A large-scale model-driven supply chain collaboration method and system

Abstract

The present invention relates to the field of supply management technology, and discloses a large-model driven supply chain collaboration method and system; S01: dividing inherent multi-source heterogeneous data into standard data and non-standard data, S02: parsing key semantics from standard data and non-standard data, S03: combining target decisions from scenario plans, S04: determining whether to execute a decision correction operation, and S05: formulating a correction strategy through an intelligent correction model, and collaboratively correcting the supply chain; the present invention avoids the inefficiency and lag of manual analysis and extraction of key semantics, circumvents the drawbacks of large models that lack logical reasoning and numerical calculation capabilities when dealing with complex supply chain problems, breaks the information island phenomenon of enterprises in the supply chain, effectively reduces the potential risks and collaboration costs of enterprises in the supply chain, improves the collaborative efficiency of the supply chain, and ensures the stability and smoothness of resource scheduling in the supply chain.

Description

Technical Field

[0001] The present invention relates to the technical field of supply management, and more particularly, to a large model-driven supply chain collaboration method and system. Background Art

[0002] With the continuous acceleration of globalization and intelligentization, enterprises are also facing huge challenges and opportunities in supply chain optimization. The traditional method of relying on manual analysis and judgment to coordinate the adjustment of the supply chain can no longer meet the collaborative optimization needs of today's enterprise supply chains. Therefore, it is necessary to combine large models to dynamically and efficiently coordinate the various supply processes of the supply chain, thereby achieving efficient collaborative optimization effects of the enterprise supply chain.

[0003] Patent application CN119599403A discloses an AI-based collaborative supply chain management method and system. This system obtains supply data for parts provided by each of N supply chains during different time periods, as well as sales data for products during different time periods. The system then fuses the supply and sales data to generate fused data, allowing the features between the data to be aligned. This fused data can then be analyzed using a neural network model to obtain the supply data for parts that each of the N supply chains will need to provide in future time periods, enabling advance prediction and thus improving the reliability and stability of collaborative supply chain management.

[0004] When existing supply chains are collaboratively processed, they can perform collaborative reasoning on the supply chain by building knowledge graphs and large models, achieving an overall analysis of multi-source heterogeneous data in the supply chain, thereby improving the accuracy of supply chain decisions. However, the overall unified analysis of multi-source heterogeneous data cannot separate and analyze standard data from non-standard data, making it impossible for large models to quickly and accurately understand the core essence of complex and diverse problems in the supply chain. This leads to a lack of targeted logical reasoning capabilities in the decision-making analysis process of the supply chain, and thus easily leads to wrong decisions in procurement sourcing or inventory allocation in the supply chain, reducing the stability and smoothness of resource scheduling in the supply chain.

[0005] In view of this, the present invention proposes a large model-driven supply chain collaboration method and system to solve the above problems. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art and achieve the above-mentioned objectives, the present invention provides the following technical solution: a large model-driven supply chain collaboration method, applied to a collaborative cloud server, comprising:

[0007] S01: Collect the inherent multi-source heterogeneous data of the supply chain, extract the data encoding, data format and key fields of the inherent multi-source heterogeneous data, and divide the inherent multi-source heterogeneous data into standard data and non-standard data;

[0008] S02: Based on the model parsing criteria, intelligently parse standard data and non-standard data by parsing the large model, and extract key semantics from the standard data and non-standard data;

[0009] The model analysis criteria are: standard data adopts an overall unified analysis method, and non-standard data adopts a single independent analysis method;

[0010] S03: Construct a decision scenario corresponding to the key semantics, simulate the scenario plan of the decision scenario through the scenario plan model, and combine the target decision from the scenario plan;

[0011] S04: Collect real-time multi-source heterogeneous data from the supply chain, extract risk features from the real-time multi-source heterogeneous data, and determine whether to perform decision correction operations based on supply risk criteria;

[0012] S05: If a decision correction operation is performed, a correction strategy is formulated through an intelligent correction model based on risk characteristics and target decisions, and the supply chain is collaboratively corrected.

[0013] Furthermore, when collecting multi-source heterogeneous data, the data storage system is recorded as a data node, and dimensional data is collected from M distributed data nodes one by one. After the collected dimensional data is aggregated, the inherent multi-source heterogeneous data of the supply chain is obtained;

[0014] The data encoding, data format and key field extraction methods are as follows:

[0015] The encoding frame and format frame of multi-source heterogeneous data are queried through the data management system, and the characters in the encoding frame are recorded as data encoding, and the characters in the format frame are recorded as data format;

[0016] The text content of multi-source heterogeneous data is identified one by one through OCR technology, the required items are marked from the text content, and the content in the required items is recorded as key fields.

[0017] Furthermore, the division method of standard data and non-standard data is as follows:

[0018] Compare the data encoding of the multi-source heterogeneous data with the standard encoding for consistency, and record the multi-source heterogeneous data whose data encoding is consistent with the standard encoding as the first verification data;

[0019] Comparing the data format of the first verification data with the standard format for consistency, and recording the first verification data whose data format is consistent with the standard format as the second verification data;

[0020] The key fields of the second verification data are compared with the standard fields for consistency, the second verification data whose key fields are consistent with the standard fields are recorded as standard data, and all remaining multi-source heterogeneous data are recorded as non-standard data, obtaining A standard data and B non-standard data respectively.

[0021] Furthermore, key semantics include standard semantics and non-standard semantics. The parsing methods for standard semantics and non-standard semantics are:

[0022] Find the mode switch box of the large model, adjust the character of the mode switch box to "ZT1", and clear the remaining parsing text in the parsing box of the large model;

[0023] Aggregate and combine A standard data into a first parsed text, and import the first parsed text into the parsing frame of the parsing model. At the same time, the parsing model performs overall intelligent parsing on the first parsed text in the parsing frame to parse out the standard semantics.

[0024] After the parsing of A standard data is completed, adjust the character of the mode switch box to "DL2" and clear the first parsing text in the parsing box of the parsing large model;

[0025] Convert B non-standard data into B second parsed texts respectively, and import the B second parsed texts into the parsing box of the parsing model in sequence. According to the order of non-standard data collection time, the parsing model performs overall intelligent parsing on the B second parsed texts in the parsing box one by one to parse out B second parsed semantics;

[0026] The repeated second parsed semantics are recorded as repeated semantics, and redundant repeated semantics are eliminated until the number of repeated semantics is 1. The remaining repeated semantics and the second parsed semantics are aggregated to generate non-standard semantics.

