Supply chain matching settlement system and method based on dynamic tolerance and vector space

Abstract

This invention relates to the field of supply chain data processing, and discloses a supply chain matching and settlement system and method based on dynamic tolerance and vector space. The system includes: acquiring heterogeneous supply chain document data, extracting business element features, and converting them into three-dimensional coordinate components containing material quantity, document net value, and tax rate features to construct document vector data; generating a dynamic tolerance limit based on the document net value combined with a logarithmic decay algorithm; extracting document vector data from inbound orders and pending settlement invoices as the baseline and target vectors, respectively, and calculating their spatial distance; determining whether the target vector falls within the boundary of a hypersphere constructed with the dynamic tolerance limit as its radius to perform accounting posting operations; and when it does not fall within the boundary, extracting deviation vector feature data, performing decision tree model calculations, and performing differentiated routing based on the anomaly attribution classification results. This invention avoids the financial risk of large-value document tolerance divergence and achieves automatic attribution and diversion of abnormal data.

Description

Technical Field

[0001] This invention relates to the field of supply chain data processing, specifically to a supply chain matching and settlement system and method based on dynamic tolerance and vector space. Background Technology

[0002] In the supply chain management and financial settlement operations of large enterprises, purchase orders, receiving slips, and invoices are typically generated and circulated by different business systems. This heterogeneous supply chain environment leads to differences in document data in terms of format and measurement standards. A core aspect of financial settlement is verifying the consistency of these three documents. Traditional settlement systems usually rely on absolute matching of discrete fields, requiring precise consistency between quantity and amount. However, in actual logistics and transportation processes, especially when dealing with bulk cargo, natural evaporation due to the physical environment or systematic errors in weighing equipment often makes it difficult for receiving characteristics and invoice characteristics to perfectly match.

[0003] To avoid frequent interruptions in the settlement process due to reasonable physical losses, existing systems generally introduce tolerance mechanisms, typically using a fixed percentage to calculate tolerance limits. When the difference in document amounts is within this fixed percentage, the system automatically allows the transaction to proceed. However, this fixed-percentage tolerance model reveals significant limitations when handling large-scale business settlements. As the total net value of the documents increases, the absolute value of the tolerance amount increases linearly, causing the tolerance limit to diverge in large-value transaction scenarios. This can easily mask serious issues such as shortchanging or compliance problems, leading to a failure of financial risk control. Simultaneously, traditional matching mechanisms lack the ability to adjust multi-dimensional feature weights when facing multiple consolidated settlements or minor discrepancies in tax-related data, easily resulting in compliance risks or false interceptions. Furthermore, when business data matching fails, existing systems lack the ability to perform dimensionality reduction and feature analysis on error data, failing to automatically determine the specific cause of the anomaly. This results in all mismatched documents being pushed into a manual review queue, significantly increasing manual intervention costs and reducing the overall efficiency of supply chain settlement. Summary of the Invention

[0004] The technical problem solved by this invention is that existing supply chain settlement matching mechanisms typically use discrete field comparisons or fixed-ratio tolerance judgments. When faced with large-scale business settlements, fixed-ratio tolerances can lead to linear amplification of the absolute value of the amount in large transaction scenarios, causing tolerance limits to diverge and thus triggering financial risk control risks. At the same time, when business data matching fails, existing systems lack the ability to reduce the dimensionality of error data and make it difficult to perform automated attribution and classification. This results in failed matching documents being directly blocked and pushed to manual review, reducing the efficiency of supply chain settlement.

[0005] To address the above problems, the present invention provides the following technical solution:

[0006] The first aspect of this invention provides a supply chain matching and settlement system based on dynamic tolerance and vector space, comprising: The data access module is used to perform standardized processing on heterogeneous supply chain document data and extract business element features. The heterogeneous supply chain document data includes the warehouse entry order and the invoice data to be settled. The core calculation module converts the business element features into three-dimensional coordinate components containing standardized material quantities, total net value of documents, and tax rate features to construct document vector data; it extracts the document vector data corresponding to the inbound order data and the pending settlement invoice data as the baseline vector and target vector, respectively; it generates a dynamic tolerance limit based on the total net value of documents and a logarithmic decay algorithm; it calculates the spatial distance between the target vector and the baseline vector, and determines whether the target vector falls within the boundary of the judgment hypersphere constructed by the dynamic tolerance limit; if it does not fall within the boundary, it extracts the deviation vector feature data of the target vector relative to the baseline vector, performs decision tree model calculations, and outputs the anomaly attribution classification result; The business routing module performs accounting posting when it determines that the target vector falls within the boundary range of the determined hypersphere, and performs differentiated routing allocation based on the anomaly attribution classification result when it determines that the target vector does not fall within the boundary range of the determined hypersphere, pushing the heterogeneous supply chain document data to the corresponding business processing queue.

[0007] Furthermore, the supply chain matching and settlement system based on dynamic tolerance and vector space also includes a distributed storage and computing module, which is communicatively connected to the data access module, core computing module, and business routing module. The distributed storage and computing module constructs a K-Dimensional Tree spatial index structure, which is used to partition the high-dimensional data space and accelerate vector retrieval. The core computing module has anti-duplication verification logic, which generates a computational fingerprint by collecting the core identifier and timestamp information of the document and combining it with a hash algorithm. This computational fingerprint is then retrieved and verified in the memory bitmap of the distributed storage and computing module, and duplicate business data computation requests are intercepted.

[0008] Furthermore, the business routing module is also used to, when performing accounting posting operations, extract the amount deviation data of the target vector relative to the reference vector for compliant micro-difference documents that fall within the boundary range of the determined hypersphere but whose coordinate components are not completely equal to those of the reference vector, generate tolerance adjustment business vouchers, and automatically post the amount deviation data to the preset tolerance profit and loss related accounts at the general ledger level.

[0009] A second aspect of this invention provides a supply chain matching and settlement method based on dynamic tolerance and vector space, comprising the following steps: Acquire heterogeneous supply chain document data and perform data format standardization operations to extract business element features. The heterogeneous supply chain document data includes at least warehouse entry order data and pending settlement invoice data. The business element features are converted into three-dimensional coordinate components containing standardized material quantities, total net value of documents, and tax rate feature values ​​to construct document vector data. Based on the total net value of the document, the basic tolerance rate, the risk coefficient of the material category, and the industry correction offset, numerical calculations are performed using the logarithmic decay algorithm to calculate and generate the dynamic tolerance limit corresponding to the current business document. Extract the document vector data corresponding to the inbound order data as the base vector, extract the document vector data corresponding to the invoice data to be settled as the target vector, and calculate the spatial distance between the target vector and the base vector. A judgment hypersphere is constructed with the reference vector as the center and the dynamic tolerance limit as the radius, and it is determined whether the target vector falls within the boundary range of the judgment hypersphere. If the target vector is determined to fall within the boundary range of the determined hypersphere, the data matching is successful and the accounting posting operation is executed; If it is determined that the target vector does not fall within the boundary range of the determination hypersphere, the deviation vector feature data of the target vector relative to the reference vector is extracted, and classification and regression decision tree model operations are performed to output the anomaly attribution classification result. Based on the anomaly attribution classification result, the heterogeneous supply chain document data is pushed to the corresponding business processing queue.

[0010] Furthermore, in the step of constructing document vector data, when it is determined that the acquired heterogeneous supply chain document data belongs to a consolidated settlement batch, multi-dimensional vector synthesis logic is executed. Multiple independent inbound document data belonging to the same business flow within the consolidated settlement batch are extracted, grouped according to material codes, and an additive synthesis operation is performed on the document vector data corresponding to multiple independent inbound document data with the same material code. During the additive synthesis operation, a numerical accumulation operation is performed on the standardized material quantity and the total net value of the document; the mode of the tax rate feature value of each independent document within the consolidated settlement batch is extracted as the synthesized unified tax rate component, and the total inbound vector data is output as the base vector.

