CN119599190A

CN119599190A - Knowledge graph and industrial model-based supply chain prediction method

Info

Publication number: CN119599190A
Application number: CN202411665681.3A
Authority: CN
Inventors: 王少华; 张廷宇; 齐光鹏; 武红强
Original assignee: Inspur Yunzhou Industrial Internet Co Ltd
Current assignee: Inspur Yunzhou Industrial Internet Co Ltd
Priority date: 2024-11-20
Filing date: 2024-11-20
Publication date: 2025-03-11

Abstract

The invention provides a supply chain prediction method based on a knowledge graph and an industrial model, which comprises the steps of collecting supply chain prediction required characteristics, wherein the required characteristics comprise historical order data, production plan data, industrial purchasing data, collection information, logistics data and storage data, carrying out data cleaning, standardization and correlation analysis on the historical order data, the production plan data, the industrial purchasing data, the collection information, the logistics data and the storage data to obtain target characteristics, identifying key entities and relations in the target characteristics, constructing the knowledge graph based on the key entities and the relations, identifying key supply chain nodes and node relations in the knowledge graph by utilizing the industrial model, converting the key supply chain key nodes and node relations into numerical feature vectors for supply chain prediction, and carrying out supply chain prediction.

Description

Knowledge graph and industrial model-based supply chain prediction method

Technical Field

The invention relates to the technical field of supply chain prediction, in particular to a supply chain prediction method based on a knowledge graph and an industrial model.

Background

In the field of supply chain management, accurate supply chain prediction is of great importance. Conventional supply chain prediction methods typically rely on historical data and simple statistical models. However, these methods have a number of limitations and drawbacks. For example, inflexible feature selection, inadequate model optimization, and poor results in coping with supply chain complexity and uncertainty. In recent years, prediction methods based on knowledge maps and industrial large models have been attracting attention. The knowledge graph is used as a structural representation method, and the defects of the traditional method in capturing the complex relation of the supply chain are overcome by integrating and correlating the multi-production data.

However, most prediction models based on knowledge patterns have insufficient adaptability in dynamic environments, and cannot effectively cope with real-time changes of a supply chain, so that the prediction accuracy of the supply chain is poor.

Disclosure of Invention

The invention provides a supply chain prediction method based on a knowledge graph and an industrial model, which is used for solving the defect of poor supply chain prediction accuracy in the prior art.

In a first aspect, the present invention provides a supply chain prediction method based on a knowledge graph and an industrial model, comprising:

Collecting supply chain forecast required characteristics, wherein the required characteristics comprise historical order data, production plan data, industrial purchase data, collection information, logistics data and warehouse data;

Performing data cleaning, standardization and correlation analysis on the historical order data, the production plan data, the industrial purchasing data, the collection information, the logistics data and the storage data to obtain target characteristics;

Identifying key entities and relations in the target features, and constructing a knowledge graph based on the key entities and the relations;

And identifying key nodes and node relations of the supply chain in the knowledge graph by utilizing an industrial model, and converting the key nodes and node relations into numerical characteristic vectors for supply chain prediction so as to carry out supply chain prediction.

The supply chain prediction method based on the knowledge graph and the industrial model provided by the invention further comprises the following steps:

acquiring actual data and forecast data of the supply chain in real time, and comparing the actual data with the forecast data;

performing deviation analysis and generating optimization suggestions based on the comparison result;

and dynamically adjusting and optimizing the industrial model through a real-time feedback loop based on the optimization suggestion.

According to the method for predicting the supply chain based on the knowledge graph and the industrial model provided by the invention, the method for collecting the historical order data and the production plan data in the characteristics required by the supply chain prediction comprises the following steps:

acquiring a to-be-historic order file and a production plan file through an enterprise resource planning platform and a manufacturing execution platform;

And carrying out weighted aggregation on the historical order files and the production plan files according to time sequences and dimensions through data aggregation to obtain historical order data and production plan data.

According to the supply chain prediction method based on the knowledge graph and the industrial model provided by the invention, the acquisition of industrial purchase data and collection information in the characteristics required by the supply chain prediction comprises the following steps:

Acquiring supplier data and material data through a supply chain management system and a purchase management system;

inputting the supplier data and the material data to a supplier and material association matrix, and outputting industrial purchase data and collection information.

According to the supply chain prediction method based on the knowledge graph and the industrial model provided by the invention, the collection of logistics data and warehouse data in the characteristics required by the supply chain prediction comprises the following steps:

Acquiring logistics information and warehouse information through a logistics management system and a warehouse management system;

Inputting the logistics information and the warehouse information into a logistics and warehouse efficiency analysis model, and outputting an efficiency index as logistics data and warehouse data.

According to the supply chain prediction method based on the knowledge graph and the industrial model provided by the invention, the correlation analysis is carried out on the historical order data and the production plan data, the industrial purchase data and the collection information, and the logistics data and the storage data to obtain target characteristics, and the method comprises the following steps:

Respectively representing the historical order data, the production plan data, the industrial purchasing data and the collection information, the logistics data and the storage data after data cleaning and standardization as corresponding characteristics;

determining a correlation between each of said features using weighted pearson correlation coefficients;

a target feature is determined based on the correlation between the features.

According to the supply chain prediction method based on the knowledge graph and the industrial model provided by the invention, the key entities and relations in the target features are identified, and the knowledge graph is constructed based on the key entities and relations, and the method comprises the following steps:

Identifying an entity in a supply chain based on the target feature, the entity including a vendor, a product, and a logistics node;

determining the relationship among the suppliers, the products and the logistics nodes to obtain an upstream-downstream relationship, a logistics path and a transportation relationship and an inventory management relationship of a supply chain;

Integrating the suppliers, the products, the logistics nodes, the upstream and downstream relations of the supply chain, the logistics paths, the transportation relations and the inventory management relations into a multi-level knowledge graph, and constructing the knowledge graph through structural integration, mathematical modeling and graph reasoning.