[0027] Furthermore, the decision-making scenarios include standard scenarios and non-standard scenarios. The construction methods of standard scenarios and non-standard scenarios are as follows:

[0028] The large language model is used to identify key fields of standard semantics and non-standard semantics respectively, and the identified key fields are summarized to generate a standard field set and a non-standard field set;

[0029] Taking two key fields as a unit, any two key fields in the standard field set and the non-standard field set are combined to generate D standard units and E non-standard units;

[0030] Constructing D first virtual scenes each having a scene position, and importing D standard units into the scene positions of the D first virtual scenes respectively, to construct D standard scenes;

[0031] E second virtual scenes with two independent unit frames are constructed, and E non-standard units are respectively imported into the two unit frames of the E second virtual scenes to construct E non-standard scenes.

[0032] Furthermore, the scenario solution model includes a first solution model and a second solution model, and the scenario solution includes a first scenario solution and a second scenario solution;

[0033] The combined approach to target decision making is:

[0034] The standard data and non-standard data are respectively aggregated and converted into feature vectors to obtain the first feature vector and the second feature vector;

[0035] Convert the scenario scheme corresponding to the standard data and the non-standard data into a label to obtain a first label and a second label;

[0036] The first eigenvector and the second eigenvector are respectively used as inputs of the reinforcement learning model, the first label and the second label corresponding to the first eigenvector and the second eigenvector are used as outputs of the reinforcement learning model, the scenario is used as the prediction target, and the sum of the prediction errors of all training data is minimized as the training target. The reinforcement learning model is trained to obtain a first scenario model and a second scenario model;

[0037] Input A standard data into the first scenario model under D standard scenarios respectively, and simulate D first scenario solutions;

[0038] Input B non-standard data into the second scenario model in each of E non-standard scenarios to simulate E second scenario solutions.

[0039] The large language model is used to identify the features of the D first scenario solutions and E second scenario solutions one by one. After removing duplicate features, all the remaining features are aggregated and combined into the target decision.

[0040] Furthermore, the risk feature extraction method is as follows:

[0041] Through natural language processing technology, the inherent data fields in the inherent multi-source heterogeneous data are identified one by one, and the data fields are imported into the supply process of the supply chain management system to determine the supply sequence number of the data fields in the supply chain;

[0042] Construct a supply map with blank serial number positions. According to the forward supply method of the supply chain, note the supply serial numbers one by one on the corresponding blank serial number positions, and import the data fields corresponding to the supply serial numbers into the blank serial number positions to generate the map positions.

[0043] Draw an indicator line between any two adjacent map positions, and draw an arrow on the indicator line pointing to the map position where the larger supply sequence number is located, to generate a search direction line;

[0044] The atlas position in the direction of the arrow of the search pointing line is recorded as the mother position, and the atlas position in the non-arrow direction of the search pointing line is recorded as the child position;

[0045] Identify the real-time data fields in the real-time multi-source heterogeneous data one by one, and match the real-time data fields one by one to the map position where the inherent data fields are located;

[0046] When the real-time data fields in the child and parent positions are consistent with the inherent data fields, there is no risk feature;

[0047] When the real-time data field in the child bit and the parent bit is inconsistent with the inherent data field, the real-time data field is recorded as a risk feature.

[0048] Furthermore, the supply risk criterion is: triggering a decision correction action when at least two risk pointing lines appear;

[0049] The method for determining whether to perform the decision correction operation is:

[0050] When the real-time data fields in the child and parent positions of the search pointing line are inconsistent with the inherent data fields, the search pointing line is recorded as a risk pointing line;

[0051] Count the number of risk-pointing lines in the supply graph. If the number of risk-pointing lines is less than 2, it is determined that no decision correction operation will be performed.

[0052] When the number of risk pointing lines is greater than or equal to 2, it is determined that a decision correction operation is performed.

[0053] Furthermore, the method for formulating the correction strategy is as follows:

[0054] Collect multiple sets of risk characteristics, target decisions and correction strategies in advance, summarize the multiple sets of risk characteristics and target decisions, and convert them into multiple sets of third feature vectors.

[0055] Convert the correction strategy into a label to obtain multiple sets of labels, and match the multiple sets of third feature vectors with the multiple sets of third labels in a one-to-one correspondence;

[0056] A third eigenvector corresponds to a third label, forming a set of training data. Multiple sets of training data constitute a training set, and the labeled training data are divided into a training set and a test set.

[0057] The third eigenvector is used as the input of the deep reinforcement learning model, and the third label corresponding to the third eigenvector is used as the output of the deep reinforcement learning model. The deep reinforcement learning model is trained using the training set and tested using the test set. An error threshold is preset. When the mean of the prediction errors of all training data in the test set is less than the preset error threshold, an intelligent correction model is obtained.

[0058] The collected risk characteristics and target decisions are converted into a third eigenvector and input into the intelligent correction model to formulate a real-time correction strategy.

[0059] A large-model-driven supply chain collaboration system, applied to a collaborative cloud server, is used to implement a large-model-driven supply chain collaboration method, including a data partitioning module, a key semantic parsing module, a target decision combination module, a decision correction and determination module, and a correction strategy formulation module, wherein each module is connected via a wired or wireless network;

[0060] The data partitioning module is used to collect the inherent multi-source heterogeneous data of the supply chain, extract the data encoding, data format and key fields of the inherent multi-source heterogeneous data, and divide the inherent multi-source heterogeneous data into standard data and non-standard data;

[0061] The key semantics parsing module is used to intelligently parse standard data and non-standard data based on the model parsing criteria by parsing the large model, and to parse key semantics from the standard data and non-standard data;

[0062] The target decision combination module is used to construct a decision scenario corresponding to the key semantics, simulate the scenario plan of the decision scenario through the scenario plan model, and combine the target decision from the scenario plan;

[0063] The decision correction and determination module is used to collect real-time multi-source heterogeneous data from the supply chain, extract risk characteristics from the real-time multi-source heterogeneous data, and determine whether to perform decision correction operations based on supply risk criteria;

[0064] The correction strategy formulation module is used to formulate correction strategies based on risk characteristics and target decisions through intelligent correction models, and to collaboratively correct the supply chain.