[0011] Furthermore, the step of calculating and generating the dynamic tolerance limit corresponding to the current business document specifically includes: obtaining the supplier's historical performance credit coefficient and mapping the supplier's historical performance credit coefficient to a supplier weight factor, extracting a preset system adjustment constant and a minimum tolerance guarantee value, and extracting the corresponding material category weight factor based on a preset material category risk matrix. The total net value of the document is subjected to a natural logarithmic operation; a comprehensive calculation is performed by combining the natural logarithmic operation result of the total net value of the document with the division term of the system adjustment constant, along with the basic tolerance rate, the material category weight factor, and the supplier weight factor, and the minimum tolerance guarantee value is superimposed on the final result. A pre-validation logic is executed before the calculation; when it is determined that the total net value of the document is less than or equal to the system's preset minimum threshold, the minimum tolerance guarantee value is directly output as the final dynamic tolerance limit. By introducing a nonlinear decay model of the logarithmic function, convergence control of the absolute value of the tolerance for large-value documents is achieved.

[0012] Further, the step of calculating the spatial distance between the target vector and the reference vector specifically includes: calculating the squared deviations between the target vector and the reference vector in the standardized material quantity dimension and the total net value of the document dimension, respectively. Calculating the squared deviation between the target vector and the reference vector in the tax rate feature value dimension, and introducing a preset tax rate sensitivity adjustment operator before the squared deviation in the tax rate feature value dimension to perform a product amplification operation. Summing the squared deviations in each dimension and performing a square root operation to obtain the corrected Euclidean distance as the spatial distance used in the boundary determination of the hypersphere. Further calculating the ratio of the vector dot product and norm product of the target vector and the reference vector to obtain the cosine similarity; when the cosine similarity is determined to be lower than a preset structural similarity threshold, a structural anomaly marker is added to the corresponding document vector data.

[0013] Further, the step of extracting the deviation vector feature data of the target vector relative to the benchmark vector and performing classification and regression decision tree model operations specifically includes: performing vector subtraction to obtain deviation vector feature data containing quantity deviation, amount deviation, and tax rate deviation, and inputting the deviation vector feature data into the classification and regression decision tree model. At the first-level verification node of the classification and regression decision tree model, it is determined whether the tax rate deviation is zero; if not, the attribution classification result for tax compliance anomalies is directly output. At the second-level verification node of the classification and regression decision tree model, the relative quantity deviation rate between the quantity deviation and the standardized material quantity in the benchmark vector is calculated; if the relative quantity deviation rate is greater than a preset reasonable loss threshold, the attribution classification result for logistics / warehouse management anomalies is output. In the third-level verification node of the classification and regression decision tree model, the target unit price and the benchmark unit price are obtained through division. When the difference between the target unit price and the benchmark unit price exceeds the preset price fluctuation threshold range, the attribution classification result of abnormal price agreement or temporary price adjustment is output. When the difference between the target unit price and the benchmark unit price does not exceed the preset price fluctuation threshold range, it is determined that the spatial distance exceeds the limit due to the superposition of multiple factors, and the attribution classification result of abnormal quantity-price coupling deviation is output. This processing mechanism eliminates the decision tree attribution dead zone and performs dimensionality reduction and feature diagnosis on non-compliant data that fails to match.

[0014] Furthermore, after outputting the anomaly attribution classification results, the method further includes: extracting the local conditional probability estimates corresponding to the deviation vector feature data as it flows through each verification node of the classification and regression decision tree model. Performing a continuous multiplication operation on the local conditional probability estimates flowing through all nodes yields a probability confidence score for the overall attribution of the heterogeneous supply chain document data. When the probability confidence score is determined to be less than a preset manual intervention threshold, automated routing allocation is stopped, the heterogeneous supply chain document data is suspended, and pushed into a manual review queue; when the probability confidence score is determined to be greater than or equal to the preset manual intervention threshold, differentiated routing is executed.

[0015] Furthermore, the supply chain matching and settlement method based on dynamic tolerance and vector space also includes a parameter self-learning step: collecting data samples generated by manual review to construct an offline training sample set, and converting the manual approval decision results into real matching labels. The difference between the spatial distance and the current dynamic tolerance limit is imported into a preset activation function to map and generate a predicted matching probability. The real matching label and the predicted matching probability are substituted into the logarithmic operation logic of the cross-entropy loss function to calculate the prediction error, and the partial derivative is obtained through an optimization algorithm to obtain the gradient vector. Combined with a preset learning rate and parameter truncation verification logic, the internal material category weight factors are iteratively updated along the opposite direction of the gradient.

[0016] This invention provides a supply chain matching and settlement system and method based on dynamic tolerance and vector space. It has the following beneficial effects: 1. This invention generates dynamic tolerance limits by extracting the net value of documents and combining it with a logarithmic decay algorithm. Utilizing the non-linear growth characteristic of the logarithmic function, it provides a basic business error tolerance space in the small transaction range, while converging the growth slope of the absolute value of the tolerance amount in the large transaction range. This calculation mechanism avoids the problem of uncontrolled financial risk exposure caused by proportionally scaling up calculations for large business documents, thus improving the security of enterprise financial risk control while ensuring the normal flow of settlements in the grassroots supply chain.

[0017] 2. This invention converts material quantity, document net value, and tax rate characteristic values ​​into three-dimensional coordinate components. A judgment hypersphere is constructed with a reference vector as the center and a dynamic tolerance limit as the radius. Discrete document text comparison is transformed into continuous spatial geometric distance measurement. When calculating spatial distance, the system introduces a tax rate sensitivity adjustment operator to perform a product amplification operation. While tolerating reasonable physical loss errors in supply chain logistics transportation, it improves the interception standard for tax-related dimension deviations, thereby reducing normal transaction interruptions caused by equipment weighing discrepancies and ensuring tax compliance requirements in the settlement process.

[0018] 3. This invention extracts deviation vector feature data containing multi-dimensional differences and inputs it into a decision tree model. Based on preset node splitting conditions, it performs attribution judgment step by step and distributes the data to the corresponding business processing queue according to the output anomaly attribution classification results. This processing logic changes the traditional settlement system's processing mode of pushing all mismatched documents into the manual review queue. It realizes automated dimensionality reduction diagnosis and diversion processing of abnormal data, reduces the cost of manual intervention, and improves the processing efficiency of abnormal data. Attached Figure Description

[0019] Figure 1 This is a diagram of the supply chain matching and settlement system architecture based on dynamic tolerance and vector space of the present invention. Figure 2 This is a macroscopic workflow diagram of the supply chain matching and settlement system based on dynamic tolerance and vector space of the present invention; Figure 3 This is a comparative experimental verification diagram showing the effects of the traditional fixed-proportion tolerance and the dynamic logarithmic decay tolerance limit of the present invention. Detailed Implementation

[0020] The technical solutions in 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.

[0021] See attached document Figure 1 The present invention provides a supply chain matching and settlement system based on dynamic tolerance and vector space. The system may include: a data access module, a core computing module, a business routing module, and a distributed storage and computing module.

[0022] The data access module is used to capture heterogeneous supply chain document data in real time and perform data format standardization preprocessing.

[0023] The core computing module communicates with the data access module to receive standardized preprocessed data. The core computing module performs feature extraction, vector mapping transformation, and spatial geometric comparison calculations.

[0024] The service routing module communicates with the core computing module, receiving the spatial geometric comparison calculation results and attribution determination results output by the core computing module. Based on the received calculation results, the service routing module performs data distribution and service process dispatch.

[0025] The distributed storage and computing module communicates with the data access module, the core computing module, and the business routing module, providing underlying computing power support and data storage support for these modules.

[0026] The distributed storage and computation module incorporates a K-Dimensional Tree spatial index structure. This structure is used to partition the high-dimensional data space, reducing the time complexity of the vector retrieval process.

[0027] To address the retrieval requirements of high-frequency settlement vectors, the distributed storage and computing module combines an approximate nearest neighbor algorithm to perform computational tasks, accelerating the calculation process of spatial distance to target data.

[0028] The core computing module is equipped with anti-duplicate calculation verification logic to ensure the idempotency of system data processing. The core computing module collects the core identifier and timestamp information of documents and generates a computational fingerprint using a hash algorithm. The system retrieves and verifies the aforementioned computational fingerprint in the memory bitmap of the distributed storage and computing module, intercepting duplicate business data computation requests.

[0029] See attached document Figure 2The overall macro-level workflow for supply chain data processing in this system may include the following steps: In step S100, the data access module receives purchase order data, warehouse receipt data, and invoice data from the heterogeneous supply chain system, and performs format standardization operations on the received business data.

[0030] In step S200, the core computing module extracts the business element features from the standardized business data output in step S100, and converts the extracted business element features into three-dimensional vector space coordinate data containing quantity parameters, amount parameters, and tax rate parameters.

