inventory element data of the invoice to be verified from the enterprise invoice pool and inventory management system, respectively. It generates a unique twin identifier for the invoice to be verified and uses the twin identifier as an index to establish an invoice-inventory coupled twin directed graph, with nodes and edges representing the invoice identifier, associated material code, associated supplier identifier, and real-time inventory level pointer. The real-time inventory level pointer is a data link pointing to the current inventory storage location of the material corresponding to the associated material code in the inventory management system. The model calculation module, connected to the twin construction module, is used to receive the directed graph of the invoice-inventory coupled twin, call at least one pre-set supply chain business mechanism model, and calculate the business urgency score of the invoice to be verified; the supply chain business mechanism model includes a material shortage risk assessment sub-model, a supplier uncertainty assessment sub-model, and an inventory health assessment sub-model; The decision scheduling module, connected to the model calculation module, is used to receive the business urgency score, determine the target processing level from the three-level verification microservice level (standard processing level, enhanced verification level, and deep verification level) based on the preset threshold range of the business urgency score, and generate a scheduling instruction carrying the identifier of the target processing level. The execution module, connected to the decision scheduling module, is used to deploy the three-level verification microservice layer in a microservice architecture, receive the scheduling instructions, call the set of microservice instances corresponding to the target processing layer to execute the invoice verification task, and output the verification result.

[0008] In a preferred embodiment, the model calculation module further includes a material shortage risk calculation submodule, a supplier uncertainty calculation submodule, and an inventory health calculation submodule; The material shortage risk calculation submodule is configured to read the current inventory and average daily consumption of the associated materials in the invoice-inventory coupled twin directed graph through the real-time inventory level pointer, normalize the ratio of the current inventory to the average daily consumption, and take the reciprocal to obtain the material shortage risk score. The supplier uncertainty calculation submodule is configured to read the historical invoicing error rate corresponding to the associated supplier identifier in the invoice-inventory coupled twin directed graph, and use the historical invoicing error rate as the supplier uncertainty score; The inventory health calculation submodule is configured to obtain the quantity of associated materials in the directed graph of the invoice-inventory coupling twin, compare the quantity of invoices to be verified with the quantity of materials already in stock, and if the quantity of invoices is greater than the quantity of materials already in stock, output a preset inventory health score indicating an inventory abnormality; wherein, the quantity of materials already in stock is the cumulative historical data value of the associated materials. The business urgency score is obtained by weighting the material shortage risk score, supplier uncertainty score, and inventory health score.

[0009] In a preferred embodiment, the material shortage risk calculation submodule, the supplier uncertainty calculation submodule, and the inventory health calculation submodule are each encapsulated as independent containerized microservices and registered through a service registry. The model calculation module dynamically discovers and calls each submodule by querying the service registry. When the container image of any submodule is updated, the service registry updates the service endpoint of that submodule, and the model calculation module automatically points to the updated service endpoint on the next call.

[0010] In a preferred embodiment, the decision scheduling module internally maintains a standard processing level threshold range, an enhanced verification level threshold range, and a deep verification level threshold range. When the business urgency score falls within the standard processing level threshold range, the target processing level is determined as the standard processing level. When the business urgency score falls within the enhanced verification level threshold range, the target processing level is determined as the enhanced verification level; When the business urgency score falls within the threshold range of the deep verification level, the target processing level is determined to be the deep verification level.

[0011] In a preferred embodiment, the decision scheduling module is further configured to: generate a scheduling instruction carrying a forced real-time synchronization flag when the target processing level is determined to be a deep verification level; The system also includes a cross-system synchronization module, which receives a scheduling instruction carrying the mandatory real-time synchronization flag and exchanges the verification status of the invoice to be verified in real time between the execution module and the external inventory management system in a synchronous communication manner until the verification task is completed.

[0012] In a preferred embodiment, the set of microservice instances corresponding to the standard processing level includes an OCR recognition microservice and a rule comparison microservice, which are deployed on CPU computing nodes; The set of microservice instances corresponding to the enhanced verification level includes OCR recognition microservice, rule comparison microservice, tax compliance verification microservice and tolerance matching microservice, which are deployed on a hybrid architecture of CPU computing nodes and GPU computing nodes; The set of microservice instances corresponding to the deep verification level includes a high-precision OCR microservice, a graph neural network analysis microservice, a full-link cross-validation microservice, and the tax compliance verification microservice. These are deployed on GPU computing nodes and configured to organize multiple verification tasks associated with the same supplier or the same material type into the same batch, sharing the preprocessed data in the GPU computing node's memory.

[0013] In a preferred embodiment, when the material shortage risk calculation submodule, the supplier uncertainty calculation submodule, and the inventory health calculation submodule perform weighted calculations, the weight coefficients of each submodule are determined by the analytic hierarchy process (AHP). Specifically, business experts construct a judgment matrix by comparing material shortage risk, supplier uncertainty, and inventory health pairwise, solve for the eigenvectors of the judgment matrix, and obtain the weight coefficients of each submodule after normalization.