[0065] The technical effects and advantages of the large-scale model-driven supply chain collaboration method and system of the present invention are as follows:

[0066] 1. Through separate intelligent parsing of standard data and non-standard data, we can achieve overall unified parsing of standardized data and independent parsing of non-standard data. This can not only improve the parsing and recognition efficiency of different types of data, but also achieve the effect of "teaching students in accordance with their aptitude" for non-standard data, thus avoiding the inefficiency and lag of manual parsing and extraction of key semantics.

[0067] 2. By building a scenario solution model to simulate scenario solutions and combining scenario solutions into target decision-making methods, it is possible to perform targeted solution simulations for different supply chain needs under standard and non-standard scenarios, enabling the large model to accurately understand the semantics and meaning of different supply demands in the supply chain, solving the problems of "black box" and insufficient logical capabilities of traditional large models. At the same time, it also avoids the shortcomings of large models in dealing with complex supply chain problems due to their lack of logical reasoning and numerical calculation capabilities, providing accurate guidance and decision-making for collaborative processing of the supply chain.

[0068] 3. By constructing a knowledge graph to identify risk characteristics, it can achieve real-time monitoring and abnormal analysis and judgment of the collaborative operation process of the supply chain. It can also combine deep reinforcement learning technology to correct and improve target decisions as soon as supply chain anomalies are discovered, so that the correction strategy can make targeted adjustments to abnormal phenomena in the supply chain to ensure the stability and smoothness of resource scheduling in the supply chain, break the information island phenomenon of enterprises in the supply chain, effectively reduce the potential risks and collaborative costs of enterprises in the supply chain, and improve the collaborative efficiency of the supply chain. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 A schematic diagram of a flow chart of a large model-driven supply chain collaboration method provided in the first embodiment of the present invention;

[0070] Figure 2 This is a schematic diagram of the architecture of the collaborative cloud server and data nodes provided in Example 1 of the present invention;

[0071] Figure 3 A module diagram of a large model-driven supply chain collaboration system provided in Example 2 of the present invention. DETAILED DESCRIPTION

[0072] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0073] Example 1: Please refer to Figure 1 and Figure 2 As shown, the large model-driven supply chain collaboration method described in this embodiment is applied to a collaborative cloud server and includes:

[0074] S01: Collect the inherent multi-source heterogeneous data of the supply chain, extract the structural features of the inherent multi-source heterogeneous data, and divide the inherent multi-source heterogeneous data into standard data and non-standard data;

[0075] The supply chain is an overall functional network structure with the enterprise as the core, starting from supporting parts, to manufacturing intermediate products and final products, and then the sales network delivers the products to consumers. Due to the large number of user types and data types involved in the supply chain, the data contained in the supply chain is diverse, which is called multi-source heterogeneous data.

[0076] Inherent multi-source heterogeneous data refers to the various data originally stored and recorded in the supply chain, which enables multi-source heterogeneous data to be used as multi-dimensional data reflecting the supply chain market performance, warehouse storage conditions, logistics and transportation conditions, enterprise production conditions, etc.

[0077] Specifically, multi-source heterogeneous data includes but is not limited to enterprise resource planning data, logistics trajectory data, market sentiment data, supplier data, inventory data, financial data, production data, etc.; it can comprehensively represent the multi-dimensional data in the supply chain.

[0078] Since multi-source heterogeneous data corresponds to different dimensions in the supply chain, the collection and acquisition channels of multi-source heterogeneous data are not unique. Data from different sources are obtained through different channels, and each channel has data of only one dimension. Specifically, when collecting multi-source heterogeneous data, different data storage systems in the supply chain are recorded as a data node separately. After obtaining the corresponding dimensional data from M distributed data nodes one by one, and summarizing the dimensional data obtained from all data nodes, the multi-source heterogeneous data of the supply chain can be obtained.

[0079] The connection between the collaborative cloud server and the data node is as follows Figure 2 As shown, at this time, the M distributed data nodes are connected to the collaborative cloud server respectively, and the collection of dimensional data in each data node is realized.

[0080] After collecting multi-source heterogeneous data from the supply chain, it is necessary to classify the large amount of multi-source heterogeneous data according to its structure and generate standard data and non-standard data;

[0081] Standard data refers to multi-source heterogeneous data that complies with the industry standards and common rules for data transmission, analysis, and storage in traditional supply chains, allowing standard data to be directly used in the supply chain;

[0082] Non-standard data refers to multi-source heterogeneous data that does not conform to the industry norms and general rules of traditional supply chains for data transmission, analysis and storage, making non-standard data unable to be directly used in the supply chain.

[0083] When dividing multi-source heterogeneous data into standard data and non-standard data, the structural characteristics of each multi-source heterogeneous data should be used as the basis for division. Among them, the structural characteristics are used to accurately represent the specific structure and content meaning of multi-source heterogeneous data in the supply chain;

[0084] Structural features include data encoding, data format, and key fields. Data encoding is used to represent the actual encoding content of multi-source heterogeneous data in the supply chain, data format is used to represent the actual storage format of multi-source heterogeneous data in the supply chain, and key fields are used to represent the meaning of mandatory fields of multi-source heterogeneous data in the supply chain.

[0085] Specifically, when extracting the structural features of multi-source heterogeneous data, the encoding box and format box of the multi-source heterogeneous data are queried through the data management system, the characters in the encoding box are recorded as data encoding, and the characters in the format box are recorded as data format. The text content of the multi-source heterogeneous data is identified one by one through OCR technology, the required items are marked from the text content, and the content in the required items is recorded as key fields.

[0086] It should be noted that the code box and format box are text boxes used to store and represent the system code and system format of multi-source heterogeneous data, and the text characters in the text boxes are used as the specific storage content of the system code and system format; the required items are used to represent the names of the indispensable supply processes in the supply chain, including but not limited to supplier names, transportation routes, company names, etc.

