A collaborative task distribution and computing power scheduling method of a supply chain edge node

Abstract

The application discloses a kind of supply chain edge node's cooperative task distribution and computing power scheduling method, comprising: collecting to be distributed task and edge node data, establishes scheduling state data;Filtering edge node satisfying resource demand, generating node context set;According to historical task execution record, establish estimation model group, generate estimated efficiency set and estimated duration set;According to node context set, estimated efficiency set and estimated duration set, execute improved LinUCB algorithm, determine target edge node and distribute task;According to estimated processing efficiency, temporarily update parameter matrix and benefit vector, and according to actual processing feedback, update benefit vector and estimation model group, in next scheduling period, call update result and output cooperative scheduling result.The application is continuously updated by estimation feedback and actual feedback, and the dynamic adjustment of task distribution parameter and node processing efficiency model is completed.

Description

Technical Field

[0001] This invention relates to the field of edge computing technology, and in particular to a method for collaborative task distribution and computing power scheduling of supply chain edge nodes. Background Technology

[0002] As supply chain operations evolve towards multi-node collaboration, real-time response, and regionalized processing, edge computing is being used to handle tasks such as order analysis, inventory verification, transportation route calculation, anomaly detection, and equipment status management. These tasks are characterized by dispersed data sources, varying computational requirements, different completion deadlines, and continuously changing network conditions. Current supply chain edge node scheduling methods typically begin by collecting data on task resource requirements, available node computing power, task queue length, communication bandwidth, and communication latency. Then, a candidate node set is formed based on screening criteria. Finally, execution nodes are selected through comprehensive scoring, priority ranking, or heuristic rules to achieve task distribution and computing resource allocation.

[0003] In the comprehensive scoring scheme, the system assigns fixed weights to available computing power, task queue length, communication bandwidth, and communication latency. After normalizing each indicator, it calculates the score of candidate nodes and determines the nodes with higher scores as task execution nodes. Some systems pre-set multiple weight templates based on task type, selecting the corresponding template to complete the scoring when a task arrives; others eliminate nodes that do not meet the conditions based on node load thresholds or communication latency thresholds, and then sort the remaining nodes.

[0004] In actual operation, edge tasks in the supply chain exhibit variations in task volume, resource requirements, and allowable latency. The same metric has inconsistent impacts on processing efficiency across different task types. Node availability, queuing status, and link conditions also change over the scheduling cycle, making it difficult to adjust the contribution of each metric based on actual processing feedback using fixed weights. When there is a delay in task processing feedback, the scheduling system cannot obtain the true processing efficiency in the current cycle, and updating scoring parameters based on lagging feedback can easily lead to a mismatch between task and node states. While preset weight templates can differentiate task types, they still rely on manual settings and cannot continuously reflect the changing relationship between the edge node's operating status and actual processing efficiency, resulting in deviations between the candidate node scoring criteria and the current task execution conditions.

[0005] Therefore, how to provide a collaborative task distribution and computing power scheduling method for supply chain edge nodes 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 collaborative task distribution and computing power scheduling method for supply chain edge nodes. This invention updates scheduling parameters collaboratively through estimated feedback and actual feedback, so that the contribution of each indicator is dynamically adjusted according to the task type and node operating status.

[0007] A collaborative task distribution and computing power scheduling method for supply chain edge nodes according to an embodiment of the present invention includes: Collect the task type, task attributes, resource requirements, task constraint parameters, and node status parameters of each edge node for the tasks to be distributed, and establish scheduling status data. Read historical task execution records, filter edge nodes that meet resource requirements based on scheduling status data, and combine the task attributes of the task to be distributed with the node status of the filtered edge nodes to obtain a set of node contexts; Establish a predictive model group based on historical task execution records; Input the node context set into the prediction model set to obtain the prediction efficiency set and the prediction duration set; The improved LinUCB algorithm is executed based on the node context set, the estimated efficiency set, and the estimated duration set to obtain the node score set. The target edge node is determined based on the node score set, the task distribution instruction is sent to the target edge node, and the LinUCB algorithm is updated and improved based on the target node context vector and the target node estimated efficiency to obtain the task execution record, temporary parameter matrix and temporary profit vector. Receive the actual processing feedback returned by the target edge node, calculate the actual processing efficiency and processing efficiency deviation based on the task execution record, and update the temporary revenue vector and the prediction model group based on the temporary parameter matrix, the actual processing efficiency and the processing efficiency deviation to obtain the updated parameter matrix, the updated revenue vector and the updated model group. Store the updated parameter matrix, updated benefit vector, and updated model group according to the task type and target edge node identifier. In the next scheduling cycle, call the updated parameter matrix, updated benefit vector, and updated model group to output the collaborative scheduling result.

[0008] Optionally, establishing the scheduling status data specifically involves: Collect the task type, task attributes, resource requirements, task processing volume, and allowed task latency of the tasks to be distributed. Write the task processing volume and allowed task latency into the task constraint parameters and associate them with the tasks to be distributed to obtain the task collection record. Collect the node status parameters of each edge node according to the edge node identifier, and associate the task collection record with each node status parameter to obtain the task node record. The node status parameters include node status, available computing power, task queue length, communication bandwidth and communication latency. Generate computing power constraint markers based on resource requirements and available computing power in the task node records; Generate a delay constraint flag based on communication bandwidth, communication latency, allowable task latency, and the amount of task data in the task attributes; By associating task node records, computing power constraint tags, and latency constraint tags according to edge node identifiers, scheduling status data is obtained.