[0014] In a preferred embodiment, the decision scheduling module is further configured to: when there are multiple invoice verification tasks at the same time, arrange them in descending order according to their business urgency scores, and generate scheduling instructions in sequence; during the scheduling process, if the business urgency score of a newly arrived task exceeds the minimum score of the currently executing task, the newly arrived task is allowed to preempt the computing resources of the task with the lowest current score, and the preempted task re-enters the queue to wait for the next scheduling.

[0015] In a preferred embodiment, the set of microservice instances corresponding to the deep verification level reserves a minimum guaranteed number of dedicated GPU resources in the GPU computing nodes. The minimum guaranteed number is set according to the maximum concurrency of deep verification level tasks in the same historical period. When the number of deep verification level tasks exceeds the minimum guaranteed number, the decision scheduling module sends a scaling request to the container orchestration system to increase the number of deep verification level microservice instances.

[0016] In a preferred embodiment, the structured invoice data of the invoice to be verified includes the invoice number, the taxpayer identification number of the buyer and seller, the product name, the product tax classification code, the amount, the tax amount, and the invoice date; the inventory element data includes the material code, real-time inventory quantity, safety stock threshold, average daily consumption, and the quantity already received.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This intelligent verification and management system for supply chain inventory invoices uses a directed graph of invoice-inventory coupled twins. With a unique twin identifier as an index, it strongly maps four core elements—invoices, materials, suppliers, and real-time inventory—through nodes and edges. The real-time inventory level pointer directly points to the current inventory storage location of the corresponding material in the inventory management system, ensuring that the inventory data called during the verification process is the latest real-time data. This reduces data latency, avoids verification deviations caused by static data synchronization, and provides an accurate and real-time data foundation for subsequent business risk assessment.

[0018] 2. In this intelligent verification and management system for supply chain inventory invoices, the business urgency score of the invoices to be verified is quantitatively calculated from three core business dimensions—material shortage risk, supplier uncertainty, and inventory health—through a supply chain business mechanism model. This breaks the limitation of traditional invoice verification, which only focuses on tax compliance, and deeply binds the verification strategy with the actual supply chain business risks of the enterprise, realizing the extension of invoice verification from compliance-oriented to business risk-oriented.

[0019] 3. This intelligent verification management system for supply chain inventory invoices adopts a three-level verification microservice hierarchy and a multi-level scheduling decision mechanism. It matches the corresponding verification processing level according to the business urgency score, matching in-depth verification capabilities for high-risk and high-urgency invoices, and matching lightweight standard processing procedures for low-risk routine invoices. This avoids the waste of resources caused by indiscriminate full verification. At the same time, the microservice architecture enables flexible deployment and invocation of verification capabilities, improving the overall processing efficiency and business adaptability of the system. Attached Figure Description

[0020] Figure 1 This is an overall system block diagram of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Small molecule peptide freeze-dried products are important products in the biopharmaceutical field. Their stability directly affects the efficacy and safety of drugs. At present, whether it is offline sampling and detection or single parameter online monitoring, there are technical problems such as high false judgment rate of stability of single parameter monitoring, delayed early warning, and the inability to achieve both system power consumption and abnormal response speed due to fixed sampling frequency. In view of this, such as Figure 1As shown, the purpose of this invention is to provide a supply chain inventory invoice intelligent verification management system. The system includes four core modules: a twin construction module, a model calculation module, a decision scheduling module, and an execution module. The four modules are connected sequentially in the direction of data flow to form a complete processing link from twin construction to verification result output. In addition, the system also includes a cross-system synchronization module, which is connected to the decision scheduling module and the execution module for real-time data exchange with an external inventory management system.

[0023] Furthermore, the various modules of the system will be made more public.

[0024] The twin construction module is the data entry point and basic modeling unit of the entire system. It is used to obtain the structured invoice data and corresponding inventory element data of the invoice to be verified from the enterprise invoice pool and inventory management system, respectively, generate a unique twin identifier for the invoice to be verified, and establish an invoice-inventory coupled twin directed graph using the twin identifier as an index.

[0025] When a new purchase invoice enters the enterprise's invoice pool, the twin construction module first extracts the structured invoice data from the invoice pool through the data interface. The structured invoice data includes the invoice number, the taxpayer identification number of the buyer and seller, the commodity name, the commodity tax classification code, the amount, the tax amount, and the invoice date. This data can be obtained directly by parsing the invoice information in XML or PDF format when the enterprise receives the electronic invoice, or it can be obtained by extracting the structured data after OCR recognition of the paper invoice.