According to the method for predicting the supply chain based on the knowledge graph and the industrial model provided by the invention, the supplier, the product, the logistics node, the upstream and downstream relation of the supply chain, the logistics path and the transportation relation and the inventory management relation are integrated into a multi-level knowledge graph, and the knowledge graph is constructed through structural integration, mathematical modeling and graph reasoning, comprising the following steps:

Formally representing the suppliers, the products, the logistics nodes, the upstream and downstream relations of the supply chains, the logistics paths and the transportation relations and the inventory management relations by a graph theory method;

introducing a cross-industry data set, and modeling probability dependency relations among the entities after formal representation through a Bayesian network to obtain a high-dimensional knowledge graph;

and embedding the high-dimensional knowledge graph into a low-dimensional space through graph embedding, and reserving the topological structure and semantic relation of the nodes to obtain the knowledge graph.

According to the supply chain prediction method based on the knowledge graph and the industrial model provided by the invention, the supply chain key nodes and node relations in the knowledge graph are identified by the industrial model and converted into the numerical feature vectors for supply chain prediction, and the method comprises the following steps:

Identifying information of each node in the knowledge graph by using a graph rolling network in the industrial model;

Aggregating the information of each node and the neighbor nodes by using the neighbor information to generate a new data representation;

identifying supply chain prediction features using the graph rolling network based on the data representation;

converting the supply chain prediction feature into a numeric feature vector.

According to the method for predicting the supply chain based on the knowledge graph and the industrial model provided by the invention, after converting the supply chain prediction characteristics into the numerical characteristic vectors, the method further comprises the following steps:

constructing a supply chain scene model based on the numerical feature vector;

Simulating predicted performance in different supply chain environments by using the supply chain scene model;

Based on the error in the predicted performance, an applicability and stability of the supply chain scene model is determined.

In a second aspect, the present invention also provides a supply chain prediction apparatus based on a knowledge graph and an industrial model, including:

The acquisition module is used for acquiring the required characteristics of the supply chain prediction, wherein the required characteristics comprise historical order data, production plan data, industrial purchase data, collection information, logistics data and storage data;

The analysis module is used for carrying out data cleaning, standardization and correlation analysis on the historical order data and the production plan data, the industrial purchasing data and the collection information, and the logistics data and the storage data to obtain target characteristics;

The construction module is used for identifying key entities and relations in the target features and constructing a knowledge graph based on the key entities and the relations;

And the prediction module is used for identifying the key nodes and the node relations of the supply chain in the knowledge graph by utilizing the industrial model, and converting the key nodes and the node relations into numerical characteristic vectors for supply chain prediction so as to carry out supply chain prediction.

In a third aspect, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, the processor implementing the knowledge-graph and industrial model-based supply chain prediction method as described in any one of the above when executing the program.

In a fourth aspect, the present invention also provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements a supply chain prediction method based on a knowledge graph and an industrial model as described in any of the above.

In a fifth aspect, the present invention also provides a computer program product comprising a computer program which, when executed by a processor, implements a supply chain prediction method based on a knowledge graph and an industrial model as described in any of the above.

The invention provides a supply chain prediction method based on a knowledge graph and an industrial model, which comprises the steps of collecting supply chain prediction required characteristics, wherein the required characteristics comprise historical order data, production plan data, industrial purchasing data, collection information, logistics data and storage data, carrying out data cleaning, standardization and correlation analysis on the historical order data, the production plan data, the industrial purchasing data, the collection information, the logistics data and the storage data to obtain target characteristics, identifying key entities and relations in the target characteristics, constructing the knowledge graph based on the key entities and the relations, identifying key nodes and node relations of the supply chain in the knowledge graph by utilizing the industrial model, and converting the key nodes and the node relations into numerical characteristic vectors for supply chain prediction so as to carry out supply chain prediction. By introducing the knowledge graph and the industrial model, the relation among all nodes in the supply chain can be better captured, so that the accuracy of the supply chain prediction is effectively improved.

Drawings

In order to more clearly illustrate the invention or the technical solutions of the prior art, the following description will briefly explain the drawings used in the embodiments or the description of the prior art, and it is obvious that the drawings in the following description are some embodiments of the invention, and other drawings can be obtained according to the drawings without inventive effort for a person skilled in the art.

FIG. 1 is a flow chart of a supply chain prediction method based on a knowledge graph and an industrial model according to the present embodiment;

FIG. 2 is a schematic diagram of the result of the predicted correlation of the supply chain according to the present embodiment;

FIG. 3 is a schematic diagram of the results of the key entity and relationship structure provided in the present embodiment;

Fig. 4 is a schematic diagram of a knowledge graph structure provided in the present embodiment;

FIG. 5 is a schematic diagram of a hybrid model structure provided in this embodiment;

FIG. 6 is a schematic diagram of MSE error provided by the present embodiment;

Fig. 7 is a schematic structural diagram of a supply chain prediction device based on a knowledge graph and an industrial model according to the present embodiment;

fig. 8 is a schematic structural diagram of an electronic device provided in the present embodiment.

Detailed Description

For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings, and it is apparent that the described embodiments are some embodiments of the present invention, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

Fig. 1 is a flow chart of a supply chain prediction method based on a knowledge graph and an industrial model according to the present embodiment.

As shown in fig. 1, the supply chain prediction method based on a knowledge graph and an industrial model provided by the embodiment of the invention mainly includes the following steps:

101. the method comprises the steps of collecting supply chain forecast required characteristics, wherein the required characteristics comprise historical order data, production plan data, industrial purchase data, collection information, logistics data and warehouse data.

In a specific implementation process, the historical order data and the production plan data are obtained to ensure that the collected data are comprehensive and accurate, and can reflect the actual operation condition of the enterprise. The acquisition of industrial purchase data and collection information is the basis of data collection, and provides a new view for supply chain management. The logistics data and the warehouse data ensure that all links from transportation to inventory have sufficient data support so as to improve the overall efficiency of the supply chain.

The method for acquiring the historical order data and the production plan data comprises the following steps:

The to-be-historic order files and the production plan files are acquired through the enterprise resource planning platform and the manufacturing execution platform. ERP (Enterprise resource planning) and MES (manufacturing execution System) platforms of the access enterprise are respectively logged into the order management module and the production planning module. In both modules, the time range (e.g., the last 12 months) and key data fields that need to be extracted, such as order number, order date, product number, planned production date, line number, etc., are selected. And then the data related to the specific product line or production area are respectively screened out through the built-in query function of the module. Multiple conditional queries, such as screening by order priority, customer type, or product type, may be provided to ensure that the extracted data more accurately reflects the actual production and order execution of the enterprise.

