A manufacturing task allocation method based on multi-agent cooperation

Abstract

The application discloses a kind of manufacturing task allocation methods based on multi-agent cooperation, it is related to intelligent manufacturing task scheduling technical field, including the following steps: S1, matching manufacturing task and manufacturing agent and binding detection result, form collaborative object;S2, convert detection result into quality deviation, calculate correctable boundary, form remaining closed field matrix;S3, establish task node and undertake edge, map source, target node embedding address, form manufacturing task graph;S4, access improved D2STGNN model, form dynamic transfer relationship;S5, spread deviation and according to un-closed residual error address cross-layer undertaking, form candidate address chain;S6, using vector weighted average algorithm optimization address pair, form optimized address combination;S7, reconstruct dynamic transfer relationship, form task allocation instruction;S8, execute instruction and according to new quality deviation update allocation result.The application improves task allocation accuracy, reduces rework risk.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing task scheduling technology, and in particular to a manufacturing task allocation method based on multi-agent collaboration. Background Technology

[0002] Existing manufacturing task allocation methods typically assign tasks to equipment or manufacturing agents with the necessary capabilities based on process routes, equipment capacity, processing time, and workload. While these methods can achieve basic process coordination and resource scheduling, the task allocation process often fails to adequately incorporate the inspection results of already executed tasks, making it difficult to reflect the actual quality deviations that occur in the workpiece during processing.

[0003] With the development of multi-agent collaborative technology, manufacturing tasks can be dynamically allocated among multiple manufacturing agents. However, existing methods mostly select manufacturing agents based on fixed node relationships or comprehensive evaluation results, failing to consider the process parameter adjustment capabilities of different equipment to determine whether quality deviations can be corrected in the remaining processes, and also making it difficult to maintain the continuous transmission of quality deviations during task acceptance.

[0004] When the test results change or the original assigned path can no longer correct the quality deviation, the existing method lacks a processing mechanism to redetermine the task assignment relationship based on the remaining correction capacity. This can easily lead to problems such as equipment having the process qualification but being unable to complete the quality correction, and the task assignment results not matching the actual processing status.

[0005] Therefore, how to provide a manufacturing task allocation method based on multi-agent collaboration is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a manufacturing task allocation method based on multi-agent collaboration. This invention employs a combination of an improved D2STGNN model and a vector weighted average algorithm to convert detection results into quality deviations and establish a residual closed domain matrix based on the quality correction capabilities of the manufacturing agents. By addressing source node embedding and target node embedding respectively, writing back unclosed residuals, and accepting target addresses across layers, a candidate address chain for manufacturing tasks is constructed. Then, vector merging and local search are performed using bound address pairs to determine the task acceptance order and the manufacturing agents, thereby realizing dynamic task allocation driven by quality deviations. This method has the advantages of high allocation accuracy, flexible task adjustment, and stable manufacturing quality.

[0007] A manufacturing task allocation method based on multi-agent cooperation according to an embodiment of the present invention includes the following steps:

[0008] S1. Match manufacturing tasks with manufacturing intelligent agents according to process relationships, bind task detection results, and form collaborative objects;

[0009] S2. Read the collaborative object, convert the detection result into quality deviation, calculate the correctable boundary and write it into the position of the unexecuted task to form the remaining closed domain matrix;

[0010] S3. Based on the remaining closed domain matrix, establish task nodes and connecting edges, and map the deviation position and boundary position to the source node embedding address and the target node embedding address, respectively, to form a manufacturing task graph.

[0011] S4. Integrate the manufacturing task graph into the improved D2STGNN model, read the node embeddings according to the source node embedding address and the target node embedding address, and form a dynamic transition relationship.

[0012] S5. Propagate the deviation along the dynamic transfer relationship, invert the unclosed residual as the correction amount and write it back. Stop accepting when the correction amount falls into the correctable boundary. If it does not fall into the boundary, convert the target address into the lower layer source address and advance to form a candidate address chain.

[0013] S6. Encode the candidate address chain into a vector weighted average algorithm solution vector, bind the target address and source address of adjacent layers into address pairs, and form an optimized address combination through vector merging and local search.

[0014] S7. Write the optimized address combination into the source address and the target address, retain the selected address to take over and clear the unselected address, rebuild the dynamic transfer relationship, and form a task allocation instruction;

[0015] S8. Execute the task allocation instruction, convert the detection result into a new quality deviation and update the remaining closed domain matrix. If there is no target address, the allocation will be terminated, and the task allocation result will be formed.

[0016] Optionally, S1 specifically includes:

[0017] S11. Read the manufacturing process route, arrange the manufacturing tasks according to the sequence of operations, and mark each manufacturing task according to its execution status to form a manufacturing task sequence;

[0018] S12. Read the range of executable processes of the equipment associated with the manufacturing agent, check the process requirements of the manufacturing tasks that have not been executed in the manufacturing task sequence one by one, associate the manufacturing agents that meet the process requirements with the corresponding manufacturing tasks, and form a task matching result.

[0019] S13. Associate the detection results with the corresponding executed manufacturing tasks in the manufacturing task sequence, and write the task matching results into the manufacturing task sequence to form a collaborative object.

[0020] Optionally, S2 specifically includes:

[0021] S21. Read the detection results and corresponding manufacturing task quality requirements within the collaborative object, unify the quality dimensions and units, subtract the target value from the detection value and retain the direction of the difference to form the quality deviation.

[0022] S22. Read the adjustable range of process parameters of the equipment associated with the unexecuted manufacturing task, multiply the endpoints of parameter changes by the corresponding process sensitivity, determine the upper and lower limits according to the product size, and accumulate them according to the quality dimension to form the quality correction range.

[0023] S23. Add the quality deviation to both ends of the quality correction range, take the intersection of the obtained range and the quality allowable range, subtract the quality deviation from both ends of the intersection to form the correctable boundary, and record the empty interval when there is no intersection.

[0024] S24. Establish a deviation matrix and a boundary matrix with the quality dimension as the rows and the unexecuted tasks as the columns. Write the quality deviation in the first column and the correctable boundary in each column to form the remaining closed domain matrix.

[0025] Optionally, S3 specifically includes:

[0026] S31. Establish task nodes according to the unexecuted manufacturing tasks corresponding to each column of the remaining closed domain matrix, and write the matching relationship of manufacturing agents in the collaborative object into the corresponding task nodes to form a set of task nodes.

[0027] S32. Read the manufacturing process sequence relationship in the collaborative object, and establish a directed connection edge between adjacent task nodes, from the previous manufacturing task to the next manufacturing task, to form a task topology.

[0028] S33. Map the first column of the deviation matrix to the source node embedding address, map the task node whose quality deviation inverted value falls into the correctable boundary to the target node embedding address, and combine the task topology to form a manufacturing task graph.

[0029] Optionally, the improved D2STGNN model includes a linear mapping layer, a dynamic graph learning module, and a decoupled spatiotemporal layer stacked sequentially.

[0030] Optionally, the formation of the dynamic transfer relationship specifically includes:

[0031] Read the manufacturing task diagram, extract node features according to the task node order, and perform linear mapping to form a node feature matrix;

[0032] Change the common node index of the source node embedding and the target node embedding to separate addressing, and read the source node embedding and the target node embedding according to the corresponding embedding address;

[0033] The corresponding rows of the node feature matrix are concatenated with the source node embedding and the target node embedding, respectively, to form the source dynamic feature matrix and the target dynamic feature matrix;

[0034] Perform query transformation and key transformation on the source dynamic feature matrix and the target dynamic feature matrix respectively, calculate the transpose product of the query result and the key result, and form the positive association score matrix and the negative association score matrix.

[0035] Write the directed receiving edges of the manufacturing task graph into the forward receiving matrix and transpose it to form the reverse receiving matrix. Multiply each of these matrices by the corresponding association score matrix that is normalized by row, and clear the rows corresponding to the target address to form a dynamic transfer relationship.

[0036] Optionally, forming the candidate address chain specifically includes:

[0037] The parameter paths of the source node embedding and the target node embedding are read synchronously according to the same task node index of the adjacent decoupled spatiotemporal layers, and the source node embedding reading address and the target node embedding reading address are set respectively.

[0038] The first decoupled spatiotemporal layer reads the source node embedding by the source node embedding address and the target node embedding by the target node embedding address, and propagates the quality deviation along the dynamic transfer relationship;

[0039] Subtract the diffusion backtracking result and the inherent backtracking result sequentially from the current layer input residual, and write the unclosed residual back to the original position of the quality deviation;

[0040] Invert the unclosed residuals to form a correction value. The process ends when the correction value falls into the corresponding correctable boundary. Otherwise, the target address of the current layer is written as the source address of the next layer.

[0041] Find the next unexecuted manufacturing task whose correction amount falls within the correctable boundary along the manufacturing process sequence, and write the corresponding address as the target address of the next layer.

[0042] The next decoupled spatiotemporal layer extracts the transfer score according to the receiving address and propagates the unclosed residual. If the next target address is not found, the residual is retained and the propagation is terminated. The receiving is repeated layer by layer to form a candidate address chain.

[0043] Optionally, S6 specifically includes:

[0044] S61. Encode the candidate address chain into a solution vector according to the quality dimension and the decoupling spatiotemporal layer order, and bind the target address and the next layer source address into an adjacent address pair.

[0045] S62. Calculate the weighted average of the solution vector using the vector weighted average algorithm, write the address component into the task address with the smallest absolute difference in sequence number and the first in order, and write adjacent address pairs into the same address to form an updated solution vector.

[0046] S63. When performing vector merging, adjacent address pairs are used as the smallest replacement unit. If the sum of the absolute values ​​of the residuals of the updated address pairs is less than that of the original address pairs, the updated address pairs are written; if they are not less than that, the original address pairs are retained to form a combined solution vector.

[0047] S64. Start the local search from the first address pair corresponding to the unclosed residual, move the target address to the next target address and rewrite the source address simultaneously to form the local search solution vector;

[0048] S65. If the next target address is not found, restore the address before the search, select the local search solution vector with the smallest sum of absolute residual values ​​and the first in order, and form an optimized address combination.

