FPC production resource scheduling method based on multi-modal data analysis

Abstract

The application discloses a kind of FPC production resource scheduling methods based on multi-modal data analysis, it is related to intelligent manufacturing technical field, including: acquisition FPC multi-modal production data, pre-processing is generated standard FPC production data;Carry out minimum rotation matching, obtain production resource association graph;Input improved Node2Vec model, obtain production resource node representation data;Determine the degree of association and calculate resource undertaking relationship, generate process resource executable relationship data;Form the resource combination corresponding to each to-be-executed process, generate initial FPC production resource scheduling result;Update production resource node representation data, generate dynamic FPC production resource scheduling result;Production execution result is written into production resource association graph, and production resource scheduling is updated.The application carries out deep association representation by improving Node2Vec model, realizes the accurate matching of FPC production resource, conflict reduction and efficient dynamic scheduling.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, and in particular to an FPC production resource scheduling method based on multimodal data analysis. Background Technology

[0002] As flexible printed circuit board (FPC) production evolves towards multi-variety, small-batch, and high-frequency switching, the production process involves various types of data, including order execution, process coordination, equipment operation, material supply, personnel operations, quality inspection, and production progress. These data types differ in their collection cycles, data structures, and correlation methods. Current FPC production resource scheduling typically relies on order plans and resource availability. It often uses independent records or simple correlations to handle production objects and their relationships within different production periods, making it difficult to fully utilize the continuous structural relationships between different types of production data. This results in incomplete representation of production resource status and insufficient accuracy in perceiving changes in process execution status and resource occupancy.

[0003] Meanwhile, in FPC production, equipment, materials, personnel, and processes have continuous and mutually influential resource allocation relationships. When equipment occupancy, material supply, personnel operation status, or production progress changes, the original resource combinations can easily become unfeasible. Existing scheduling methods typically rearrange production tasks based on local resource status, making it difficult to continuously match the processes to be executed with available production resources by considering the structural relationships between production objects. This can easily lead to duplicate occupancy of the same resource, interruption of process allocation, and increased waiting time. Furthermore, it is difficult to promptly correct the resource combinations of affected processes after changes in production status, thereby reducing the efficiency and continuity of FPC production resource scheduling.

[0004] Therefore, how to provide an FPC production resource scheduling method based on multimodal data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an FPC production resource scheduling method based on multimodal data analysis. This invention comprehensively utilizes multimodal production data processing, production resource association graph construction, and improved Node2Vec model technology to complete the entire process from production data acquisition and preprocessing, resource association construction, node structure representation, process resource matching to dynamic scheduling updates. In terms of model structure, it sets up a combination number forward and reverse recursive calculation module, a window factor expansion module, and a four-node full permutation multiplication and addition module to enhance the representation capability of complex associations between production resource nodes, and combines the executable relationships of process resources to form resource combinations and dynamic scheduling results. Compared with existing technologies, this invention can improve the utilization degree of multimodal production data association and the accuracy of production status perception, reduce resource conflicts and process waiting, and improve the dynamic matching capability and scheduling efficiency of FPC production resources.

[0006] An FPC production resource scheduling method based on multimodal data analysis according to an embodiment of the present invention includes: Collect multimodal production data during the FPC production process, preprocess the multimodal production data, and generate standard FPC production data; Based on standard FPC production data, a two-layer splicing matrix is ​​constructed with production time period as the horizontal axis and object category as the vertical axis. Minimum rotation matching is performed on the two-layer splicing matrix to obtain the production resource association diagram. The production resource association graph is input into the improved Node2Vec model to perform deep representation learning on the structural associations between nodes in the production resource association graph, and the production resource node representation data is obtained. The degree of association between each process to be executed and available production resources is determined by the production resource node representation data, the resource acceptance relationship of each process to be executed is calculated, and process resource executable relationship data is generated. Based on the executable relationship data of process resources, resource combinations are formed for each process to be executed. The production start status, production completion status and process waiting status corresponding to different resource combinations are calculated to generate the initial FPC production resource scheduling result. Based on the initial FPC production resource scheduling results, continuously acquire changes in production status during the production execution process, update the production resource node representation data, and generate dynamic FPC production resource scheduling results; Production is executed according to the dynamic FPC production resource scheduling results, and the production execution results are written into the production resource association graph to update the production resource scheduling.

[0007] Optionally, obtaining the production resource association diagram includes: Establish a horizontal time period sequence according to the production time period order in the standard FPC production data, and establish a vertical object sequence according to the production object category. Fill the corresponding matrix positions with the production objects that appear in the same production time period to generate an object occurrence layer matrix. Fill the corresponding object occurrence layer matrix positions with production objects that have a continuous execution relationship in adjacent production periods to generate the object inheritance layer matrix. Then, concatenate the object occurrence layer matrix and the object inheritance layer matrix according to the same production period and the same object category to generate a two-layer concatenation matrix. Minimum rotation matching is performed on the double-layer splicing matrix. The adjacent time period matrix columns in the double-layer splicing matrix are rotated cyclically. The number of consistent positions of objects of the same type, objects of different types, and objects of receiving objects before and after the rotation is compared one by one. The rotation result corresponding to the largest number of consistent positions is taken as the minimum rotation matching result. Production object nodes and node connections are established based on the minimum rotation matching result to obtain the production resource association graph.

[0008] Optionally, the minimum rotation matching of the two-layer splicing matrix includes: The matrix column of the previous production period in the double-layer splicing matrix is ​​selected as the reference column, and the matrix column of the next production period is selected as the column to be matched. The object occurrence layer and the object receiving layer in the column to be matched are rotated cyclically position by position to generate multiple rotation candidate columns. Read the number of times each rotation candidate column matches the object number in the same object category position, the number of times the successor number matches in different object category positions, and the number of times the execution relationship matches in the corresponding positions of the upper and lower layers, to obtain the matching record for each rotation candidate column; All matching records are compared sequentially based on the number of matching object IDs, the number of matching successor IDs, and the number of matching execution relationships. The first rotation candidate column that reaches the maximum consistency result is determined as the target matching column. Production object node connections are generated based on the corresponding positions between the baseline column and the target matching column.