And carrying out weighted aggregation on the historical order files and the production plan files according to the time sequence and the dimension through data aggregation to obtain historical order data and production plan data. The formula of the weighted aggregation is (1):

Where Y _t,d,k represents the weighted aggregate result, i.e., the weighted total order volume or production plan quantity, within time window t (e.g., monthly), dimension d (e.g., a particular product number), and category k (e.g., high priority orders). X _i,t,d represents the ith data point, e.g., the number of a certain order or the number of a certain production plan, within time window t, dimension d, and class k. w _i,k is the weighting factor of the ith data point under category k for adjusting the weights of the different data points according to importance. For example, a high priority order may be given a higher weight. n _t, d represents the total number of data points in the time window t and dimension d, i.e., the total number of orders or production plans under that time period and product dimension. I (C _i,k) is an indicator function that determines whether data point I satisfies a particular condition C _i,k (e.g., whether a high priority order is being placed). If the condition is satisfied, I (C _i,k) =1, otherwise I (C _i,k) =0.

Through data aggregation, weighted time series data reflecting the fluctuation of order demands and the change trend of production load can be generated for subsequent trend analysis and prediction. The aggregation algorithm can dynamically adjust the time window (e.g., weekly or monthly) and weight and filter according to specific conditions (e.g., order priority or customer type) to accommodate data analysis on different time scales and business requirements.

As shown in table 1, the order history and production plan data after screening, integration and aggregation are shown, including information such as order number, order date, product number, planned production date and line number.

TABLE 1

Therefore, not only is the accuracy and integrity of the historical order data and the production plan data ensured, but also a high-quality data basis is provided for subsequent supply chain predictions.

The acquisition mode of the industrial purchase data and collection information is as follows:

The supplier data and the material data are obtained through a supply chain management system and a purchase management system. A supply chain management System (SCM) or a purchase management system (Procurement System) of the enterprise is accessed, a purchase order management module and a collection item management module are entered, and a time range (for example, the past 12 months) and key data fields required to be extracted, such as a purchase order number, a purchase date, a material number, a purchase quantity, supplier information, a purchase price, a delivery date, a collection item number, and the like, are selected. Through the built-in query function, data related to specific suppliers, material classes or collection items are screened out. Multiple conditional queries, such as screening by priority of materials, vendor ratings, purchase amounts, or the scale of collection items, may be set to ensure that the extracted data accurately reflects the actual purchasing behavior and collection activities of the enterprise.

In order to reveal a deeper supply chain relationship in purchase data, supplier data and material data are input into a supplier and material correlation matrix, and industrial purchase data and collection information are output. The vendor and materials association matrix is as in equation (2):

Where A _s,m represents the strength of the association between the supplier s and the material m, i.e. the supply ratio of the supplier to a certain type of material. P _i,s,m represents the quantity or amount of material m in the ith order provided by supplier s. Q _m represents the total number or amount of material m purchased by the business over the entire time frame. n _s represents the total number of orders offered by the supplier s. I (V _i,s) is an indicator function that determines whether the order belongs to the premium provider V _i,s. If so, I (V _i,s) =1, otherwise I (V _i,s) =0.

As shown in table 2, examples of data calculated from the supplier and material correlation matrix analysis are shown, including information of supplier number, material number, supplier-material correlation strength, total purchase quantity, quality supplier, etc.

TABLE 2

Through this correlation matrix analysis, businesses can identify which suppliers have a higher supply proportion in a particular material category, as well as the importance of good suppliers in critical material supply. The method not only can reflect the supply stability of the suppliers, but also can provide data support for the optimization of the collection strategy.

The logistics data and the warehouse data are obtained in the following modes:

And acquiring logistics information and warehouse information through a logistics management system and a warehouse management system. A logistics management system (TMS, transportation management system) and a warehouse management system (WMS, warehouse management system) of the visiting enterprise log in to the transportation scheduling module and the inventory management module, respectively. The time range (e.g., the last 12 months) and key data fields to be extracted are selected, including the shipping lot number, shipping date, shipping route, shipping time, shipping cost, warehouse location, inventory quantity, inventory turnover rate, etc. Through the built-in query function, data related to a specific transportation route, storage position or product type are screened out. Multiple condition queries, such as screening according to transportation frequency, warehouse utilization rate, or product characteristics, can be set to ensure that the extracted data can accurately reflect the operational efficiency and potential bottlenecks of enterprises in logistics and warehouse links.

In order to further analyze the logistics and storage data, a logistics and storage efficiency analysis formula can be utilized to calculate the comprehensive index of the transportation efficiency and the storage efficiency, namely, the logistics information and the storage information are input into a logistics and storage efficiency analysis model, and the efficiency index is output as the logistics data and the storage data, as shown in formula (3):

Where E _l,w,t represents the overall efficiency index over the transport route l, the warehouse location w and the time window t. T _l represents a unit transportation time of the transportation route l, for example, an average transportation time per batch. C _l represents the unit transportation cost of the transportation route i, such as the average transportation cost per batch. R _w represents the inventory turnover rate of the warehouse location w, reflecting the use efficiency of the warehouse. K _w,t represents an ideal stock level or target stock level of the warehouse location w within the time window t.

Through formula (3), the enterprise can calculate the comprehensive efficiency under different time periods, different transportation routes and storage positions, and then identify the link with the most efficiency in the supply chain. As shown in table 3, an example of data calculated from the logistics and warehousing efficiency analysis formulas is shown.

TABLE 3 Table 3

Therefore, enterprises can ensure the accuracy and the integrity of logistics data and warehouse data, and can also identify key links needing improvement in a supply chain through comprehensive efficiency analysis, so that the overall logistics and warehouse management level is improved.

102. And carrying out data cleaning, standardization and correlation analysis on the historical order data, the production plan data, the industrial purchasing data, the collection information, the logistics data and the storage data to obtain target characteristics.

Firstly, carrying out data cleaning and standardization processing on historical order data, production plan data, industrial purchasing data, collection information, logistics data and storage data. The data is checked, abnormal values, repeated values and missing values are identified and processed, and the integrity and the accuracy of the data are ensured. The unified standardization processing is carried out on the data from different sources, so that the data are consistent in scale and format, and the subsequent analysis and comparison are convenient. And the consistency check of the cleaned and standardized data is completed, so that the data quality is ensured, and a foundation is laid for subsequent analysis.