[0049] Optionally, S7 specifically includes:

[0050] S71. Write the optimized address combination into the source address and the target address according to the quality dimension, keeping the current layer target address bound to the next layer source address to form an optimized address;

[0051] S72. Read the optimized address, retain the selected address and clear the unselected address in the order of the decoupling spatiotemporal layer, and form the node embedding reading result;

[0052] S73. Based on the node embedding reading results, recalculate the forward correlation score matrix and the reverse correlation score matrix, normalize them by row, and multiply them element by element with the receiving edge matrix and the transpose matrix respectively to form a reconstructed dynamic transition relationship.

[0053] S74. Determine the order of unexecuted manufacturing tasks based on the optimized address combination, read the manufacturing intelligent agent associated with the task node, and form a task allocation instruction.

[0054] Optionally, S8 specifically includes:

[0055] S81. Send manufacturing tasks to the manufacturing agent corresponding to the task allocation instruction, control the associated equipment to complete the processing and collect the test results to form the current test results;

[0056] S82. Unify the quality dimensions and units of the current test results with the quality requirements of the manufacturing task, and form a new quality deviation by subtracting the target value from the test value and retaining the direction of the difference.

[0057] S83. Mark the column corresponding to the executed manufacturing task as completed and stop address reading. Write the new quality deviation into the first unexecuted task column. Check the correctable boundary column by column to form the updated remaining closed domain matrix.

[0058] S84. Compare the current test results with the manufacturing task quality requirements dimension by dimension. If the test value falls within the corresponding allowable range, the test is confirmed as completed. If it does not fall within the range, a result to be accepted is formed.

[0059] S85. Read the results to be accepted. If a target address exists, retain the corresponding manufacturing task and manufacturing agent and form a redistribution result. If no target address exists, terminate the allocation and form a termination result.

[0060] The beneficial effects of this invention are:

[0061] This invention links manufacturing tasks, manufacturing agents, and task detection results, incorporating quality deviations from executed tasks into the task allocation process, and establishing a remaining closed-domain matrix based on the quality correction range corresponding to unexecuted manufacturing tasks. Compared to existing methods that allocate tasks solely based on equipment capacity, processing time, or load conditions, this invention can select manufacturing agents from the perspective of whether quality deviations can be corrected in the remaining processes, improving the matching degree between task allocation results and actual processing status.

[0062] Furthermore, this invention improves the D2STGNN model by changing the source node embedding and target node embedding from reading a common node index to addressing them separately, and utilizes unclosed residuals to control the cross-layer acceptance of the target address. When quality deviations propagate between adjacent manufacturing tasks, the correspondence between residual positions, task addresses, and correction boundaries can be maintained, thereby continuously judging the correction capability of each unexecuted manufacturing task for the current deviation and reducing erroneous acceptance caused by fixed task relationships or static node connections.

[0063] This invention further binds adjacent layer target addresses and source addresses into address pairs using a vector weighted averaging algorithm, and maintains cross-layer address relationships during vector merging and local search processes to obtain optimized address combinations that meet quality correction requirements. Based on the optimization results, dynamic transfer relationships are reconstructed and task allocation instructions are formed, enabling timely adjustments to the task acceptance order and manufacturing agents when detection results change, thereby improving the accuracy, continuity, and manufacturing quality stability of multi-agent collaborative allocation. Attached Figure Description

[0064] 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:

[0065] Figure 1 This is an overall flowchart of a manufacturing task allocation method based on multi-agent collaboration proposed in this invention;

[0066] Figure 2 This is a schematic diagram of the structure of an improved D2STGNN model for a manufacturing task allocation method based on multi-agent collaboration proposed in this invention.

[0067] Figure 3 This is a schematic diagram of the unclosed residual addressing cross-layer assignment of a manufacturing task allocation method based on multi-agent collaboration proposed in this invention. Detailed Implementation

[0068] 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.

[0069] refer to Figures 1-3 A manufacturing task allocation method based on multi-agent collaboration includes the following steps:

[0070] S1. Match manufacturing tasks with manufacturing intelligent agents according to process relationships, bind task detection results, and form collaborative objects;

[0071] S2. Read the collaborative object, convert the detection result into quality deviation, calculate the correctable boundary and write it into the position of the unexecuted task to form the remaining closed domain matrix;

[0072] S3. Based on the remaining closed domain matrix, establish task nodes and connecting edges, and map the deviation position and boundary position to the source node embedding address and the target node embedding address, respectively, to form a manufacturing task graph.

[0073] S4. Integrate the manufacturing task graph into the improved D2STGNN model, read the node embeddings according to the source node embedding address and the target node embedding address, and form a dynamic transition relationship.

[0074] S5. Propagate the deviation along the dynamic transfer relationship, invert the unclosed residual as the correction amount and write it back. Stop accepting when the correction amount falls into the correctable boundary. If it does not fall into the boundary, convert the target address into the lower layer source address and advance to form a candidate address chain.

[0075] S6. Encode the candidate address chain into a vector weighted average algorithm solution vector, bind the target address and source address of adjacent layers into address pairs, and form an optimized address combination through vector merging and local search.

[0076] S7. Write the optimized address combination into the source address and the target address, retain the selected address to take over and clear the unselected address, rebuild the dynamic transfer relationship, and form a task allocation instruction;

[0077] S8. Execute the task allocation instruction, convert the detection result into a new quality deviation and update the remaining closed domain matrix. If there is no target address, the allocation will be terminated, and the task allocation result will be formed.

[0078] In this embodiment, S1 specifically includes:

[0079] The manufacturing process route is read from the process document in the manufacturing execution system. The process document records the process name, dependencies, and process sequence number. First, the task sequence is established according to the dependencies. Then, the tasks are arranged in ascending order of process sequence number. If the sequence numbers are the same, the task with the earlier planned start time is given priority, thus forming a manufacturing task sequence.

[0080] The execution status of manufacturing tasks is obtained from the equipment execution record and the manufacturing execution system reporting record. Tasks that have received a start signal but have not received a completion signal are marked as being executed. Tasks that have received a completion signal and have written an actual end time are marked as being executed. Tasks that have not received a start signal are marked as not being executed. In case of a status conflict, the simultaneous existence of an equipment completion signal and an actual end time is used as the condition for determining whether a task has been executed.

[0081] The relationship between manufacturing agents and equipment is determined by the equipment access configuration record. Each manufacturing agent is bound to a controllable device and reads the device's capability parameters. The device capability parameters include the types of executable processes, the range of processing objects, and the adjustable range of process parameters. The reading results form the range of executable processes for the manufacturing agent.

[0082] The process requirements for manufacturing tasks that have not been executed are read from the process documents corresponding to the manufacturing tasks. The process category, processing object range, and process parameter range are compared with the range of executable processes of the manufacturing agent. When all fall within the equipment's capability range, the association between the task and the manufacturing agent is established. If any item exceeds the range, no association is established.

[0083] When multiple manufacturing agents simultaneously meet the process requirements, all relationships are retained and candidate manufacturing agent sets are formed according to the manufacturing tasks. When no manufacturing agent meets the process requirements, the position of the manufacturing task in the original sequence is retained, the candidate manufacturing agent set is recorded as an empty set, and a task matching result containing empty matching results is formed.

[0084] The inspection results carry the workpiece identifier and the executed process identifier. The two identifiers are compared with the executed manufacturing tasks in the manufacturing task sequence one by one. When a unique matching record exists, the association is completed directly. When multiple matching records exist, the manufacturing task with the closest actual end time and earlier than the inspection time is selected. When no matching record is found, the inspection result is recorded as pending verification and writing stops.

[0085] The collaborative object is a set of records arranged in the sequence of manufacturing tasks. Each record stores the manufacturing task, execution status, candidate manufacturing agent set, and associated detection results. After all records are written, a collaborative object is formed, providing a data foundation with a clear order and consistent source for quality deviation calculation.

[0086] In this embodiment, S2 specifically includes:

[0087] The quality dimensions are extracted from the quality requirements carried by the manufacturing tasks within the collaborative object. The dimension names and order follow the order of the test items in the inspection specifications. When the test results are inconsistent with the units of the quality requirements, the conversion ratio in the calibration document of the inspection equipment is used to unify the units. Test results that lack units or conversion ratios are marked as invalid and stop participating in the calculation of quality deviation.

[0088] The quality target value and the acceptable inspection range are read from the corresponding inspection items in the process document. When the process document records the target value, the lower limit of acceptance, and the upper limit of acceptance at the same time, the target value is subtracted from the lower limit of acceptance and the target value is subtracted from the upper limit of acceptance, respectively, to obtain the lower boundary and the upper boundary of the quality allowable deviation. When the process document only records the upper and lower limits of acceptance, the arithmetic mean of the two ends is taken as the target value, and then the same subtraction is performed to form the quality allowable deviation range.

[0089] Quality deviations are calculated independently within each quality dimension. The target value is subtracted from the measured value after unification of dimensions. A positive value indicates that the measured value is in the direction of increasing target value, and a negative value indicates that the measured value is in the direction of decreasing target value. The calculation results are arranged in order of quality dimensions to form a quality deviation vector that corresponds one-to-one with the inspection item.

[0090] The adjustable range of process parameters is calculated using a single candidate manufacturing agent that has not performed a manufacturing task as the calculation object. The controller allowable range, tool compensation allowable range, and current machining safety range of the associated equipment are read respectively. The maximum value of the lower boundary and the minimum value of the upper boundary of the three closed intervals are taken as the effective range. The parameter decrease endpoint is obtained by subtracting the current parameter value from the lower boundary of the effective range, and the parameter increase endpoint is obtained by subtracting the current parameter value from the upper boundary of the effective range. When the lower boundary is greater than the upper boundary, the quality correction range corresponding to the current manufacturing agent is recorded as an empty interval.