[0026] Meanwhile, the module parses the material code associated with the invoice by mapping the product name or product tax classification code to the enterprise's material master data. Based on this material code, the module extracts the corresponding inventory element data from the inventory management system. The inventory element data includes the material code, real-time inventory quantity, safety stock threshold, average daily consumption, and quantity already received. The average daily consumption is calculated by the inventory management system based on the material's outbound records over a preset time period, such as the total consumption over the past 30 calendar days divided by 30. The safety stock threshold is the minimum inventory level preset by the enterprise based on procurement lead time and demand fluctuations. The quantity already received is the cumulative quantity of the material that has been received under the purchase order item corresponding to this invoice.

[0027] After obtaining the above data, the twin construction module generates a globally unique twin identifier for the invoice to be verified, such as a unique string generated by UUID or Snowflake algorithm. Using this twin identifier as an index, the module builds a directed graph of invoice-inventory coupled twins.

[0028] A directed graph is a data structure composed of nodes and directed edges. In this invention, the nodes of the directed graph are divided into four categories: invoice identifier nodes, which store the invoice number, amount, tax amount and other invoice attributes; associated material code nodes, which store the material code and its corresponding safety stock threshold, average daily consumption, quantity already received and other material attributes; associated supplier identifier nodes, which store the seller's taxpayer identification number and the supplier's historical behavior data; and real-time inventory level pointer nodes, which store a data link pointing to the current inventory level of the material in the inventory management system, rather than the inventory level value itself.

[0029] The directed graph was chosen as the data structure for the twin because there are inherent many-to-many relationships between invoices, materials, and suppliers. For example, an invoice may contain multiple material rows, each corresponding to a different material code; multiple materials may be supplied by the same supplier; and the real-time inventory of a material is simultaneously compared with multiple invoices awaiting verification. The edges of the directed graph can accurately express the direction and type of these relationships. For instance, a directed edge from the invoice identifier node to the associated material code node indicates that the invoice contains that material; a directed edge from the associated material code node to the real-time inventory level pointer node indicates that the real-time inventory of that material is obtained through that pointer. This graph structure provides a unified and flexible data access foundation for subsequent business mechanism model calculations.

[0030] Setting a real-time inventory level pointer instead of directly storing a copy of the inventory value is a key technical consideration in this invention. The entire invoice verification process can last from several seconds to several minutes. If an inventory value is copied during the twin's construction, this value may have changed due to other inbound or outbound operations during the verification process, causing subsequent business urgency calculations to be based on outdated data. By using a pointer mechanism, the model calculation module dereferences the pointer in real time and reads the inventory from the inventory management system when needed, ensuring that every calculation is based on the latest inventory status. This design decouples the static attribute part and the dynamic real-time part of the twin, balancing the completeness of the modeling with the timeliness of the data.

[0031] The model calculation module, connected to the twin construction module, is used to receive the directed graph of the invoice-inventory coupled twin, call at least one pre-built supply chain business mechanism model, and calculate the business urgency score of the invoice to be verified. The model calculation module includes a material shortage risk calculation submodule, a supplier uncertainty calculation submodule, and an inventory health calculation submodule.

[0032] Once the twin construction module completes the construction of the directed graph, it sends the complete graph data structure to the model calculation module. This module, as the pre-calculation unit for scheduling decisions, is responsible for quantifying the multi-dimensional business information associated with the invoice into a single, comparable business urgency score.

[0033] The calculation principles and implementation methods of the three sub-modules are explained below.

[0034] The material shortage risk calculation submodule is configured to read the current inventory and average daily consumption of related materials in the invoice-inventory coupled twin directed graph through the real-time inventory level pointer, normalize the ratio of the current inventory to the average daily consumption, and take the reciprocal to obtain the material shortage risk score.

[0035] The core problem addressed by this submodule is: how many days of production consumption can the materials associated with the invoice support? The fewer days they can support, the more urgent the business needs to verify the invoice. Directly using the number of days that can be consumed as a metric presents two problems: First, the number of days that can be consumed by different materials may vary greatly in magnitude, making it inconvenient to use the same weighting scale as supplier uncertainty scores and inventory health scores; second, a longer number of days that can be consumed means a lower risk, while what we need is a higher risk score, which is consistent with the direction of business urgency.

[0036] Therefore, the execution steps of the material shortage risk calculation submodule are as follows: First, a query request is sent to the inventory management system through the real-time inventory level pointer in the directed graph to obtain the current inventory quantity of the material at this moment, denoted as S; at the same time, the average daily consumption is read from the associated material code node in the directed graph, denoted as D; the number of days that can be consumed is calculated as S divided by D; when S is zero or negative, it indicates that the inventory has been exhausted or there is a negative inventory, and at this time the material shortage risk score is directly set to the preset maximum value.

[0037] Secondly, T is normalized. The normalization formula can use an S-shaped function to map T to the interval (0,1). Specifically, the normalized value is equal to 1 divided by 1 plus e raised to the power of (T minus the threshold parameter). The threshold parameter can be set according to the number of days of safety stock for the enterprise. For example, when the safety stock coverage days are 7 days, the threshold parameter is 7. In this way, when the number of days that can be consumed is much greater than the number of days that can be consumed, the normalized value approaches 1; when the number of days that can be consumed is equal to the number of days that can be consumed, the normalized value is about 0.5; and when the number of days that can be consumed approaches zero, the normalized value approaches 0.