[0031] In step S300, the core calculation module obtains the document amount parameters and business system attribute parameters, performs numerical calculations based on multidimensional weighting and logarithmic decay algorithms, and calculates and outputs the dynamic tolerance limit data corresponding to the current business document.

[0032] In step S400, the core calculation module calculates the spatial distance between the three-dimensional vector spatial coordinate data of different documents generated in step S200. The core calculation module constructs a hypersphere with the spatial coordinate data corresponding to the inbound order as the center and the dynamic tolerance limit data output in step S300 as the radius, and determines whether the spatial coordinate data corresponding to the invoice falls within the boundary range of this hypersphere.

[0033] In step S500, for the spatial coordinate data corresponding to the invoices that fall within the boundary range of the hypersphere, the business routing module determines that the data match is successful. The business routing module then initiates a three-phase distributed transaction mechanism of reservation, confirmation, and rollback to complete the accounting posting and voucher generation operations.

[0034] In step S600, for the spatial coordinate data corresponding to invoices that do not fall within the boundary of the hypersphere, the core calculation module extracts the deviation vector feature data of the corresponding documents. The core calculation module performs classification and regression decision tree model calculations and outputs the anomaly attribution classification results. The business routing module receives the above anomaly attribution classification results and routes the abnormal data to the corresponding business processing queue.

[0035] In step S700, the system performs self-learning of dynamic tolerance model parameters and adaptive update of model weights based on the data samples returned by manual intervention.

[0036] The supply chain matching and settlement system based on dynamic tolerance and vector space provided by this invention, when executing the aforementioned step S300, specifically has its core calculation module complete the basic assessment and core calculation operations of the dynamic tolerance limit. To implement the calculation logic of this dynamic tolerance limit, the aforementioned step S300 may specifically include the following sub-steps.

[0037] First, the core calculation module obtains the basic business parameters of the business documents. The core calculation module extracts the total net value of the documents and the basic tolerance rate from the standardized business data output by the data access module. The aforementioned total net value is the actual business settlement amount after deducting relevant taxes and fees. As a preferred method, the aforementioned basic tolerance rate is usually initially set by business personnel based on the company's annual audit standards, and its conventional value range can be set between 0.005 and 0.02 to establish the basic benchmark for tolerance calculation. For the process of parsing and extracting specific business fields from various standardized format data, those skilled in the art can use conventional data parsing algorithms or regular expression tools. The file parsing and data mapping extraction operations are well-known technologies in this field and will not be elaborated upon here.

[0038] Simultaneously, the core calculation module determines the risk coefficient of the material category in the current business data. Different supply chain materials exhibit different price fluctuation attributes and physical loss characteristics during actual circulation. To transform this business experience into calculable data, the core calculation module uses the material coding information carried in the standardized business data... The system matches the corresponding material category in the material master data table and assigns a corresponding numerical weight based on the category. Specifically, for standard parts with fixed specifications and not prone to physical deformation, the system configures a smaller material category risk coefficient, for example, limiting its value range to between 0.1 and 0.5; conversely, for bulk cargo that is prone to weight or volume loss due to environmental factors during transportation and storage, the system configures a larger material category risk coefficient, for example, limiting its value range to between 1.5 and 3.0.

[0039] Subsequently, the core calculation module obtains the industry correction offset. Supply chain transactions exhibit systematic deviations caused by inherent equipment errors across different vertical industries. These deviations are often expressed as fixed values ​​rather than fluctuating percentages. Based on this, the core calculation module retrieves the corresponding industry correction offset value from the system's stored industry characteristic parameter table according to the industry attribute code of the current trading entity. This industry correction offset, as a constant value, participates in the core calculations to offset fixed deviations caused by the physical settlement environment of a specific industry. For example, in a coal weighing settlement scenario, the industry correction offset can be set to the pre-statistically calculated average allowable error value of the weighbridge equipment.

[0040] Based on the aforementioned multidimensional parameters, the core calculation module performs numerical calculations to generate dynamic tolerance limits according to the obtained total net value of the documents, the basic tolerance rate, the material category risk coefficient, and the industry correction offset. In this embodiment, the core calculation module is internally configured with a basic dynamic calculation mathematical model. By inputting the extracted multidimensional parameters into this model, the corresponding calculation results can be obtained.

[0041] The formula for the above basic dynamic calculation mathematical model is as follows: ; In the formula, This represents the calculated dynamic tolerance limit; This indicates the total net value of the document; Indicates the basic tolerance rate; Indicates the risk coefficient of the material category; Indicates industry-corrected offset; This represents the logarithm of the total net value of the documents, base 10.

[0042] In practical supply chain settlement scenarios, from a general perspective, if a fixed percentage is used to calculate tolerance for business documents, the absolute value of the tolerance amount will increase linearly as the total net value of the document increases. For example, a fixed tolerance of 0.01% might only amount to 10 yuan in a 100,000 yuan transaction, which is a reasonable slight error; however, in a 1 billion yuan transaction, it amplifies to 100,000 yuan, which can easily lead to the failure of fund risk control. To improve the above technical problem, the core calculation module configures a logarithmic function structure in the basic dynamic calculation mathematical model, utilizing the non-linear growth characteristic of the logarithmic function to execute the calculation logic. When the total net value of the document is small, the system calculates and generates a relatively broad tolerance calculation base; when the total net value of the document moves into a larger range, the system uses the logarithmic function curve to converge the growth slope of the absolute value of the tolerance.

[0043] Furthermore, to ensure the completeness of the algorithm logic and avoid dead zones during execution, the core calculation module includes pre-validation logic before substituting data into the formula. This takes into account the total net value of the document. Smaller (e.g.) When logarithmic terms If the value approaches zero or becomes negative, it will distort the entire calculation. Therefore, as a preferred approach, the system presets a minimum net value threshold and a minimum tolerance constant for the document amount. When the system determines that the total net value of the input document is lower than the minimum net value threshold, the core calculation module will skip the logarithmic operation and directly output the preset minimum tolerance constant value. The value of .

[0044] Through the above multi-dimensional parameter combination and algorithm structure design, the dynamic tolerance limit output by the core calculation module can adapt to the transaction scale, which to a certain extent improves the problem that the traditional fixed amount tolerance mechanism is difficult to adapt to the settlement of transactions of different scales.

[0045] To further enhance the adaptability and accuracy of the dynamic tolerance algorithm, the core calculation module is equipped with a more optimized anti-divergence logarithmic decay model on top of the basic dynamic calculation mathematical model. This model incorporates more business dimensions and system adjustment parameters when calculating dynamic tolerance limits to adapt to the complex and ever-changing commercial settlement environment.

[0046] As a specific implementation method, the core calculation module extracts the corresponding material category weight factors based on a pre-set material category risk matrix. Meanwhile, the core calculation module further incorporates a supplier's historical performance reputation coefficient. This coefficient can be quantitatively evaluated based on the supplier's historical transaction data, such as by weighting indicators like on-time delivery rate, quality pass rate, and historical settlement discrepancy rate. The result is mapped to a supplier weighting factor. The supplier weighting factor This is used to quantify the risk level of cooperation with specific trading partners. In this embodiment, the system typically maps the original reputation score, on a 100-point or 10-point scale, to a floating-point range around a central value using a normalization function. Preferably, this mapping range can be set between 0.8 and 1.2. When the supplier has good reputation, the supplier weighting factor... Values ​​less than 1 are used to tighten the tolerance; when the reputation is poor, the supplier weighting factor... Use values ​​greater than 1 to raise the risk warning threshold.

[0047] To finely control the shape of the model curve and the influence weights of each factor, the core calculation module in this embodiment has a built-in system adjustment constant. With system adjustment constant Among them, the system adjustment constant The system adjustment constant is used to control the growth slope of the logarithmic function curve. This is used to adjust the degree of influence of the supplier weighting factor. In typical supply chain implementation scenarios, the system adjustment constant... The value range is usually set between 2.0 and 5.0, while the system adjustment constant... The value is usually set between 0.5 and 1.5, and the specific value is preset by the implementing company based on its own sensitivity to capital risk control.

[0048] In this embodiment, the complete dynamic tolerance calculation is performed by the core calculation module, and the core calculation formula used is as follows: ; In the formula, Indicates the dynamic tolerance limit; This indicates the total net value of the document; Indicates the basic tolerance rate; Indicates the weighting factor for the material category; This represents the supplier weighting factor; Represents the system regulation constant; Represents the system regulation constant; This represents the minimum tolerance threshold. This represents the natural logarithm of the total net value of the document.