[0091] Process sensitivity is calibrated separately for candidate manufacturing agents, process parameters, quality dimensions, and parameter increase / decrease directions. The parameter disturbance amount is taken as an integer multiple of the minimum executable increment of the associated equipment controller and kept within the effective range. The single sensitivity is obtained by subtracting the detection value before the disturbance from the detection value after the disturbance and then dividing by the parameter disturbance amount. In one embodiment, 30 samples are collected in each direction, the average sensitivity and standard deviation are calculated, and samples with an absolute value of difference from the average value exceeding 3 times the standard deviation are removed. The arithmetic mean of the retained samples is then taken. If there are fewer than 24 retained samples, the sensitivity of the current direction is recorded as invalid.

[0092] The parameter decreasing endpoint is multiplied by the decreasing direction process sensitivity, and the parameter increasing endpoint is multiplied by the increasing direction process sensitivity. The smaller of the two products is used as the lower limit of the quality correction for the current parameter, and the larger value is used as the upper limit of the quality correction. When the sensitivity in one direction is invalid, the correction interval is formed by multiplying 0 and the other direction. When the sensitivity in both directions is invalid, the current parameter does not participate in the accumulation of the quality change.

[0093] Multiple process parameters of the same candidate manufacturing agent are accumulated one by one according to the order of the equipment process parameter table. The lower limit is added to the lower limit and the upper limit is added to the upper limit. The current interval is updated after each parameter is completed. After all valid parameters are processed, the quality correction range of the current manufacturing agent under the corresponding quality dimension is formed. When there are no valid parameters, an empty interval is formed.

[0094] The lower and upper boundaries of the quality correction range are added to the current quality deviation to form the predicted deviation range after the current manufacturing task is executed. The larger value between the lower boundary of the predicted deviation and the lower boundary of the quality allowable deviation is taken as the lower boundary of the intersection, and the smaller value between the upper boundary of the predicted deviation and the upper boundary of the quality allowable deviation is taken as the upper boundary of the intersection. When the lower boundary of the intersection is greater than the upper boundary of the intersection, an empty interval is formed.

[0095] When the intersection is not empty, the current quality deviation is subtracted from the lower boundary and upper boundary of the intersection to form a correctable boundary record with the manufacturing agent identifier. When the same manufacturing task is associated with multiple candidate manufacturing agents, their respective boundary records are saved in ascending order according to the manufacturing agent identifier. Empty intervals are only marked with empty intervals and are not written to the numerical boundary.

[0096] The deviation matrix and boundary matrix are arranged with the quality dimension as the row and the manufacturing tasks that have not yet been executed when the matrix is ​​built as the column. The first unexecuted task column of the deviation matrix is ​​written with the corresponding quality deviation, and the remaining columns are written with 0. Each intersection position of the boundary matrix is ​​written with the set of correctable boundary records of the current manufacturing task in the corresponding quality dimension. The record set includes the manufacturing agent identifier, the corrected lower boundary, the corrected upper boundary, and the empty interval marker. The two types of matrices together form the remaining closed domain matrix.

[0097] In this embodiment, S3 specifically includes:

[0098] Each column of the remaining closed domain matrix corresponds one-to-one with an unexecuted manufacturing task in the manufacturing task sequence. The column order follows the process order in the manufacturing process route. When the task status changes to executed, the original column position is retained and a completion mark is written, without moving the positions of other columns.

[0099] Task nodes are generated from corresponding matrix columns. Each task node contains a manufacturing task identifier, process sequence, and a set of candidate manufacturing agents. The set of candidate manufacturing agents directly reads the task matching results from the collaborative object. The task nodes corresponding to the empty set are retained in the manufacturing task graph, but no target node embedding address is established.

[0100] The manufacturing process sequence is read from the manufacturing task sequence within the collaborative object. When the previous task can directly enter the next task after the previous task is completed, a directed connection edge is established from the previous task node to the next task node. When there are parallel processes, connection edges are established separately. When there are mutually exclusive processes, only the branches allowed to enter in the process file are retained.

[0101] The task topology consists of a set of task nodes and directed connecting edges. The node index uses the position number of the unexecuted manufacturing task in the manufacturing task sequence, and the edge index records the starting node number and the ending node number. Nodes without predecessor tasks are used as starting nodes, and nodes without next tasks that can be executed are used as ending nodes.

[0102] In the first column of the deviation matrix, the quality deviation position corresponding to each quality dimension is mapped to the source node embedding address. The source node embedding address is determined by the quality dimension row number and the column number of the first unexecuted manufacturing task. When the first column of quality deviation is empty, no corresponding source node embedding address is generated.

[0103] When there is at least one record at the intersection of the boundary matrix where the lower correction boundary is not greater than the upper correction boundary, it is determined that the current position has a non-empty correctable boundary. The quality deviation is multiplied by -1 to obtain the current correction amount. The correction amount is checked one by one to see if it is not less than the lower correction boundary and not greater than the upper correction boundary. When at least one record is hit, the corresponding task node index is written into the target node embedding address, and the manufacturing agent identifier of the hit record is written into the target candidate set. When all records are not hit, no target node embedding address is generated.

[0104] The manufacturing task graph is organized using a node table, a directed edge table, and an embedding address table. The node table stores task nodes and sets of candidate manufacturing agents, the edge table stores task acceptance directions, and the embedding address table stores the embedding addresses of source nodes and target nodes. These three types of data are linked through task node indexes to form a manufacturing task graph that can be read by the dynamic graph learning module.

[0105] In this embodiment, the improved D2STGNN model includes a linear mapping layer, a dynamic graph learning module, and a decoupled spatiotemporal layer stacked sequentially.

[0106] The original D2STGNN model first maps the node observations in continuous time periods to hidden features through a linear layer. The dynamic graph learning module generates a dynamic transition matrix between nodes. The decoupled spatiotemporal layers stacked in sequence separate the diffused signal from the intrinsic signal. After the prediction branches of each layer are summarized, the multi-step prediction results are output through the regression layer.

[0107] The original D2STGNN model reads source node embeddings and target node embeddings in pairs under the same node number. The dynamic graph learning module and estimation gate generate graph relationships and gate values ​​in a fixed node order. Each decoupled spatiotemporal layer continues to pass residuals along the same node order. In the manufacturing task scenario, the source task of quality deviation and the target task that undertakes the correction can be located in different nodes. Fixed co-position pairing cannot express the address connection between the two.

[0108] The improved D2STGNN model retains the linear mapping layer, dynamic graph learning module, and decoupled spatiotemporal layers stacked sequentially. The linear mapping layer connects the dynamic graph learning module and the first decoupled spatiotemporal layer. The dynamic transition relationship generated by the dynamic graph learning module connects each decoupled spatiotemporal layer. The unclosed residual output by the previous decoupled spatiotemporal layer is connected to the next decoupled spatiotemporal layer.

[0109] The decoupled spatiotemporal layer includes a decoupling block, a diffusion block, and an intrinsic block. The decoupling block includes an estimation gate and a residual decomposition structure. The diffusion block includes a spatiotemporal local convolutional layer, a prediction branch, and a backtracking branch. The intrinsic block includes a gated recurrent unit, a multi-head self-attention layer, a prediction branch, and a backtracking branch.

[0110] The linear mapping layer transforms the node features in the manufacturing task graph into a unified hidden dimension. The dynamic graph learning module calculates the forward and backward dynamic transition relationships. The decoupled spatiotemporal layer separates the deviation features that can propagate along the task nodes from the node's own change features, and retains the residuals that have not yet been closed.

[0111] The improvements are located at the node embedding reading end of the dynamic graph learning module and the residual propagation end of the adjacent decoupled spatiotemporal layer. The source node embedding and target node embedding are changed from reading the common node index to addressing them separately. The target node embedding reading address of the current layer is written as the source node embedding reading address of the next layer. Unclosed residuals are fixedly written back to the original quality dimension position.

[0112] Addressing the node pairing objects that change the dynamic transfer relationship separately, cross-layer address writing changes the embedded reading range of the next decoupled spatiotemporal layer, fixed residual write-back maintains the correspondence between quality deviation, correctable boundary and task address, and closing position stops generating new target node reading addresses.

[0113] Each detection result is written back and the remaining closed domain matrix is ​​updated to form a task status slice. The task status slices of the same workpiece are arranged in ascending order according to the actual completion time. Each quality dimension is encapsulated as a model sample. The model input adopts a four-dimensional floating-point tensor of batch, task status slice, task node and feature channel, so that different quality dimensions are calculated independently within the batch and maintain their respective dimensions.

[0114] Each task node has 5 feature channels, which are written sequentially to the current quality dimension's quality deviation, the activity-correctable lower boundary, the activity-correctable upper boundary, the task order normalization value, and the manufacturing agent's valid flag. The activity-correctable boundary is the manufacturing agent record that hits the current correction amount and has the largest boundary margin. If the boundary margins are the same, the record with the manufacturing agent's identifier in ascending order is taken. If no record is hit, the upper and lower boundary channels are written to 0 and the manufacturing agent's valid flag is written to 0.

[0115] In one implementation, the number of task state slices can be 8, the maximum number of task nodes can be 32, the model batch size can be 32, and the input size is 32×8×32×5. When there are fewer than 8 task state slices, zeros are padded at the beginning of the time axis, and when there are fewer than 32 task nodes, zeros are padded at the end of the node axis. The node mask size is 32×8×32×1, with 1 written in the valid position and 0 written in the zero-padding position. The zeros are broadcast and multiplied during the calculation of each feature channel.

[0116] The linear mapping layer multiplies each 5-dimensional node feature with a 5×32 weight matrix, accumulates the products item by item, superimposes a 32-dimensional bias, and then processes it with ReLU to form a 32×8×32×32 node feature tensor; the hidden dimension of 32 remains unchanged throughout all decoupled spatiotemporal layers.

[0117] The dynamic graph learning module generates a forward dynamic transition matrix and a reverse dynamic transition matrix for each quality dimension sample within a batch. The size of a single matrix is ​​the number of task nodes multiplied by the number of task nodes. The two types of matrices and the node feature tensor are input into the first decoupled spatiotemporal layer. The unclosed residual features between layers maintain the four dimensions of batch, task state slice, task node, and hidden channel unchanged.