[0009] Optionally, obtaining the production resource node representation data includes: An improved Node2Vec model is constructed. The original Node2Vec model includes random walk processing, context training order generation, and node vector training. A combination number forward and reverse recursion module is set at the random walk processing position of the original Node2Vec model, a window factor expansion module is set at the context training order generation position, and a four-node permutation multiplication and addition module is set at the node vector training position. The production resource association graph is input into the improved Node2Vec model. The nodes in the production resource association graph are used as the starting point of the walk. The multi-level connection relationship between the current node and the candidate node is processed by the combination number forward and reverse recursion module to determine the next walk node and generate the production resource walk sequence according to the node arrival order. The production resource walk sequence is input into the window factor expansion module, and the central node and context nodes are factored. The training order of the context corresponding to the central node is determined based on the factor expansion result. Input the central node and its corresponding context training order into the four-node permutation multiplication and addition module to perform permutation multiplication and addition training, and obtain the production resource node representation data.

[0010] Optionally, generating the production resource walk sequence includes: Starting from the current node, read the direct and multi-level connection relationships between each candidate node and the current node in the candidate node set to form a candidate node connection record. The candidate node connection records are processed in a forward recursive manner, and then the forward recursive processing results are processed in a reverse recovery manner to obtain the recovery results of each candidate node under the multi-level connection relationship. The direct connections between each candidate node are removed, and the removed connection records are recalculated in the forward direction and recovered in the reverse direction. The walking order of the candidate nodes is determined based on the difference between the recovery results before and after the removal, and the walking sequence of production resources is generated.

[0011] Optionally, the step of performing full permutation multiplication and addition training includes: From the context training order, select one central node and three context nodes to form a four-node training group, and divide the node vectors of the four nodes into multiple coordinate groups according to four consecutive coordinates. For each coordinate group, a four-row, four-column numerical table is created. All permutations are formed by selecting one coordinate in each row and ensuring that the selected coordinate columns are not repeated. The four coordinates in each permutation are multiplied consecutively, and the results of all permutations are summed to obtain the four-node multiplication result. Based on the results of the four-node multiplication and addition, the vectors of each node in the four-node training group are updated sequentially. When updating the current coordinate, the row and column where the current coordinate is located are deleted. The multiplication and addition process is performed on all the permutations of the remaining three rows and three columns of values, and the update result of the current coordinate is determined. After completing the training of all coordinate groups, the production resource node representation data is obtained.

[0012] Optionally, the generated process resource executable relationship data includes: Locate the process nodes corresponding to each process to be executed according to the production resource association diagram, filter the production resource nodes that are currently available along the node connections of the process nodes, and read the production resource node representation data corresponding to the process nodes and production resource nodes. The node representation data of the process node and each production resource node are multiplied and accumulated according to the same vector position. The accumulated result is divided by the product of the square root of the square of each node representation data to obtain the degree of association between the process to be executed and each production resource. The starting point for the next process is determined by the completion position of the preceding process. The current end position of each production resource is sequentially connected to the starting point. The relationship between production resources that can continuously accept the next process is determined based on the degree of correlation. The production resource relationships corresponding to each process to be executed are matched according to the starting point of acceptance and the end position of resource occupation. Production resource relationships that do not have co-occupancy and meet the requirements for continuous execution of the process are retained, and process resource executable relationship data is generated.

[0013] Optionally, generating the initial FPC production resource scheduling result includes: According to the process to be executed, extract the production resources that can jointly undertake the current process from the process resource executable relationship data, and combine the corresponding production resources according to the same process to be executed to generate candidate resource combinations; The production start position is determined by the end position of each production resource in the candidate resource combination and the previous completion position of the process to be executed. The production start position is obtained by continuously superimposing the process execution length. The process waiting length is obtained by the interval between the production start position and the previous completion position. Candidate resource combinations corresponding to the same process to be executed are embedded into the current production task in sequence according to the production completion position. The resource occupation positions of the production tasks before and after embedding are checked one by one. Candidate resource combinations that are occupied repeatedly are removed from their corresponding positions and then embedded into the next available position to obtain the resource combination arrangement result. Based on the sequential relationship of each pending process, the resource combination arrangement results are connected, and the production start position, production completion position, and process waiting length are written into the corresponding production task to generate the initial FPC production resource scheduling result.

[0014] Optionally, obtaining the resource combination arrangement result includes: Using the production start position of the process to be executed as the insertion position, candidate resource combinations are inserted into the current production task arrangement respectively. The resource occupation positions corresponding to each production task before and after the insertion position are read to form the front occupation segment, current occupation segment and back occupation segment of each candidate resource combination. Cross-check the current occupied segment with the previous occupied segment and the next occupied segment position by position. Record the continuous overlap length at the position where the same production resource overlaps. Add the continuous overlap lengths on the front and back sides to obtain the occupation conflict amount of the candidate resource combination. The candidate resource combinations are moved in order of increasing position of the conflict, and the front, current and rear occupied segments are re-formed after each move until the conflict is zero. The corresponding insertion position is then determined as the effective arrangement position of the candidate resource combination, and the resource combination arrangement result is obtained.

[0015] Optionally, generating dynamic FPC production resource scheduling results includes: The real-time production status is matched with the initial FPC production resource scheduling results for each task, and the production resource nodes and pending process nodes corresponding to the status changes are located to form a set of changed nodes. The connection path between the set of changing nodes and the corresponding process node to be executed is extracted along the production resource association graph. The connection path is input into the improved Node2Vec model to re-represent the structural association and replace the original production resource node representation data of the corresponding node to obtain the updated production resource node representation data. Based on the updated production resource node representation data, the resource acceptance relationship and resource combination of the corresponding process to be executed are re-determined, and the newly formed resource combination replaces the corresponding part in the initial FPC production resource scheduling result to generate a dynamic FPC production resource scheduling result.

[0016] The beneficial effects of this invention are: This invention improves the utilization of relationships between different production periods and different production objects by uniformly processing multimodal production data in the FPC production process and establishing a production resource association graph using a two-layer splicing matrix and minimum rotation matching. By improving the Node2Vec model to perform deep representation learning on the production resource association graph, and using forward and reverse recursion of combination numbers, window factor expansion, and four-node full permutation multiplication and addition processing, the representation ability of multi-level structural relationships and combination associations between nodes is enhanced, improving the completeness and accuracy of production status perception, and providing a reliable data foundation for matching between the processes to be executed and production resources.