[0038] Finally, the reciprocal of the normalized value is taken to obtain the material shortage risk score. Since the normalized value is in the (0,1) interval, the reciprocal falls within the range of 1 to positive infinity. The more scarce the material, the higher the risk score, thus establishing a positive correlation with the urgency of the business. Taking the reciprocal after normalization instead of directly using T ensures that even if T fluctuates between 1 and 100 days due to material differences, the final score is still constrained within a reasonable range that can be weighted and averaged with the scores of other sub-models. For example, when the available consumption days are only 1 day, the normalized value is close to 0.27, and the risk score after taking the reciprocal is about 3.7; when the available consumption days are 30 days, the normalized value is close to 1.0, and the risk score after taking the reciprocal is about 1.0. It can be seen that the risk score of urgent materials is significantly amplified, which is completely consistent with the intuition that the risk of production line stoppage must be dealt with first in business operations.

[0039] The supplier uncertainty calculation submodule is configured to read the historical invoicing error rate corresponding to the associated supplier identifier in the invoice-inventory coupled twin directed graph, and use the historical invoicing error rate as the supplier uncertainty score.

[0040] The design motivation for this submodule is that a supplier’s historical behavior patterns are an effective predictor of its future behavior probability. If a supplier’s past invoices frequently contain issues such as inconsistencies between product names and tax classification codes, incorrect amount calculations, and inappropriate tax rate applications, then its newly submitted invoices also have a high probability of error and should be subject to more verification resources for careful review.

[0041] The historical invoicing error rate data comes from the historical verification records accumulated in the enterprise invoice management system. Whenever an invoice is verified and marked as incorrect or returned, the system records the supplier's identifier for that invoice. The supplier's historical invoicing error rate is equal to the number of invoices marked as incorrect by that supplier in the past preset time period divided by the total number of invoices submitted by that supplier. The preset time period can be the most recent 12 months to balance statistical stability and behavioral timeliness. For new suppliers, if there is no historical data, the average invoicing error rate of the supplier's industry or region can be used as the initial value.

[0042] The error rate can be directly used as the supplier uncertainty score without additional normalization or transformation, because the error rate itself naturally falls within the range of 0 to 1, which is comparable to the magnitude of the reciprocal of the material shortage risk score after normalization, making it convenient for subsequent weighting.

[0043] The inventory health calculation submodule is configured to obtain the quantity of materials already received in the directed graph of the invoice-inventory coupled twin, compare the quantity of invoices to be verified with the quantity already received, and if the quantity of invoices is greater than the quantity already received, output a preset inventory health score indicating an inventory abnormality.

[0044] This submodule detects anomalies where the invoice quantity exceeds the inventory quantity. In a standard procurement and warehousing process, goods arrive at the warehouse first, undergo quality inspection and warehousing, and then the supplier issues an invoice. The quantity already received should be greater than or equal to the invoice quantity. If the invoice quantity exceeds the quantity already received, it may mean that the invoice was issued before the goods were received, or there are issues such as missing warehousing records or insufficient quantities received. Such anomalies require close attention from verification personnel because they not only involve the accuracy of the invoices themselves but may also involve the risk of discrepancies between the recorded and actual physical assets.

[0045] In the specific implementation, the inventory health calculation submodule reads the value of the quantity already received from the associated material code node of the directed graph and the invoice quantity from the invoice identifier node, and compares them. If the invoice quantity is less than or equal to the quantity already received, it means that the inventory record is consistent with the invoice information or that there is a surplus of inventory, which is a healthy state. This submodule outputs a lower inventory health score, such as 1.0. If the invoice quantity is greater than the quantity already received, it is an abnormal state, and a preset higher score is output. The setting of this preset value needs to consider its matching with the magnitude of the material shortage risk score and the supplier uncertainty score. According to calculations, the material shortage risk score can reach 5 to 10 in extreme material shortage cases, while the upper limit of the supplier uncertainty score is 1. In order to give appropriate weight to the inventory abnormality, the preset value can be a fixed value between 3 and 5.

[0046] After each of the three sub-modules outputs a score, the model calculation module performs a weighted average to obtain the final business urgency score. The weighting coefficients for the weighted average are determined using the analytic hierarchy process (AHP).