[0118] Model parameters include linear mapping weights, source node embeddings, target node embeddings, task state slice position embeddings, quality dimension embeddings, query transformation parameters, key transformation parameters, estimation gate parameters, spatiotemporal local convolution kernels, gated loop parameters, self-attention parameters, prediction branch parameters, backtracking branch parameters, prediction readout parameters, and residual readout parameters; readout address, task topology, boundary record set, and node mask belong to the running index and do not participate in gradient updates.

[0119] Linear mapping weights, query transformation parameters, key transformation parameters, prediction branch weights, backtracking branch weights, prediction readout weights, and residual readout weights are initialized using a Xavier uniform distribution, and all biases are initialized to 0. Source node embeddings, target node embeddings, task state slice position embeddings, and quality dimension embeddings are initialized using a normal distribution with a mean of 0 and a standard deviation of 0.02. Gated loop parameters are orthogonally initialized, and all trainable parameters are updated through backpropagation of the loss.

[0120] Training data is extracted from the historical records of the manufacturing execution system, equipment controller, and quality inspection system, and is associated with workpiece identification, manufacturing task identification, and actual completion sequence. Each valid record simultaneously includes task node characteristics, manufacturing task diagram, quality deviation before inspection, quality deviation after inspection, and the actual execution equipment.

[0121] The training samples are constructed from the workpiece records that have completed all inspections. Records that lack process inspection results, have conflicting task sequences, or whose equipment identifiers cannot be associated are directly removed. The continuous task records of the same workpiece are divided into fixed-length windows according to the actual completion sequence. If the window length is insufficient, zeros are added to the end and a node mask is written.

[0122] Training labels are constructed separately for workpiece and quality dimensions. Each model sample saves only one quality dimension. The label value is the actual quality deviation after the corresponding manufacturing task is executed. The label tensor corresponds one-to-one with the task state slice and task node position. Positions without actual detection results are written with 0 and 0 is written in the label mask. Effective detection positions are written with 1 in the label mask.

[0123] The task succession relationship is obtained from the actual historical execution order. The next task that is executed directly after the previous task is completed is written into the directed succession edge. The source node embedding address is taken from the task node where the quality deviation was located before detection, and the target node embedding address is taken from the task node that actually undertakes the next deviation correction. The two addresses are used as reading indices for training samples to control the embedding rows and dynamic transition rows and columns, and are not used as regression labels for continuous value loss.

[0124] The training set, validation set, and test set are divided according to the workpiece identifier, and in one implementation, the proportions can be 70%, 15%, and 15%, respectively. Each continuous value channel is standardized using the mean and standard deviation of the training set. First, the mean is subtracted from the original value, and then the value is divided by the standard deviation. When the standard deviation is 0, the divisor is recorded as 1. The validation set, test set, and deployment input use the statistical values ​​of the training set.

[0125] The outputs of the diffusion prediction branch and the intrinsic prediction branch of each decoupled spatiotemporal layer are added one by one according to the level, and then mapped to the standardized quality deviation prediction value by 32×1 prediction readout weight and 1-dimensional bias. The prediction error is obtained by subtracting the prediction value from the actual label position by position and taking the absolute value, then multiplying it by the label mask, and finally dividing by the sum of the label mask values.

[0126] The unclosed residual features of each decoupled spatiotemporal layer are standardized into residual values ​​through residual readout parameters, and the masked mean absolute error is calculated with the actual quality deviation after the corresponding task is executed. In one embodiment, the training data loss can be the sum of 70% of the prediction error and 30% of the residual error, and samples with a sum of label mask values ​​of 0 are removed from the current batch.

[0127] The parameter regularization term is obtained by squaring and summing the elements of the linear mapping weight, query transformation weight, key transformation weight, convolution kernel, prediction branch weight, backtracking branch weight, prediction readout weight, and residual readout weight, and then multiplying it by the regularization coefficient and adding it to the training data loss to form the total training loss. In one embodiment, the regularization coefficient can be 0.00001, and the bias, node embedding readout address, and node mask are not included in the regularization term.

[0128] Training uses the Adam optimizer to update trainable parameters. In one implementation, the initial learning rate can be 0.001 and the batch size can be 32. When the L2 norm of all parameter gradients exceeds 5, each gradient component is multiplied by 5 and divided by the current L2 norm to make the scaled L2 norm equal to 5.

[0129] After each training round, the training data loss on the validation set is calculated. When the current error is lower than the historical minimum, the parameters are saved and the historical minimum is updated. If no new historical minimum is formed for 5 consecutive rounds, the learning rate is multiplied by 0.5 and is not lower than 0.00001. The relative improvement rate for a single round is calculated by subtracting the error of the current round from the historical minimum before the update and then dividing by the historical minimum before the update. When the relative improvement rate is lower than 0.1% for 15 consecutive rounds, training is stopped and the parameters corresponding to the historical minimum error are restored.

[0130] After the model completes training, the linear mapping parameters, dynamic graph learning parameters, estimation gate parameters, parameters of each decoupled spatiotemporal layer, prediction readout parameters, and residual readout parameters are fixed. The task graph is then created and encapsulated with floating-point node features, integer readout addresses, boundary record sets, and Boolean node masks. The model outputs dynamic transition relationships, unclosed residual features, unclosed residuals restored to their original mass units, and node embedding readout results, providing input for candidate address chain generation.

[0131] In this embodiment, forming a dynamic transfer relationship specifically includes:

[0132] When forming a dynamic transfer relationship, the original feature matrix of the node is extracted from the latest task state slice of the current quality dimension sample. Each row corresponds to an unexecuted manufacturing task, and the five columns correspond to the quality deviation, the activity correctable lower boundary, the activity correctable upper boundary, the task order normalization value, and the effective label of the manufacturing agent, respectively.

[0133] When the number of task nodes is N, the size of the original feature matrix of the node is N×5; the task order normalization value is obtained by subtracting 1 from the current task number and then dividing by the current maximum task number minus 1. When there is only 1 task node, write 0.

[0134] The linear mapping multiplies the original N×5 node feature matrix with the 5×32 mapping weight matrix, accumulates the product of each feature component and its corresponding weight row by row, and then superimposes a 32-dimensional bias and performs ReLU processing to form an N×32 node feature matrix.

[0135] The original D2STGNN model reads both source node embeddings and target node embeddings simultaneously at the entry point of the dynamic graph learning module using task node indices. The same task node position simultaneously serves as both the source end reading position and the target end reading position, making it impossible to write the task where the current quality deviation is located and another task responsible for quality correction into two separate embedding reading ends.

[0136] The rewriting settings for separate addressing are located between the source node embedding table, the target node embedding table, and the dynamic feature splicing entry. The common node index interface is split into the source node embedding address interface and the target node embedding address interface. The two addresses come from the source node embedding address and the target node embedding address in the manufacturing task graph, respectively.

[0137] The source node embedding address is first entered into the source node embedding table, and the embedding vector corresponding to the task where the current quality deviation is located is read. The target node embedding address is then entered into the target node embedding table, and the embedding vector corresponding to the manufacturing task whose correctable boundary covers the current correction amount is read. Invalid addresses do not trigger embedding reading.

[0138] Both the source node embedding table and the target node embedding table use an N×32 floating-point parameter matrix. The row number is consistent with the task node index, and the column direction stores 32 trainable embedding components. The embedding parameters are updated by backpropagation of the loss during the model training phase, and the parameters saved in the convergence round are read when the model is deployed.

[0139] For the current quality dimension sample, the source dynamic feature matrix places the 32-dimensional node features of the task node corresponding to the source address in the first 32 columns and the 32-dimensional source node embedding in the last 32 columns, forming an N×64 matrix, and writes 0 in the remaining rows; the target dynamic feature matrix writes the node features and target node embedding in the rows corresponding to each valid target address in the same column order, and writes 0 in the rows not pointed to by the target address.

[0140] The source dynamic feature matrix is ​​entered into the source query transformation and source key transformation respectively, and the target dynamic feature matrix is ​​entered into the target query transformation and target key transformation respectively. Each transformation uses a 64×32 weight matrix and a 32-dimensional bias. The input row vector is multiplied by the weight matrix and the bias is superimposed to form an N×32 query result and a key result.

[0141] The query transform parameters and key transform parameters read the parameters corresponding to the historical minimum validation error saved during the model training phase. During training, the gradient is backpropagated from the loss predicted by the quality deviation and the loss read out by the residual through the dynamic transfer relationship. After the model is deployed, the parameter values ​​are fixed and are not updated again during the task allocation process.

[0142] The positive association score matrix is ​​obtained by multiplying the transpose of the source query result and the target key result and then dividing by the square root of 32. The negative association score matrix is ​​obtained by multiplying the transpose of the target query result and the source key result and then dividing by the square root of 32. Each matrix intersection corresponds to an association score between a source task and a target task. Both matrices are of size N×N.

[0143] Before row-wise normalization, an address mask is first created. Positions formed by the source address and the valid destination address are written with 1, and all other positions are written with 0. In one implementation, positions with a mask value of 0 are written with -10. 9The original associated score is retained at the position with a mask value of 1. Then, exponential normalization is performed based on the maximum value of the current row. If there is no valid target address in the current row, the exponential calculation is skipped and the entire row is written to 0.

[0144] In the manufacturing task graph, the directed receiving edges are written into an N×N positive receiving matrix. When task node i can directly enter task node j, the corresponding position is written as 1, and the other positions are written as 0. The positive receiving matrix serves as the receiving edge matrix, and after transposition, it forms the reverse receiving matrix.

[0145] The normalized positive correlation score matrix is ​​multiplied position by position with the positive inheritance matrix, and the normalized negative correlation score matrix is ​​multiplied position by position with the negative inheritance matrix. The corresponding correlation score is cleared when the position of 0 in the inheritance matrix is ​​written, and the correlation score is retained when the position of 1 is written, thus forming the dynamic transfer relationship of the current quality dimension sample.

[0146] When no target node embedding address is generated for the current quality dimension sample, the entire row corresponding to the source address in the forward dynamic transition matrix and the reverse dynamic transition matrix is ​​cleared to zero; when the source address is also invalid, both matrices are cleared to zero, and the clearing result is directly input into the decoupled spatiotemporal layer and the propagation of the current quality dimension is stopped.