[0017] This invention establishes executable relationships for process resources based on production resource node representation data, and combines production start status, production completion status, and process waiting status to form resource combinations and initial scheduling results. When the production status changes, the representation data of the affected nodes are updated and the resource acceptance relationship is redefined, realizing dynamic adjustment of FPC production resources. Compared with fixed scheduling or partial resource adjustment methods, this invention can reduce resource occupation conflicts between equipment, materials, and personnel, reduce process waiting and ineffective resource occupation, and improve the continuous acceptance capacity between production tasks, resource utilization efficiency, and FPC production scheduling efficiency. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 The flowchart shows an FPC production resource scheduling method based on multimodal data analysis proposed in this invention. Figure 2 This is a structural block diagram of an improved Node2Vec model for an FPC production resource scheduling method based on multimodal data analysis proposed in this invention. Figure 3 This is a functional diagram illustrating the updating of production resource scheduling in an FPC production resource scheduling method based on multimodal data analysis proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1 , Figure 2 and Figure 3A method for FPC production resource scheduling based on multimodal data analysis, comprising: Collect multimodal production data during the FPC production process, preprocess the multimodal production data, and generate standard FPC production data; Based on standard FPC production data, a two-layer splicing matrix is ​​constructed with production time period as the horizontal axis and object category as the vertical axis. Minimum rotation matching is performed on the two-layer splicing matrix to obtain the production resource association diagram. The production resource association graph is input into the improved Node2Vec model to perform deep representation learning on the structural associations between nodes in the production resource association graph, and the production resource node representation data is obtained. The degree of association between each process to be executed and available production resources is determined by the production resource node representation data, the resource acceptance relationship of each process to be executed is calculated, and process resource executable relationship data is generated. Based on the executable relationship data of process resources, resource combinations are formed for each process to be executed. The production start status, production completion status and process waiting status corresponding to different resource combinations are calculated to generate the initial FPC production resource scheduling result. Based on the initial FPC production resource scheduling results, continuously acquire changes in production status during the production execution process, update the production resource node representation data, and generate dynamic FPC production resource scheduling results; Production is executed according to the dynamic FPC production resource scheduling results, and the production execution results are written into the production resource association graph to update the production resource scheduling.

[0021] The generation of standard FPC production data includes: Collect data records corresponding to orders, processes, equipment, materials, personnel, quality inspection, and production progress during the FPC production process. Organize the collection time and execution time of each data record according to a unified production time benchmark. Establish the correspondence between different data records based on the production object number to obtain multimodal production initial data. Duplicate records in the initial data of multimodal production are merged, missing records are marked with missing positions, abnormal values ​​are marked with abnormal positions, and the units of measurement for production quantity, processing time, inventory quantity and resource occupation time are standardized to obtain the sorted FPC production data. The FPC production data is arranged and organized according to the order execution sequence and the order of process occurrence. The status records of the same production object in continuous production periods are matched, and the data field names, data formats and object identifiers are unified to generate standard FPC production data.

[0022] In this embodiment, obtaining the production resource association diagram includes: A horizontal time period sequence is established according to the production time period order in the standard FPC production data, and a vertical object sequence is established according to the production object category. Production objects appearing in the same production time period are filled into the corresponding matrix positions to generate an object occurrence layer matrix. The specific details of the generated object occurrence layer matrix are as follows: Using the start and end times of each production record in the standard FPC production data as the time period division positions, continuous production time periods are formed according to the time sequence. Each production time period is then used as a matrix column, and matrix rows are established according to the production object category. For each object category, the number of production objects appearing simultaneously in each production time period is counted. The number of sub-positions in the corresponding row is determined by the maximum number of simultaneous occurrences in that category. For each production time period, production objects whose execution time overlaps with that production time period are selected and arranged in order of first appearance time. If the first appearance time is the same, they are arranged according to the object number. The arranged object numbers are then filled into the sub-positions of the corresponding categories. Positions where no production object appears are filled with zero values. After filling all production time periods and production object categories, each column in the matrix represents a production time period, and each row and its sub-positions represent the occurrence of production objects of the corresponding category in that time period, generating an object occurrence layer matrix. Production objects with continuous execution relationships in adjacent production periods are filled into the corresponding object occurrence layer matrix positions to generate an object inheritance layer matrix. The object occurrence layer matrix and the object inheritance layer matrix are then concatenated according to the same production period and the same object category to generate a two-layer concatenation matrix. Specifically, production objects with continuous execution relationships in adjacent production periods are filled into the corresponding object occurrence layer matrix positions. Using the matrix columns corresponding to two adjacent production periods in the object occurrence layer matrix as the processing range, the object number, order number, process number, and end time of each production object in the previous production period, and the object number, order number, process number, and start time of each production object in the next production period are read respectively. First, production objects with the same order number are filtered out, and then the process inheritance relationship is checked according to the process route. When the process corresponding to the next production object is immediately followed by the process corresponding to the previous production object, and the start time of the next production object is not earlier than the end time of the previous production object, the two production objects are established as a continuous execution relationship. When there are multiple inheritable subsequent production objects for the same preceding production object, the interval between the start time of each subsequent production object and the end time of the preceding production object is calculated. The production object with the smallest interval is retained. According to the production period, object category, and sub-position of the two production objects in the object occurrence layer matrix, the continuous execution relationship is written to the same position. Positions where no continuous execution relationship is formed are filled with 0. After completing the processing of all adjacent production periods, the object inheritance layer matrix is ​​generated. Minimum rotation matching is performed on the double-layer splicing matrix. The adjacent time period matrix columns in the double-layer splicing matrix are rotated cyclically. The number of consistent positions of objects of the same type, objects of different types, and objects of receiving objects before and after the rotation is compared one by one. The rotation result corresponding to the largest number of consistent positions is taken as the minimum rotation matching result. Production object nodes and node connections are established based on the minimum rotation matching result to obtain the production resource association graph.

[0023] In this embodiment, the minimum rotation matching of the double-layer splicing matrix includes: In the double-layer concatenation matrix, the column from the previous production period is selected as the baseline column, and the column from the next production period is selected as the column to be matched. The object occurrence layer and the object receiving layer in the column to be matched are simultaneously rotated position-by-position in a cyclical manner to generate multiple rotation candidate columns. Specifically, the generation of multiple rotation candidate columns is as follows: Using the matrix column of the next production period in the double-layer splicing matrix as the column to be matched, the total number of rotatable positions is determined from top to bottom according to the object category and its sub-position. Data in the same position of the object appearance layer and the object receiving layer is taken as one synchronous rotation unit. First, the original arrangement of the column to be matched is retained as the first candidate result. All synchronous rotation units are moved down one position as a whole, and the data in the original end position is returned to the beginning position to form the second candidate result. Each time, the column is moved down one position, and the data that has crossed the end is returned to the beginning position in turn until the cumulative number of moved positions is equal to the total number of rotatable positions minus 1. During each movement, the object appearance layer and the object receiving layer maintain the same movement direction and the same number of moved positions. The complete matrix column after each rotation is saved to generate multiple rotation candidate columns. Read the number of times each rotation candidate column matches the object number in the same object category position, the number of times the successor number matches in different object category positions, and the number of times the execution relationship matches in the corresponding positions of the upper and lower layers, to obtain the matching record for each rotation candidate column; All matching records are compared sequentially based on the number of matching object IDs, the number of matching successor IDs, and the number of matching execution relationships. The first rotation candidate column that reaches the maximum consistency result is determined as the target matching column. Production object node connections are generated based on the corresponding positions between the baseline column and the target matching column.