[0049] From a general business perspective, the above formula aims to address the issue of non-proportional scaling of tolerance in large-value fund settlements. In actual business operations, for a transaction of 100,000 yuan, an error of tens of yuan can often be considered a reasonable loss; however, for a transaction of 1 billion yuan, if scaled up proportionally, the tolerance would reach hundreds of thousands of yuan, which is generally unacceptable in financial auditing. Therefore, the system utilizes the mathematical property that the growth rate of the natural logarithm gradually slows down as the independent variable increases to smooth large-value data.

[0050] At the implementation level, this formula incorporates the total net value of the documents and a system adjustment constant. The division operation term controls the rate of increase in tolerance. This is achieved by adjusting the system adjustment constant. The system can control the growth slope of the logarithmic curve by adjusting the value of the logarithmic constant. When the value is large, the growth of the logarithmic function will be more gradual, thus more significantly suppressing the growth of the absolute value of the tolerance amount in large transactions, and playing a role in preventing divergence to a certain extent.

[0051] Minimum tolerance value in the formula As a supplementary fixed deviation, its value is usually set in the range of 0.01 to 1.00, for example, set to the system's minimum accounting unit of 0.01 yuan. The introduction of this parameter can usually prevent extremely small amount documents from losing basic error tolerance when performing proportional calculations.

[0052] It should be noted that, considering the domain and range characteristics of mathematical functions, when inputting the total net value of the document... When the value is less than or equal to 1, the logarithm will naturally be less than or equal to zero. Directly substituting this into the above formula will cause the product term within the square brackets to have a negative value, leading the system to calculate an abnormal dynamic tolerance limit. To block this algorithmic dead zone, the core calculation module reads the total net value of the document before executing the above formula. The value is then used to determine the threshold. When determining the total net value of the document... Less than or equal to the natural constant When the minimum value threshold (or other minimum value threshold set by the system) is reached, the core calculation module will directly block the calculation path of the formula and instead directly output the minimum tolerance guarantee value. This serves as the final dynamic tolerance limit.

[0053] Through this anti-divergence logarithmic decay model, the core calculation module extends the tolerance calculation from two dimensions to three dimensions and uses the system adjustment constant for flexible configuration, thereby generating dynamic tolerance limits that are highly matched with specific business scenarios.

[0054] To enhance the adaptability of the dynamic tolerance calculation model in complex business environments, the supply chain matching and settlement system based on dynamic tolerance and vector space provided by this invention is equipped with a parameter self-learning mechanism. Through this mechanism, the system utilizes historical data generated by manual secondary confirmation to perform closed-loop optimization of various weight parameters in the tolerance calculation model. The manual intervention data feedback and adaptive weight update operations in the aforementioned macro-workflow are specifically completed by the core computing module in collaboration with the distributed storage and computation module. In this embodiment, the specific implementation process of this parameter self-learning mechanism may include the following steps.

[0055] Specifically, the business routing module collects data samples generated by manual review and constructs an offline training sample set. In actual supply chain settlement operations, when the feature vector of an invoice fails to fall within the aforementioned hypersphere boundary, the system intercepts the settlement transaction and pushes it into the manual review queue. Business personnel, combining offline communication or paper document verification, make a final settlement approval decision for the suspended documents. The business routing module captures this manual approval decision result, converts it into a true matching label, and binds it together with the original document business feature data. The system continuously stores the bound data records as incremental data samples with supervisory signals in the feature library of the distributed storage and computing module.

[0056] Next, the core computing module extracts a specified batch of training samples from the distributed storage and computing module and calculates the predicted matching probability for each sample. The core computing module reads feature dimension data such as document amount, material category, and supplier reputation from the samples and performs forward calculations based on the weighting factors currently in effect within the system. To convert discrete document deviation features into a continuous probability distribution, the core computing module uses the difference between the calculated absolute deviation distance of the document space and the current dynamic tolerance limit as an input variable, importing it into a preset activation function. As a preferred approach, the preset activation function is the Sigmoid function. The system utilizes the characteristics of this function to map the aforementioned difference, which includes the business physical state, into probability values ​​distributed between zero and one. This probability value represents the statistical probability that the current computing model considers the document to be allowed for automatic matching.

[0057] Subsequently, the core calculation module calculates the prediction error based on the true matching labels and the predicted matching probabilities output by the model. To transform the differences in matching decisions within the business system into a mathematical objective that can be optimized by the computer, the core calculation module uses the cross-entropy loss function to measure the output of the current model. In this embodiment, the core calculation module substitutes the true matching labels and predicted matching probabilities from the input sample batch into the logarithmic operation logic of the cross-entropy loss function, and calculates the overall cross-entropy loss function value by accumulating the errors of each sample. As a specific implementation, the aforementioned true matching labels are assigned binary Boolean values ​​based on the manual approval decision. When the manual review result is "allowed and settled," the system assigns the true matching label a value of 1; when the manual review result is "rejected and settled," the system assigns the true matching label a value of 0.

[0058] Finally, the core computing module executes parameter backpropagation logic based on the gradient descent algorithm. After obtaining the cross-entropy loss function value, the core computing module uses an optimization algorithm to find the parameter combination that minimizes the cross-entropy loss function value. The system calculates the partial derivatives of each parameter involved in the dynamic tolerance calculation to obtain the gradient vector in the current parameter space. Subsequently, based on the system's preset learning rate, the core computing module iteratively updates the material category weight factors and system adjustment constants within the system along the reverse direction of the gradient.

[0059] In specific model training configurations, the learning rate is typically set between 0.001 and 0.01 to avoid excessively large step sizes that could cause model parameters to cross the optimal solution region. To ensure the stability of the update process and prevent the algorithm from falling into a dead zone, the core calculation module incorporates parameter truncation verification logic before performing parameter update coverage. The system sets update threshold boundaries for each weight factor. If the weight correction magnitude calculated by a single gradient descent exceeds the preset threshold boundary, the core calculation module will forcibly perform data truncation to prevent abrupt weight drops caused by abnormal samples. The specific mathematical operations of backpropagation and partial derivative calculation in the gradient descent algorithm can be implemented using conventional machine learning optimizer modules by those skilled in the art. The underlying calculus operations and parameter assignment coverage process are well-known technologies in this field and will not be elaborated upon here.

[0060] From the perspective of specific business logic mechanisms, the aforementioned self-learning mechanism based on cross-entropy and gradient descent has clear physical guiding significance. For example, in the transaction accounting for a specific category of materials, if the system repeatedly determines that the business data is out of tolerance and rejects the match based on existing rules, but business experts approve a large number of matches during manual review, then a large number of true matches will have a value of 1, while the model's predicted matching probability will be low. When the system calculates the cross-entropy loss function value, it will accumulate a large prediction error.

[0061] In the subsequent gradient descent update step, the core calculation module calculates the positive increment based on the direction of the partial derivative and increases the weight factor of the corresponding material category. This fine-tuning of the parameter space manifests at the system level as a targeted relaxation of the dynamic tolerance limit for that specific material category, gradually bringing the system's automated decision boundary closer to the actual decision boundary of business experts. Through this iterative parameter method based on cyclical backflow samples, the core calculation module, without relying on manual modification of the underlying system rules, helps the system adapt to normal fluctuations in commodity prices due to seasonal or market changes in the supply chain environment to a certain extent.

[0062] To address the issue of matching failures that occur in complex document settlement using traditional text-based or discrete field-based comparison methods, this invention provides a supply chain matching and settlement system based on dynamic tolerance and vector space. During step S200, the core computing module transforms discrete business documents into continuous high-dimensional vectors. In this embodiment, the core computing module performs the continuous vectorization and synthesis operation of this business element in the following specific manner.

[0063] In the initial stage of data processing, the core computing module extracts core business element features for spatial matching from the standardized business data output by the data access module. Based on the financial attributes of supply chain settlement, the core computing module extracts material quantity, document net value, and tax rate characteristic values ​​from purchase order data, warehouse receipt data, and invoice data, respectively. Considering the often inconsistent units of measurement between heterogeneous supply chain systems, the core computing module is configured with a unit conversion mapping table. For the material quantities transmitted from different systems, the core computing module converts them uniformly into standard basic units based on the aforementioned unit conversion mapping table, generating standardized material quantities. The data format verification and basic unit conversion processes can be implemented by those skilled in the art using conventional data cleaning and type conversion scripts; the data type parsing and basic unit mapping calculations are well-known technologies in the field and will not be elaborated upon here.