[0147] Compared to the original D2STGNN model, the improvement of this invention lies in the node embedding reading entry of the dynamic graph learning module. The common node index is split into independent source and target addresses, and the source task position and target task position are changed from a co-position binding relationship to a disjoint pairing relationship. This structural change enables the dynamic transfer relationship to be established between the current deviation task and another manufacturing task with correction capabilities, avoiding the common index from restricting the propagation to the same task node, and improving the correspondence between the task succession relationship and the actual quality correction capability.

[0148] In this embodiment, forming a candidate address chain specifically includes:

[0149] The original D2STGNN model first generates a dynamic transition matrix in a fixed node order by a dynamic graph learning module. Each decoupled spatiotemporal layer then receives the residual tensor of the previous layer according to the same node arrangement. The source node embedding and the target node embedding are paired according to the same node number during the graph relationship generation stage. The inter-layer residual itself does not carry source and target addresses that can be independently rewritten.

[0150] Fixed node arrangement is suitable for continuous signal prediction on the same group of nodes, but the manufacturing task requires the current quality deviation to be transferred from the source task to another task that undertakes the correction; when only fixed position residuals are transferred between layers, the current target task cannot become the new source location of deviation in the next layer.

[0151] The truncation point of the synchronous reading parameter path is set between the common node index output of the dynamic graph learning module and the embedded reading entry of each decoupled spatiotemporal layer. The synchronous broadcast connection of the common index to the next layer is truncated, and an independent source address register and target address register are set for each layer. The connection from the unclosed residual features of the previous layer to the input of the next layer remains unchanged. The training parameters of convolution, gated loop, self-attention, prediction branch and backtracking branch continue to use the convergence parameters.

[0152] Each decoupled spatiotemporal layer after truncation is configured with a source node embedding read address and a target node embedding read address. The two addresses are stored using integer indices. The address value corresponds to the task node row number in the manufacturing task graph. When there is no valid read position, -1 is written and the corresponding embedding read channel is closed.

[0153] The source node embedding and the target node embedding are stored in an N×H parameter matrix, where N corresponds to the number of task nodes and H corresponds to the dimension of the hidden features. In one implementation, H can be 32, and each valid address reads a 32-dimensional row vector from the parameter matrix.

[0154] The source node embedding address of the first decoupled spatiotemporal layer comes from the task node where the quality deviation is located, and the target node embedding address comes from the unexecuted task node where the inverse value of the quality deviation falls into the correctable boundary; each quality dimension stores a set of source addresses and target addresses to avoid different quality dimensions sharing the same task location.

[0155] The first decoupled spatiotemporal layer first reads the source node embedding according to the source node embedding address, and then reads the target node embedding one by one according to the target node embedding address; each quality dimension sample is read independently, and the corresponding embedding reading channel is closed when the address value is −1.

[0156] The estimated gate embeddings the current task state slice position, current quality dimension, source node, and target node embeddings sequentially to form a 128-dimensional vector. This vector is then multiplied by a 128×32 first gate weight matrix and a 32-dimensional bias is added. After ReLU processing, the vector is multiplied by a 32×1 second gate weight matrix and a 1-dimensional bias is added. Finally, the vector is compressed by Sigmoid to obtain a gate value between 0 and 1. The gate value is broadcast on the 32 hidden channels and multiplied element-wise with the residual features of the current layer input to form a diffuse block input.

[0157] Starting from the row corresponding to the source address, the diffusion block multiplies the gated residual features by the positive dynamic transfer scores from the source address to each target address and accumulates them at the target node position. Backpropagation reads the reverse dynamic transfer matrix and performs the same calculation. The two-way results are concatenated in the hidden channel direction. When the target address is empty, both the diffusion input and the two-way propagation results are written to 0.

[0158] The spatiotemporal local convolutional layer performs task state slice direction convolution and node diffusion result fusion on the splicing result. The convolution window covers the adjacent task state slice. The convolution kernel multiplies each position in the window one by one and then accumulates and superimposes the bias. The convolution result is processed by ReLU and then enters the prediction branch and backtracking branch of the diffusion block.

[0159] The diffusion backtracking result is mapped to the same four-dimensional size as the input residual of the current layer by the backtracking branch. Each output position is obtained by multiplying the convolution channel and the backtracking weight item by item and accumulating them. If a valid target address cannot be obtained, the quality dimension directly outputs a diffusion backtracking result of all zeros.

[0160] The residual decomposition structure first subtracts the diffusion backtracking result from the corresponding position of the input residual in the current layer to obtain the first residual residual; the subtraction is performed position by position using the same batch, same task state slice, same task node and same hidden channel, without performing cross-position mixing.

[0161] The intrinsic block reads the first remaining residual, the gated loop unit updates the hidden state of the node sequentially along the task state slice, the multi-head self-attention layer calculates the association result of the same task node between each state slice, and the intrinsic backtracking branch maps the association result to an intrinsic backtracking result with the same size as the first remaining residual.

[0162] The residual decomposition structure then subtracts the inherent backtracking result from the corresponding position of the first residual residual to obtain the unclosed residual; the calculation order is fixed as first subtracting the diffusion backtracking result, then subtracting the inherent backtracking result, and the two backtracking results do not replace each other and do not change the quality dimension index.

[0163] After obtaining the 32-dimensional unclosed residual features from the residual decomposition structure, the vector corresponding to the current source address is read, multiplied by the 32×1 residual readout weights, and a 1-dimensional bias is added to obtain the standardized residual value. Then, the standardized residual value is multiplied by the current quality dimension training standard deviation and the training mean is added to restore the unclosed residual with the same dimension as the quality deviation, and written into the residual state table with the quality dimension and the current source address as the keys.

[0164] The correction amount is obtained by multiplying the unclosed residual by -1. The manufacturing agent boundary records of the current target task are read one by one. When the correction amount is not less than the lower correction boundary and not greater than the upper correction boundary, the corresponding manufacturing agent identifier is written into the hit set. When the hit set is not empty, it is confirmed that the current quality dimension has been closed and the next layer source address and target address are written as -1.

[0165] When the hit set is empty, the open state is maintained. First, the target address of the current layer is copied to the source address of the next layer. Then, the target address of the next layer is searched from the unexecuted manufacturing tasks after the current task. When the current target address is -1, the copy is not performed and the current quality dimension propagation is terminated directly.

[0166] The next-level target address starts from the next process position of the current target task and reads the task nodes one by one along the manufacturing process route. When the task is in an unexecuted state, the candidate manufacturing agent set is not empty, at least one boundary record contains the current correction amount and the dynamic transfer score from the source address to the current task address is greater than 0, the hit task address at the beginning of the manufacturing process sequence is written to the next-level target address.

[0167] The next decoupled spatiotemporal layer reads the new source address and the new target address, extracts the corresponding cross score from the dynamic transfer relationship, multiplies the 32-dimensional unclosed residual feature at the current layer source address position with the cross score and writes it into the new source address position to form the input residual feature of the next layer; the unclosed residuals of the same dimension in the residual state table only participate in boundary judgment and address optimization, and are not directly written into the next layer as hidden feature tensors.

[0168] The next layer continues to perform diffusion backtracking deduction, inherent backtracking deduction, residual in-situ write-back, and correctable boundary judgment; after each layer is completed, only the source address, target address, unclosed residual, and closed state of the current quality dimension are updated, while other quality dimensions maintain their own independent inheritance process.

[0169] If no valid dynamic transfer score is found for the task node, the node is considered a miss and the next task node in the manufacturing process route is read. If all remaining tasks are misses, the next level target address is written as -1, the unclosed residual retains its current value and the propagation of this quality dimension is stopped.

[0170] If there are still unclosed residuals after all decoupling spatiotemporal layers have been used, the source address of the last layer, the unclosed residuals, and the miss state are retained, and the first decoupling spatiotemporal layer is no longer called in a loop; the results are retained and entered into the candidate address chain for further processing of address combination optimization.

[0171] The candidate address chain starts with each valid initial target address. When there are multiple initial target addresses in the same quality dimension, the source address is fixed and each initial target address is independently executed layer by layer to form multiple candidate address chains. The candidate address chain uses the quality dimension as the record line and saves the source address, target address, unclosed residual and address state in a decoupled spatiotemporal layer order. The location where the closure is completed is recorded with a termination mark, and the location without a target address is recorded with a miss mark.

[0172] Compared with the original D2STGNN model, the core improvement of this invention lies in the state inheritance structure between decoupled spatiotemporal layers. The original model passes residual tensors without positional semantics along fixed node indices, while this invention transforms inter-layer inheritance into an address-based recursive structure jointly controlled by source address, target address, unclosed residuals, and closed states.

[0173] This change means that model depth no longer represents only the number of repeated feature extractions, but corresponds to the continuous bearing depth of quality deviations in the remaining manufacturing tasks. Closed deviations exit the propagation in the current layer, while unclosed deviations carry the task position into the next layer. This eliminates the bearing breaks and invalid deep propagation caused by fixed node indexes from within the model, and ensures that each inter-layer recursion corresponds to a clear manufacturing task transfer relationship.

[0174] In this embodiment, S6 specifically includes:

[0175] The candidate address chain uses quality as the record row and decoupling spatiotemporal layer order as the record column. Each intersection position stores the current layer source address, current layer target address, unclosed residual, and address status. Empty positions after the termination mark do not participate in the devector encoding.

[0176] The solution vectors are first arranged in order of quality dimension, and then in order of decoupled spatiotemporal layers within the same quality dimension. Source address and target address are written to each layer. The target address of the current layer and the source address of the next layer are continuously written to adjacent components to form inseparable adjacent address pairs.

[0177] If the first-layer source address does not exist and is bound to the previous-layer target address, it is reserved separately as the starting component of the solution vector. If the last-layer target address does not exist and is bound to the next-layer source address, the component corresponding to the no-target-address is written with -1 and the remaining position encoding of that quality dimension is stopped.