A data correlation analysis is then performed, which aims to identify the linear relationship between the different features, and thus find the variables that have the most impact on supply chain predictions. To improve the accuracy and practicality of the analysis, weighted pearson correlation coefficients and weighted bias correlation coefficients are introduced to better reflect the actual relationship between the different features when controlling the effects of other variables.

Performing correlation analysis to obtain target characteristics, wherein the target characteristics comprise:

And respectively representing the historical order data, the production plan data, the industrial purchasing data, the collection information, the logistics data and the storage data after data cleaning and standardization as corresponding characteristics. The correlation between the individual features is determined using a weighted pearson correlation coefficient. The weighted pearson correlation coefficient can give different weights to different data points, thereby making the analysis result more consistent with the actual situation in the supply chain. For example, in analyzing the relationship between order history data and industrial procurement data, critical orders or high volume purchases may be given higher weights to highlight the importance of these data points in supply chain management, as in equation (4):

In the formula, Is the weighted pearson correlation coefficient between features x and y, taking into account that the weight w _i.w_i of each data point is the weight of the i-th data point, reflecting the importance of that data point. The greater the weight, the greater the impact of the data point on the correlation coefficient. x _i and y _i are observations of feature x and feature y, respectively, in the ith sample.AndIs a weighted average of features x and y, defined as (5) and (6), respectively:

Wherein the molecular moiety AndThe weighted covariance is represented, which reflects the joint variation of feature x and feature y under weight assignment. The denominator part consists of the square root of two weighted variances, respectivelyAndThey are used to normalize the weighted covariance so that the correlation coefficient is between-1 and 1.

However, the relationship between different features in the supply chain may be affected by other variables, such as logistics and warehouse data. In this case, the true relationship between the two features may not be fully understood by only weighting the pearson correlation coefficients. The weighted bias correlation coefficient is used to measure the characteristic relationship after controlling the influence of other variables such as logistics data. It helps reveal the relationship between the two main features by controlling the effects of other variables (e.g., logistic data). The weighted bias correlation coefficient can be calculated to determine how strong the relationship between the historical order data and the production plan data and the industrial purchase data and collection information is after the influence of other variables (such as logistics data) is eliminated. As in formula (7):

in the following, the following steps are taken: The method is characterized in that after the influence of logistics and warehouse data (z) is controlled, the weighted bias correlation coefficient between historical order data and production plan data (x) and industrial purchase data and collection information (y) is controlled. Is a weighted pearson correlation coefficient between historical order data and production plan data and industrial purchase data and collection information. This factor takes into account the importance of critical orders and high volume purchases to supply chain forecast.AndThe historical order data, the production plan data, the logistics data, the warehouse data, the industrial purchasing data, the collection information, the logistics data and the warehouse data are weighted correlation coefficients respectively.

Finally, a target feature is determined based on the correlation between the features. The features most critical to supply chain prediction are selected as target features. Fig. 2 is a schematic diagram of a predicted correlation result of a supply chain according to the present embodiment, as shown in fig. 2, wherein the numerical value in each square represents a weighted correlation coefficient between two characteristics, i.e. the correlation between the historical order data and the production plan data and the industrial purchase data and the collection information is 1, and a very strong positive correlation is shown, which means that the interaction between the history of the order and the production plan and the purchase of the industrial in the supply chain is very close.

The correlation between the historical order data and the production plan data and the logistics data and the warehouse data is 1, and also shows high positive correlation, which indicates that a strong dependency relationship exists between the order and the production plan and the logistics management. The correlation between the purchasing data and collecting information of the industrial products, the logistics data and the storage data is 1, the strong correlation between purchasing and logistics is displayed, and the purchasing quantity and logistics arrangement are closely related.

103. And identifying key entities and relations in the target features, and constructing a knowledge graph based on the key entities and the relations.

Specifically, key entities and relationships are first defined. Based on target characteristics from the results of correlation analysis, identifying entities in the supply chain, namely core elements, and defining the correlation between the entities, thereby providing a solid foundation for knowledge graph construction. The entities include suppliers, products and logistics nodes, specifically:

(1) The importance of suppliers in the supply chain, such as their supply capacity, response time, and correlation with order history data, is determined by weighted pearson correlation coefficient and weighted partial correlation coefficient analysis.

(2) Product-definition of product entities key attributes such as production cycle, yield requirements, etc., reflected in production planning data, which have been identified as highly supply chain performance related features, need to be considered.

(3) Logistics nodes include warehouses, transportation routes, distribution centers, etc., and by correlation analysis, the relationship of these nodes to suppliers, products, and their impact on the overall supply chain efficiency are clarified.

And determining the relationship among the suppliers, the products and the logistics nodes to obtain the upstream and downstream relationship, the logistics path and the transportation relationship and the inventory management relationship of the supply chain, wherein the relationship comprises the following specific steps:

(1) And (3) the upstream and downstream relation of the supply chain, namely, the upstream and downstream relation among suppliers, products and logistics nodes is defined based on the correlation analysis result, such as the supply relation of the suppliers and the products, the circulation paths of the products at different logistics nodes and the like.

(2) And defining a transportation path among the logistics nodes, the products and suppliers, and determining the efficiency and the bottleneck of each transportation route by analyzing logistics and warehouse data.

(3) Inventory management relationships by analysis of inventory data, inventory management relationships of products among various logistics nodes are defined, including inventory turnover rate, storage cost, etc., which will help optimize inventory management policies in the supply chain.

FIG. 3 is a schematic diagram of the results of a key entity and relationship structure in the supply chain, as shown in FIG. 3, illustrating the four core entities (suppliers, products, logistics nodes, inventory) in the supply chain and their key relationships to each other. These relationships include supply relationships between suppliers and products, transport paths between products and logistics nodes, inventory management relationships between logistics nodes and inventory, and logistics distribution relationships between suppliers and logistics nodes. Intuitively express the core links and the interdependencies thereof in the supply chain, and provide a structural basis for further analysis and optimization.

