Supply chain accounting clearing and settlement system and method based on deep learning

Abstract

The invention discloses a supply chain accounting clearing and settlement system and method based on deep learning, and the method comprises the steps: S1, collecting accounting data in a supply chain, and carrying out the standardization processing; s2, processing the accounting data through a capsule network, and generating advanced feature representation of each bill element; s3, modeling a complex relationship between the bills by using a relationship reasoning module, establishing a relationship graph between the bills, and obtaining a final settlement sequence and an amount distribution scheme of the bills; s4, adopting historical accounting data as a training set, and training the capsule network and the relation reasoning module; s5, performing real-time automatic clearing and settlement processing on the newly generated bill data, and updating the settlement state in the supply chain; and S6, generating a detailed settlement report, and returning the settlement report to the supply chain node. According to the method, an efficient and scientific optimization scheme can be provided in supply chain accounting clearing and settlement, and remarkable technical values and economic benefits are brought to practical application.

Description

Technical Field

[0001] The present invention relates to the technical field of account clearing and settlement, and in particular to a supply chain account clearing and settlement system and method based on deep learning. Background Art

[0002] With the increasing complexity of global trade and supply chain management, supply chain account clearing and settlement has become a crucial part of corporate operations. Traditional account clearing and settlement usually relies on manual operations or rule-based automated systems to determine the payment amount and payment order of each bill by checking the bill data one by one. Existing technologies mainly rely on simple rule matching and data comparison. Although these methods can meet basic account settlement needs, they are inefficient and prone to errors when processing large-scale supply chain data. In addition, these methods have difficulty in dealing with complex relationships between bills, such as dynamic factors such as multi-level payment sequences, amount allocation, and transaction interactions between suppliers, resulting in a cumbersome account settlement process and prone to deviations.

[0003] Currently, accounting data analysis methods based on traditional neural networks have been proposed and applied in some account settlement systems. However, when processing account data, traditional convolutional neural networks focus primarily on extracting local features, especially features from image data, and fail to effectively capture the global dependencies between bills and the complex dynamic interactions between them. For example, while convolutional neural networks can process the visual and numerical features of bills, they have significant limitations in modeling complex relationships such as payment order and amount distribution between bills. Furthermore, existing account settlement systems struggle to automatically infer long-range dependencies and spatial relationships between bills, which restricts settlement accuracy and efficiency.

[0004] In this context, the introduction of capsule networks and relational reasoning provides a new solution for account clearing and settlement. Compared with traditional convolutional neural networks, capsule networks have stronger capabilities and can effectively capture the spatial and posture relationships in account data, solving the shortcomings of traditional neural networks in expressing long-range dependencies and spatial features. Through capsule units, the system can better represent the complex relationships and structured dependencies between bill elements, which provides a theoretical basis for more accurate account clearing and settlement. However, despite the advantages of capsule networks in processing account data, existing account clearing systems still face two major challenges: on the one hand, complex relationships between bills, such as payment order and amount distribution, are still difficult to effectively model using traditional methods; on the other hand, existing technologies lack reasoning mechanisms that can effectively utilize the dynamic relationships between bills, resulting in poor adaptability of settlement models in real-time data and complex transaction scenarios. In addition, although the introduction of relational reasoning modules can effectively capture the complex relationships between bills, in practical applications, how to reasonably design and optimize the modeling methods of these complex relationships to improve the accuracy and efficiency of the settlement process remains an urgent problem to be solved.

[0005] Based on the above background, this paper proposes a supply chain account clearing and settlement method based on capsule networks and relational reasoning. It aims to use deep learning and dynamic reasoning mechanisms to optimize the account clearing process, accurately model the complex relationships between bills, and improve the intelligence level of supply chain settlement through multi-level reasoning, thereby addressing the shortcomings of existing technologies. Summary of the Invention

[0006] One purpose of the present invention is to propose a supply chain account clearing and settlement system and method based on deep learning. The present invention can provide an efficient and scientific optimization solution in supply chain account clearing and settlement, bringing significant technical value and economic benefits to practical applications.

[0007] A supply chain account clearing and settlement method based on deep learning according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect accounting data from the supply chain and standardize it;

[0009] S2. Process the account data through a capsule network, assign and update features based on a dynamic routing mechanism, capture the spatial and pose relationships between account data elements, and generate high-level feature representations for each bill element.