[0087] After obtaining the structural features of multi-source heterogeneous data, each multi-source heterogeneous data can be divided based on the structural features, thereby distinguishing between standard data and non-standard data;

[0088] The division method between standard data and non-standard data is:

[0089] Compare the data encoding of the multi-source heterogeneous data with the standard encoding one by one for consistency, and record the multi-source heterogeneous data whose data encoding is consistent with the standard encoding as the first verification data;

[0090] Comparing the data format of the first verification data with the standard format for consistency, and recording the first verification data whose data format is consistent with the standard format as the second verification data;

[0091] The key fields of the second verification data are compared with the standard fields for consistency, the second verification data whose key fields are consistent with the standard fields are recorded as standard data, and all remaining multi-source heterogeneous data are recorded as non-standard data, obtaining A standard data and B non-standard data respectively.

[0092] It should be noted that standard coding, standard format and standard fields are standards for the structural features corresponding to multi-source heterogeneous data identified as standard data. Only when the data coding, data format and key fields of multi-source heterogeneous data are consistent with the standard coding, standard format and standard fields respectively, can the multi-source heterogeneous data be recorded as standard data; among them, when setting the standard coding, standard format and standard fields, they are often set according to the corresponding industry general specifications and industry codes of conduct in the supply chain.

[0093] S02: Based on the model parsing criteria, intelligently parse standard data and non-standard data, and extract key semantics from them;

[0094] After dividing standard data and non-standard data, further intelligent analysis of the standard data and non-standard data is required to accurately and intuitively identify relevant content that meets the needs of subsequent supply chain collaborative decision-making, namely: key semantics, from the standard data and non-standard data to ensure that the key semantics can provide a data basis for subsequent supply chain collaborative decision-making.

[0095] Since multi-source heterogeneous data includes standard data and non-standard data, the key semantics corresponding to standard data and non-standard data are different. Therefore, the key semantics include standard semantics and non-standard semantics, and standard semantics and non-standard semantics can correspond to the key semantics of standard data and non-standard data respectively.

[0096] When intelligently parsing standard and non-standard data and extracting both standard and non-standard semantics, this needs to be done within the constraints of model parsing criteria to ensure that all different types of multi-source heterogeneous data can be intelligently parsed by the corresponding large model.

[0097] The model analysis criteria are: standard data adopts an overall unified analysis method, and non-standard data adopts a single independent analysis method; this ensures that standard data can be analyzed as a whole and quickly, and non-standard data can be analyzed independently and specifically in accordance with the "teaching students in accordance with their aptitude" approach.

[0098] When extracting key semantics, the means relied upon and adopted are artificial intelligence big model technology, which can quickly and intelligently parse and extract data of different types and meanings in the supply chain, avoiding the inefficiency and lag of manual parsing and extraction of key semantics.

[0099] The parsing methods for standard semantics and non-standard semantics are:

[0100] Query the mode switch box for parsing the large model, adjust the characters in the mode switch box to "ZT1", and clear the remaining parsing text in the parsing box of the large model; the mode switch box is used to switch and control the parsing state of the large model, and achieve the parsing effect of standard data and non-standard data by adjusting and switching characters;

[0101] Aggregate and combine A pieces of standard data into a first parsing text, and import the first parsing text into the parsing frame of the parsing model;

[0102] At the same time, the large parsing model performs an overall intelligent parsing of the first parsed text in the parsing box to parse out the standard semantics;

[0103] After the parsing of A standard data is completed, adjust the character of the mode switch box to "DL2" and clear the first parsing text in the parsing box of the parsing large model;

[0104] Convert B non-standard data into B second parsing texts respectively, and import the B second parsing texts into the parsing frame of the parsing model in sequence;

[0105] According to the order of non-standard data collection time, the large parsing model performs overall intelligent parsing on the B second parsed texts in the parsing box one by one, and parses out B second parsed semantics;

[0106] The repeated second parsed semantics are recorded as repeated semantics, and redundant repeated semantics are eliminated until the number of repeated semantics is 1. All the remaining repeated semantics and second parsed semantics are aggregated to generate non-standard semantics.

[0107] The big parsing model is based on deep learning technology and is obtained after intelligent parsing training with a large number of different types of standard data and non-standard data. The big parsing model enables the big parsing model to quickly and accurately identify and extract key semantics in standard data and non-standard data, thereby achieving independent parsing effects for standard data and non-standard data, avoiding the inefficiency of overall hybrid parsing.

[0108] It should be noted that the number of extracted standard semantics and non-standard semantics is usually greater than 1, which enables standard semantics and non-standard semantics to represent multi-dimensional data information in the supply chain in a diversified way, thereby avoiding the limitations of single-dimensional data information, and laying a necessary and comprehensive foundation for subsequent collaborative decision-making in the supply chain.

[0109] S03: Construct a decision scenario corresponding to the key semantics, simulate the scenario plan of the decision scenario through the scenario plan model, and combine the target decision from the scenario plan;

[0110] A decision scenario is a virtual scenario with specific supply requirements that is constructed based on different key semantics in the supply chain, and serves as a scenario corresponding to different supply demand solutions in the supply chain.

[0111] Since the supply chain's supply demands during collaborative processing are not unique, the number and characteristics of supply demands are affected by the number and meaning of key semantics. Therefore, decision scenarios need to be constructed based on the different specific meanings of standard semantics and non-standard semantics.

[0112] Decision scenarios include standard scenarios and non-standard scenarios; among them, standard scenarios are virtual scenarios used to represent supply and demand corresponding to standard semantics, and non-standard scenarios are virtual scenarios used to represent supply and demand corresponding to non-standard semantics.

[0113] Specifically, the construction methods of standard scenarios and non-standard scenarios are as follows:

[0114] The large language model is used to identify key fields of standard semantics and non-standard semantics respectively, and the identified key fields are summarized to generate a standard field set and a non-standard field set;

[0115] Taking two key fields as a unit, any two key fields in the standard field set and the non-standard field set are combined to generate D standard units and E non-standard units;

[0116] Construct D first virtual scenes each having a scene position, and import D standard units into the scene positions of the D first virtual scenes respectively to construct D standard scenes; the scene position is used to provide a location limit for importing standard semantics to ensure the accuracy of constructing the standard scenes;

[0117] Construct E second virtual scenes with two independent cell frames, and import E non-standard cells into the two cell frames of the E second virtual scenes, respectively, to construct E non-standard scenes. The cell frames are used to provide location constraints for importing non-standard semantics, ensuring the accuracy of constructing non-standard scenes.