And integrating the supplier, the product, the logistics node, the upstream and downstream relation of the supply chain, the logistics path, the transportation relation and the inventory management relation into a multi-level knowledge graph, and constructing the knowledge graph through structural integration, mathematical modeling and graph reasoning. After the definition of key entities and relations is completed, integrating the elements into a multi-level knowledge graph, and constructing a knowledge framework capable of comprehensively capturing the complexity of a supply chain by introducing a trans-industry data set. Not only includes the structured integration of data, but also involves complex mathematical modeling and graph reasoning algorithms to ensure that the knowledge graph can perform its maximum function in subsequent analysis and optimization. The method comprises the following steps:

In the first step, core entities, namely suppliers, products, logistics nodes, upstream and downstream relations of a supply chain, logistics paths, transportation relations and inventory management relations are expressed in a formalized mode through a graph theory method. Each entity E _i can be represented as a node, and each relationship R _ij corresponds to a directed edge between entities, and the relationships in the graph can be represented using the following formula, such as formula (8):

G=(V,E),V={E₁,E₂,....,E_n},E＝{R_ij∣i,j∈V} (8)

Wherein G represents the overall structure of the knowledge graph, V is the set of all entity nodes, E is the set of all relationships, and R _ij represents the relationship between the entity E _i and the entity E _j.

Secondly, in order to further enrich the expression capability of the knowledge graph, a cross-industry data set is introduced, and probability dependency relations among all entities after formal representation are modeled through a Bayesian network, so that the high-dimensional knowledge graph is obtained. Bayesian networks can handle uncertainties and complex dependencies efficiently, which is particularly important for randomness and uncertainty factors that often occur in the supply chain. The conditional probability distribution P (E _i∣E_j) in a bayesian network can be represented by the following formula (9):

In practical application, the probability influence on other related nodes when the state of a certain key node in the supply chain changes can be calculated through an inference algorithm, such as a maximum Expectation (EM) algorithm, of the Bayesian network. For example, how the supplier's supply capacity E ₁ changes will affect the product production plan E ₂ and inventory level E ₄.

In order to ensure the high efficiency of the knowledge graph, a Random Walk (Random Walk) algorithm is further introduced for analyzing the importance of each node in the graph. This algorithm can be described by the following Markov Chain (Markov Chain) model, as in equation (10):

P_t+1＝P_t*M (10)

Where P _t is the state vector of each node in the graph at time t, and M is the transition matrix describing the probability of transitioning from one node to another.

The random walk algorithm can help identify the most critical entities and paths in the supply chain, thereby guiding subsequent optimization decisions. For example, by analyzing random walks of inventory nodes, it may be determined which of the logistics nodes are most critical to the overall efficiency of the supply chain.

And thirdly, embedding the high-dimensional knowledge graph into a low-dimensional space through a graph embedding technology (Graph Embedding), so that a subsequent machine learning model is convenient to use, and the topological structure and semantic relation of the nodes in the original graph are reserved to obtain the knowledge graph. The embedding process can be expressed as formula (11):

f:V→R^d (11)

Where f is the embedding function, V is the node set, and R ^d is the target low-dimensional space.

Fig. 4 is a schematic diagram of a knowledge graph structure, and as shown in fig. 4, a knowledge graph is finally constructed, and in fig. 4, the nodes are given different weights according to importance of the nodes in a supply chain, and are distinguished by node sizes. The solid lines in FIG. 4 represent direct relationships in the supply chain (e.g., supply relationships, transport paths, inventory management, etc.) that form the core framework of the supply chain, while the dashed lines represent associations formed by the introduction of cross-industry data sets (e.g., market demand forecast, supply chain financial data, environmental impact, etc.), demonstrating the supply chain's interaction with a broader economic environment. Through the complicated and multi-level map, key paths and nodes in the supply chain are intuitively reflected, and enterprises are helped to identify the most important supply chain links, so that powerful support is provided for optimizing supply chain management and improving prediction accuracy.

104. And identifying key nodes and node relations of the supply chain in the knowledge graph by using the industrial model, and converting the key nodes and node relations into numerical characteristic vectors for supply chain prediction.

Specifically, the first step performs key feature extraction as follows:

Namely, using a graph rolling network (GCN) as an industrial model to conduct deep analysis on the constructed knowledge graph and extract key features highly relevant to supply chain prediction. The GCN model will help identify the most important information from key nodes in the supply chain (e.g., suppliers, products, logistics nodes, inventory, etc.) and complex relationships between them, and convert this information into a numerical feature vector that can be used for supply chain prediction. The method comprises the following specific steps:

First, information of each node in the knowledge-graph is identified using a graph-convolution network in the industrial model. In the supply chain knowledge graph, each node represents an important element in the supply chain, such as a supplier, a product, a logistics node or an inventory. The GCN model helps identify the importance of each node to supply chain predictions by analyzing its initial characteristics (e.g., supplier's supply capacity, product demand, logistics node transport efficiency, etc.). The GCN model first inputs an initial feature vector for each node. For example, the feature vectors of the provider nodes may include historical supply volume, delivery time, and product quality of the provider, etc. The GCN then integrates the initial characteristics of these nodes with the information of their neighboring nodes (e.g., product nodes associated with the vendor) through a convolution operation. Node analysis is as in equation (12):

In the formula, The feature representation of the provider node integrated with the neighbor information is that N (i) is a neighbor node set of the provider node, for example, its related product node, W ⁽⁰⁾ is a weight matrix of the model, and the optimal parameters are obtained through learning.

Through this process, the GCN model not only analyzes the basic information of the provider, but also combines the information of other nodes directly related to the provider (such as the product demand, the transportation capacity of the logistics nodes, etc.), thereby generating a richer node representation.

And secondly, aggregating the information of each node and the neighbor nodes by using the neighbor information to generate a new data representation. The process of neighbor information aggregation refers to integrating each node in the supply chain with the information of its associated node (neighbor node) to generate a new data representation. For example, for a provider node, its neighbor nodes may include all of its associated product nodes. The GCN model combines the information of these product nodes with the information of the supplier nodes themselves to generate a new, more complex representation of the feature. For each supplier node, the GCN will gather information about all of the product nodes connected to it, such as market demand fluctuations, production plans, etc. for the product. The GCN then integrates this information with the provider node's own information by means of weighted summation or averaging to generate a new feature representation. The polymerization formula is (13):

In the formula, Is a new feature representation of the provider node that is generated after the second level convolution operation.