[0064] After acquiring the features, the core computing module extracts and maps the standardized business element features into a Euclidean three-dimensional space. The core computing module constructs a three-dimensional coordinate system in the system's memory, using the standardized material quantity, document net value, and tax rate feature values ​​as independent coordinate components.

[0065] In practical engineering implementation, the three coordinate components mentioned above exhibit significant differences in physical magnitude. For example, the net value of a document may reach hundreds of thousands, while the tax rate characteristic value typically exhibits a decimal distribution less than 1. To avoid the problem of large numerical dimensions completely masking the weights of small numerical dimensions during spatial calculations, the core calculation module retains the original standardized business dimensions when constructing the vector. Furthermore, in the subsequent spatial distance calculation algorithm, a control factor of a specific dimension is introduced to balance the spatial geometric weights between different dimensions.

[0066] After the above mapping and coordinate transformation processes, any independent business document is abstracted by the core calculation module into a continuous document vector data. The mathematical expression for the above document vector data is: ; In the formula, This represents document vector data; Indicates the quantity of materials after standardization. Indicates the net value of the document; This represents the characteristic value of the tax rate; This represents the matrix transpose operation. By establishing this document vector data, the core calculation module can, to some extent, help transform the subsequent document matching and judgment process from traditional Boolean logic judgment to a topological calculation process based on spatial geometric relationships.

[0067] Furthermore, the core computing module performs vector composition operations in complex business scenarios. In actual supply chain settlement operations, there are often scenarios where multiple inbound orders are merged to correspond to a single settlement invoice. For this type of settlement merging, the core computing module has built-in multi-dimensional vector composition logic. When the system determines that the current business belongs to a merged settlement batch, the core computing module extracts the document vector data corresponding to all independent inbound orders belonging to the same business flow within that batch based on the document's associated serial number. To avoid feature distortion caused by the addition of scalars from different material dimensions, the core computing module groups the aforementioned independent inbound orders according to their material codes before performing the composition. Subsequently, within the aforementioned Euclidean three-dimensional space, the core computing module performs an additive composition operation on the document vector data corresponding to multiple independent inbound orders with the same material code.

[0068] The calculation formula for vector synthesis under the above-mentioned consolidated settlement batch is as follows: ; In the formula, This represents the total vector data of the merged data entering the database; This indicates the total number of inbound orders involved in the merger; This indicates the sequence number index of the inbound order currently participating in the merger; Indicates the first The document vector data corresponding to Zhang's independent entry order.

[0069] It is important to note that, to ensure the synthesized vector possesses genuine physical meaning at the business level, the core calculation module employs differentiated calculation rules for different coordinate components when performing the aforementioned summation operation. For the standardized material quantity and net value dimensions of the document vector data, the system strictly performs numerical accumulation to reflect the total scale of multiple inbound documents. However, for the tax rate feature value dimension, since the system stipulates that mixed tax rates are not allowed within the same invoice batch, the core calculation module no longer performs tax rate accumulation. Instead, it extracts the mode of the tax rate feature values ​​from each independent document within the batch as the unified tax rate component in the synthesized vector. Through this differentiated merging mechanism, the system can retain the basic total settlement scale while avoiding distortion of the synthesized vector in terms of compliance.

[0070] To ensure algorithm stability, the core calculation module executes algorithm completeness verification and abnormal data filtering logic. As a preferred method to ensure algorithm stability, the core calculation module is configured with a numerical polarity pre-interception mechanism after generating or synthesizing the aforementioned document vector data. Since upstream links in the supply chain often involve reversal documents or return red-ink invoices, this can cause the net value of the documents extracted by the data access module or the standardized material quantity to appear as a negative number. Although Euclidean space mathematically supports negative coordinates, in certain spatial distance measurement algorithms, directly participating in calculations with untreated negative coordinates may cause errors in the system's comparison logic. Therefore, after completing vector mapping, the core calculation module verifies the numerical polarity of each coordinate component. When a negative characteristic representing a reversal transaction is detected in a coordinate component, the core calculation module adds a specific isolation identifier to the document vector data. After capturing this isolation identifier, the business routing module routes the document vector data to the red-ink reverse processing branch. This pre-data isolation mechanism typically avoids boundary misjudgments in the vector matching algorithm when processing reverse transactions.

[0071] The supply chain matching and settlement system based on dynamic tolerance and vector space provided by this invention performs geometric calculation logic on the spatially mapped business data through the core computing module during the aforementioned step S400. After completing the continuous vectorization of business elements, the system needs to perform quantitative analysis on the spatial topological relationships between multi-dimensional coordinates to replace the traditional character comparison mode. In this embodiment, the core computing module performs spatial distance measurement and hypersphere boundary determination operations as follows.

[0072] Before the comparison is executed, the core computing module obtains the baseline and target vectors for spatial comparison. Based on the settlement instructions from the supply chain business system, the core computing module extracts the document vector data corresponding to the invoice to be settled from memory, using it as the target vector for spatial comparison. Based on data association features, the core computing module extracts the document vector data corresponding to the inbound order associated with the invoice, establishing it as the baseline vector for spatial comparison. In handling complex business scenarios involving multiple order merging, this baseline vector uses the total inbound vector data output by the system's pre-processing synthesis operation.

[0073] Subsequently, the core calculation module performs a modified Euclidean distance calculation by introducing a tax rate sensitivity adjustment operator. In conventional three-dimensional spatial distance measurement, the contribution weight of each coordinate component to the final distance is basically equal. However, in the actual environment of supply chain settlement business, tax rate differences in invoice data involve strict tax compliance constraints. Even a slight deviation in the tax rate figure may lead to the invalidation of the entire document. To accurately reflect this business compliance constraint A in the spatial geometric model, the core calculation module introduces a specific numerical adjustment factor into the tax rate characteristic value dimension of the standard Euclidean distance formula.

[0074] The formula for calculating the corrected Euclidean distance is as follows: ; In the formula, Indicates the corrected Euclidean distance; This represents the standardized quantity of materials after processing the baseline vector. This represents the standardized quantity of materials in the target vector. Represents the net value of the document in the benchmark vector; Represents the net value of the target vector; The tax rate eigenvalues ​​represent the base vector. The characteristic value of the tax rate representing the target amount; This represents the tax rate sensitivity adjustment operator.

[0075] As a preferred approach, the system has a pre-defined tax rate sensitivity adjustment operator. This parameter is typically set to a value much greater than the constant 1. In a typical deployment environment, technicians can adjust this tax rate sensitivity operator. The value is set within the range of 100 to 1000. This tax rate sensitivity adjustment operator adjusts the tax rate even if there is a tax rate deviation of less than a decimal point between the target vector and the reference vector. This will have a significant multiplicative amplification effect on the squared term of the deviation, causing the final calculated corrected Euclidean distance value to increase dramatically. This algorithm structure with weighted bias greatly improves the system's sensitivity to deviation data in specific compliance dimensions, thereby triggering corresponding interception controls in the subsequent boundary determination stage.

[0076] Simultaneously, the core calculation module calculates the structural proportional consistency between the target vector and the reference vector in space. In addition to calculating the corrected Euclidean distance for the absolute dimension, the core calculation module further configures a cosine similarity algorithm to evaluate the proportional relationship of features within a single document. This operation is primarily used to quantitatively verify whether, under the premise that the unit price of the goods remains unchanged, the quantity of materials and the net value of the document maintain a synchronous scaling with the same proportion.

[0077] The formula for calculating the cosine similarity is as follows: ; In the formula, This represents the cosine similarity between the reference vector and the target vector. Represents the reference vector; Represents the target vector; The vector norm represents the reference vector; The vector norm of the target vector; This represents the angle between the reference vector and the target vector in space. Before performing the dot product and norm division calculations for this vector, the core calculation module reads each coordinate component of the target vector and the reference vector and performs a non-zero value verification operation to avoid triggering the algorithm logic error dead zone when the denominator is zero.

[0078] After obtaining the cosine similarity, the core calculation module incorporates it as an auxiliary constraint for spatial topology comparison. As a preferred approach, the system pre-sets a structural similarity threshold. This threshold is typically set between 0.95 and 0.99. The core calculation module compares the calculated cosine similarity with this preset value. When the cosine similarity is lower than the preset threshold, the system determines that the transaction has a risk of price tampering and adds a structural anomaly marker to the corresponding document vector data.