[0010] S3: Based on the high-level feature representation of bill elements, the relational reasoning module is used to model the complex relationships between bills, establish a relationship graph between bills, and deduce the priority of each bill through the reasoning mechanism to obtain the final settlement order and amount distribution plan of the bills;

[0011] S4. Use historical account data as a training set and train the capsule network and relational reasoning module through backpropagation and gradient descent methods;

[0012] S5. Apply the trained capsule network and relational reasoning module to automatically sort and settle newly generated bill data in real time. Based on the reasoning results, the amounts between bills are matched, the payment sequence is determined, and the settlement status in the supply chain is automatically updated.

[0013] S6. Generate a detailed settlement report based on the results of automatic clearing and settlement, and return the settlement report to the supply chain node for payment processing.

[0014] Optionally, the S1 includes the following steps:

[0015] S11. Collect accounting data in the supply chain, including invoices, bills, payment records, and supplier information data;

[0016] S12. Pre-process the collected accounting data, including data cleaning, removing invalid and duplicate records, filling in missing data, and standardizing data in different formats;

[0017] S13. Normalize the accounting data.

[0018] Optionally, the S2 includes the following steps:

[0019] S21. Perform preliminary feature extraction on the collected accounting data using YOLOv10, and input the preliminary feature vector extracted by YOLOv10 into the capsule unit in the capsule network for further processing;

[0020] S22. In the capsule unit, features are allocated and updated through a dynamic routing mechanism to capture the spatial and posture relationships between accounting data elements;

[0021] Let the output vector of the i-th capsule be Through linear transformation and nonlinear activation function processing, the transformed output vector is obtained

[0022] s ij =R ij ·(v i +b i )+b ij ;

[0023] Among them, R ij is the rotation matrix from capsule i to capsule j, b i is the bias term of capsule i, b ij is the bias term from capsule i to capsule j. Through the rotation matrix and affine transformation, the model can capture the spatial transformation relationship between bills;

[0024] S23. Use the routing mechanism to establish the relationship between capsules and calculate the routing coefficient c between capsules. ij To get the influence of each capsule:

[0025]

[0026] in, is the weight vector of capsule j, s ij is the output vector from capsule i to capsule j, is the set of adjacent nodes of capsule i, exp(.) is the natural exponential function, c ij It is the normalized routing coefficient. It dynamically adjusts the weight transfer between capsules according to the impact of each capsule's output on the adjacent capsules, so that the network can dynamically adjust the information transfer according to the output strength of each capsule.

[0027] S24, by transforming each capsule into a vector s ij and routing coefficient c ij Perform weighted summation to update the output vector of each capsule:

[0028]

[0029] Among them, ||.|| represents the vector modulo operation, ∈ is a small constant to prevent division by zero errors, ||.|| 2 represents the Euclidean norm;

[0030] S25. According to the output result v′ of the capsule network j , generating high-level feature representations for each bill element.

[0031] Optionally, S3 includes the following steps:

[0032] S31. Based on the high-level feature representation of each bill element obtained from the capsule network, the relational reasoning module is used to model the complex relationships between bills and analyze the relevance, payment sequence, and amount distribution information between bills.

[0033] S32, by calculating the similarity S between bill i and bill j ij To build a relationship graph between bills, each node represents a bill element, and the edge represents the relationship between bill elements:

[0034]

[0035] Among them, sim(.) represents the similarity function, v i is the feature vector of bill i, v j is the feature vector of bill j, v ik and vjk are the kth element in the eigenvector, and d is the node degree;

[0036] S33. The feature update formula for each bill node i in the graph neural network is:

[0037]

[0038] in, A represents the feature vector of bill i after the l+1 layer graph convolution, ij is the relationship matrix element between bill i and bill j, is the degree of node i, represents the set of neighboring nodes adjacent to node i, σ is the sigmoid activation function;

[0039] S34. Based on graph convolution, we further use multi-layer perceptron to perform nonlinear modeling on the relationship between bill nodes. and Input into the multi-layer perceptron to calculate the payment order and amount distribution relationship between bills:

[0040]

[0041] in, and are the output feature vectors of bill i and bill j at the Lth layer of the graph convolutional network, respectively. MLP(·) represents a multi-layer perceptron, which learns the complex relationships between bills through multi-layer nonlinear mapping. This model is used to predict the settlement rules and priorities between bills.