Through this aggregation operation, the feature vector of the provider node not only contains its own attributes, but also reflects its relationship and interaction with all relevant product nodes.

Important features are then identified as follows:

After completion of the multi-layer convolution and information aggregation, the GCN model generates a final feature vector for each node in the supply chain. These feature vectors contain the information of the node itself and the comprehensive information of all the neighboring nodes related to it. Through training and optimization of the model, the GCN model can identify which features have the greatest impact on supply chain predictions. The GCN model minimizes the prediction error by a back propagation algorithm, thereby automatically adjusting the weight parameter W. The feature vectors for each node are continually optimized during the process so that features most relevant to supply chain predictions are weighted higher, while irrelevant or secondary features are weakened. The final objective is equation (14):

min_W∑_{(i,j)∈edges}(y_ij-f_W(h_i,h_j))² (14)

where y _ij is the actual observed behavior in the supply chain (e.g., supplier supply versus product demand). f _W(h_i,h_j) is the behavior that the model predicts based on the feature vectors h _i and h _j.

Through continuous optimization, the GCN model ultimately identifies the most important features for supply chain prediction, namely supply chain prediction features, for example, it may be found that the correlation of supplier supply time fluctuations and product demand fluctuations is most critical for predicting supply chain performance. These optimized feature vectors will serve as the core inputs to the supply chain prediction model.

Finally, the complex relationship in the knowledge graph is converted into a numerical feature vector as follows:

After identifying key features by the GCN model, the next step is to convert these complex relationships and information in the atlas, i.e., supply chain predicted features, into numeric feature vectors. The purpose is to represent complex relationships between nodes (e.g., relationships between suppliers and products, and between products and logistics nodes) as computer processable numerical values that will be input into the supply chain predictive model. The conversion process is as follows:

(1) Node feature integration the GCN model generates a feature vector for each node containing its own and neighbor information after the multi-layer convolution operation. For example, the feature vector of the provider node contains not only its supply capacity but also information about the market demand of the product node and the transport efficiency of the logistics node related thereto.

(2) Generating a digitized feature vector, the integrated information being ultimately represented as a digitized feature vector as in equation (15):

Wherein each numerical value Representing the performance of the provider node on a particular attribute, such as provisioning capability, response time, etc.

(3) Meaning of feature vector each feature vectorRepresenting the complex characteristics of node i in the supply chain, the quantized vector reduces these complex relationships to a set of values that can be used in a machine learning model. These vectors provide a unified input format for the model, facilitating subsequent analysis and prediction.

The results are shown in Table 4:

TABLE 4 Table 4

Table 4 shows the specific representation of the numerical features, and each node (e.g., supplier A, product X, logistics node 1, inventory Y) has a corresponding feature value. These characteristic values represent key attributes such as supply capacity, market demand fluctuation, transport efficiency, and inventory management stability, respectively. These values will be used as inputs to the supply chain prediction model for more accurately predicting various types of behaviors and trends in the supply chain.

Secondly, scene modeling and simulation are carried out, as follows:

After the key features are successfully extracted and converted into the numerical feature vectors, a supply chain scene model needs to be built based on the features, and the applicability and stability of the model are verified by simulating the performance under different supply chain environments. The purpose is to ensure that the model can still make accurate predictions and reasonable decisions under complex and variable supply chain scenarios. The method comprises the following specific steps:

First, a supply chain scene model is created:

And constructing a supply chain scene model based on the numerical feature vector. This model will simulate the dynamic behavior of the supply chain under different conditions in conjunction with each key node in the supply chain (e.g., suppliers, products, logistics nodes, inventory) and its interrelationships. The construction of the scene model comprises the following steps:

(1) Input features are defined by taking feature vectors extracted from the GCN as the input of the model, wherein the feature vectors comprise the supply capacity of suppliers, market demand fluctuation of products, transportation efficiency of logistics nodes and the like.

(2) The application proposes to use a dynamic prediction model fused by an attention mechanism to predict the future performance of each node in a supply chain, as shown in fig. 5, which is a schematic diagram of a mixed model structure. This model combines a multi-headed self-attention mechanism with a Recurrent Neural Network (RNN) to capture complex dependencies and time dynamics in the supply chain.

The input feature vector is processed by a plurality of attention heads, each responsible for capturing a different dependency between features. The outputs of all the attention heads are then combined to generate a rich representation. The multi-headed self-attention mechanism output is denoted (16):

Z=Concat(head₁,head₂,…,head_h)W_o (16)

Where Z is the output representation of the multi-headed self-attention mechanism, each attention head _i is generated by the input features weighted by the attention, and W _o is the output transform matrix.

The output of the attention mechanism is then input into the RNN, which processes the time series dynamics of these features, predicting future node states (e.g., demand, inventory levels, etc.). The RNN processing time series is expressed as (17):

h_t＝RNN(Z,h_t-1) (17)

Where h _t is the hidden state of the RNN at time step t. h _t-μ is the hidden state of the last time step.

Finally, the prediction results of the supply chain are generated by linear combination or multi-layer perceptron (MLP) processing of the RNN output. The combined model output is represented as (18):

In the formula, Is the final supply chain forecast. h _t is the hidden state of the RNN at the last time step.

Second, the performance under different supply chain environments is simulated:

And simulating the predicted performance under different supply chain environments by using a supply chain scene model. Once the supply chain scene model is established, simulation of performance under different supply chain environments may begin. This includes modeling various scenarios:

Simulation scenario = { supply chain break, demand fluctuation, transport delay }

The model then generates a set of predictions for each simulation scenarioAnd compared with the actual data to evaluate the applicability of the model. The data during the simulation are shown in table (5):

Context type	Actual value	Predictive value	Prediction error
				Supply chain interruption	1000	980	20
Demand fluctuations	1500	1525	-25
				Transport delay	800	790	10
Extreme events	1200	1150	50

TABLE 5

As can be seen from table 5, the model is able to keep small prediction errors under different scenarios, indicating its high stability in coping with various supply chain challenges.