[0079] Based on the above spatial parameters, the core computing module performs hypersphere boundary determination calculations according to the calculated spatial distance parameters. The core computing module constructs a dynamic topology determination model within the system's computational space. In this model, the core computing module establishes the geometric center of the sphere using the coordinate position of the reference vector and the sphere radius using the dynamic tolerance limit output in step S300, thus constructing a closed determination hypersphere.

[0080] From a general perspective, this hypersphere model transforms traditional fixed threshold comparison into spatial tolerance calculation. In real-world business settlements, due to the allowable errors of physical weighing equipment or the slight discrepancies in tax allocation, invoice characteristics often do not perfectly align with inventory characteristics. This hypersphere model defines a reasonable deviation geometric space. As long as the offset of the target vector relative to the reference vector in the three-dimensional coordinate system does not exceed the dynamically calculated reasonable radius, the system geometrically determines that the business deviation is commercially reasonable.

[0081] The system's state function formula for determining whether the target vector lies within the boundary of the hypersphere is: ; In the formula, Indicates the matching determination status value; Indicates the corrected Euclidean distance; This indicates the dynamic passenger difference limit.

[0082] The core calculation module compares the calculated positive Euclidean distance with the dynamic tolerance limit. When the calculated positive Euclidean distance is less than or equal to the dynamic tolerance limit, the system determines that the target vector falls within the boundary of the hypersphere. In this case, the core calculation module will match the determination state value. The value is assigned to 1, and this status data is transmitted to the business routing module to trigger the subsequent invoicing and posting business instruction. When the calculation result shows that the corrected Euclidean distance is greater than the dynamic tolerance limit, the system determines that the target vector is outside the hypersphere boundary. In this deviation case, the core calculation module will match the determination status value. The value is assigned to 0 to block the automatic posting process, and the document vector data is extracted to prepare for subsequent attribution classification analysis of abnormal deviation parameters.

[0083] The supply chain matching and settlement system based on dynamic tolerance and vector space provided by this invention will trigger an exception handling process when the target vector fails to fall within the boundary range of the hypersphere during the aforementioned step S600. To perform refined diagnosis of the business data with matching failures and avoid pushing all abnormal data into the manual review queue, the core computing module is configured with a decision-making model based on the CART algorithm. In this embodiment, the specific logic of the core computing module performing deviation vector feature extraction and CART tree splitting node logical operations is as follows.

[0084] First, the core calculation module extracts the deviation vector feature data of the target vector relative to the reference vector. When the system determines that the matching state has failed, the core calculation module extracts the specific coordinate offset that caused the matching failure in the system's Euclidean three-dimensional space. Based on the reference vector and target vector generated in the previous calculation, the core calculation module performs vector subtraction to generate a set of independent deviation vectors.

[0085] The formula for calculating the above deviation vector is: ; In the formula, Represents the deviation vector; Represents the target vector; Represents the reference vector; Indicates the quantity deviation; Indicates the deviation in amount; This indicates the tax rate deviation. This represents the matrix transpose operation. By extracting this deviation vector, the system reduces the overall mismatch between business documents into specific deviation indicators across three independent dimensions: material quantity, document net value, and tax rate feature value, after standardization. These specific deviation indicators will serve as input feature data for the subsequent decision tree model.

[0086] After obtaining the deviation vector, the core calculation module inputs the extracted deviation vector into the first-level verification node of the CART decision tree to perform tax compliance attribute determination. In the anomaly attribution logic of supply chain settlement, the priority of tax anomaly determination is usually higher than physical loss and commercial price fluctuation. The core calculation module reads the tax rate deviation difference from the above deviation vector and inputs it into the first-level verification node of the decision tree. In this first-level verification node, the system's preset classification splitting condition is to determine whether the tax rate deviation difference is equal to zero. When the system determines that the tax rate deviation difference is not equal to zero, the business data will be guided to the first branch of the current node, and the core calculation module will directly output the attribution classification result of "tax compliance anomaly". When the system determines that the tax rate deviation difference is equal to zero, it indicates that the document's tax characteristics are compliant, and the business data will be guided to the second branch of the node to enter the subsequent level of determination process.

[0087] Next, the core calculation module uses the secondary verification node of the decision tree to perform physical attribute constraint judgment. For deviation vectors that pass the primary tax compliance verification, the core calculation module extracts the quantity deviation difference and calculates the relative quantity deviation rate. The relative quantity deviation rate is obtained by dividing the absolute value of the quantity deviation difference by the standardized material quantity in the baseline vector. To avoid triggering the algorithm logic error dead zone of zero denominator in the division, as a preferred method, the core calculation module performs a non-zero verification on the standardized material quantity feature of the baseline vector before performing the division. For special inbound documents such as pure expense documents where the material quantity is zero, the system will directly skip this secondary verification node.

[0088] After obtaining the relative quantity deviation rate, the system inputs it into a secondary verification node for conditional comparison. In the secondary verification node, the core calculation module reads the system's preset reasonable loss threshold for the specific material. This reasonable loss threshold is typically set based on the natural evaporation rate of the material during physical transportation or the inherent systematic error of the weighing equipment; for example, its value range can be set between 0.001 and 0.005. When the system determines that the calculated relative quantity deviation rate is greater than this reasonable loss threshold, it indicates that the quantity deviation of the document exceeds the reasonable range of natural loss, and the core calculation module outputs an attribution classification result of "logistics / warehouse management anomaly." If the system determines that the relative quantity deviation rate does not exceed the reasonable loss threshold, the data flows to the next level.

[0089] Subsequently, the core calculation module performs business attribute constraint judgment at the third-level verification node of the decision tree. For documents where the quantity deviation is within a reasonable range in the previous judgment but still triggers a tolerance warning overall, the system identifies the source of the deviation as the document's net value amount feature. The core calculation module extracts the amount deviation difference from the deviation vector and performs a comprehensive evaluation at the third-level verification node. As a specific implementation method, the core calculation module performs this evaluation by comparing unit price fluctuations through reverse calculation. The system extracts the document's net value amount from the target vector and performs a division operation with the standardized material quantity to obtain the target unit price. Similarly, it extracts the corresponding parameters from the benchmark vector to obtain the benchmark unit price. The core calculation module compares the difference between the target unit price and the benchmark unit price. When it is determined that the unit price difference exceeds the system's preset price fluctuation threshold range, it indicates that the amount deviation difference is mainly caused by unauthorized modification of the unit price or system pricing failure. The core calculation module outputs the attribution classification result of "price agreement abnormality" or "temporary price adjustment". When the unit price difference is determined to be within the preset price fluctuation threshold range, it indicates that the overall deviation of the document is caused by the combined resonance of multiple compliance dimensions, such as minor quantity losses and minor unit price discrepancies, resulting in the final vector space distance exceeding the hypersphere boundary. At this point, the core calculation module outputs the attribution classification result of "quantity-price coupling deviation anomaly," thereby eliminating the algorithm's decision dead zone. Through the above hierarchical progression and the CART tree splitting node logic covering multi-dimensional coupling branches, the core calculation module, to a certain extent, helps to automatically classify complex document mismatch phenomena into clear business anomaly categories such as taxation, warehousing and logistics, procurement commercial agreements, and comprehensive quantity-price deviations.

[0090] The supply chain matching and settlement system based on dynamic tolerance and vector space provided by this invention, after performing the aforementioned CART decision tree attribution, further achieves a closed loop for differentiated processing of abnormal data through evaluation and routing mechanisms. In this embodiment, the core computing module and the business diversion and routing module collaboratively execute attribution confidence and differentiated routing posting operations. The specific implementation method of this process is as follows.

[0091] To constrain the evaluation results, the core calculation module executes attribution confidence calculation logic based on multiplicative probabilities. Although the aforementioned CART decision tree can output clear attribution classification results, in complex supply chain operations, the judgment result of a single node often carries a certain risk of misjudgment. To impose secondary constraints on the attribution results output by the computer, the core calculation module equips each attribution classification result generated by the system with a confidence scoring mechanism.