[0118] A large language model is a deep learning model trained using large amounts of text data. This allows the model to generate natural language text or understand the meaning of language text, thereby accurately identifying standard and non-standard semantics.

[0119] After constructing standard and non-standard scenarios, you can use these scenarios to simulate solutions for standard and non-standard supply demands in the supply chain. When simulating solutions for different decision-making scenarios, you need to combine the scenario solution model to achieve fast and accurate simulation results for the corresponding solutions for standard and non-standard scenarios.

[0120] The scenario solution model is based on standard scenarios and non-standard scenarios, with standard data and non-standard data as input. It uses reinforcement learning technology to simulate a model of solutions corresponding to the input standard data and non-standard data, so that the scenario solution model can simulate scenario solutions that are compatible with standard scenarios and non-standard scenarios, ensuring that the scenario solution model can perform reinforcement learning processing on the corresponding supply demand in the input standard data and non-standard data.

[0121] Scenario solutions are used to address different supply chain requirements in standard and non-standard scenarios. These solutions can only represent a single or specific supply demand, resulting in certain limitations for each scenario. To comprehensively represent the supply chain and provide a direct basis for subsequent supply chain collaboration, scenario solutions need to be combined to generate target decisions.

[0122] Specifically, the scenario solution model includes a first solution model and a second solution model, wherein the first solution model is used to simulate the solution for the standard scenario, and the second solution model is used to simulate the solution for the non-standard scenario;

[0123] The scenario scheme includes a first scenario scheme and a second scenario scheme; wherein the first scenario scheme and the second scenario scheme are used to represent the outputs of the first scheme model and the second scheme model respectively.

[0124] Target decision-making is a supply chain collaborative processing solution based on multiple scenarios, which can comprehensively and holistically represent standard and non-standard data in the supply chain and serve as a guide for the specific implementation steps of supply chain collaborative processing;

[0125] The combined approach to target decision making is:

[0126] The standard data and non-standard data are respectively aggregated and converted into feature vectors to obtain the first feature vector and the second feature vector;

[0127] Convert the scenario scheme corresponding to the standard data and the non-standard data into a label to obtain a first label and a second label;

[0128] The first eigenvector and the second eigenvector are respectively used as inputs of the reinforcement learning model, the first label and the second label corresponding to the first eigenvector and the second eigenvector are used as outputs of the reinforcement learning model, the scenario is used as the prediction target, and the sum of the prediction errors of all training data is minimized as the training target. The reinforcement learning model is trained to obtain a first scenario model and a second scenario model;

[0129] Input A standard data into the first scenario model under D standard scenarios respectively, and simulate D first scenario solutions;

[0130] Input B non-standard data into the second scenario model in each of E non-standard scenarios to simulate E second scenario solutions.

[0131] The large language model identifies the features of the D first- and E second-scenario solutions one by one, removes duplicate features, and aggregates all remaining features to form the target decision. These features concisely represent the true meaning of the first and second-scenario solutions, providing direct and clear content support for the subsequent target decision combination.

[0132] It should be noted that the target decision is a decision that can accurately and comprehensively represent the key real meanings in the first scenario and the second scenario, that is, it can comprehensively represent the different supply demands of standard data and non-standard data in the supply chain, thereby providing a basis for the overall collaborative operation of the supply chain.

[0133] S04: Collect real-time multi-source heterogeneous data from the supply chain, extract risk features from the real-time multi-source heterogeneous data, and determine whether to perform decision correction operations based on supply risk criteria;

[0134] Real-time multi-source heterogeneous data is used to analyze the multi-dimensional collaborative change process and change trends during the specific execution of target decisions. It can also represent the supply chain's current market performance, warehouse storage conditions, logistics and transportation conditions, and enterprise production conditions.

[0135] Real-time multi-source heterogeneous data is consistent with the specific data corresponding to the inherent multi-source heterogeneous data, that is, real-time multi-source heterogeneous data includes but is not limited to real-time enterprise resource planning data, logistics trajectory data, market sentiment data, supplier data, inventory data, financial data, production data, etc.

[0136] After acquiring real-time multi-source heterogeneous data, it is necessary to analyze the real-time multi-source heterogeneous data and extract risk features from the real-time multi-source heterogeneous data. The risk features can represent the difference between the real-time multi-source heterogeneous data and the inherent multi-source heterogeneous data, and serve as the basis for subsequent judgment on whether real-time correction and collaborative operation of the target decision are needed.

[0137] The risk feature extraction method is:

[0138] Using natural language processing technology, we identify the inherent data fields in the inherent multi-source heterogeneous data one by one, and then import the data fields into the supply process of the supply chain management system to determine the supply sequence number of the data fields in the supply chain. The supply sequence number refers to the number of the supply process from upstream to downstream in the supply chain. Since the numbers of each process are different, all supply processes can be effectively distinguished.

[0139] Construct a supply graph with blank sequence number positions. According to the forward supply chain method, the supply sequence numbers are annotated one by one in the corresponding blank sequence number positions, and the data fields corresponding to the supply sequence numbers are imported into the blank sequence number positions to generate graph positions. The supply graph refers to a knowledge graph without substantial content. The sequence number position is the smallest unit that constitutes the supply graph and can provide a location for importing data fields.

[0140] Draw an indicator line between any two adjacent map positions, and draw an arrow on the indicator line pointing to the map position where the larger supply sequence number is located, to generate a search direction line;

[0141] The map bit located in the direction of the arrow of the search pointing line is recorded as the parent bit, and the map bit located in the non-arrow direction of the search pointing line is recorded as the child bit. By distinguishing between the child bit and the parent bit, the map bits at both ends of the search pointing line can be accurately distinguished, preventing the map bits at both ends of the same search pointing line from being confused and crossed during identification, and providing a basis for subsequent analysis and judgment of whether there are abnormal situations in the supply chain.