[0024] In this embodiment, obtaining the production resource node representation data includes: An improved Node2Vec model is constructed. The original Node2Vec model includes random walk processing, context training order generation, and node vector training. A combination recursion module is added to the random walk processing position of the original Node2Vec model; a window factorization module is added to the context training order generation position; and a four-node permutation multiplication and addition module is added to the node vector training position. Specifically, the improved Node2Vec model is constructed as follows: This invention structurally improves the original Node2Vec model by setting up a combination number forward and reverse recursion module, a window factor expansion module, and a four-node permutation multiplication and addition module in the random walk processing, context training order generation, and node vector training positions, respectively. This enables the model to continuously process multi-level connections, context combinations, and multi-node joint relationships, thereby improving the structural representation capability of the production resource association graph. During the random walk processing phase, the combination number forward and reverse recursion module performs forward recursion and reverse recovery on the multi-level connections between the current node and the candidate nodes. After removing direct connections, it processes them again and determines the walking order of the candidate nodes based on the difference between the two results, so that the generated production resource walking sequence can retain more stable continuous connection relationships. In the context training order generation stage, the window factor expansion module establishes factors based on the positional interval and co-occurrence relationship between the central node and the context nodes. By comparing the changes in the expansion results through continuous expansion, single node removal, and double node removal, the training order of each context node is determined, thereby enhancing the model's ability to identify combined associations in the wandering sequence. During the node vector training phase, the four-node permutation multiplication and addition module forms a training group with one central node and three context nodes, groups and multiplies the vector coordinates of the four nodes, and updates the node vectors based on the arrangement results of the remaining coordinates. This expands the node training from pairwise relationships to a four-node joint relationship, improving the structural bearing capacity of the node representation. The improved Node2Vec model can represent the relationship between production resources from three levels: multi-level connection, context combination, and multi-node union. This improves the completeness and distinguishability of the data represented by production resource nodes, and provides a more accurate data foundation for judging the relationship between the process to be executed and the available production resources and for subsequent resource scheduling. The production resource association graph is input into the improved Node2Vec model. The nodes in the production resource association graph are used as the starting point of the walk. The multi-level connection relationship between the current node and the candidate node is processed by the combination number forward and reverse recursion module to determine the next walk node and generate the production resource walk sequence according to the node arrival order. The production resource walk sequence is input into the window factorization module. Factorization is performed on the center node and context nodes. Based on the factorization result, the training order of the context corresponding to the center node is determined. Specifically, the factorization of the center node and context nodes involves: Using the current central node in the production resource walk sequence as a reference, read the context nodes within a set range before and after it, record the positional interval between each context node and the central node and the number of times they co-occur in the entire production resource walk sequence, and use the positional interval of each context node as the factor and the number of times they co-occur as the constant term of the factor to form the corresponding node factor. Multiply each node factor in turn according to the order in which the context nodes appear in the walk sequence. For each factor added, multiply each of the existing terms by the two terms in the new factor. Add up the results with the same degree until the complete expansion result is obtained. One context node factor is removed sequentially and expanded again. The total change in the value of each exponent term before and after removal is compared. Two adjacent context nodes are removed simultaneously and expanded repeatedly. The change generated when the two nodes coexist is compared. The correlation strength between each context node and the central node is determined according to the change of a single node and the change of two nodes together, thus forming the context training order corresponding to the central node. Input the central node and its corresponding context training order into the four-node permutation multiplication and addition module to perform permutation multiplication and addition training, and obtain the production resource node representation data.

[0025] Specifically, training and improving the Node2Vec model involves: The improved Node2Vec model uses the production resource association graph as the training input and connects them in the following order: combination number forward and reverse recursion module - window factor expansion module - four-node full permutation multiplication and addition module - node vector output. A node vector table is established according to the number of nodes in the production resource association graph. Each node is given a 128-dimensional trainable vector, and the initial values ​​of each dimension are randomly generated in the range of -0.01 to 0.01. Using each node as the starting point for the walk, the four-level connection relationship from the current node to the candidate node is read through the combination number forward and reverse recursion module. Forward recursion and reverse recovery are performed on the number of connections at each level. After removing direct connections, the process is repeated. The walk order of the candidate nodes is determined based on the results of the two recovery operations. The walk is performed 10 times from each node, with 40 nodes each time, to generate a production resource walk sequence. For each walk sequence, the center node is selected in sequence. The five positions before and after the center node are used as the context range. The position interval and the number of times they co-occur between the context node and the center node are recorded through the window factor expansion module. The node factors are established in sequence and continuously expanded. Then, one and two context nodes are removed and expanded again. The context training order is determined based on the changes in the expansion results. Following the context training order, a positive four-node training group is formed by selecting one central node and the first three context nodes. Five nodes are selected from outside the context range of the central node to replace one context node in turn, forming five negative four-node training groups. The 128-dimensional vector of the four nodes is divided into 32 groups according to four consecutive coordinates using the four-node permutation multiplication and addition module. A 4-row, 4-column numerical table is created for each group, resulting in 24 permutations with non-repeating column positions. The four coordinates in each permutation are multiplied consecutively and then summed to obtain the four-node multiplication and addition result. When updating any coordinate, delete the row and column containing that coordinate. Form 6 permutations of the remaining 3 rows and 3 columns of data and multiply and add them respectively. Use the results to determine the update amount of the corresponding coordinate. Mark the positive four-node training group as 1 and the negative four-node training group as 0. Use binary classification cross-entropy to determine the training error. Set the batch size to 256, the initial learning rate to 0.001, and the maximum number of training rounds to 100. Stop training when the average training error change is less than 0.00001 for 5 consecutive rounds. Save the 128-dimensional node vector corresponding to each node to generate production resource node representation data.