[0178] The number of valid candidate address chains generated across all quality dimensions of the same workpiece is denoted as K. Each candidate address chain is encoded to form an equal-length solution vector. Shorter address chains are padded with -1 after the termination position. The padded component does not participate in the weighted average, vector merging, or residual evaluation. When K is 0, the output is a state without a target address. When K is 1, the corresponding solution vector weight is directly written as 1.

[0179] Before weight calculation, the unclosed residuals of each quality dimension are divided by the corresponding quality allowable deviation range width to obtain dimensionless residuals; the quality allowable deviation range width is the upper boundary of the quality allowable deviation minus the lower boundary of the quality allowable deviation. When the width is not greater than 0, the current quality dimension is marked as invalid and the address combination optimization is stopped.

[0180] The evaluation value of a single solution vector is obtained by summing the absolute values ​​of the dimensionless residuals of all valid positions. When the evaluation value is 0, it means that all quality dimensions involved in the evaluation have been closed. An increase in the evaluation value indicates that the degree of non-closure has increased.

[0181] Divide the minimum resolution of each participating testing device by the width of the corresponding permissible quality deviation range, and take the minimum value among all positive values ​​as the zero-prevention value; add the evaluation value of each solution vector to the zero-prevention value and take the reciprocal, then divide the single reciprocal by the sum of all reciprocals to obtain the corresponding weight, so that the sum of K weights equals 1.

[0182] Each valid address component is multiplied by the weight of its corresponding solution vector, and the products of the same component position are accumulated to form the address weighted average; when all component positions are −1, they remain −1 and do not participate in address mapping.

[0183] When the weighted average of addresses is a continuous value and cannot be directly used as an index for task nodes, the absolute difference between each address and the weighted average is calculated from the candidate task addresses of the same quality dimension; the address with the smallest absolute difference is written into the updated solution vector, and if the absolute differences are the same, the address with the earlier manufacturing process order is written into the updated solution vector.

[0184] Adjacent address pairs are updated synchronously. Once the target address is determined, the same address is written to both the target component of the current layer and the source component of the next layer. If a valid address cannot be obtained for either component, the entire adjacent address pair retains its original value to prevent the target address from losing its correspondence with the source address of the next layer.

[0185] Vector merging processes each pair of address pairs starting from the first adjacent address pair in the solution vector. The address pairs in the updated solution vector temporarily replace the corresponding address pairs in the original solution vector. The replacement range includes both the target address of the current layer and the source address of the next layer. It does not replace any component in the address pair individually.

[0186] After the address pair is temporarily replaced, dynamic transfer extraction and residual propagation are re-executed starting from the unclosed residual corresponding to the address pair, while the model parameters remain fixed; the recalculated results are written to the temporary residual record, without overwriting the residuals saved in the original solution vector.

[0187] The sum of the absolute values ​​of the residuals of the original address pairs is obtained by accumulating the absolute values ​​of the dimensionless residuals of the corresponding quality dimensions in the replacement position and the subsequent effective layer. The updated address pairs are calculated using the same layer range. If the updated result is less than the original result, it is written into the updated address pair as a whole; if it is not less than the original result, the original address pair is restored as a whole.

[0188] After all adjacent address pairs are compared, a combined solution vector is formed. The combined solution vector retains the source address, target address, and recalculated residual after each effective replacement. For positions where no effective replacement has occurred, the addresses and residuals in the original candidate address chain are retained.

[0189] The local search starts with the first unclosed adjacent address pair with a non-zero residual in the combined solution vector. The first position is determined by the order of the quality dimension and the order of the decoupling spatiotemporal layer. Address pairs that have already been closed are skipped and do not enter the search range.

[0190] The next target address is searched starting from the next manufacturing process of the current target task. The search scope is limited to task nodes in the manufacturing task graph that are still in an unexecuted state, have a non-empty set of candidate manufacturing agents, have non-empty correctable boundaries, and have valid dynamic transfer scores.

[0191] After finding the next target address, the target address of the current layer is moved to the task address, and the source address of the next layer is synchronously rewritten to the same address; after the address is rewritten, the unclosed residual is repropagated from the current layer to form a local search solution vector.

[0192] When there are multiple available target addresses for the same adjacent address pair, local search solution vectors are generated one by one along the manufacturing process sequence. Each time, only the current address pair and the lower layer address affected by it are changed; the unaffected front layer address, residual and closure state remain unchanged.

[0193] If the found task node cannot propagate the residual, or if no valid target address exists after recalculation, the current rewriting of the target address and source address is cancelled, and the address pairs, unclosed residuals, and lower-level address states saved before the search are restored.

[0194] After all candidate addresses for the current adjacent address pair have been searched, the next unclosed address pair will be processed until all unclosed positions have been searched. Address pairs without a next closable task will retain their original addresses and the search termination status will be recorded.

[0195] The evaluation value of each local search solution vector is calculated by accumulating the absolute values ​​of the same dimensionless residuals. The evaluation scope covers all quality dimensions and all effective decoupling spatiotemporal layers. The solution vector with the smallest evaluation value is determined as the optimal address combination. When the evaluation values ​​are the same, the solution vector with the earlier formation order is retained.

[0196] The optimized address combination saves the source address, target address, cross-layer binding relationship, and unclosed residuals after evaluation for each quality dimension; the saved results directly enter the source address and target address write-back process without changing the address order again.

[0197] In this embodiment, S7 specifically includes:

[0198] The optimized address combination is written to the address record area with the quality dimension as the row and the decoupled spatiotemporal hierarchical order as the column. The address record area shares the quality dimension row number with the remaining closed domain matrix. The source address and the target address respectively store the column index of the corresponding task node.

[0199] The address recording area does not cover the values ​​in the deviation matrix and boundary matrix. The source address points to the task column where the current quality deviation is located, and the target address points to the unexecuted task column that bears the correction amount. The address value is consistent with the position of the task column in the remaining closed domain matrix.

[0200] The current layer target address and the next layer source address are written using the same address unit copy. First, the current layer target address is written, and then the same task column index is copied to the next layer source address. If either write fails, the write to both positions is rolled back and the address before optimization is restored.

[0201] The node embedding read index is read layer by layer to optimize address combinations. The selected source and target addresses retain the corresponding task node indexes, and the unselected addresses are written with -1. The node embedding table is stopped at the position where -1 is written and the corresponding dynamic feature row is cleared.

[0202] After the index retention and clearing are completed, the source node embedding and the target node embedding are concatenated with the hidden features of the corresponding task nodes to form new source dynamic feature matrices and target dynamic feature matrices. The matrix rows still correspond to the task nodes, and the column direction keeps the hidden features first and the node embeddings second.

[0203] The positive association score is obtained by multiplying the source query result by the transpose of the target key result, and the negative association score is obtained by multiplying the target query result by the transpose of the source key result; the cumulative value of the product of each component of the corresponding vector is written at each intersection position.

[0204] For row-level normalization, first read all valid associated scores of the current row, subtract the maximum value of the current row from each score, calculate the index, and then divide the individual index value by the sum of the index values ​​of the current row; if there is no valid target address in the entire row, write 0 directly.

[0205] The directed receiving edge is stored in a square matrix with the number of task nodes multiplied by the number of task nodes. When the previous task can directly enter the next task, 1 is written at the intersection of the row of the previous task and the column of the next task. 0 is written at the position where the process sequence relationship is not satisfied.

[0206] The forward connection matrix is ​​transposed to form the reverse connection matrix. The normalized correlation score is multiplied position by position with the corresponding connection matrix. When the product is greater than 0 and both the source address and the destination address are valid, it is determined to be a non-zero connection, and the remaining positions are cleared to zero.

[0207] Non-zero connection locations and their corresponding products together form the dynamic transfer relationship for reconstruction. The product value serves as the connection weight for quality deviations to propagate from the source task to the target task. When there are multiple non-zero connections for the same source task, only the target task connection pointed to by the optimized address combination is retained.

[0208] The manufacturing task acceptance sequence starts from the first source address of the optimized address combination, and then reads the target address of the current layer and the source address of the next layer in sequence. When the two addresses are the same, a cross-layer acceptance is confirmed. Reading stops when a -1 address, a termination marker, or no non-zero connection is encountered.

[0209] Each target address corresponds to an unexecuted manufacturing task in the remaining closed domain matrix. For all unclosed quality dimensions pointing to the same target address in the optimized address combination, the manufacturing agent identifiers that hit the current correction amount are read from the boundary records. When only one quality dimension points to the current target address, a candidate manufacturing agent set is directly formed. When multiple quality dimensions point to the current target address, the intersection of the manufacturing agent identifier sets of each dimension is taken. When the intersection is empty, the current target address is cleared and the state of no common manufacturing agent is recorded.

[0210] For each manufacturing agent in the common candidate set, calculate the smaller value of the distance from the two ends of the corresponding boundary for each unclosed quality dimension correction, and then take the minimum value among all quality dimension results as the overall boundary margin of the current manufacturing agent; select the manufacturing agent with the largest overall boundary margin, and if the values ​​are the same, select the former in ascending order of the manufacturing agent identifier in the device access configuration record.

[0211] The task allocation instruction is written sequentially to the manufacturing task identifier, the selected manufacturing agent identifier, the acceptance order, the source address, the target address, the quality dimension involved in the correction, the corresponding correction amount, and the boundary record of the selected manufacturing agent; after all the contents are written, they are encapsulated into a task allocation instruction and the instruction issuance process begins.

[0212] In this embodiment, S8 specifically includes:

[0213] The task allocation instruction is issued by the manufacturing execution system through the manufacturing agent communication interface. The communication interface configuration table saves the confirmation time limit according to the manufacturing agent identifier and in milliseconds. In one embodiment, the confirmation time limit can be 3000ms. When the manufacturing agent returns a receipt confirmation containing the instruction identifier and verification result before the timer ends, it enters the pending execution state. If no confirmation is returned before the timer ends, the current instruction is cancelled, the target address is retained, the current manufacturing agent is marked as communication unavailable and removed from the common candidate set. When the common candidate set is empty, a termination result is formed.