[0042] S35. Based on the similarity between bills and the output of the graph neural network, a settlement priority matrix P is constructed:

[0043]

[0044] Among them, w is the weight vector, and are the outputs of bill i and bill j at the Lth layer in the graph convolutional network, P ij represents the settlement priority between bill i and bill j, exp(.) is the natural exponential function, and α is the adjustment factor;

[0045] S36. Combine the derived settlement priority matrix with the amount allocation strategy to obtain the final settlement order and amount allocation plan of the bill.

[0046] Optionally, the S4 includes the following steps:

[0047] S41, using historical account data as a training set for training;

[0048] S42, use the gradient descent method to optimize the loss function and calculate each parameter θ j Gradient The formula is as follows:

[0049]

[0050] Among them, η is the learning rate, which controls the step size of each parameter update. Represents the loss function L(θ) for the parameter θ j gradient;

[0051] S43, use the back propagation algorithm to calculate the gradient of each layer of the network, update the parameters of the capsule model layer by layer, and calculate the output error δ of each layer (k) , the formula is as follows:

[0052] δ (k) =(W (k+1) ) T δ (k+1) ·f′(z (k) );

[0053] Among them, δ (k) is the error term of the kth layer, W (k+1) is the weight matrix of the k+1th layer, z (k) is the weighted input of the kth layer, f′(z (k) ) is the derivative of the activation function.

[0054] Optionally, the S5 includes the following steps:

[0055] S51. Perform real-time automatic clearing and settlement processing on the newly generated bill data. Input the new bill data into the trained capsule network and relational reasoning model to obtain the clearing and settlement information of each bill.

[0056] S52. Based on the inference results, calculate the amount matching between bills, determine the payment order of each bill, assign a payment time sequence to each bill, and automatically update the settlement status in the supply chain.

[0057] The supply chain account clearing and settlement system based on deep learning according to an embodiment of the present invention includes the following modules:

[0058] Data collection module, used to collect accounting data in the supply chain and perform standardized processing;

[0059] The feature extraction module uses YOLOv10 to perform preliminary feature extraction on the collected account data, obtain basic visual and numerical information of the account data, and further extract the spatial and posture relationships between bills through the capsule network;

[0060] The relationship modeling module is used to model the complex relationships between bills based on the relationship reasoning module, analyze the payment sequence and amount distribution factors between bills, and generate the priority and settlement rules between bills;

[0061] The model training module uses historical account data as a training set to train the capsule network and relational reasoning module through backpropagation and gradient descent optimization methods;

[0062] The real-time clearing and settlement module is used to automatically clear and settle newly generated bill data in real time. It matches the amounts between bills and determines the payment sequence based on the inference results, and automatically updates the settlement status in the supply chain.

[0063] The settlement report generation module is used to generate a detailed settlement report based on the results of automatic clearing and settlement, and return the settlement report to the supply chain node for payment processing.

[0064] The beneficial effects of the present invention are:

[0065] (1) The present invention effectively improves the spatial relationship modeling capability in the account clearing and settlement process by introducing capsule networks and relational reasoning modules. Compared with traditional convolutional neural networks, the present invention can capture the spatial and posture relationships between elements in bill data and accurately represent the complex dependencies between bills. This enables the system to more accurately understand and process long-distance dependencies and global relationships in account data, thereby improving the accuracy and reliability of account clearing.

[0066] (2) By introducing a relational reasoning module, the present invention can accurately model the dynamic relationships between bills, such as the payment order and amount distribution. On this basis, the system can automatically deduce the settlement priority and rules of the bills, thereby avoiding the limitations of traditional methods that rely on static rules and simple matching. The present invention significantly improves the efficiency and accuracy of the account clearing and settlement process by analyzing and reasoning about the complex relationships between bills.