[0142] Identify the real-time data fields in the real-time multi-source heterogeneous data one by one, and match the real-time data fields one by one to the map position where the inherent data fields are located;

[0143] Compare the real-time data fields in the child and parent positions of all search-directed lines with the inherent data fields one by one;

[0144] When the real-time data fields in the child and parent positions are consistent with the inherent data fields, there is no abnormality in the real-time multi-source heterogeneous data corresponding to the search pointing line. At this time, no unexpected events have occurred in the supply chain process, and there is no risk feature.

[0145] When the real-time data fields in the child position and the parent position are inconsistent with the inherent data fields, an abnormality exists in the real-time multi-source heterogeneous data corresponding to the search pointing line. At this time, an unexpected event occurs in the supply process in the supply chain, and a risk feature exists. The real-time data field is recorded as a risk feature.

[0146] After extracting the risk characteristics, we can use them as a basis to determine whether an unexpected event has occurred in the supply chain and decide whether to make corrections to the target decision. This allows us to monitor and respond to real-time changes in the supply chain and discover unexpected risk events in the supply chain as soon as possible.

[0147] When determining whether to execute a decision correction operation, the supply risk criteria should be used as the basis to ensure the accuracy and rationality of the subsequent judgment on the execution of decision correction operations. This will prevent the occurrence of extremely rare phenomena in the supply chain that do not affect the target decision and lead to the false triggering of the execution of decision correction operations, thereby raising the threshold for executing decision correction operations.

[0148] The supply risk criterion is to trigger a decision correction operation when at least two risk-pointing lines appear; this can avoid false triggering caused by inaccurate individual data in the supply chain.

[0149] The method for determining whether to perform the decision correction operation is:

[0150] When the real-time data fields in the child and parent positions of the search direction line are inconsistent with the inherent data fields, the search direction line is recorded as a risk direction line;

[0151] Count the number of risk-pointing lines in the supply map. When the number of risk-pointing lines is less than 2, it means that there are no phenomena in the supply chain that are sufficient to have a negative impact on the target decision. The supply chain process can be executed as planned, and the decision correction operation is not performed.

[0152] When the number of risk pointing lines is greater than or equal to 2, a phenomenon has occurred in the supply chain that is sufficient to have a negative impact on the target decision, and the supply process of the supply chain cannot be executed as originally planned, then it is determined to execute a decision correction operation.

[0153] S05: If a decision correction operation is executed, a correction strategy is developed through the intelligent correction model based on risk characteristics and target decisions, and the supply chain is collaboratively corrected;

[0154] If it is decided to perform a decision correction operation, it is necessary to base it on risk characteristics, take the target decision as the correction object, and use a pre-established intelligent correction model to correct the target decision. In this way, corresponding correction measures can be provided for unexpected events and supply risk phenomena that occur in the supply chain, and the correction measures can be recorded as correction strategies.

[0155] The correction strategy is based on risk characteristics and target decisions, and is combined with the intelligent correction model to intelligently formulate decisions to correct and optimize the target decisions. It also serves as a measure to correct and adjust the real-time supply situation of the supply chain, thereby effectively responding to unexpected risk events in the supply chain and avoiding the negative impact of drastic price fluctuations, transportation delays, supplier loss of contact, and other phenomena on the stability and reliability of the supply chain.

[0156] The intelligent correction model combines risk characteristics and target decisions, and is trained using deep reinforcement learning technology to construct a model for correcting and optimizing the target decisions of the supply chain. The intelligent correction model can use risk characteristics and target decisions as input and correction strategies as output, thereby achieving real-time correction and adjustment effects on the supply chain.

[0157] The revision strategy is formulated as follows:

[0158] Collect multiple sets of risk characteristics, target decisions and correction strategies in advance, summarize the multiple sets of risk characteristics and target decisions, and convert them into multiple sets of third feature vectors.

[0159] Convert the correction strategy into a label to obtain multiple sets of labels, and match the multiple sets of third feature vectors with the multiple sets of third labels in a one-to-one correspondence;

[0160] A third eigenvector corresponds to a third label, forming a set of training data. Multiple sets of training data constitute a training set. The labeled training data are divided into a training set and a test set. 70% of the training data is used as the training set, and 30% of the training data is used as the test set.

[0161] The third eigenvector is used as the input of the deep reinforcement learning model, and the third label corresponding to the third eigenvector is used as the output of the deep reinforcement learning model. The deep reinforcement learning model is trained using the training set and tested using the test set. An error threshold is preset. When the mean of the prediction errors of all training data in the test set is less than the preset error threshold, an intelligent correction model is obtained.

[0162] The collected risk characteristics and target decisions are converted into a third eigenvector and input into the intelligent correction model to formulate a real-time correction strategy.

[0163] After formulating the correction strategy, it is necessary to carry out corresponding collaborative correction processing on the supply chain process according to the content of the correction strategy;

[0164] Specifically, when collaboratively correcting the supply chain, the policy semantics in the correction strategy are first identified through the large language model, and the supply process consistent with the policy semantics is queried in the supply chain management system. Then, the correction process that matches the supply process is marked from the correction strategy, and the supply process in the target decision is eliminated to form a process blank. Finally, the correction processes are imported one by one into the corresponding process blanks in the target decision to achieve the collaborative correction effect of the supply chain's target decision.

[0165] For example, the content of the correction strategy includes but is not limited to automatic scheduling of supply chain resources, automatic optimization of logistics transportation routes, automatic adjustment of inventory turnover rate, etc., which can ensure that the company's supply chain can maintain real-time monitoring, analysis and correction and adjustment operations, and ensure that all supply processes in the supply chain can maintain dynamic coordination effects, thereby improving the controllability of supply chain management and scheduling.

[0166] In this embodiment, by performing separate intelligent parsing processing on standard data and non-standard data, the overall unified parsing of standardized data and the independent parsing effect of non-standard data can be achieved, which can not only improve the parsing and recognition efficiency of different types of data, but also achieve the effect of "teaching students in accordance with their aptitude" for non-standard data, thereby avoiding the inefficiency and lag of manual parsing and extraction of key semantics.