Finally, verifying the applicability and stability of the model:

Based on the error in the predicted performance, the applicability and stability of the supply chain scene model are determined. Calculating MSE errors of the model under different situations, and verifying whether the model can keep stable performance under complex and changeable supply chain situations, as shown in a formula (19):

Where n is the number of samples, Is the predicted value of the model, and y _i is the actual observed value.

FIG. 6 is a graph of MSE error, as shown in FIG. 6, which shows that the MSE is lower in supply chain outage and demand fluctuation scenarios, 0.02 and 0.03, respectively, indicating that the model has higher prediction accuracy in coping with these common supply chain scenarios. In the case of a transport delay, the MSE is 0.025, and the display model is still able to predict the performance of the supply chain better. Even at extreme events (e.g., extreme weather, sudden policy changes, etc.), the MSE of the model is 0.04, although rising slightly, remains at a low level, demonstrating the stability of the model in the face of the extreme.

Through simulation tests, the high applicability and stability of the model under different situations are verified, so that the model can provide reliable prediction support for supply chain management, and the model can be ensured to show higher accuracy and stability under different supply chain situations, so that the reliable prediction support is provided for actual supply chain management.

Further, since the supply chain environment is dynamically changing, various conditions in the market demand, supply chain, and the like may change over time. Therefore, on the basis of the embodiment, the method further comprises the step of introducing a dynamic adjustment and optimization mechanism and continuously optimizing the feature selection and parameter setting of the model by utilizing an improved group strategy algorithm, so that the long-term accuracy and stability are maintained. The method comprises the following steps:

Firstly, real data and forecast data of a supply chain are collected in real time for comparison. During operation, the model continuously collects actual data (such as actual demand, actual stock level, etc.) in the supply chain, and compares the actual data with the predicted results of the model. Comprising the following steps:

(1) And acquiring real-time data, namely acquiring key data such as actual order quantity, inventory change, transportation delay and the like in real time through each node of a supply chain.

(2) Calculating a predicted value for each time step t by using a deviation calculation formulaThe deviation Δy _t from the actual value y _t is as in equation (20):

Where y _t is the actual value of time step t, Is the predicted value of time step t, Δy _t is the predicted deviation of time step t.

And secondly, carrying out deviation analysis and generating optimization suggestions based on the comparison result. Deviation analysis and optimization suggestion generation are performed, and once deviations are identified, the next step is to analyze the sources of these deviations and generate corresponding optimization suggestions. Comprising the following steps:

(1) And (3) analyzing the deviation, namely analyzing the performance of the model under the specific supply chain situation according to the deviation size and the change trend. The emphasis on which features or parameters may lead to a deviation. For example, if the deviation is large when predicting the demand of a certain product, it may be because the model fails to sufficiently take into account seasonal fluctuations of the market.

(2) Optimization advice generation formula, calculating feature weight adjustment advice aw according to the deviation, as in formula (21):

Δw=η*Δy_t*x_t (21)

Where η is a learning rate, x _t is an input feature vector, and Δw is an adjustment amount of a feature weight.

By equation (21), the system can generate suggestions for adjusting the feature weights based on the bias and feed these suggestions back to the optimization algorithm to adjust the model in the next step.

And finally, dynamically adjusting and optimizing the industrial model through a real-time feedback loop based on the optimization suggestion. And the real-time feedback mechanism finally forms a closed-loop system to continuously monitor, analyze and optimize the prediction result of the model. Comprising the following steps:

(1) Closed loop flow:

and monitoring, namely continuously monitoring the difference between the predicted result and the actual result of the model.

Analyzing the deviation in real time and generating optimization suggestions.

And adjusting feature weights and parameter configurations according to the suggestions.

(2) Feedback closed loop formula (22):

New_Parameters=Old_Parameters+Δw (22)

Where New_ Paramete is the parameter value of the model after receiving feedback and making adjustments. The new parameter is obtained by adding the old parameter to the adjustment amount aw. The old_parameters are the parameter values currently used by the model. The model uses these parameters to make predictions prior to feedback. Δw is a feature weight adjustment amount generated based on real-time feedback, and represents the correction made by the model based on actual data and prediction bias. The specific calculation is generally based on the prediction bias Δy _t and the corresponding feature vector x _t, multiplied by the learning rate η to control the magnitude of the adjustment.

Through the closed loop system, the model can be continuously self-adjusted and optimized to adapt to the continuously changing conditions in the supply chain, and the accuracy and timeliness of prediction are ensured.

The supply chain prediction method has the following advantages:

And 3, improving the prediction accuracy, namely, better capturing the complex relation among all nodes in the supply chain by introducing a knowledge graph and an industrial model algorithm, so that the prediction accuracy of the supply chain is obviously improved.

The dynamic adaptability is strong, the dynamic adjustment and optimization mechanism is combined, the model can automatically adjust parameter configuration according to real-time feedback, and the model has strong self-adaptive capacity and can cope with dynamic changes in the supply chain environment.

Optimizing feature selection, namely, the features most relevant to supply chain prediction can be flexibly selected through improved feature combination and a link optimizing method, so that the defect of feature selection is overcome.

And integrating the multi-industry data, namely, by constructing a knowledge graph covering the multi-industry data, different link data in the supply chain can be effectively integrated, and more comprehensive data support is provided for supply chain prediction.

Based on the same general inventive concept, the present invention also protects a supply chain prediction device based on a knowledge graph and an industrial model, and the supply chain prediction device based on a knowledge graph and an industrial model provided by the present invention is described below, and the supply chain prediction device based on a knowledge graph and an industrial model described below and the supply chain prediction method based on a knowledge graph and an industrial model described above can be correspondingly referred to each other.

Fig. 7 is a schematic structural diagram of a supply chain prediction device based on a knowledge graph and an industrial model according to the present embodiment.

As shown in fig. 7, the supply chain prediction apparatus based on a knowledge graph and an industrial model provided in this embodiment includes:

The collection module 701 is configured to collect supply chain prediction required features, where the required features include historical order data and production plan data, industrial purchase data, collection information, logistics data and warehouse data;

The analysis module 702 is configured to perform data cleaning, standardization and correlation analysis on the historical order data and production plan data, the industrial purchase data and collection information, and the logistics data and storage data to obtain target features;

A construction module 703, configured to identify key entities and relationships in the target feature, and construct a knowledge graph based on the key entities and relationships;

The prediction module 704 is configured to identify supply chain key nodes and node relationships in the knowledge graph by using an industrial model, and convert the supply chain key nodes and node relationships into a numerical feature vector for supply chain prediction to perform supply chain prediction.