[0067] (3) The present invention achieves higher accuracy and flexibility in settlement priority calculation by combining multi-dimensional information such as similarity between bills, payment order, and amount distribution. Compared with the existing technology, the present invention can automatically adjust the settlement order and amount distribution of bills in real time according to dynamic changes in account data, thereby improving the adaptability and global optimization capability of settlement decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0069] Figure 1This is a flow chart of the supply chain account clearing and settlement method based on deep learning proposed by the present invention;

[0070] Figure 2 This is a structural diagram of the supply chain account clearing and settlement system based on deep learning proposed by the present invention. DETAILED DESCRIPTION

[0071] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0072] refer to Figure 1 , a supply chain account clearing and settlement method based on deep learning includes the following steps:

[0073] S1. Collect accounting data from the supply chain and standardize it;

[0074] S2. Process the account data through a capsule network, assign and update features based on a dynamic routing mechanism, capture the spatial and pose relationships between account data elements, and generate high-level feature representations for each bill element.

[0075] S3: Based on the high-level feature representation of bill elements, the relational reasoning module is used to model the complex relationships between bills, establish a relationship graph between bills, and deduce the priority of each bill through the reasoning mechanism to obtain the final settlement order and amount distribution plan of the bills;

[0076] S4. Use historical account data as a training set and train the capsule network and relational reasoning module through backpropagation and gradient descent methods;

[0077] S5. Apply the trained capsule network and relational reasoning module to automatically sort and settle newly generated bill data in real time. Based on the reasoning results, the amounts between bills are matched, the payment sequence is determined, and the settlement status in the supply chain is automatically updated.

[0078] S6. Generate a detailed settlement report based on the results of automatic clearing and settlement, and return the settlement report to the supply chain node for payment processing.

[0079] This invention collects accounting data from the supply chain, including invoices, bills, and payment records, and standardizes this data to ensure data consistency and integrity during subsequent analysis. This step enables the system to collect and preprocess accounting information from multiple sources, thereby improving data quality and providing a reliable foundation for subsequent feature extraction and inference modeling. This standardized processing avoids errors caused by inconsistent data formats and ensures the efficiency and accuracy of the clearing and settlement process.

[0080] In this embodiment, S1 includes the following steps:

[0081] S11. Collect accounting data in the supply chain, including invoices, bills, payment records, and supplier information data;

[0082] S12. Pre-process the collected accounting data, including data cleaning, removing invalid and duplicate records, filling in missing data, and standardizing data in different formats;

[0083] S13. Normalize the accounting data.

[0084] This invention further processes account data using capsules to capture the spatial and gestural relationships between bill elements. This step enhances the representation of long-range dependencies in account data through capsule units, addressing the inability of traditional neural network methods to effectively handle complex spatial relationships. The introduction of capsule networks effectively improves the system's ability to model global relationships between bills, enhancing the accuracy and flexibility of the settlement process. In particular, when processing complex bill data, it can accurately identify and preserve the hierarchical features between bills.

[0085] In this embodiment, S2 includes the following steps:

[0086] S21. Perform preliminary feature extraction on the collected accounting data using YOLOv10, and input the preliminary feature vector extracted by YOLOv10 into the capsule unit in the capsule network for further processing;

[0087] S22. In the capsule unit, features are allocated and updated through a dynamic routing mechanism to capture the spatial and posture relationships between accounting data elements;

[0088] Let the output vector of the i-th capsule be Through linear transformation and nonlinear activation function processing, the transformed output vector is obtained

[0089] s ij =R ij ·(v i +b i )+b ij ;

[0090] Among them, R ij is the rotation matrix from capsule i to capsule j, b i is the bias term of capsule i, b ij is the bias term from capsule i to capsule j. Through the rotation matrix and affine transformation, the model can capture the spatial transformation relationship between bills;

[0091] S23. Use the routing mechanism to establish the relationship between capsules and calculate the routing coefficient c between capsules. ij To get the influence of each capsule:

[0092]

[0093] in, is the weight vector of capsule j, s ij is the output vector from capsule i to capsule j, is the set of adjacent nodes of capsule i, exp(.) is the natural exponential function, c ij It is the normalized routing coefficient. It dynamically adjusts the weight transfer between capsules according to the impact of each capsule's output on the adjacent capsules, so that the network can dynamically adjust the information transfer according to the output strength of each capsule.