[0167] By building a scenario solution model to simulate scenario solutions and combining scenario solutions into target decision-making methods, it is possible to perform targeted solution simulations for different supply chain needs under standard scenarios and non-standard scenarios, enabling the big model to accurately understand the semantics and meaning of different supply demands in the supply chain, solving the problems of "black box" and insufficient logical capabilities of traditional big models. At the same time, it also avoids the shortcomings of big models in lacking logical reasoning and numerical calculation capabilities when dealing with complex supply chain problems, and provides accurate guidance and decision-making for collaborative processing of the supply chain.

[0168] By constructing a knowledge graph to identify risk characteristics, it is possible to monitor the collaborative operation process of the supply chain in real time and analyze and judge anomalies. In addition, when supply chain anomalies are discovered, deep reinforcement learning technology can be combined to correct and improve target decisions, so that the correction strategy can make targeted adjustments to anomalies in the supply chain to ensure the stability and smoothness of resource scheduling in the supply chain, break the information island phenomenon of enterprises in the supply chain, effectively reduce the potential risks and collaboration costs of enterprises in the supply chain, and improve the collaborative efficiency of the supply chain.

[0169] Example 2: Please refer to Figure 3 As shown, for parts not described in detail in this embodiment, please refer to the description of embodiment 1. A large model-driven supply chain collaboration system is provided, which is applied to a collaborative cloud server and is used to implement a large model-driven supply chain collaboration method, including a data partitioning module, a key semantic parsing module, a target decision combination module, a decision correction determination module and a correction strategy formulation module, wherein each module is connected via a wired or wireless network;

[0170] The data partitioning module is used to collect the inherent multi-source heterogeneous data of the supply chain, extract the data encoding, data format and key fields of the inherent multi-source heterogeneous data, and divide the inherent multi-source heterogeneous data into standard data and non-standard data;

[0171] The key semantics parsing module is used to intelligently parse standard data and non-standard data based on the model parsing criteria by parsing the large model, and to parse key semantics from the standard data and non-standard data;

[0172] The target decision combination module is used to construct a decision scenario corresponding to the key semantics, simulate the scenario plan of the decision scenario through the scenario plan model, and combine the target decision from the scenario plan;

[0173] The decision correction and determination module is used to collect real-time multi-source heterogeneous data from the supply chain, extract risk characteristics from the real-time multi-source heterogeneous data, and determine whether to perform decision correction operations based on supply risk criteria;

[0174] The correction strategy formulation module is used to formulate correction strategies based on risk characteristics and target decisions through intelligent correction models, and to collaboratively correct the supply chain.

[0175] In a large-scale model-driven supply chain collaboration system, multi-source heterogeneous data includes enterprise resource planning data, logistics trajectory data, market sentiment data, supplier data, inventory data, financial data, and production data;

[0176] The model analysis criteria are: standard data adopts an overall unified analysis method, and non-standard data adopts a single independent analysis method;

[0177] The supply risk criterion is: triggering decision correction actions when at least two risk pointing lines appear.

[0178] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A large model-driven supply chain collaboration method, applied to a collaborative cloud server, characterized by: include: S01: Collect the inherent multi-source heterogeneous data of the supply chain, extract the data encoding, data format and key fields of the inherent multi-source heterogeneous data, and divide the inherent multi-source heterogeneous data into standard data and non-standard data; S02: Based on the model parsing criteria, intelligently parse the standard data and non-standard data by parsing the large model, and parse the key semantics from the standard data and non-standard data. The key semantics include standard semantics and non-standard semantics. The model analysis criteria are: standard data adopts an overall unified analysis method, and non-standard data adopts a single independent analysis method; S03: Construct a decision scenario corresponding to the key semantics, simulate the scenario plan of the decision scenario through the scenario plan model, and combine the target decision from the scenario plan; Decision scenarios include standard scenarios and non-standard scenarios. The construction methods of standard scenarios and non-standard scenarios are as follows: The large language model is used to identify key fields of standard semantics and non-standard semantics respectively, and the identified key fields are summarized to generate a standard field set and a non-standard field set; Taking two key fields as a unit, any two key fields in the standard field set and the non-standard field set are combined to generate D standard units and E non-standard units; Constructing D first virtual scenes each having a scene position, and importing D standard units into the scene positions of the D first virtual scenes respectively, to construct D standard scenes; Constructing E second virtual scenes with two independent unit frames, and importing E non-standard units into the two unit frames of the E second virtual scenes respectively, to construct E non-standard scenes; The scene position is used to define the position of the import for standard semantics, and the cell frame is used to define the position of the import for non-standard semantics. The scenario solution model includes the first solution model and the second solution model, and the scenario solution includes the first scenario solution and the second scenario solution: The combined approach to target decision making is: The standard data and non-standard data are respectively aggregated and converted into feature vectors to obtain the first feature vector and the second feature vector; Convert the scenario scheme corresponding to the standard data and the non-standard data into a label to obtain a first label and a second label; The first eigenvector and the second eigenvector are respectively used as inputs of the reinforcement learning model, the first label and the second label corresponding to the first eigenvector and the second eigenvector are used as outputs of the reinforcement learning model, the scenario is used as the prediction target, and the sum of the prediction errors of all training data is minimized as the training target. The reinforcement learning model is trained to obtain a first scenario model and a second scenario model; Input A standard data into the first scenario model under D standard scenarios respectively, and simulate D first scenario solutions; Input B non-standard data into the second scenario model in each of E non-standard scenarios to simulate E second scenario solutions. The large language model is used to identify the features of the D first scenario solutions and the E second scenario solutions one by one. After removing duplicate solution features, all remaining solution features are aggregated and combined into the target decision. S04: Collect real-time multi-source heterogeneous data from the supply chain, extract risk features from the real-time multi-source heterogeneous data, and determine whether to perform decision correction operations based on supply risk criteria; S05: If a decision correction operation is performed, a correction strategy is formulated through an intelligent correction model based on risk characteristics and target decisions, and the supply chain is collaboratively corrected.

2. A large model-driven supply chain collaboration method according to claim 1, characterized in that: When collecting multi-source heterogeneous data, the data storage system is recorded as a data node, and dimensional data is collected from M distributed data nodes one by one. After the collected dimensional data is aggregated, the inherent multi-source heterogeneous data of the supply chain is obtained; The data encoding, data format and key field extraction methods are as follows: The encoding frame and format frame of multi-source heterogeneous data are queried through the data management system, and the characters in the encoding frame are recorded as data encoding, and the characters in the format frame are recorded as data format; The text content of multi-source heterogeneous data is identified one by one through OCR technology, the required items are marked from the text content, and the content in the required items is recorded as key fields.