[0214] The manufacturing agent identifies associated devices through the device access configuration record and converts the manufacturing task identifier into the processing program number approved in the process document. The device controller starts processing when it confirms that the program number exists and the program check code is consistent, the current workpiece identifier is consistent with the instruction workpiece identifier, and all safety interlock bits are 1. If any condition is not met, it returns to the execution failure state, retains the target address, deletes the current manufacturing agent from the common candidate set, and reselects the remaining manufacturing agents.

[0215] Once the current manufacturing task is completed, the equipment controller writes the processing completion signal and the actual end time. When the workpiece reaches the corresponding detection position, the manufacturing agent triggers the detection equipment to collect the data. The detection result carries the workpiece identifier, manufacturing task identifier, and collection time. Results with inconsistent identifiers are entered into the verification record and stop participating in the current matrix update.

[0216] The names and order of the quality dimensions in the current test results are consistent with the inspection specifications corresponding to the manufacturing task. The unit conversion is read from the conversion ratio in the calibration file of the testing equipment. The test value is converted into the unit used in the quality requirements after being multiplied by the conversion ratio. If the conversion ratio is missing or the quality dimensions cannot be matched, the current test result will be marked as invalid.

[0217] The new quality deviation is calculated separately for each quality dimension. The quality target value is subtracted from the measured value after unification of dimensions. A positive value indicates that the measured value is higher than the target value, and a negative value indicates that the measured value is lower than the target value. The calculation results are written into the new quality deviation vector in the order of the quality dimensions in the inspection specification.

[0218] The quality target value and the acceptable inspection range continue to read the inspection requirements in the corresponding process document of the manufacturing task; when the process document records the target value, the lower acceptable limit and the upper acceptable limit at the same time, it is read directly; when the process document only records the upper and lower acceptable limits, the arithmetic mean of the two ends is taken as the quality target value, and the quality target value is subtracted from the upper and lower acceptable limits respectively to update the quality allowable deviation range.

[0219] The remaining closed-domain matrix columns corresponding to the completed manufacturing task are written to the completion marker. The completion marker is stored together with the task column index. The matrix column is not deleted and the position of the remaining task columns is not changed. After the node embedding reading module finds the completion marker, it writes the corresponding source address and target address to -1 and stops accessing the embedding parameters of the task node.

[0220] New quality deviations are written into the first unexecuted task column in the manufacturing process sequence, with the original quality dimension row number remaining unchanged. If there are no unexecuted task columns, the new quality deviations are retained and directly entered into the quality allowable range comparison.

[0221] The correctable boundary is checked column by column along the manufacturing process sequence, starting from the column where the new quality deviation is written. For each quality dimension, the boundary record with the manufacturing agent identifier is read one by one. The new quality deviation is multiplied by -1 to obtain the correction amount. When the correction amount is not less than the lower boundary of the record and not greater than the upper boundary of the record, the current manufacturing agent identifier is written into the candidate set corresponding to the current quality dimension.

[0222] When the same task column corresponds to only one unclosed quality dimension, the effective manufacturing agent is determined by the candidate set of the current quality dimension; when the same task column corresponds to multiple unclosed quality dimensions, the intersection of the candidate sets of each quality dimension is taken. If the intersection is not empty, the current task column is retained and the target address is generated. If the intersection is empty, the current task column does not form a target address.

[0223] The current test value is compared with the acceptable range for each dimension. If the test value is not less than the lower acceptable limit and not greater than the upper acceptable limit, the current quality dimension is recorded as qualified. When all quality dimensions are qualified, a completed result is formed. If any quality dimension is unqualified, the corresponding new quality deviation is saved and a result to be accepted is formed.

[0224] When the target address exists for the receiving result, read the set of common candidate manufacturing agents corresponding to the target address, calculate the overall boundary margin for all manufacturing agents participating in the quality dimension correction, select the manufacturing agent with the largest overall boundary margin and the first one in ascending order of the identifier, retain the target task address and form the redistribution result.

[0225] When no valid target address is found in the received result, the unclosed quality dimension, new quality deviation, checked task column and no target address status are written to the termination record, the remaining source address and target address are cleared and the allocation is stopped; the termination record is output as the task allocation result for manual process review or rework.

[0226] Example 1:

[0227] To verify the feasibility of this invention in practice, it was applied to a precision bore machining unit for a hydraulic valve body. The valve body sequentially undergoes semi-finish boring, finish boring, honing, and final inspection. The finish boring and honing equipment are connected to a manufacturing intelligence system. The finished diameter of the valve body's main bore is controlled between 20.000 mm and 20.012 mm, the cylindricity is no greater than 0.006 mm, and the surface roughness Ra is no greater than 0.4 μm. The original allocation system schedules machining based on equipment idle time. After a valve body completes semi-finish boring, the online measuring instrument measures the main bore diameter as 19.973 mm and the cylindricity as 0.011 mm. At that time, the finish boring equipment B, which was idle, was affected by tool wear, resulting in insufficient remaining compensation. If it continues to be combined with the honing equipment D for machining, the main bore diameter can only reach 19.994 mm to 19.999 mm, which cannot enter the acceptable range for finished products. The original allocation method would still prioritize equipment B, easily leading to rework.

[0228] When applying this invention, the system associates the above detection results with the semi-finish boring task, subtracts the quality target value from the detected value to obtain the hole diameter deviation and cylindricity deviation, and reads the tool compensation range, spindle temperature rise correction value, and process sensitivity of the corresponding equipment of each manufacturing agent. The precision boring equipment A can currently enlarge the hole diameter by 0.020mm to 0.026mm, and the honing equipment C can further enlarge it by 0.006mm to 0.010mm, and reduce the cylindricity by 0.004mm to 0.007mm; the quality range after the two tasks are executed consecutively overlaps with the allowable range of the finished product. Based on this, the system establishes a residual closed domain matrix, assigns the semi-finish boring deviation position as the source node embedding address, and sets the task corresponding to precision boring equipment A as the target node embedding address. After precision boring, online retesting showed that the main hole diameter was 19.997 mm and the cylindricity was 0.007 mm. The unclosed residual still needs to be corrected by the honing task. The improved D2STGNN model takes the target address of the precision boring task as the source address of the next layer and writes the corresponding task of honing equipment C into the next target address.

[0229] The vector weighted average algorithm binds the target address corresponding to the precision boring equipment A with the source address of the next layer into an address pair. During vector merging and local search, the two addresses change synchronously, ultimately forming a task allocation instruction that is continuously received by the precision boring equipment A and the honing equipment C. After honing is completed, the final measured main hole diameter is 20.005 mm, cylindricity is 0.004 mm, and surface roughness Ra is 0.32 μm. All three results are within the acceptable quality range, and the system stops address reception and marks the valve body as completed.

[0230] To verify the stability of the method in continuous production, 720 valve bodies of the same model were selected and divided into three groups according to the batch of blanks, the degree of semi-finish boring deviation, and the equipment condition. Each method tested 240 pieces, and each group included 68 valve bodies that exceeded the inspection pass range in at least one quality dimension after semi-finish boring. The three groups used the same equipment, processing procedures, testing equipment, inspection pass range, and production shift, and the testing equipment was calibrated before testing.

[0231] The rule-based idle device allocation method selects the device that first enters the idle state from the devices that meet the process category and processing object range. The traditional D2STGNN allocation method uses the same training data, node features and training parameters as this invention, but retains the same node index reading method of source node embedding and target node embedding, and passes inter-layer residuals according to fixed task node positions, without performing address pair binding, vector merging and local search.

[0232] The initial deviation closure rate is calculated by dividing the number of valve bodies with all initial quality deviations that ultimately fall within the acceptable inspection range by 68. The final inspection first-pass rate is calculated by dividing the number of valve bodies that passed the first final inspection by 240. The number of erroneous task assignments is calculated by counting the number of instructions whose correction amount did not fall within the correctable boundary of the selected manufacturing agent when the task was issued. The average redistribution time is calculated by taking the arithmetic mean of the time difference between writing the inspection result into the manufacturing execution system and issuing the next task allocation instruction. The rework rate is calculated by dividing the number of valve bodies that enter the rework process by 240. The statistical results are shown in Table 1.

[0233] Table 1 Comparison of Hydraulic Valve Body Precision Hole Quality Deviation Closure and Task Allocation Effect

[0234] Initial deviation closure rate 75.0% 86.8% 95.6% Final inspection pass rate 90.8% 94.6% 97.9% Error task acceptance count 16 times 8 times 2 times Average redistribution time 2.46s 1.71s 1.13s Return rate 7.1% 3.8% 1.7%

[0235] As shown in Table 1, the method of this invention outperforms both the rule-based idle equipment allocation method and the traditional D2STGNN allocation method in five indicators: initial deviation closure rate, final inspection first-pass yield, number of erroneous task assignments, average redistribution time, and rework rate. The rule-based method has an initial deviation closure rate of 75.0%, a final inspection first-pass yield of 90.8%, 16 erroneous task assignments, an average redistribution time of 2.46 seconds, and a rework rate of 7.1%, indicating that assigning tasks solely based on equipment idle status makes it difficult to ensure that equipment correction capabilities match workpiece deviations. The traditional D2STGNN method increases the initial deviation closure rate to 86.8%, the final inspection first-pass yield to 94.6%, reduces the number of erroneous task assignments to 8, shortens the average redistribution time to 1.71 seconds, and reduces the rework rate to 3.8%, but cross-layer task assignment remains unstable. The initial deviation closure rate of the method of this invention reaches 95.6%, which is 20.6 percentage points higher than the rule-based method and 8.8 percentage points higher than the traditional D2STGNN method; the first-pass yield of the final inspection reaches 97.9%, the number of erroneous tasks is only 2, the average redistribution time is shortened to 1.13s, and the rework rate is reduced to 1.7%. This shows that the present invention can more accurately identify manufacturing tasks with correction capabilities and improve redistribution efficiency.