[0094] S24, by transforming each capsule into a vector s ij and routing coefficient c ij Perform weighted summation to update the output vector of each capsule:

[0095]

[0096] Among them, ||.|| represents the vector modulo operation, ∈ is a small constant to prevent division by zero errors, ||.|| 2 represents the Euclidean norm;

[0097] S25. According to the output result v′ of the capsule network j , generating high-level feature representations for each bill element.

[0098] The present invention models the complex relationships between bills through a relational reasoning module, and can dynamically analyze factors such as the payment order and amount distribution between bills. This step effectively constructs a relationship diagram between bills and derives the settlement rules and priorities of each bill through a reasoning mechanism. By considering the complex relationships between bills, the present invention avoids the simplified processing of account settlement in traditional methods, ensures accurate matching and reasonable distribution in the settlement process, and greatly improves the decision-making ability of the system.

[0099] In this embodiment, S3 includes the following steps:

[0100] S31. Based on the high-level feature representation of each bill element obtained from the capsule network, the relational reasoning module is used to model the complex relationships between bills and analyze the relevance, payment sequence, and amount distribution information between bills.

[0101] S32, by calculating the similarity S between bill i and bill j ijTo build a relationship graph between bills, each node represents a bill element, and the edge represents the relationship between bill elements:

[0102]

[0103] Among them, sim(.) represents the similarity function, v i is the feature vector of bill i, v j is the feature vector of bill j, v ik and v jk are the kth element in the eigenvector, and d is the node degree;

[0104] S33. The feature update formula for each bill node i in the graph neural network is:

[0105]

[0106] in, A represents the feature vector of bill i after the l+1 layer graph convolution, ij is the relationship matrix element between bill i and bill j, is the degree of node i, represents the set of neighboring nodes adjacent to node i, σ is the sigmoid activation function;

[0107] S34. Based on graph convolution, we further use multi-layer perceptron to perform nonlinear modeling on the relationship between bill nodes. and Input into the multi-layer perceptron to calculate the payment order and amount distribution relationship between bills:

[0108]

[0109] in, and are the output feature vectors of bill i and bill j at the Lth layer of the graph convolutional network, respectively. MLP(·) represents a multi-layer perceptron, which learns the complex relationships between bills through multi-layer nonlinear mapping. This model is used to predict the settlement rules and priorities between bills.

[0110] S35. Based on the similarity between bills and the output of the graph neural network, a settlement priority matrix P is constructed:

[0111]

[0112] Among them, w is the weight vector, and are the outputs of bill i and bill j at the Lth layer in the graph convolutional network, P ijrepresents the settlement priority between bill i and bill j, exp(.) is the natural exponential function, and α is the adjustment factor;

[0113] S36. Combine the derived settlement priority matrix with the amount allocation strategy to obtain the final settlement order and amount allocation plan of the bill.

[0114] The present invention trains a model using historical account data and uses optimization methods such as backpropagation and gradient descent to adjust the model's parameters to improve the accuracy and efficiency of the account clearing and settlement process. Through training, the present invention can continuously optimize the model's adaptability in different account data environments and enhance the system's performance in practical applications. This step provides the system with self-learning capabilities, enabling it to continuously improve settlement accuracy as data accumulates, avoiding the limitations of manual intervention and hard-coded rules.

[0115] In this embodiment, S4 includes the following steps:

[0116] S41, using historical account data as a training set for training;

[0117] S42, use the gradient descent method to optimize the loss function and calculate each parameter θ j Gradient The formula is as follows:

[0118]

[0119] Among them, η is the learning rate, which controls the step size of each parameter update. Represents the loss function L(θ) for the parameter θ j gradient;

[0120] S43, use the back propagation algorithm to calculate the gradient of each layer of the network, update the parameters of the capsule model layer by layer, and calculate the output error δ of each layer (k) , the formula is as follows:

[0121] δ (k) =(W (k+1) ) T δ (k+1) ·f′(z (k) );

[0122] Among them, δ (k) is the error term of the kth layer, W (k+1) is the weight matrix of the k+1th layer, z (k) is the weighted input of the kth layer, f′(z (k) ) is the derivative of the activation function.