[0026] In this embodiment, generating the production resource walk sequence includes: Starting from the current node, read the direct and multi-level connection relationships between each candidate node and the current node in the candidate node set to form a candidate node connection record. The candidate node connection records are processed in a forward recursive manner, and then the results of the forward recursive processing are processed in a reverse recovery manner to obtain the recovery results of each candidate node under the multi-level connection relationship. Specifically, the recovery results of each candidate node under the multi-level connection relationship are as follows: The number of connections from level 0 to level 4 for each candidate node is used as a set of connection records. The number of connections from level 0 is the flag value indicating whether the current node and the candidate node are the same node. The number of connections from level 1 to level 4 represent the number of connections that can reach the candidate node through the connections from level 1 to level 4 nodes, respectively. During forward recursion, the first level recursion value is the number of level 1 connections minus the number of level 0 connections; the second level recursion value is the number of level 2 connections minus twice the number of level 1 connections plus the number of level 0 connections; the third level recursion value is the number of level 3 connections minus three times the number of level 2 connections plus three times the number of level 1 connections minus the number of level 0 connections; and the fourth level recursion value is the number of level 4 connections minus four times the number of level 3 connections plus six times the number of level 2 connections minus four times the number of level 1 connections plus the number of level 0 connections. Reverse recovery is performed. The first-level recovery result is the sum of the recursive value of level 0 and the recursive value of level 1. The second-level recovery result is the sum of the recursive value of level 0, twice the recursive value of level 1 and the recursive value of level 2. The third-level recovery result is the sum of the recursive value of level 0, three times the recursive value of level 1, three times the recursive value of level 2 and the recursive value of level 3. The fourth-level recovery result is the sum of the recursive value of level 0, four times the recursive value of level 1, six times the recursive value of level 2, four times the recursive value of level 3 and the recursive value of level 4. The recovery results of each candidate node under the connection relationship of level 1 to level 4 are obtained. The direct connections of each candidate node are extracted, and the extracted connection records are recalculated in a forward and reverse manner. The traversal order of the candidate nodes is determined based on the difference in the traversal results before and after extraction, generating a production resource traversal sequence. Specifically, the generation of the production resource traversal sequence is as follows: For each candidate node, the number of direct connections at level 1 is first set to 0, while the number of connections at levels 2 to 4 remains unchanged. Forward recursion and reverse recovery are performed to obtain the recovery results at levels 1 to 4 after removing direct connections. The recovery results at the same level before and after removal are subtracted level by level, and the absolute values ​​are summed to obtain the connection change of each candidate node. At the same time, the recovery results at levels 2 to 4 after removal are summed to obtain the remaining connection quantity. Candidate nodes are arranged in descending order of remaining connection quantity. When the remaining connection quantities are the same, the candidate node with the smaller connection change quantity is selected as the next walking node, and this node is used as the new current node for repeated processing. When the number of consecutively recorded nodes reaches 40, the current walk ends. The 40 nodes are arranged in the order of arrival to generate the production resource walk sequence.

[0027] In this embodiment, the step of performing full permutation multiplication and addition training includes: From the context training order, select one central node and three context nodes to form a four-node training group, and divide the node vectors of the four nodes into multiple coordinate groups according to four consecutive coordinates. For each coordinate group, a four-row, four-column numerical table is created. All permutations are formed by selecting one coordinate in each row and ensuring that the selected coordinate columns are not repeated. The four coordinates in each permutation are multiplied consecutively, and the results of all permutations are summed to obtain the four-node multiplication result. Based on the results of the four-node multiplication and addition, the vectors of each node in the four-node training group are updated sequentially. When updating the current coordinate, the row and column where the current coordinate is located are deleted. The multiplication and addition process is performed on all the permutations of the remaining three rows and three columns of values, and the update result of the current coordinate is determined. After completing the training of all coordinate groups, the production resource node representation data is obtained.

[0028] In this embodiment, the generated process resource executable relationship data includes: Locate the process nodes corresponding to each process to be executed according to the production resource association diagram, filter the production resource nodes that are currently available along the node connections of the process nodes, and read the production resource node representation data corresponding to the process nodes and production resource nodes. The node representation data of the process node and each production resource node are multiplied and accumulated according to the same vector position. The accumulated result is divided by the product of the square root of the square of each node representation data to obtain the degree of association between the process to be executed and each production resource. The starting point for the next process is determined based on the completion position of the preceding process. The current end position of each production resource is then sequentially connected to this starting point. The relationships between production resources that can continuously support the next process are determined based on their degree of association. Specifically, the determination of these relationships involves: For the same pending process, the correlation degree and current end position of each candidate production resource are read. The completion position of the preceding process is used as the starting point for acceptance. The time difference between the end position of resource occupancy and the starting point is recorded as the acceptance interval. The acceptance interval that is earlier than or equal to the starting point is recorded as 0. Candidate production resources are compared pairwise. When the correlation degree of a production resource is not lower than that of another production resource and the acceptance interval is not greater than that of another production resource, the former is retained and the latter is deleted. This process continues until there are no candidate resources that are surpassed by other resources in both correlation degree and acceptance interval. The retained resources are then linked to the pending process to establish a continuous acceptance relationship, thus obtaining the production resource relationship. The production resource relationships corresponding to each process to be executed are matched according to the starting point of acceptance and the end position of resource occupation. Production resource relationships that do not have co-occupancy and meet the requirements for continuous execution of the process are retained, and process resource executable relationship data is generated.

[0029] In this embodiment, generating the initial FPC production resource scheduling result includes: According to the process to be executed, extract the production resources that can jointly undertake the current process from the process resource executable relationship data, and combine the corresponding production resources according to the same process to be executed to generate candidate resource combinations; The production start position is determined by the end position of each production resource in the candidate resource combination and the previous completion position of the process to be executed. The production start position is obtained by continuously superimposing the process execution length. The process waiting length is obtained by the interval between the production start position and the previous completion position. Candidate resource combinations corresponding to the same process to be executed are embedded into the current production task in sequence according to the production completion position. The resource occupation positions of the production tasks before and after embedding are checked one by one. Candidate resource combinations that are occupied repeatedly are removed from their corresponding positions and then embedded into the next available position to obtain the resource combination arrangement result. Based on the sequential relationship of each pending process, the resource combination arrangement is connected, and the production start position, production completion position, and process waiting length are written into the corresponding production task to generate the initial FPC production resource scheduling result. Specifically, the generation of the initial FPC production resource scheduling result is as follows: The resource combination arrangement results are read sequentially according to the succession relationship of each pending process. The production start position is determined by the effective arrangement position of the first pending process. The production start position is added to the process execution length to obtain the production completion position. For all processes, the production completion position of the preceding process is compared with the available position of the corresponding resource combination. The later position is taken as the new production start position. The process waiting length is obtained by subtracting the production completion position of the preceding process from the production start position. The production completion positions are then determined sequentially. If the corresponding position is occupied by resources, the current process is moved to the next available position of that resource and the position of the subsequent process is re-determined. After all pending processes are processed, the resource combination, production start position, production completion position and process waiting length of each process are recorded according to the production task order to generate the initial FPC production resource scheduling result.