[0047] The Analytic Hierarchy Process (AHP) is a method that mathematically quantifies expert judgment to determine the weights of multiple factors. In practical applications, business experts compare the importance of three factors—material shortage risk, supplier uncertainty, and inventory health—pair by pairwise, constructing a 3x3 judgment matrix. The elements aij in the judgment matrix represent the importance of factor i relative to factor j, using a scale of 1 to 9, where 1 indicates equal importance, 3 indicates slightly important, 5 indicates significantly important, 7 indicates strongly important, and 9 indicates extremely important. For example, if business experts believe that in a manufacturing scenario, material shortage risk is significantly more important than supplier uncertainty, the corresponding matrix element is 5; material shortage risk is slightly more important than inventory health, so the corresponding element is 3; and supplier uncertainty is equally important as inventory health, so the corresponding element is 1. This constitutes a complete judgment matrix.

[0048] To find the eigenvectors of the judgment matrix, we can calculate the geometric mean of each row's elements. Then, the vector formed by these geometric means is normalized so that the sum of the three weights equals 1. The three coefficients obtained after normalization are the weight coefficients for the material shortage risk score, supplier uncertainty score, and inventory health score. For example, the weights calculated in the above example might be 0.64, 0.10, and 0.26, indicating that material shortage risk dominates the business urgency assessment, followed by inventory health, with supplier uncertainty serving as a secondary adjustment factor. The advantage of using the analytic hierarchy process (AHP) is that the weights are not arbitrarily set but derived through structured expert evaluation and mathematical calculation, ensuring traceability and interpretability. When the business focus changes, the weights can be updated simply by revising the judgment matrix, giving the system good adaptability.

[0049] The final business urgency score is equal to the shortage risk score multiplied by its weight, plus the supplier uncertainty score multiplied by its weight, plus the inventory health score multiplied by its weight. This score is the sole quantitative input for subsequent multi-level scheduling decisions.

[0050] The material shortage risk calculation submodule, the supplier uncertainty calculation submodule, and the inventory health calculation submodule are each encapsulated as independent containerized microservices and registered through the service registry center; the model calculation module dynamically discovers and calls each submodule by querying the service registry center.

[0051] The advantages of this design lie in its maintainability and scalability. The computational logic of each submodule is relatively independent. Containerizing it means that improvements to the normalization formula for material shortage risk, adjustments to the statistical caliber of supplier historical invoicing error rates, or corrections to the preset values ​​of inventory health scores can all be accomplished by updating the container image of the corresponding submodule without affecting the operation of other submodules. Before each call, the model calculation module dynamically discovers the current service endpoints of each submodule through the service registry. When the container image of any submodule is updated and redeployed, the service registry updates the service endpoint of that submodule, and the model calculation module automatically points to the updated service endpoint on the next call. This enables online hot updates of submodules, and the system can complete algorithm iterations without overall shutdown, which is especially important for supply chain systems that operate 24 / 7.

[0052] The decision scheduling module, connected to the model calculation module, receives the business urgency score. Based on the preset threshold range of the business urgency score, it determines the target processing level from the three-level verification microservice hierarchy that is pre-divided into standard processing level, enhanced verification level and deep verification level, and generates a scheduling instruction carrying the target processing level identifier.

[0053] The decision-making and scheduling module internally maintains standard processing level threshold ranges, enhanced verification level threshold ranges, and deep verification level threshold ranges. When the business urgency score falls into the standard processing level threshold range, the target processing level is determined to be the standard processing level. When the business urgency score falls into the enhanced verification level threshold range, the target processing level is determined to be the enhanced verification level. When the business urgency score falls into the deep verification level threshold range, the target processing level is determined to be the deep verification level.

[0054] Determining the threshold range is a crucial parameter configuration step that directly impacts the system's hierarchical scheduling behavior. The determination method can combine the experience and judgment of business experts with statistical analysis of historical data. A specific example of determination is as follows: Collect all fully verified invoice data from the past quarter or six months, including the business urgency score calculated by the model calculation module for each invoice, and whether the invoice was ultimately confirmed to contain a material error. Sort the scores from low to high and analyze the relationship between the scores and the probability of invoice errors.

[0055] A pattern is typically observed: when the score is low, the probability of invoice errors is also low and changes gradually; after the score crosses a certain threshold, the probability of errors begins to rise significantly; as the score continues to rise, the probability of errors remains at a high level or even accelerates; these two thresholds can be used as the dividing lines between the standard processing level and the enhanced verification level, and between the enhanced verification level and the deep verification level.

[0056] For example, analysis shows that when the business urgency score is below 1.5, the error probability of the invoice is less than 2%, which falls into the low-risk category and can be assigned to standard processing level verification; when the score is between 1.5 and 3.0, the error probability rises to between 2% and 10%, which falls into the medium-risk category and should be assigned to enhanced verification level verification; when the score exceeds 3.0, the error probability exceeds 10%, and often involves urgently needed materials or habitual offenders, which falls into the high-risk category and should be assigned to deep verification level verification.

[0057] Therefore, the threshold range for standard processing level is set to 0 to 1.5, the threshold range for enhanced verification level is 1.5 to 3.0, and the threshold range for deep verification level is above 3.0. The threshold range can be dynamically adjusted according to business needs. For example, the threshold can be lowered overall during peak season to allow more invoices to undergo higher-level verification to avoid risks, while it can be appropriately raised during off-season to save computing resources.