[0236] This embodiment uses the precision hole machining of hydraulic valve bodies as an application. It correlates the hole diameter deviation and cylindricity deviation obtained from online detection with the adjustable range of process parameters and process sensitivity of the equipment corresponding to the unexecuted manufacturing task, forming a correctable boundary and a residual closed domain matrix. The improved D2STGNN model changes the source node embedding and target node embedding from reading a common index to addressing them separately. Based on the unclosed residual, the target address is controlled to inherit the next layer's source address, allowing the unclosed quality deviation to be passed along the manufacturing process sequence to tasks with corrective capabilities. The vector weighted average algorithm binds the current layer's target address and the next layer's source address as adjacent address pairs, maintaining cross-layer correspondence during vector merging and local search, avoiding optimization results that disrupt task inheritance continuity. After machining, the system updates the residual closed domain matrix based on the new detection results and determines whether to continue or terminate the allocation based on the existence of the target address. This embodiment demonstrates that the present invention can continuously adjust the manufacturing task inheritance relationship and manufacturing agent allocation results based on the actual quality deviation and the remaining process correction capability, reducing the risk of incorrect allocation and rework caused by unclosed quality deviations.

[0237] The above are merely preferred embodiments 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 manufacturing task allocation method based on multi-agent collaboration, characterized in that, Includes the following steps: S1. Match manufacturing tasks with manufacturing intelligent agents according to process relationships, bind task detection results, and form collaborative objects; S2. Read the collaborative object, convert the detection result into quality deviation, calculate the correctable boundary and write it into the position of the unexecuted task to form the remaining closed domain matrix; S3. Based on the remaining closed domain matrix, establish task nodes and connecting edges, and map the deviation position and boundary position to the source node embedding address and the target node embedding address, respectively, to form a manufacturing task graph. S4. Integrate the manufacturing task graph into the improved D2STGNN model, read the node embeddings according to the source node embedding address and the target node embedding address, and form a dynamic transition relationship. S5. Propagate the deviation along the dynamic transfer relationship, invert the unclosed residual as the correction amount and write it back. Stop accepting when the correction amount falls into the correctable boundary. If it does not fall into the boundary, convert the target address into the lower layer source address and advance to form a candidate address chain. S6. Encode the candidate address chain into a vector weighted average algorithm solution vector, bind the target address and source address of adjacent layers into address pairs, and form an optimized address combination through vector merging and local search. S7. Write the optimized address combination into the source address and the target address, retain the selected address to take over and clear the unselected address, rebuild the dynamic transfer relationship, and form a task allocation instruction; S8. Execute the task allocation instruction, convert the detection result into a new quality deviation and update the remaining closed domain matrix. If there is no target address, the allocation will be terminated, and the task allocation result will be formed.

2. The manufacturing task allocation method based on multi-agent collaboration according to claim 1, characterized in that, S1 specifically includes: S11. Read the manufacturing process route, arrange the manufacturing tasks according to the sequence of operations, and mark each manufacturing task according to its execution status to form a manufacturing task sequence; S12. Read the range of executable processes of the equipment associated with the manufacturing agent, check the process requirements of the manufacturing tasks that have not been executed in the manufacturing task sequence one by one, associate the manufacturing agents that meet the process requirements with the corresponding manufacturing tasks, and form a task matching result. S13. Associate the detection results with the corresponding executed manufacturing tasks in the manufacturing task sequence, and write the task matching results into the manufacturing task sequence to form a collaborative object.

3. The manufacturing task allocation method based on multi-agent collaboration according to claim 1, characterized in that, S2 specifically includes: S21. Read the detection results and corresponding manufacturing task quality requirements within the collaborative object, unify the quality dimensions and units, subtract the target value from the detection value and retain the direction of the difference to form the quality deviation. S22. Read the adjustable range of process parameters of the equipment associated with the unexecuted manufacturing task, multiply the endpoints of parameter changes by the corresponding process sensitivity, determine the upper and lower limits according to the product size, and accumulate them according to the quality dimension to form the quality correction range. S23. Add the quality deviation to both ends of the quality correction range, take the intersection of the obtained range and the quality allowable range, subtract the quality deviation from both ends of the intersection to form the correctable boundary, and record the empty interval when there is no intersection. S24. Establish a deviation matrix and a boundary matrix with the quality dimension as the rows and the unexecuted tasks as the columns. Write the quality deviation in the first column and the correctable boundary in each column to form the remaining closed domain matrix.

4. The manufacturing task allocation method based on multi-agent collaboration according to claim 1, characterized in that, S3 specifically includes: S31. Establish task nodes according to the unexecuted manufacturing tasks corresponding to each column of the remaining closed domain matrix, and write the matching relationship of manufacturing agents in the collaborative object into the corresponding task nodes to form a set of task nodes. S32. Read the manufacturing process sequence relationship in the collaborative object, and establish a directed connection edge between adjacent task nodes, from the previous manufacturing task to the next manufacturing task, to form a task topology. S33. Map the first column of the deviation matrix to the source node embedding address, map the task node whose quality deviation inverted value falls into the correctable boundary to the target node embedding address, and combine the task topology to form a manufacturing task graph.

5. The manufacturing task allocation method based on multi-agent collaboration according to claim 1, characterized in that, The improved D2STGNN model includes a linear mapping layer, a dynamic graph learning module, and a decoupled spatiotemporal layer stacked sequentially.

6. The manufacturing task allocation method based on multi-agent collaboration according to claim 1, characterized in that, The formation of dynamic transfer relationships specifically includes: Read the manufacturing task diagram, extract node features according to the task node order, and perform linear mapping to form a node feature matrix; Change the common node index of the source node embedding and the target node embedding to separate addressing, and read the source node embedding and the target node embedding according to the corresponding embedding address; The corresponding rows of the node feature matrix are concatenated with the source node embedding and the target node embedding, respectively, to form the source dynamic feature matrix and the target dynamic feature matrix; Perform query transformation and key transformation on the source dynamic feature matrix and the target dynamic feature matrix respectively, calculate the transpose product of the query result and the key result, and form the positive association score matrix and the negative association score matrix. Write the directed receiving edges of the manufacturing task graph into the forward receiving matrix and transpose it to form the reverse receiving matrix. Multiply each of these matrices by the corresponding association score matrix that is normalized by row, and clear the rows corresponding to the target address to form a dynamic transfer relationship.

7. The manufacturing task allocation method based on multi-agent collaboration according to claim 1, characterized in that, The formation of the candidate address chain specifically includes: The parameter paths of the source node embedding and the target node embedding are read synchronously according to the same task node index of the adjacent decoupled spatiotemporal layers, and the source node embedding reading address and the target node embedding reading address are set respectively. The first decoupled spatiotemporal layer reads the source node embedding by the source node embedding address and the target node embedding by the target node embedding address, and propagates the quality deviation along the dynamic transfer relationship; Subtract the diffusion backtracking result and the inherent backtracking result sequentially from the current layer input residual, and write the unclosed residual back to the original position of the quality deviation; Invert the unclosed residuals to form a correction value. The process ends when the correction value falls into the corresponding correctable boundary. Otherwise, the target address of the current layer is written as the source address of the next layer. Find the next unexecuted manufacturing task whose correction amount falls within the correctable boundary along the manufacturing process sequence, and write the corresponding address as the target address of the next layer. The next decoupled spatiotemporal layer extracts the transfer score according to the receiving address and propagates the unclosed residual. If the next target address is not found, the residual is retained and the propagation is terminated. The receiving is repeated layer by layer to form a candidate address chain.

8. The manufacturing task allocation method based on multi-agent collaboration according to claim 1, characterized in that, S6 specifically includes: S61. Encode the candidate address chain into a solution vector according to the quality dimension and the decoupling spatiotemporal layer order, and bind the target address and the next layer source address into an adjacent address pair. S62. Calculate the weighted average of the solution vector using the vector weighted average algorithm, write the address component into the task address with the smallest absolute difference in sequence number and the first in order, and write adjacent address pairs into the same address to form an updated solution vector. S63. When performing vector merging, adjacent address pairs are used as the smallest replacement unit. If the sum of the absolute values ​​of the residuals of the updated address pairs is less than that of the original address pairs, the updated address pairs are written; if they are not less than that, the original address pairs are retained to form a combined solution vector. S64. Start the local search from the first address pair corresponding to the unclosed residual, move the target address to the next target address and rewrite the source address simultaneously to form the local search solution vector; S65. If the next target address is not found, restore the address before the search, select the local search solution vector with the smallest sum of absolute residual values ​​and the first in order, and form an optimized address combination.

9. A manufacturing task allocation method based on multi-agent collaboration according to claim 1, characterized in that, Specifically, S7 includes: S71. Write the optimized address combination into the source address and the target address according to the quality dimension, keeping the current layer target address bound to the next layer source address to form an optimized address; S72. Read the optimized address, retain the selected address and clear the unselected address in the order of the decoupling spatiotemporal layer, and form the node embedding reading result; S73. Based on the node embedding reading results, recalculate the forward correlation score matrix and the reverse correlation score matrix, normalize them by row, and multiply them element by element with the receiving edge matrix and the transpose matrix respectively to form a reconstructed dynamic transition relationship. S74. Determine the order of unexecuted manufacturing tasks based on the optimized address combination, read the manufacturing intelligent agent associated with the task node, and form a task allocation instruction.

10. A manufacturing task allocation method based on multi-agent collaboration according to claim 1, characterized in that, S8 specifically includes: S81. Send manufacturing tasks to the manufacturing agent corresponding to the task allocation instruction, control the associated equipment to complete the processing and collect the test results to form the current test results; S82. Unify the quality dimensions and units of the current test results with the quality requirements of the manufacturing task, and form a new quality deviation by subtracting the target value from the test value and retaining the direction of the difference. S83. Mark the column corresponding to the executed manufacturing task as completed and stop address reading. Write the new quality deviation into the first unexecuted task column. Check the correctable boundary column by column to form the updated remaining closed domain matrix. S84. Compare the current test results with the manufacturing task quality requirements dimension by dimension. If the test value falls within the corresponding allowable range, the test is confirmed as completed. If it does not fall within the range, a result to be accepted is formed. S85. Read the results to be accepted. If a target address exists, retain the corresponding manufacturing task and manufacturing agent and form a redistribution result. If no target address exists, terminate the allocation and form a termination result.

Images