[0123] The present invention uses a trained model to perform real-time clearing and settlement processing on newly generated bill data. This step reduces the need for manual participation by automating the processing of new bill data, and improves the real-time performance and efficiency of account clearing and settlement. The model's inference results can not only match the amounts between bills and determine the payment sequence, but also automatically update the settlement status in the supply chain, thereby improving the system's responsiveness and accuracy in a rapidly changing supply chain environment.

[0124] In this embodiment, S5 includes the following steps:

[0125] S51. Perform real-time automatic clearing and settlement processing on the newly generated bill data. Input the new bill data into the trained capsule network and relational reasoning model to obtain the clearing and settlement information of each bill.

[0126] S52. Based on the inference results, calculate the amount matching between bills, determine the payment order of each bill, assign a payment time sequence to each bill, and automatically update the settlement status in the supply chain.

[0127] refer to Figure 2 , a supply chain account clearing and settlement system based on deep learning, including the following modules:

[0128] Data collection module, used to collect accounting data in the supply chain and perform standardized processing;

[0129] The feature extraction module uses YOLOv10 to perform preliminary feature extraction on the collected account data, obtain basic visual and numerical information of the account data, and further extract the spatial and posture relationships between bills through the capsule network;

[0130] The relationship modeling module is used to model the complex relationships between bills based on the relationship reasoning module, analyze the payment sequence and amount distribution factors between bills, and generate the priority and settlement rules between bills;

[0131] The model training module uses historical account data as a training set to train the capsule network and relational reasoning module through backpropagation and gradient descent optimization methods;

[0132] The real-time clearing and settlement module is used to automatically clear and settle newly generated bill data in real time. It matches the amounts between bills and determines the payment sequence based on the inference results, and automatically updates the settlement status in the supply chain.

[0133] The settlement report generation module is used to generate a detailed settlement report based on the results of automatic clearing and settlement, and return the settlement report to the supply chain node for payment processing.

[0134] Example:

[0135] In a certain large electronic product manufacturing enterprise, the enterprise maintains cooperative relationships with multiple parts suppliers around the world. A large amount of accounting data is added every day. The invoice content provided by each supplier is different, including multiple information fields such as invoice amount, payment terms, discounts, taxes, and payment method. Since multiple suppliers provide invoices at the same time and there are complex payment sequences and amount distribution relationships between the invoices, traditional rule-based account clearing systems have difficulty handling these complex situations. High manual involvement, slow processing speed, and easy errors are common problems in traditional systems.

[0136] In this context, enterprises have introduced deep learning methods based on capsule networks and relational reasoning to achieve automation and intelligence in account clearing and settlement, and optimize the supply chain settlement process. The system needs to be able to automatically identify the payment order, amount distribution and other relationships between bills, and provide the optimal clearing and settlement plan.

[0137] After the enterprise applies the account clearing and settlement method of the present invention, it first collects bill information from different suppliers through the data acquisition module and standardizes the bill data. The system converts various information in the bill (such as amount, invoice number, supplier information, etc.) into numerical feature vectors for subsequent deep learning model processing.

[0138] YOLOv10 then performs preliminary feature extraction on this data, extracting local features from the bill data, such as bill amount, supplier information, and payment terms. Next, the capsule network module models the spatial relationships between bills, enabling the system to capture complex and long-distance dependencies between bills, such as the payment order and amount distribution across multiple bills from the same supplier.

[0139] The relational reasoning module then infers the complex relationships between bills, analyzes factors such as the payment order and amount distribution between bills, and generates priorities and settlement rules between bills. In this way, the system can not only identify the direct relationships between bills, but also handle the implicit relationships between bills, greatly improving the accuracy and flexibility of settlement.

[0140] Through training with historical accounting data, the system can automatically clear and settle new bills in real-time, reducing manual intervention and human errors. After each settlement, the system will also automatically generate a detailed settlement report, including bill payment details, amount distribution, settlement status and other information, for further audit and payment by relevant supply chain nodes.