[0058] After receiving the business urgency score, the decision scheduling module compares it with the three threshold ranges mentioned above. For example, if the business urgency score of an invoice is 2.8, which falls into the enhanced verification level threshold range, the target processing level is determined to be the enhanced verification level. The module generates a scheduling instruction, which includes the twin identifier of the invoice and the target processing level identifier, and sends it to the execution module.

[0059] When the target processing level is determined to be the deep verification level, the decision scheduling module generates a scheduling instruction carrying a mandatory real-time synchronization flag. The system also includes a cross-system synchronization module, which receives the scheduling instruction carrying the mandatory real-time synchronization flag and exchanges the verification status of the invoice to be verified in real time between the execution module and the external inventory management system in a synchronous communication manner until the verification task is completed.

[0060] The motivation behind this design is that invoices rated as deep verification often involve urgently needed materials or serious inventory anomalies. During the verification process, the execution module may suggest modifications to the invoice information or mark it as abnormal. Real-time synchronization of these statuses is crucial for subsequent operational decisions by the inventory management system. For example, if a significant discrepancy is found between the invoice amount and the received amount during verification, requiring manual intervention, the inventory management system needs to be immediately aware of this status and suspend the settlement process for that batch of materials. Synchronous communication ensures that every change in the verification status is pushed to the external inventory management system in a request-confirmation handshake mode, rather than due to the potential delays or data loss caused by asynchronous messages. After the verification task is completed, the cross-system synchronization module notifies the external system to unlock the status.

[0061] When multiple invoice verification tasks exist simultaneously, the decision scheduling module sorts them in descending order of their business urgency scores and generates scheduling instructions in sequence. During the scheduling process, if the business urgency score of a newly arrived task exceeds the minimum score of the currently executing task, the newly arrived task is allowed to preempt the computing resources of the task with the lowest current score, and the preempted task re-enters the queue to wait for the next scheduling.

[0062] This preemptive priority scheduling strategy is a response mechanism for sudden high-risk events in supply chain scenarios. For example, the system is currently executing the verification task of three invoices with scores of 1.2, 1.8 and 2.5 respectively. At this time, the company suddenly receives a notification from the supplier that a key raw material will be delayed due to transportation delays, and the current inventory can only support one day of production. The system triggers the verification of the invoices pre-issued by the supplier related to the raw material. According to the business mechanism model, the business urgency score of this invoice is as high as 4.7. At this time, the preemption logic is executed: 4.7 exceeds the lowest score of 1.2 of the currently executing task. Therefore, the new task preempts the computing resources occupied by the task with a score of 1.2. The state of the low-scoring task is saved and put back to the head of the waiting queue. When the idle resources are available, it can continue to execute. When the preempted task regains resources, it can continue to execute from the last saved state instead of starting from the beginning, so as to reduce the overhead of repeated calculations.

[0063] The set of microservice instances corresponding to the deep verification level has a minimum guaranteed number of dedicated GPU resources reserved in the GPU computing nodes. The minimum guaranteed number is set according to the maximum concurrency of deep verification level tasks in the same period in history. For example, historical data shows that the peak number of invoices that need to be deeply verified on each working day in the past 6 months reached 15 on Mondays. Therefore, the minimum guaranteed number can be set to 15. This ensures that under normal operating conditions, deep verification level tasks always have GPU resources available in real time and will not be completely squeezed out by standard processing level and enhanced verification level tasks.

[0064] When the number of deep verification tasks exceeds the minimum guaranteed number, it indicates a concentrated outbreak of high-risk invoices, and the pre-set dedicated resources are insufficient to meet real-time processing demands. At this point, the decision-making and scheduling module sends a scaling request to the container orchestration system to increase the number of deep verification microservice instances. Upon receiving the request, the container orchestration system allocates new container instances from the idle resource pool of GPU compute nodes, launching replicas of the high-precision OCR microservice and graph neural network analysis microservice. Once the number of deep verification tasks returns to normal levels, the excess instances can be reclaimed, releasing GPU resources. This elastic scaling mechanism enables the system to rapidly increase computing power in the face of sudden high-risk events without requiring a large amount of expensive GPU hardware to handle extreme peak loads over a long period.

[0065] The execution module is used to deploy a three-level verification microservice hierarchy in a microservice architecture, receive scheduling instructions, call the set of microservice instances corresponding to the target processing level to execute the invoice verification task, and output the verification results.