[0030] In this embodiment, obtaining the resource combination and arrangement result includes: Using the production start position of the process to be executed as the insertion position, candidate resource combinations are inserted into the current production task arrangement respectively. The resource occupation positions corresponding to each production task before and after the insertion position are read to form the front occupation segment, current occupation segment and back occupation segment of each candidate resource combination. Cross-check the current occupied segment with the previous occupied segment and the next occupied segment position by position. Record the continuous overlap length at the position where the same production resource overlaps. Add the continuous overlap lengths on the front and back sides to obtain the occupation conflict amount of the candidate resource combination. The candidate resource combinations are moved in ascending order of their placeholder conflict levels. Each move re-forms the preceding, current, and following occupied segments until the placeholder conflict level is zero. The corresponding insertion position is then determined as the valid arrangement position of the candidate resource combination, resulting in the resource combination arrangement. Specifically, the resource combination arrangement result is as follows: Starting with the candidate resource combination with the smallest placeholder conflict, its insertion position is moved one production position backward each time. The previous occupied segment, the current occupied segment, and the next occupied segment are reread after the movement. The continuous length of the current occupied segment and the previous and next occupied segments that use the same production resource and whose positions overlap is counted position by position. The overlap lengths are added to obtain the new placeholder conflict. If the placeholder conflict is still greater than 0, it continues to move one position backward. When the placeholder conflict drops to 0, the current insertion position is fixed, and the occupied position corresponding to the resource combination is written into the production task arrangement. The remaining candidate resource combinations are processed in the same way. After all the valid insertion of the pending processes is completed, the resource combinations and their occupied positions are summarized according to the production task arrangement order to obtain the resource combination arrangement result.

[0031] In this embodiment, generating dynamic FPC production resource scheduling results includes: The real-time production status is matched with the initial FPC production resource scheduling results for each task, and the production resource nodes and pending process nodes corresponding to the status changes are located to form a set of changed nodes. The connection path between the set of changing nodes and the corresponding process node to be executed is extracted along the production resource association graph. The connection path is input into the improved Node2Vec model to re-represent the structural association and replace the original production resource node representation data of the corresponding node to obtain the updated production resource node representation data. Based on the updated production resource node representation data, the resource acceptance relationship and resource combination for the corresponding processes to be executed are re-determined. The newly formed resource combination replaces the corresponding part in the initial FPC production resource scheduling result, generating a dynamic FPC production resource scheduling result. Specifically, the generation of the dynamic FPC production resource scheduling result is as follows: Read the updated production resource node representation data, multiply and accumulate the results of the same vector positions, and then divide by the product of the square root of the sum of the squares of the two node vectors to obtain the degree of association between the affected pending processes and each available production resource. Based on the updated degree of association, the end position of resource occupation and the completion position of the preceding process, redetermine the resource acceptance relationship and combine them to form a new candidate resource combination. Redetermine the production start position, production completion position and process waiting length of each combination. Select the resource combination with 0 occupation conflict to replace the original resource combination of the corresponding pending process in the initial FPC production resource scheduling result. At the same time, adjust the execution position of the affected subsequent processes in sequence according to the new production completion position, keep the completed processes and unaffected processes unchanged, and generate the dynamic FPC production resource scheduling result after all affected processes have been replaced.

[0032] This includes updating production resource scheduling, including: Execute each production task according to the dynamic FPC production resource scheduling results, record the actual production start position, production completion position, resource occupation position and process waiting length, and match the actual execution results with the dynamic FPC production resource scheduling results one by one to obtain production execution deviation data. Update the status and node connection relationships of the corresponding nodes in the production resource association diagram based on the production execution deviation data, replace the original planned completion position with the actual completion position, and write the actual resource occupancy status into the corresponding resource node to obtain the updated production resource association diagram; Based on the updated production resource association graph, the production resource node representation data, resource acceptance relationships, and resource combinations of processes that have not yet been executed are redefined, and the unexecuted parts in the dynamic FPC production resource scheduling results are replaced to complete the production resource scheduling update.

[0033] Example 1: To verify the feasibility of this invention in practice, it was applied to a continuous FPC production scheduling cycle. Multimodal data was constructed, including orders, processes, equipment, materials, personnel, quality inspection, and production progress, totaling 1680 production records. Of these, 1460 were pending processes, 1320 were equipment status records, 1280 were material status records, and 1240 were personnel operation records. Of the original data, 1368 records showed directly identifiable object relationships, while the remaining 312 records exhibited unclear cross-time period relationships and asynchronous resource statuses. These records were used to simulate production scenarios with insufficient multimodal data association and difficulties in dynamic resource matching.

[0034] For 1680 production records, a unified time base, object number, and data format were established. Duplicate records were merged, and missing and abnormal locations were marked, resulting in 1652 standard FPC production records. Based on start and end positions, 1280 consecutive production periods were defined, and an object occurrence layer matrix was created. In a group of adjacent periods, the preceding period contained 69 production objects, and the following period contained 69 production objects. After checking order attribution, process sequence, and start and end positions, 46 groups of continuous execution relationships were identified and written into the object inheritance layer matrix.

[0035] After concatenating the object occurrence layer matrix and the object inheritance layer matrix, the adjacent matrix columns are synchronously rotated cyclically. A certain column to be matched contains 18 rotatable positions, generating 18 candidate rotation columns. At the original position, the number of matching object numbers is 9, the number of matching inheritance numbers is 7, and the number of matching execution relationships is 8. After rotating 4 positions, these numbers increase to 14, 12, and 13 respectively, thus identifying it as the target matching column. After processing all time periods, the production resource association graph contains 1548 valid nodes and 1826 node connections, of which 418 connections are confirmed by minimum rotation matching.

[0036] The model was trained using 1600 sets of historical production resource graph samples and 1200 sets of independent test samples. Each node was assigned a 128-dimensional vector, with initial values ​​for each dimension between -0.01 and 0.01. Ten walks were performed from each node, recording 40 nodes each time, with the five positions before and after the central node serving as the context range. The training batch size was set to 256, the initial learning rate was 0.001, and the maximum training duration was 100 epochs. Training ended when the average training error change was below 0.00001 for five consecutive epochs.