As shown in FIG. 8, the electronic device may include a processor 810, a communication interface (Communications Interface) 820, a memory 830, and a communication bus 840, where the processor 810, the communication interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 may invoke logic instructions in the memory 830 to perform a supply chain prediction method based on a knowledge graph and an industrial model, the method comprising collecting supply chain prediction required features including historical order data and production plan data, industrial procurement data and collection information, logistics data and warehouse data, performing data cleaning, normalization and correlation analysis on the historical order data and production plan data, the industrial procurement data and collection information, the logistics data and warehouse data to obtain target features, identifying key entities and relationships in the target features, and constructing a knowledge graph based on the key entities and relationships, identifying supply chain key nodes and node relationships in the knowledge graph using the industrial model, and converting into a numerical feature vector for supply chain prediction to perform supply chain prediction.

Further, the logic instructions in the memory 830 described above may be implemented in the form of software functional units and may be stored in a computer-readable storage medium when sold or used as a stand-alone product. Based on this understanding, the technical solution of the present invention may be embodied essentially or in a part contributing to the prior art or in a part of the technical solution, in the form of a software product stored in a storage medium, comprising several instructions for causing a computer device (which may be a personal computer, a server, a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present invention. The storage medium includes a U disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, an optical disk, or other various media capable of storing program codes.

In another aspect, the present invention further provides a computer program product, where the computer program product includes a computer program, and the computer program can be stored on a non-transitory computer readable storage medium, and when the computer program is executed by a processor, the computer is capable of executing the supply chain prediction method based on a knowledge graph and an industrial model provided by the above methods, where the method includes collecting supply chain prediction required features, where the required features include historical order data and production plan data, industrial purchase data and collection information, logistics data and warehouse data, performing data cleaning, standardization and correlation analysis on the historical order data and production plan data, industrial purchase data and collection information, logistics data and warehouse data to obtain target features, identifying key entities and relationships in the target features, constructing knowledge based on the key entities and relationships, identifying key nodes and node relationships of a supply chain in the knowledge graph by using an industrial model, and converting the key nodes and node relationships into numerical feature vectors for supply chain prediction, so as to perform supply chain prediction.

In yet another aspect, the present invention further provides a non-transitory computer readable storage medium, on which a computer program is stored, which when executed by a processor, is implemented to perform the supply chain prediction method based on a knowledge graph and an industrial model provided by the above methods, the method comprising collecting supply chain prediction required features including historical order data and production plan data, industrial purchase data and collection information, logistics data and warehouse data, performing data cleaning, standardization and correlation analysis on the historical order data and production plan data, the industrial purchase data and collection information, the logistics data and warehouse data to obtain target features, identifying key entities and relationships in the target features, constructing a knowledge graph based on the key entities and relationships, identifying supply chain key nodes and node relationships in the knowledge graph by using the industrial model, and converting into numeric feature vectors for supply chain prediction to perform supply chain prediction.

The apparatus embodiments described above are merely illustrative, wherein the elements illustrated as separate elements may or may not be physically separate, and the elements shown as elements may or may not be physical elements, may be located in one place, or may be distributed over a plurality of network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art will understand and implement the present invention without undue burden.

From the above description of the embodiments, it will be apparent to those skilled in the art that the embodiments may be implemented by means of software plus necessary general hardware platforms, or of course may be implemented by means of hardware. Based on this understanding, the foregoing technical solution may be embodied essentially or in a part contributing to the prior art in the form of a software product, which may be stored in a computer readable storage medium, such as ROM/RAM, a magnetic disk, an optical disk, etc., including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the method described in the respective embodiments or some parts of the embodiments.

It should be noted that the above-mentioned embodiments are merely for illustrating the technical solution of the present invention, and not for limiting the same, and although the present invention has been described in detail with reference to the above-mentioned embodiments, it should be understood by those skilled in the art that the technical solution described in the above-mentioned embodiments may be modified or some technical features may be equivalently replaced, and these modifications or substitutions do not make the essence of the corresponding technical solution deviate from the spirit and scope of the technical solution of the embodiments of the present invention.

Claims

1. A supply chain prediction method based on a knowledge graph and an industrial model is characterized by comprising the following steps:

2. The knowledge-graph and industrial model-based supply chain prediction method of claim 1, further comprising:

3. The knowledge-graph and industrial model based supply chain prediction method of claim 1, wherein the collecting historical order data and production plan data in the supply chain predicted desired features comprises:

4. The supply chain prediction method based on a knowledge graph and an industrial model according to claim 1, wherein the collecting industrial purchase data and collection information in the characteristics required for the supply chain prediction comprises:

5. The supply chain prediction method based on a knowledge graph and an industrial model according to claim 1, wherein the collecting logistics data and warehouse data in the characteristics required for the supply chain prediction comprises:

6. The supply chain prediction method based on a knowledge graph and an industrial model according to claim 1, wherein the performing correlation analysis on the historical order data and production plan data, the industrial purchase data and collection information, the logistics data and the warehouse data to obtain target features comprises:

a target feature is determined based on the correlation between the features.

7. The knowledge-graph and industry model-based supply chain prediction method of claim 1, wherein the identifying key entities and relationships in the target features and constructing a knowledge graph based on the key entities and relationships comprises:

8. The knowledge-graph and industrial model-based supply chain prediction method of claim 7, wherein integrating the suppliers, the products, the logistics nodes, the supply chain upstream and downstream relationships, the logistics paths and transportation relationships, and the inventory management relationships into a multi-level knowledge graph through structured integration, mathematical modeling, and graph reasoning comprises:

9. The supply chain prediction method based on a knowledge graph and an industrial model according to any one of claims 1-8, wherein the identifying supply chain key nodes and node relationships in the knowledge graph using the industrial model and converting into a digitized feature vector for supply chain prediction comprises:

converting the supply chain prediction feature into a numeric feature vector.

10. The knowledge-graph and industrial model based supply chain prediction method of claim 9, further comprising, after said converting the supply chain prediction feature into a digitized feature vector:

constructing a supply chain scene model based on the numerical feature vector;