[0141] In one real-world application, a company used the system to clear and settle accounts between January 1, 2025, and January 31, 2025, involving 2,200 invoices from 50 suppliers. Traditional manual clearing took seven days to complete, often with significant errors in the results. After the system was put into use, the settlement time was shortened to two days for the same number of invoices, and the settlement error rate was reduced from 15% with the traditional system to 2%. The following is a comparison of the specific data:

[0142] Table 1 Performance comparison between the present invention and traditional methods in account data clearing and settlement

[0143] project Traditional systems System of the present invention Number of bills 2200 sheets 2200 sheets Total settlement time 7 days 2 days lump sum About 1.2 billion RMB About 1.2 billion RMB Settlement error rate 15％ 2％ Number of manual interventions 150 times 30 times Automatically generate settlement reports Unable to generate automatically 30 minutes Number of supplier settlement corrections 50 times 5 times

[0144] The data comparison clearly shows that the application of the present invention's method significantly shortens settlement time, significantly reduces error rates, and reduces the number of manual interventions, thereby improving overall settlement efficiency. Furthermore, the system automatically generates settlement reports, reducing the workload of manual report generation and avoiding potential errors caused by manual operation.

[0145] This example fully demonstrates the application effect of the supply chain account clearing and settlement method based on capsule networks and relational reasoning in practical scenarios. Through intelligent account clearing and settlement, enterprises have significantly improved the efficiency and accuracy of settlement, reduced the need for manual intervention, and optimized the supply chain settlement process. The application of the system not only improves settlement accuracy, but also enhances the overall efficiency and transparency of supply chain management by automatically generating settlement reports and processing account data in real time.

[0146] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A supply chain account clearing and settlement method based on deep learning, characterized by: The steps include: S1. Collect accounting data from the supply chain and standardize it; S2. Process the account data through a capsule network, assign and update features based on a dynamic routing mechanism, capture the spatial and pose relationships between account data elements, and generate high-level feature representations for each bill element. S3: Based on the high-level feature representation of bill elements, the relational reasoning module is used to model the complex relationships between bills, establish a relationship graph between bills, and deduce the priority of each bill through the reasoning mechanism to obtain the final settlement order and amount distribution plan of the bills; S4. Use historical account data as a training set and train the capsule network and relational reasoning module through backpropagation and gradient descent methods; S5. Apply the trained capsule network and relational reasoning module to automatically sort and settle newly generated bill data in real time. Based on the reasoning results, the amounts between bills are matched, the payment sequence is determined, and the settlement status in the supply chain is automatically updated. S6. Generate a detailed settlement report based on the results of automatic clearing and settlement, and return the settlement report to the supply chain node for payment processing.

2. The supply chain account clearing and settlement method based on deep learning according to claim 1 is characterized in that: Said S1 comprises the following steps: S11. Collect accounting data in the supply chain, including invoices, bills, payment records, and supplier information data; S12. Pre-process the collected accounting data, including data cleaning, removing invalid and duplicate records, filling in missing data, and standardizing data in different formats; S13. Normalize the accounting data.

3. The supply chain account clearing and settlement method based on deep learning according to claim 1 is characterized in that: The S2 comprises the following steps: S21. Perform preliminary feature extraction on the collected accounting data using YOLOv10, and input the preliminary feature vector extracted by YOLOv10 into the capsule unit in the capsule network for further processing; S22. In the capsule unit, features are allocated and updated through a dynamic routing mechanism to capture the spatial and posture relationships between accounting data elements; Let the output vector of the i-th capsule be Through linear transformation and nonlinear activation function processing, the transformed output vector is obtained s ij =R ij ·(v i +b i )+b ij ; Among them, R ij is the rotation matrix from capsule i to capsule j, b i is the bias term of capsule i, b ij is the bias term from capsule i to capsule j. Through the rotation matrix and affine transformation, the model can capture the spatial transformation relationship between bills; S23. Use the routing mechanism to establish the relationship between capsules and calculate the routing coefficient c between capsules. ij To get the influence of each capsule: in, is the weight vector of capsule j, s ij is the output vector from capsule i to capsule j, N(i) is the set of adjacent nodes of capsule i, exp(.) is the natural exponential function, c ij is the normalized routing coefficient; S24, by transforming each capsule into a vector s ij and routing coefficient c ij Perform weighted summation to update the output vector of each capsule: Among them, ||.|| represents the vector modulo operation, ∈ is a small constant to prevent division by zero errors, ||.|| 2 represents the Euclidean norm; S25, according to the output result v of the capsule network ′ j , generating high-level feature representations for each bill element.