[0066] The standard processing level corresponds to a set of microservice instances including an OCR recognition microservice and a rule comparison microservice. The OCR recognition microservice is responsible for recognizing text in invoice images, extracting and structuring the text information on the invoice. The rule comparison microservice is responsible for matching and comparing the extracted structured data with a preset set of verification rules. The verification rules include basic rules such as the legality of the invoice number length, the correctness of the taxpayer identification number check digit of the buyer and seller, and the consistency of the amount and tax rate. Both of these microservices are computationally low tasks and can be deployed on CPU computing nodes to meet performance requirements. CPU computing nodes are low-cost, and enterprises typically have a large number of CPU server resources available.

[0067] The enhanced verification level adds a tax compliance verification microservice and a tolerance matching microservice to the standard processing level. The tax compliance verification microservice is responsible for verifying the applicability and correctness of commodity tax classification codes, such as verifying whether the commodity name matches the selected tax classification code. This involves natural language processing or rule matching and requires a certain amount of computing power. The tolerance matching microservice is responsible for automatically matching the invoice amount with the amount deposited with a tolerance threshold. It can tolerate small discrepancies while capturing the real amount deviation. The task complexity of the enhanced verification level is increased. Therefore, it is deployed in a hybrid architecture of CPU computing nodes and GPU computing nodes. OCR recognition and rule comparison are still performed on the CPU, while the model inference involved in tax compliance verification and tolerance matching is accelerated on the GPU.

[0068] The set of microservice instances corresponding to the deep verification level includes high-precision OCR microservices, graph neural network analysis microservices, end-to-end cross-validation microservices, and tax compliance verification microservices. The high-precision OCR microservice, compared to standard OCR, uses a deeper neural network model to perform character-by-character recognition and confidence assessment, making it particularly suitable for complex invoices such as those with blurred text or obscured seals. The graph neural network analysis microservice utilizes the relationships between invoices and enterprise transaction networks to construct a heterogeneous graph between invoices, enterprises, and goods. Through the node embedding and edge prediction capabilities of graph neural networks, it discovers abnormal patterns in invoices throughout the transaction network, such as suspicious invoices issued by the same supplier to multiple related enterprises simultaneously. The end-to-end cross-validation microservice performs end-to-end field-by-field comparisons of invoice information with documents from multiple stages, such as purchase orders, warehouse receipts, and quality inspection reports, capturing inconsistencies across documents. These tasks all involve deep learning model inference, resulting in high computational demands; therefore, they are all deployed on GPU computing nodes.

[0069] Furthermore, the deep verification-level microservice instance set is also configured to organize multiple verification tasks associated with the same supplier or the same material type into the same batch, sharing the preprocessed data in the GPU computing node's memory. The basis for this batch processing optimization strategy is that when there are multiple invoices from the same supplier to be verified in the system at the same time, their operations such as image preprocessing before OCR recognition, supplier identifier embedding vector calculation in graph neural networks, and weight loading of high-precision OCR models are repetitive. By aggregating these tasks into the same batch and allocating them to the same GPU instance, the preprocessed data only needs to be loaded into the GPU memory once, and subsequent multiple invoices share this data, reducing the number of times data is repeatedly moved between host memory and GPU memory, improving GPU utilization and overall processing throughput.

[0070] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A supply chain inventory invoice intelligent verification management system, characterized in that, include: The twin construction module is used to obtain the structured invoice data and corresponding inventory element data of the invoice to be verified from the enterprise invoice pool and inventory management system, respectively. It generates a unique twin identifier for the invoice to be verified and uses the twin identifier as an index to establish an invoice-inventory coupled twin directed graph, with nodes and edges representing the invoice identifier, associated material code, associated supplier identifier, and real-time inventory level pointer. The real-time inventory level pointer is a data link pointing to the current inventory storage location of the material corresponding to the associated material code in the inventory management system. The model calculation module, connected to the twin construction module, is used to receive the directed graph of the invoice-inventory coupled twin, call at least one pre-set supply chain business mechanism model, and calculate the business urgency score of the invoice to be verified; the supply chain business mechanism model includes a material shortage risk assessment sub-model, a supplier uncertainty assessment sub-model, and an inventory health assessment sub-model; The decision scheduling module, connected to the model calculation module, is used to receive the business urgency score, determine the target processing level from the three-level verification microservice level (standard processing level, enhanced verification level, and deep verification level) based on the preset threshold range of the business urgency score, and generate a scheduling instruction carrying the identifier of the target processing level. The execution module, connected to the decision scheduling module, is used to deploy the three-level verification microservice layer in a microservice architecture, receive the scheduling instructions, call the set of microservice instances corresponding to the target processing layer to execute the invoice verification task, and output the verification result.