[0092] In this embodiment, the core calculation module extracts the local conditional probability estimates corresponding to the current deviation vector as it passes through each decision tree node, and performs a continuous multiplication operation on these probability values ​​flowing through all nodes. This transforms the evaluation data of the local nodes into a probability confidence score for the overall attribution of the business transaction. As a specific implementation, the aforementioned local conditional probability estimates are derived from the frequency statistics of the system's historical manual judgment sample library. For each decision tree node, the core calculation module calculates the total number of samples flowing through that node in the most recent period and the number of samples manually judged as correct, using this to calculate the historical accuracy of that node as the conditional probability estimate. Considering the algorithmic pitfall of rapid numerical decay that may result from continuous multiplication, the system uses Laplace smoothing technology to set a smoothing lower limit for each conditional probability estimate before performing the multiplication operation, to avoid the extremely low probability of a single point directly causing the overall evaluation result to approach zero.

[0093] Subsequently, the core computing module performs a downgrade suspension decision based on the attribution confidence score. After obtaining the overall attribution probability confidence score, the core computing module compares this score with a preset human intervention threshold. As a preferred approach, this human intervention threshold is typically set within the range of 0.80 to 0.90. When the core computing module determines that the probability confidence score is less than this preset human intervention threshold, it indicates that the current machine-determined attribution classification result has high uncertainty. At this point, the system triggers a downgrade strategy, suspends automated routing, suspends the abnormal document, and directly pushes it into the human review queue. The actual judgment result after human review will be used as incremental sample feature data and fed back into the aforementioned parameter self-learning mechanism. When the judgment probability confidence score is greater than or equal to the human intervention threshold, the core computing module sends the attribution classification result to the business routing module to execute the subsequent dispatch process.

[0094] For high-confidence data, the business routing module performs differentiated routing operations for abnormal data. The business routing module receives the attribution classification results of high-confidence data and performs data distribution based on the distributed message queue engine mechanism. The business routing module has an internal routing mapping rule table. When the received attribution classification result is "price agreement abnormal" or "temporary price adjustment," the business routing module pushes the corresponding business data route to the procurement domain message queue, triggering the procurement staff to intervene in the price change process. When the attribution classification result is "logistics / warehouse management abnormal," the business routing module pushes the data to the warehouse domain message queue, and the system automatically connects to the warehouse management system or weighbridge monitoring module to verify the actual shipment data. When the attribution classification result is "tax compliance anomaly," the business routing module pushes the data to a dedicated queue at the financial compliance desk for processing by financial experts. For the underlying communication process of establishing cross-domain distributed message queues and consuming topics, those skilled in the art can use conventional message middleware such as Kafka or RabbitMQ. Their publish-subscribe mechanisms and asynchronous message transmission operations are well-known technologies in the field and will not be elaborated upon here.

[0095] Furthermore, the business routing module executes tolerance correction posting closed-loop logic for compliant minor deviation documents. Within the system's judgment framework, there exists a specific type of business scenario: the target vector falls within the hypersphere boundary (satisfying the condition that the corrected Euclidean distance is less than or equal to the dynamic tolerance limit), but the coordinate components of the deviation vector are not entirely equal to zero. These documents fall under the category of safe minor deviation documents within a reasonable tolerance range. For such documents, the business routing module extracts the amount deviation data from the deviation vector when triggering the formal posting instruction.

[0096] The business routing module generates a specific tolerance adjustment business voucher based on the amount deviation data. The system calls the distributed financial interface to automatically post the amount deviation value carried by the tolerance adjustment business voucher to the relevant "Tolerance Profit and Loss" account at the general ledger level. Through this automated posting logic, the system ensures that the supply chain settlement process is not interrupted by minor deviations, while to a certain extent meeting the macro-level balancing requirements of financial debit and credit accounting data.

[0097] Specific application examples and effect verification: Experimental scenario setup: Suppose a large manufacturing company handles settlement transactions for bulk commodities (such as coal). These materials experience normal physical evaporation during transportation, resulting in slight variations in weighbridge weight.

[0098] The system extracts the features of current business elements: (1) The material category is bulk cargo, and the system matches its material category risk coefficient. ; (2) For coal weighing in the trading industry, the industry correction offset for the average permissible error of the equipment. (Yuan); (3) The basic tolerance rate set by the enterprise (i.e., 0.05%).

[0099] A traditional fixed-proportion tolerance model (uniformly set at 1%) was selected as a comparative example, and its effectiveness was compared and verified with the dynamic tolerance model of this invention. The calculation formula for the dynamic tolerance limit of this invention is as follows: ; Effect verification and data comparison: For orders ranging from small purchases (10,000 yuan) to ultra-large settlements (10 million yuan), the system performs tolerance calculations. Experimental results are then compared. Figure 3 , Figure 3 The horizontal axis (X-axis) represents the total net value of the documents. (RMB), representing the actual net amount of a single transaction or combined batch in the settlement of supply business, showing a range from RMB 10,000 to RMB 10,000,000; the vertical axis (Y-axis) represents the tolerance limit. (Yuan), which is the maximum absolute deviation between the invoice feature and the warehouse entry feature calculated by the system. This value represents the hypersphere judgment radius constructed above in the three-dimensional vector space matching model.

[0100] Scenario A (Small-amount routine settlement, total net value of documents) Yuan): Traditional fixed tolerance (1%): limit is Yuan.

[0101] Dynamic tolerance of this invention: Calculated Yuan.

[0102] Performance Analysis: In small-amount settlements, the dynamic tolerance calculated by this invention (74 yuan) is on the same order of magnitude as the traditional tolerance (100 yuan). As shown in the initial interval of the experimental performance comparison chart, the dynamic tolerance limit curve of this invention can provide a reasonable tolerance base in the small-amount range, fully guaranteeing the basic tolerance space for grassroots business when facing reasonable weighing tail differences, and preventing business process bottlenecks.

[0103] Scenario B (Large-scale settlement, total net value of documents) Yuan): Traditional fixed tolerance (1%): limit is yuan (one hundred thousand yuan) Dynamic tolerance of this invention: Calculated Yuan.

[0104] Effect analysis combined with appendix Figure 3 Note: When dealing with massive invoices worth tens of millions, traditional proportional scaling can easily conceal malicious alterations or serious short-weighting by suppliers. See the experimental results for comparison. Figure 3 As shown by the black dashed line (representing the traditional fixed-ratio tolerance curve), it is a straight upward ray, which causes the absolute amount of the tolerance to diverge out of control in the range of large transactions. Figure 3 The solid black dots in the image mark this extremely large transaction involving tens of millions of yuan, visually demonstrating the risk exposure that even a huge deviation of up to 100,000 yuan would still be automatically allowed by the traditional system.

[0105] In contrast, this invention effectively reduces risk by utilizing the smooth decay characteristics of the logarithmic function. (See experimental results comparison.) Figure 3 As shown by the black solid line (representing the dynamic tolerance limit curve of the present invention), the slope of the curve becomes significantly gentler and smoother in the large range (right side of the X-axis). Figure 3 The black hollow squares in the diagram mark the actual computation points of the algorithm in this invention in the tens of millions scenario, intuitively demonstrating the effect of the system's anti-divergence technology in strictly converging the tolerance safety boundary and reducing it to about 42,000 yuan.

[0106] Experimental conclusion: The above embodiments, experimental data, and corresponding experimental effect comparison charts verify that the vector space hypersphere radius (i.e., the tolerance limit shown on the ordinate) of the present invention is within acceptable limits. This is not a rigid linear divergence. The dynamic tolerance model of this invention can provide a relatively lenient automatic matching boundary when the document amount is small; while when the document amount moves to a large amount, it can automatically use the logarithmic slope to reduce the relative tolerance ratio. This solution reduces the exposure risk of large fund settlements by more than 50% without increasing the cost of manual intervention, perfectly balancing settlement efficiency and fund risk control security.

Claims

1. A supply chain matching and settlement system based on dynamic tolerance and vector space, characterized in that, include: The data access module is used to perform standardized processing on heterogeneous supply chain document data and extract business element features. The heterogeneous supply chain document data includes warehouse entry orders and pending settlement invoice data. The core calculation module converts the business element features into three-dimensional coordinate components containing standardized material quantities, total net value of documents, and tax rate feature values ​​to construct document vector data. Extract the document vector data corresponding to the warehouse entry order data and the invoice data to be settled as the base vector and the target vector, respectively; A dynamic tolerance limit is generated based on the total net value of the document and the logarithmic decay algorithm. Calculate the spatial distance between the target vector and the reference vector, and determine whether the target vector falls within the boundary range of the determination hypersphere constructed by the dynamic tolerance limit; If the target does not fall within the target range, extract the deviation vector feature data of the target relative to the baseline vector, perform decision tree model calculations, and output the anomaly attribution classification result. The business routing module performs accounting posting when it determines that the target vector falls within the boundary range of the determined hypersphere, and performs differentiated routing allocation based on the anomaly attribution classification result when it determines that the target vector does not fall within the boundary range of the determined hypersphere, pushing the heterogeneous supply chain document data to the corresponding business processing queue.