[0037] Taking one training sample as an example, the current process node has 5 candidate nodes, with the number of level 1 to level 4 connections being "12, 31, 58, 86", "9, 27, 49, 72", "11, 24, 39, 55", "7, 25, 52, 81", and "10, 29, 61, 94", respectively. After forward and reverse calculation of the combination number and removal of direct connections, the 5th candidate node still retains 184 level 2 to level 4 recovery connections, while the 1st candidate node retains 151. Therefore, the 5th candidate node is selected first, and processing continues until a production resource walk sequence containing 40 nodes is formed.

[0038] In the window factorization process, a central node has 10 context nodes before and after it, with a common occurrence frequency ranging from 18 to 47. Based on positional intervals and common occurrence frequencies, 10 node factors are established and continuously expanded. After removing the 3rd context node, the total change is 126; after removing the 7th node, it is 73. Simultaneously, the common change after removing the 3rd and 5th nodes is 168. Therefore, the 3rd and 5th nodes are moved to earlier training positions. Subsequently, one central node and three context nodes are selected for four-node permutation multiplication-addition training. The output of a positive training group increases from 0.612 to 0.903, while the average output of five negative training groups decreases from 0.487 to 0.094.

[0039] After training, 1200 test samples were input into both the method of this invention and the traditional Node2Vec scheduling method. A certain process to be executed corresponds to three groups of candidate equipment resources, with correlation degrees of 0.86, 0.74, and 0.69, and processing intervals of 3, 0, and 7 production positions, respectively. After comparison, the first two groups of resources were retained, and these were combined with materials and personnel to form six candidate resource combinations. The position conflict rates for these six combinations were 8, 5, 3, 2, 0, and 0, respectively. Finally, the resource combination with 0 position conflict rate and the earliest production completion position was selected to form the initial schedule.

[0040] During production execution, the simulation of extended equipment occupancy and changes in material status resulted in alterations to the status of 1248 nodes, affecting 1316 pending processes. Based on the changed nodes, associated paths were re-extracted and node representation data updated. Resource acceptance relationships and candidate resource combinations were then redefined. Of these, 1186 affected processes underwent resource combination replacement, while 130 processes maintained their original scheduling because their original resources remained executable. This resulted in a dynamic FPC production resource scheduling outcome.

[0041] Under the same conditions of 1600 training samples and 1200 test samples, the method of this invention was compared with the traditional Node2Vec node representation combined with the earliest available resource arrangement method. The resource combination matching accuracy of the traditional method was 81.6%, while that of this invention was 92.8%; the resource occupation conflict rate was reduced from 14.7% to 5.9%; the average process waiting length was reduced from 18.4 production positions to 10.7; the success rate of resource re-matching after state change was increased from 76.3% to 91.2%; the continuous acceptance rate of production tasks was increased from 79.8% to 93.1%; and the average scheduling processing time of 1200 test samples was reduced from 1.46 seconds to 1.18 seconds.

[0042] The simulation results show that the present invention, starting from 1680 multimodal production records, forms a complete scheduling process through a double-layer splicing matrix, minimum rotation matching, improved Node2Vec representation, process resource acceptance, resource combination arrangement, and dynamic updates. It outperforms traditional methods in resource association identification, resource conflict control, process waiting time reduction, and resource re-matching after state changes, thus verifying the feasibility and effectiveness of the present invention for FPC production resource scheduling.

[0043] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for FPC production resource scheduling based on multimodal data analysis, characterized in that, include: Collect multimodal production data during the FPC production process, preprocess the multimodal production data, and generate standard FPC production data; Based on standard FPC production data, a two-layer splicing matrix is ​​constructed with production time period as the horizontal axis and object category as the vertical axis. Minimum rotation matching is performed on the two-layer splicing matrix to obtain the production resource association diagram. The production resource association graph is input into the improved Node2Vec model to perform deep representation learning on the structural associations between nodes in the production resource association graph, and the production resource node representation data is obtained. The degree of association between each process to be executed and available production resources is determined by the production resource node representation data, the resource acceptance relationship of each process to be executed is calculated, and process resource executable relationship data is generated. Based on the executable relationship data of process resources, resource combinations are formed for each process to be executed. The production start status, production completion status and process waiting status corresponding to different resource combinations are calculated to generate the initial FPC production resource scheduling result. Based on the initial FPC production resource scheduling results, continuously acquire changes in production status during the production execution process, update the production resource node representation data, and generate dynamic FPC production resource scheduling results; Production is executed according to the dynamic FPC production resource scheduling results, and the production execution results are written into the production resource association graph to update the production resource scheduling.

2. The FPC production resource scheduling method based on multimodal data analysis according to claim 1, characterized in that, The obtained production resource association diagram includes: Establish a horizontal time period sequence according to the production time period order in the standard FPC production data, and establish a vertical object sequence according to the production object category. Fill the corresponding matrix positions with the production objects that appear in the same production time period to generate an object occurrence layer matrix. Fill the corresponding object occurrence layer matrix positions with production objects that have a continuous execution relationship in adjacent production periods to generate the object inheritance layer matrix. Then, concatenate the object occurrence layer matrix and the object inheritance layer matrix according to the same production period and the same object category to generate a two-layer concatenation matrix. Minimum rotation matching is performed on the double-layer splicing matrix. The adjacent time period matrix columns in the double-layer splicing matrix are rotated cyclically. The number of consistent positions of objects of the same type, objects of different types, and objects of receiving objects before and after the rotation is compared one by one. The rotation result corresponding to the largest number of consistent positions is taken as the minimum rotation matching result. Production object nodes and node connections are established based on the minimum rotation matching result to obtain the production resource association graph.

3. The FPC production resource scheduling method based on multimodal data analysis according to claim 2, characterized in that, The minimum rotation matching of the double-layer splicing matrix includes: The matrix column of the previous production period in the double-layer splicing matrix is ​​selected as the reference column, and the matrix column of the next production period is selected as the column to be matched. The object occurrence layer and the object receiving layer in the column to be matched are rotated cyclically position by position to generate multiple rotation candidate columns. Read the number of times each rotation candidate column matches the object number in the same object category position, the number of times the successor number matches in different object category positions, and the number of times the execution relationship matches in the corresponding positions of the upper and lower layers, to obtain the matching record for each rotation candidate column; All matching records are compared sequentially based on the number of matching object IDs, the number of matching successor IDs, and the number of matching execution relationships. The first rotation candidate column that reaches the maximum consistency result is determined as the target matching column. Production object node connections are generated based on the corresponding positions between the baseline column and the target matching column.