2. The intelligent verification management system for supply chain inventory invoices according to claim 1, characterized in that, The model calculation module further includes a material shortage risk calculation submodule, a supplier uncertainty calculation submodule, and an inventory health calculation submodule; The material shortage risk calculation submodule is configured to read the current inventory and average daily consumption of the associated materials in the invoice-inventory coupled twin directed graph through the real-time inventory level pointer, normalize the ratio of the current inventory to the average daily consumption, and take the reciprocal to obtain the material shortage risk score. The supplier uncertainty calculation submodule is configured to read the historical invoicing error rate corresponding to the associated supplier identifier in the invoice-inventory coupled twin directed graph, and use the historical invoicing error rate as the supplier uncertainty score; The inventory health calculation submodule is configured to obtain the quantity of associated materials in the directed graph of the invoice-inventory coupling twin, compare the quantity of invoices to be verified with the quantity of materials already in stock, and if the quantity of invoices is greater than the quantity of materials already in stock, output a preset inventory health score indicating an inventory abnormality; wherein, the quantity of materials already in stock is the cumulative historical data value of the associated materials. The business urgency score is obtained by weighting the material shortage risk score, supplier uncertainty score, and inventory health score.

3. The intelligent verification management system for supply chain inventory invoices according to claim 2, characterized in that, The material shortage risk calculation submodule, supplier uncertainty calculation submodule, and inventory health calculation submodule are each encapsulated as independent containerized microservices and registered through a service registry center; The model calculation module dynamically discovers and calls each sub-module by querying the service registry; when the container image of any sub-module is updated, the service registry updates the service endpoint of that sub-module, and the model calculation module automatically points to the updated service endpoint on the next call.

4. The intelligent verification management system for supply chain inventory invoices according to claim 1, characterized in that, The decision scheduling module internally maintains standard processing level threshold ranges, enhanced verification level threshold ranges, and deep verification level threshold ranges. When the business urgency score falls within the standard processing level threshold range, the target processing level is determined as the standard processing level. When the business urgency score falls within the enhanced verification level threshold range, the target processing level is determined as the enhanced verification level; When the business urgency score falls within the threshold range of the deep verification level, the target processing level is determined to be the deep verification level.

5. The intelligent verification management system for supply chain inventory invoices according to claim 4, characterized in that, The decision scheduling module is also configured to generate a scheduling instruction carrying a forced real-time synchronization flag when the target processing level is determined to be the deep verification level. The system also includes a cross-system synchronization module, which receives a scheduling instruction carrying the mandatory real-time synchronization flag and exchanges the verification status of the invoice to be verified in real time between the execution module and the external inventory management system in a synchronous communication manner until the verification task is completed.

6. The intelligent verification management system for supply chain inventory invoices according to claim 4, characterized in that, The set of microservice instances corresponding to the standard processing level includes OCR recognition microservice and rule comparison microservice, which are deployed on CPU computing nodes; The set of microservice instances corresponding to the enhanced verification level includes OCR recognition microservice, rule comparison microservice, tax compliance verification microservice and tolerance matching microservice, which are deployed on a hybrid architecture of CPU computing nodes and GPU computing nodes; The set of microservice instances corresponding to the deep verification level includes a high-precision OCR microservice, a graph neural network analysis microservice, a full-link cross-validation microservice, and the tax compliance verification microservice. These are deployed on GPU computing nodes and configured to organize multiple verification tasks associated with the same supplier or the same material type into the same batch, sharing the preprocessed data in the GPU computing node's memory.

7. The intelligent verification management system for supply chain inventory invoices according to claim 2, characterized in that, When performing weighted calculations, the weight coefficients of each submodule—the material shortage risk calculation submodule, the supplier uncertainty calculation submodule, and the inventory health calculation submodule—are determined using the analytic hierarchy process (AHP). Specifically, business experts construct a judgment matrix by comparing material shortage risk, supplier uncertainty, and inventory health in pairs, solve for the eigenvectors of the judgment matrix, and obtain the weight coefficients of each submodule after normalization.

8. The intelligent verification management system for supply chain inventory invoices according to claim 1, characterized in that, The decision-making and scheduling module is also configured to: when there are multiple invoice verification tasks at the same time, sort them in descending order according to the business urgency score of each task, and generate scheduling instructions in sequence. During the scheduling process, if the business urgency score of a newly arrived task exceeds the minimum score of the currently executing task, the newly arrived task is allowed to preempt the computing resources of the task with the lowest current score, and the preempted task re-enters the queue to wait for the next scheduling.

9. The intelligent verification management system for supply chain inventory invoices according to claim 8, characterized in that: The set of microservice instances corresponding to the deep verification level has a minimum guaranteed number of dedicated GPU resources reserved in the GPU computing nodes. The minimum guaranteed number is set according to the maximum concurrency of deep verification level tasks in the same period in history. When the number of deep verification level tasks exceeds the minimum guaranteed number, the decision scheduling module sends a scaling request to the container orchestration system to increase the number of deep verification level microservice instances.

10. The intelligent verification management system for supply chain inventory invoices according to claim 1, characterized in that, The structured invoice data to be verified includes the invoice number, buyer and seller taxpayer identification number, product name, product tax classification code, amount, tax amount, and invoice date; the inventory data includes the material code, real-time inventory quantity, safety stock threshold, average daily consumption, and quantity already received.

Images