2. The supply chain matching and settlement system based on dynamic tolerance and vector space according to claim 1, characterized in that, Also includes: The distributed storage and computing module is communicatively connected to the data access module, the core computing module, and the service routing module. The distributed storage and computing module is equipped with a K-Dimensional Tree spatial index structure, which is used to partition the high-dimensional data space and accelerate vector retrieval. The core computing module has anti-duplicate calculation verification logic. It generates a calculation fingerprint by collecting the core identifier and timestamp information of the document and combining it with a hash algorithm. The calculation fingerprint is then retrieved and verified in the memory bitmap of the distributed storage and computing module, and duplicate business data calculation requests are blocked.

3. The supply chain matching and settlement system based on dynamic tolerance and vector space according to claim 1, characterized in that: The business routing module is also used to, when performing accounting posting operations, extract the amount deviation data of the target vector relative to the benchmark vector for compliant micro-difference documents that fall within the boundary range of the determined hypersphere but whose coordinate components are not completely equal to those of the benchmark vector, generate tolerance adjustment business vouchers, and automatically post the amount deviation data to the preset tolerance profit and loss related accounts at the general ledger level.

4. A supply chain matching and settlement method based on dynamic tolerance and vector space, applied to the supply chain matching and settlement system based on dynamic tolerance and vector space as described in any one of claims 1 to 3, characterized in that, Includes the following steps: Acquire heterogeneous supply chain document data and perform data format standardization operations to extract business element features. The heterogeneous supply chain document data includes at least warehouse entry order data and pending settlement invoice data. The business element features are converted into three-dimensional coordinate components containing standardized material quantities, total net value of documents, and tax rate feature values ​​to construct document vector data. Based on the total net value of the document, the basic tolerance rate, the risk coefficient of the material category, and the industry correction offset, numerical calculations are performed using the logarithmic decay algorithm to calculate and generate the dynamic tolerance limit corresponding to the current business document. Extract the document vector data corresponding to the inbound order data as the base vector, extract the document vector data corresponding to the invoice data to be settled as the target vector, and calculate the spatial distance between the target vector and the base vector. A judgment hypersphere is constructed with the reference vector as the center and the dynamic tolerance limit as the radius. It is then determined whether the target vector falls within the boundary range of the judgment hypersphere. If the target vector is determined to fall within the boundary range of the determined hypersphere, the data matching is successful and the accounting posting operation is executed; If it is determined that the target vector does not fall within the boundary range of the determination hypersphere, the deviation vector feature data of the target vector relative to the reference vector is extracted, and classification and regression decision tree model operations are performed to output the anomaly attribution classification result. Based on the anomaly attribution classification result, the heterogeneous supply chain document data is pushed to the corresponding business processing queue.

5. The supply chain matching and settlement method according to claim 4, characterized in that, In the process of constructing document vector data, when it is determined that the acquired heterogeneous supply chain document data belongs to a consolidated settlement batch, multi-dimensional vector synthesis logic is executed: Extract multiple independent inbound order data belonging to the same business flow within the merged settlement batch, group them according to material code, and perform addition synthesis operation on the document vector data corresponding to multiple independent inbound order data with the same material code; When performing the addition synthesis operation, the standardized material quantity and the total net value of the document are added together. The mode of the tax rate feature values ​​of each independent document in the merged settlement batch is extracted for the tax rate feature value as the synthesized unified tax rate component, and the total inbound vector data is output as the base vector.

6. The supply chain matching and settlement method according to claim 4, characterized in that, The steps for calculating and generating the dynamic tolerance limit corresponding to the current business document include: Obtain the supplier's historical performance credit coefficient and map the supplier's historical performance credit coefficient to the supplier weight factor, and extract the preset system adjustment constant and minimum tolerance guarantee value, as well as the preset material category risk matrix; Perform a natural logarithmic calculation on the total net value of the aforementioned documents; The calculation is performed by dividing the natural logarithm of the total net value of the document with the division term of the system adjustment constant, combined with the basic tolerance rate, material category weight factor and supplier weight factor, and the minimum tolerance guarantee value is added to the final result. Before the calculation, a pre-verification logic is executed. When it is determined that the total net value of the document is less than or equal to the minimum value threshold preset by the system, the minimum tolerance guarantee value is directly output as the final dynamic tolerance limit.

7. The supply chain matching and settlement method according to claim 4, characterized in that, The step of calculating the spatial distance between the target vector and the reference vector specifically includes: Calculate the squared deviations between the target vector and the benchmark vector in the dimensions of material quantity and total net value of documents after standardization. Calculate the squared deviation between the target vector and the reference vector in the tax rate feature value dimension, and introduce a preset tax rate sensitivity adjustment operator before the squared deviation in the tax rate feature value dimension to perform a product amplification operation; The square root of the summation of the squared deviations in each dimension is then performed to obtain the corrected Euclidean distance, which is used as the spatial distance to participate in the boundary determination of the hypersphere. Further calculate the ratio of the dot product and norm product of the target vector and the reference vector to obtain the cosine similarity. When the cosine similarity is determined to be lower than the preset structural similarity threshold, add a structural anomaly marker to the corresponding document vector data.

8. The supply chain matching and settlement method according to claim 4, characterized in that, The steps of extracting the deviation vector feature data of the target vector relative to the baseline vector and performing classification and regression decision tree model operations specifically include: Perform vector subtraction to obtain deviation vector feature data containing quantity deviation, amount deviation, and tax rate deviation, and input the deviation vector feature data into the classification and regression decision tree model; At the first-level verification node of the classification and regression decision tree model, it is determined whether the tax rate deviation is zero. If it is not zero, the attribution classification result of the tax compliance abnormality is directly output. At the second-level verification node of the classification and regression decision tree model, the relative quantity deviation rate between the quantity deviation difference and the standardized material quantity in the benchmark vector is calculated. If the relative quantity deviation rate is greater than the preset reasonable loss threshold, the attribution classification result of logistics / warehouse management anomaly is output. At the third-level verification node of the classification and regression decision tree model, the target unit price and the benchmark unit price are obtained by division. When the difference between the target unit price and the benchmark unit price exceeds the preset price fluctuation threshold range, the attribution classification result of abnormal price agreement or temporary price adjustment is output. When the difference between the target unit price and the benchmark unit price does not exceed the preset price fluctuation threshold range, it is determined that the spatial distance exceeds the limit due to the superposition of multiple factors, and the attribution classification result of abnormal quantity-price coupling deviation is output.

9. The supply chain matching and settlement method according to claim 8, characterized in that, After outputting the anomaly attribution classification results, the following is also included: Extract the local conditional probability estimates corresponding to the deviation vector feature data as it flows through each verification node of the classification and regression decision tree model; Perform a continuous multiplication operation on the estimated local conditional probabilities flowing through all nodes to obtain the probability confidence score of the overall attribution of the heterogeneous supply chain document data; When the probability confidence score is determined to be less than the preset manual intervention threshold, the automated routing allocation is stopped, the heterogeneous supply chain document data is suspended and pushed into the manual review queue. When the probability confidence score is determined to be greater than or equal to the preset human intervention threshold, the differentiated routing allocation is executed.

10. The supply chain matching and settlement method according to claim 6, characterized in that, The supply chain matching and settlement method based on dynamic tolerance and vector space also includes a parameter self-learning step: Collect data samples generated by manual review to construct an offline training sample set, and convert the manually approved decision results into real matching labels; The difference between the spatial distance and the current dynamic tolerance limit is imported into a preset activation function to generate a predicted matching probability. The prediction error is calculated by substituting the true matching label and the predicted matching probability into the logarithmic operation logic of the cross-entropy loss function, and the gradient vector is obtained by taking the partial derivative through the optimization algorithm. Combining the preset learning rate and parameter truncation verification logic, the internal material category weight factor is iteratively updated along the opposite direction of the gradient.

Images