4. The supply chain account clearing and settlement method based on deep learning according to claim 1 is characterized in that: The S3 includes the following steps: S31. Based on the high-level feature representation of each bill element obtained from the capsule network, the relational reasoning module is used to model the complex relationships between bills and analyze the relevance, payment sequence, and amount distribution information between bills. S32, by calculating the similarity S between bill i and bill j ij To build a relationship graph between bills, each node represents a bill element, and the edge represents the relationship between bill elements: Among them, sim(.) represents the similarity function, v i is the feature vector of bill i, v j is the feature vector of bill j, v ik and v jk are the kth element in the eigenvector, and d is the node degree; S33. The feature update formula for each bill node i in the graph neural network is: in, A represents the feature vector of bill i after the l+1 layer graph convolution, ij is the relationship matrix element between bill i and bill j, is the degree of node i, represents the set of neighboring nodes adjacent to node i, σ is the sigmoid activation function; S34. Based on graph convolution, we further use multi-layer perceptron to perform nonlinear modeling on the relationship between bill nodes. and Input into the multi-layer perceptron to calculate the payment order and amount distribution relationship between bills: in, and are the output feature vectors of bill i and bill j at the Lth layer of the graph convolutional network, respectively. MLP(·) represents a multi-layer perceptron, which learns the complex relationships between bills through multi-layer nonlinear mapping. This model is used to predict the settlement rules and priorities between bills. S35. Based on the similarity between bills and the output of the graph neural network, a settlement priority matrix P is constructed: Among them, w is the weight vector, and are the outputs of bill i and bill j at the Lth layer in the graph convolutional network, P ij represents the settlement priority between bill i and bill j, exp(.) is the natural exponential function, and α is the adjustment factor; S36. Combine the derived settlement priority matrix with the amount allocation strategy to obtain the final settlement order and amount allocation plan of the bill.

5. The supply chain account clearing and settlement method based on deep learning according to claim 1 is characterized in that: The S4 comprises the following steps: S41, using historical account data as a training set for training; S42, use the gradient descent method to optimize the loss function and calculate each parameter θ j Gradient The formula is as follows: Among them, η is the learning rate, which controls the step size of each parameter update. Represents the loss function L(θ) for the parameter θ j gradient; S43, use the back propagation algorithm to calculate the gradient of each layer of the network, update the parameters of the capsule model layer by layer, and calculate the output error δ of each layer (k) , the formula is as follows: δ (k) =(W (k+1) ) T δ (k+1) ·f′(z (k) ); Among them, δ (k) is the error term of the kth layer, W (k+1) is the weight matrix of the k+1th layer, z (k) is the weighted input of the kth layer, f ′ (z (k) ) is the derivative of the activation function.

6. The supply chain account clearing and settlement method based on deep learning according to claim 1 is characterized in that: The S5 comprises the following steps: S51. Perform real-time automatic clearing and settlement processing on the newly generated bill data. Input the new bill data into the trained capsule network and relational reasoning model to obtain the clearing and settlement information of each bill. S52. Based on the inference results, calculate the amount matching between bills, determine the payment order of each bill, assign a payment time sequence to each bill, and automatically update the settlement status in the supply chain.

7. A supply chain account clearing and settlement system based on deep learning, which implements the supply chain account clearing and settlement method based on deep learning according to any one of claims 1 to 6, characterized in that: Includes the following modules: Data collection module, used to collect accounting data in the supply chain and perform standardized processing; The feature extraction module uses YOLOv10 to perform preliminary feature extraction on the collected account data, obtain basic visual and numerical information of the account data, and further extract the spatial and posture relationships between bills through the capsule network; The relationship modeling module is used to model the complex relationships between bills based on the relationship reasoning module, analyze the payment sequence and amount distribution factors between bills, and generate the priority and settlement rules between bills; The model training module uses historical account data as a training set to train the capsule network and relational reasoning module through backpropagation and gradient descent optimization methods; The real-time clearing and settlement module is used to automatically clear and settle newly generated bill data in real time. It matches the amounts between bills and determines the payment sequence based on the inference results, and automatically updates the settlement status in the supply chain. The settlement report generation module is used to generate a detailed settlement report based on the results of automatic clearing and settlement, and return the settlement report to the supply chain node for payment processing.

Images