3. A large model driven supply chain collaboration method according to claim 2, characterized in that: The division method between standard data and non-standard data is: Compare the data encoding of the multi-source heterogeneous data with the standard encoding for consistency, and record the multi-source heterogeneous data whose data encoding is consistent with the standard encoding as the first verification data; Comparing the data format of the first verification data with the standard format for consistency, and recording the first verification data whose data format is consistent with the standard format as the second verification data; The key fields of the second verification data are compared with the standard fields for consistency, the second verification data whose key fields are consistent with the standard fields are recorded as standard data, and all remaining multi-source heterogeneous data are recorded as non-standard data, obtaining A standard data and B non-standard data respectively.

4. A large model driven supply chain collaboration method according to claim 3, characterized in that: The parsing methods for standard semantics and non-standard semantics are: Find the mode switch box for parsing the large model, adjust the character of the mode switch box to "ZT1", and clear the remaining parsing text in the parsing box of the large model; Aggregate and combine A standard data into a first parsed text, and import the first parsed text into the parsing frame of the parsing model. At the same time, the parsing model performs overall intelligent parsing on the first parsed text in the parsing frame to parse out the standard semantics. After the parsing of A standard data is completed, the character of the mode switch box is adjusted to "DL2", and the first parsed text in the parsing box of the parsing large model is cleared; Convert B non-standard data into B second parsed texts respectively, and import the B second parsed texts into the parsing box of the parsing model in sequence. According to the order of non-standard data collection time, the parsing model performs overall intelligent parsing on the B second parsed texts in the parsing box one by one to parse out B second parsed semantics; The repeated second parsed semantics are recorded as repeated semantics, and redundant repeated semantics are eliminated until the number of repeated semantics is 1. The remaining repeated semantics and the second parsed semantics are aggregated to generate non-standard semantics.

5. A large model driven supply chain collaboration method according to claim 4, characterized in that: The risk feature extraction method is: Through natural language processing technology, the inherent data fields in the inherent multi-source heterogeneous data are identified one by one, and the data fields are imported into the supply process of the supply chain management system to determine the supply sequence number of the data fields in the supply chain; Construct a supply map with blank serial number positions. According to the forward supply method of the supply chain, note the supply serial numbers one by one on the corresponding blank serial number positions, and import the data fields corresponding to the supply serial numbers into the blank serial number positions to generate the map positions. Draw an indicator line between any two adjacent map positions, and draw an arrow on the indicator line pointing to the map position where the larger supply sequence number is located, to generate a search direction line; The atlas position in the direction of the arrow of the search pointing line is recorded as the mother position, and the atlas position in the non-arrow direction of the search pointing line is recorded as the child position; Identify the real-time data fields in the real-time multi-source heterogeneous data one by one, and match the real-time data fields one by one to the map position where the inherent data fields are located; When the real-time data fields in the child and parent positions are consistent with the inherent data fields, there is no risk feature; When the real-time data field in the child bit and the parent bit is inconsistent with the inherent data field, the real-time data field is recorded as a risk feature.

6. A large model driven supply chain collaboration method according to claim 5, characterized in that: The supply risk criterion is: triggering a decision correction action when at least two risk pointing lines appear; The method for determining whether to perform the decision correction operation is: When the real-time data fields in the child and parent positions of the search pointing line are inconsistent with the inherent data fields, the search pointing line is recorded as a risk pointing line; Count the number of risk-pointing lines in the supply graph. If the number of risk-pointing lines is less than 2, it is determined that no decision correction operation will be performed. When the number of risk pointing lines is greater than or equal to 2, it is determined that a decision correction operation is performed.

7. A large model driven supply chain collaboration method according to claim 6, characterized in that: The revision strategy is formulated as follows: Collect multiple sets of risk characteristics, target decisions and correction strategies in advance, summarize the multiple sets of risk characteristics and target decisions, and convert them into multiple sets of third feature vectors. Convert the correction strategy into a label to obtain multiple sets of labels, and match the multiple sets of third feature vectors with the multiple sets of third labels in a one-to-one correspondence; A third eigenvector corresponds to a third label, forming a set of training data. Multiple sets of training data constitute a training set, and the labeled training data are divided into a training set and a test set. The third eigenvector is used as the input of the deep reinforcement learning model, and the third label corresponding to the third eigenvector is used as the output of the deep reinforcement learning model. The deep reinforcement learning model is trained using the training set and tested using the test set. An error threshold is preset. When the mean of the prediction errors of all training data in the test set is less than the preset error threshold, an intelligent correction model is obtained. The collected risk characteristics and target decisions are converted into a third eigenvector and input into the intelligent correction model to formulate a real-time correction strategy.

8. A large model driven supply chain collaboration system, applied to a collaborative cloud server, for implementing a large model driven supply chain collaboration method according to any one of claims 1 to 7, characterized in that: It includes a data partitioning module, a key semantic parsing module, a target decision combination module, a decision correction determination module and a correction strategy formulation module, wherein each module is connected via a wired or wireless network; The data partitioning module is used to collect the inherent multi-source heterogeneous data of the supply chain, extract the data encoding, data format and key fields of the inherent multi-source heterogeneous data, and divide the inherent multi-source heterogeneous data into standard data and non-standard data; The key semantics parsing module is used to intelligently parse standard data and non-standard data based on the model parsing criteria by parsing the large model, and to parse key semantics from the standard data and non-standard data; The target decision combination module is used to construct a decision scenario corresponding to the key semantics, simulate the scenario plan of the decision scenario through the scenario plan model, and combine the target decision from the scenario plan; The decision correction and determination module is used to collect real-time multi-source heterogeneous data from the supply chain, extract risk characteristics from the real-time multi-source heterogeneous data, and determine whether to perform decision correction operations based on supply risk criteria; The correction strategy formulation module is used to formulate correction strategies based on risk characteristics and target decisions through intelligent correction models, and to collaboratively correct the supply chain.

Images