4. The FPC production resource scheduling method based on multimodal data analysis according to claim 1, characterized in that, The obtained production resource node representation data includes: An improved Node2Vec model is constructed. The original Node2Vec model includes random walk processing, context training order generation, and node vector training. A combination number forward and reverse recursion module is set at the random walk processing position of the original Node2Vec model, a window factor expansion module is set at the context training order generation position, and a four-node permutation multiplication and addition module is set at the node vector training position. The production resource association graph is input into the improved Node2Vec model. The nodes in the production resource association graph are used as the starting point of the walk. The multi-level connection relationship between the current node and the candidate node is processed by the combination number forward and reverse recursion module to determine the next walk node and generate the production resource walk sequence according to the node arrival order. The production resource walk sequence is input into the window factor expansion module, and the central node and context nodes are factored. The training order of the context corresponding to the central node is determined based on the factor expansion result. Input the central node and its corresponding context training order into the four-node permutation multiplication and addition module to perform permutation multiplication and addition training, and obtain the production resource node representation data.

5. The FPC production resource scheduling method based on multimodal data analysis according to claim 4, characterized in that, The generation of the production resource walk sequence includes: Starting from the current node, read the direct and multi-level connection relationships between each candidate node and the current node in the candidate node set to form a candidate node connection record. The candidate node connection records are processed in a forward recursive manner, and then the forward recursive processing results are processed in a reverse recovery manner to obtain the recovery results of each candidate node under the multi-level connection relationship. The direct connections between each candidate node are removed, and the removed connection records are recalculated in the forward direction and recovered in the reverse direction. The walking order of the candidate nodes is determined based on the difference between the recovery results before and after the removal, and the walking sequence of production resources is generated.

6. The FPC production resource scheduling method based on multimodal data analysis according to claim 4, characterized in that, The training of multiplying and adding all permutations includes: From the context training order, select one central node and three context nodes to form a four-node training group, and divide the node vectors of the four nodes into multiple coordinate groups according to four consecutive coordinates. For each coordinate group, a four-row, four-column numerical table is created. All permutations are formed by selecting one coordinate in each row and ensuring that the selected coordinate columns are not repeated. The four coordinates in each permutation are multiplied consecutively, and the results of all permutations are summed to obtain the four-node multiplication result. Based on the results of the four-node multiplication and addition, the vectors of each node in the four-node training group are updated sequentially. When updating the current coordinate, the row and column where the current coordinate is located are deleted. The multiplication and addition process is performed on all the permutations of the remaining three rows and three columns of values, and the update result of the current coordinate is determined. After completing the training of all coordinate groups, the production resource node representation data is obtained.

7. The FPC production resource scheduling method based on multimodal data analysis according to claim 1, characterized in that, The executable relationship data for the generated process resources includes: Locate the process nodes corresponding to each process to be executed according to the production resource association diagram, filter the production resource nodes that are currently available along the node connections of the process nodes, and read the production resource node representation data corresponding to the process nodes and production resource nodes. The node representation data of the process node and each production resource node are multiplied and accumulated according to the same vector position. The accumulated result is divided by the product of the square root of the square of each node representation data to obtain the degree of association between the process to be executed and each production resource. The starting point for the next process is determined by the completion position of the preceding process. The current occupancy end position of each production resource is sequentially connected to the starting point. The relationship between production resources that can continuously accept the next process is determined based on the degree of correlation. The production resource relationships corresponding to each process to be executed are matched according to the starting point of acceptance and the end position of resource occupation. Production resource relationships that do not have co-occupancy and meet the requirements for continuous execution of the process are retained, and process resource executable relationship data is generated.

8. The FPC production resource scheduling method based on multimodal data analysis according to claim 1, characterized in that, The generation of the initial FPC production resource scheduling result includes: According to the process to be executed, extract the production resources that can jointly undertake the current process from the process resource executable relationship data, and combine the corresponding production resources according to the same process to be executed to generate candidate resource combinations; The production start position is determined by the end position of each production resource in the candidate resource combination and the previous completion position of the process to be executed. The production start position is obtained by continuously superimposing the process execution length. The process waiting length is obtained by the interval between the production start position and the previous completion position. Candidate resource combinations corresponding to the same process to be executed are embedded into the current production task in sequence according to the production completion position. The resource occupation positions of the production tasks before and after embedding are checked one by one. Candidate resource combinations that are occupied repeatedly are removed from their corresponding positions and then embedded into the next available position to obtain the resource combination arrangement result. Based on the sequential relationship of each pending process, the resource combination arrangement results are connected, and the production start position, production completion position, and process waiting length are written into the corresponding production task to generate the initial FPC production resource scheduling result.

9. The FPC production resource scheduling method based on multimodal data analysis according to claim 8, characterized in that, The obtained resource combination and arrangement results include: Using the production start position of the process to be executed as the insertion position, candidate resource combinations are inserted into the current production task arrangement respectively. The resource occupation positions corresponding to each production task before and after the insertion position are read to form the front occupation segment, current occupation segment and back occupation segment of each candidate resource combination. Cross-check the current occupied segment with the previous occupied segment and the next occupied segment position by position. Record the continuous overlap length at the position where the same production resource overlaps. Add the continuous overlap lengths on the front and back sides to obtain the occupation conflict amount of the candidate resource combination. The candidate resource combinations are moved in order of increasing position of the conflict, and the front, current and rear occupied segments are re-formed after each move until the conflict is zero. The corresponding insertion position is then determined as the effective arrangement position of the candidate resource combination, and the resource combination arrangement result is obtained.

10. The FPC production resource scheduling method based on multimodal data analysis according to claim 1, characterized in that, The generation of dynamic FPC production resource scheduling results includes: The real-time production status is matched with the initial FPC production resource scheduling results for each task, and the production resource nodes and pending process nodes corresponding to the status changes are located to form a set of changed nodes. The connection path between the set of changing nodes and the corresponding process node to be executed is extracted along the production resource association graph. The connection path is input into the improved Node2Vec model to re-represent the structural association and replace the original production resource node representation data of the corresponding node to obtain the updated production resource node representation data. Based on the updated production resource node representation data, the resource acceptance relationship and resource combination of the corresponding process to be executed are re-determined, and the newly formed resource combination replaces the corresponding part in the initial FPC production resource scheduling result to generate a dynamic FPC production resource scheduling result.

Images