[0009] Optionally, obtaining the node context set specifically involves: Extract edge node records that meet resource requirements from the scheduling status data, and associate the task attributes of the tasks to be distributed with the node status of the edge nodes to obtain candidate combination records; Based on the task type, extract the historical values ​​of each data indicator from the historical task execution records, and combine them with candidate combination records to determine the value range of each data indicator, thus obtaining the indicator range set; Each data indicator in the candidate combination record is matched with the indicator range set, and the value range of the corresponding data indicator is adjusted according to the matching result to obtain the updated range set. Based on the update range set, perform dimensionless mapping on the task processing volume, resource requirements, allowed task latency, available computing power, task queue length, communication bandwidth and communication latency respectively to obtain the initial mapping set; Extract the diagonal elements of the improved LinUCB parameter matrix corresponding to the task type, excluding the task type encoding dimension, and scale the data indicators in the initial mapping set according to the diagonal elements to obtain the matrix scaling set. Set the task type codes according to the order of task types, and concatenate the task type codes with the matrix scaling set to obtain the node context vector; The data location and vector dimension of each node's context vector are checked according to the edge node identifier, and the node context vectors are summarized to obtain the node context set.

[0010] Optionally, the step of establishing a prediction model group based on historical task execution records specifically includes: Historical task execution records are divided according to task type and edge node identifier, and then arranged to obtain a historical sample sequence. Historical node context vectors and historical actual processing efficiency are extracted from historical sample sequences, and a correspondence is established to obtain the recognition sample set; The forgetting coefficients are determined by the order of the identified sample set and the context vectors of adjacent historical nodes, thus obtaining the forgetting coefficient sequence. Based on the forgetting coefficient sequence, recursive least squares identification is performed on the identification sample set to obtain the efficiency parameter set; The efficiency parameter set is associated with the corresponding task type and edge node identifier. The historical node context vector is processed according to the efficiency parameter set, and the processing results are stored in correspondence with the historical actual processing efficiency to obtain the lag-free efficiency model. The feedback time length is determined based on the historical task completion time and the historical actual feedback time in the historical sample sequence, and the sequence offset is determined based on the feedback time length. The position of the hysteresis-free efficiency model is adjusted based on the sequence offset and correlated with the corresponding historical actual processing efficiency to obtain the pure hysteresis output model. Based on task type and edge node identifier, the lag-free efficiency model and the pure lag output model are combined to obtain the prediction model group.

[0011] Optionally, inputting the node context set into the prediction model group specifically involves: Based on the task type and edge node identifier, the context vector of each node in the node context set is matched with the predicted model group to obtain the model matching record; Based on the model matching record processing node context vector, the current efficiency prediction value is obtained; Based on the pure lag output model in the model matching record, process the current efficiency prediction value to obtain the delayed efficiency prediction value and the estimated feedback time; Based on the task type and edge node identifier, the most recent task that has been responded to is found from the historical task execution records and matched with the pure lag output model to obtain the feedback matching record; Calculate the feedback correction value based on the feedback matching record; The estimated processing efficiency is calculated based on the current efficiency prediction value and the feedback correction value, and the estimated processing efficiency is summarized according to the edge node identifier to obtain the estimated efficiency set. The estimated feedback duration is obtained by summarizing the edge node identifiers.

[0012] Optionally, the improved LinUCB algorithm is specifically as follows: According to the edge node identifier, the node context vector in the node context set is matched with the estimated processing efficiency in the estimated efficiency set and the estimated feedback time in the estimated time set to obtain the node input record. Extract the improved LinUCB parameter matrix, improved LinUCB reward vector, and basic exploration coefficient according to the task type, and determine the reward parameter vector; The base exploration coefficient is adjusted based on the estimated feedback duration and allowable task delay recorded in the input records of each node to obtain the adjusted exploration coefficient. Calculate the efficiency correction coefficient based on each estimated processing efficiency in the estimated efficiency set; Calculate the estimated node return based on the node context vector, return parameter vector, and efficiency correction coefficient; Calculate the node confidence increment based on the node context vector, the improved LinUCB parameter matrix, and the corrected exploration coefficients; The estimated return for each node is added to the corresponding node confidence increment to obtain the node selection score. The node selection scores are then aggregated according to the edge node identifier to obtain the node score set.

[0013] Optionally, obtaining the task execution record, temporary parameter matrix, and temporary profit vector specifically involves: The target edge nodes are determined based on the node score set, and the target node records are obtained. Extract the target node context vector and target node prediction efficiency from the target node record, and establish the target input record; Based on the target input record, send task distribution instructions to the target edge node and establish a task execution record; Calculate the temporary parameter matrix based on the target node context vector and the improved LinUCB parameter matrix corresponding to the task type; Calculate the temporary benefit vector based on the target node context vector, the target node estimated efficiency, and the improved LinUCB benefit vector for the corresponding task type. Optionally, obtaining the updated parameter matrix, the updated profit vector, and the updated model set specifically involves: Receive the actual processing feedback returned by the target edge node, and associate the actual processing feedback with the task execution record to obtain the feedback association record; The actual processing efficiency and actual feedback time are calculated by linking feedback records. Calculate the processing efficiency deviation based on the actual processing efficiency and the estimated efficiency of the target node; The updated profit vector is calculated based on the target node context vector, processing efficiency deviation, and temporary profit vector. Establish parameter update records based on the temporary parameter matrix and the updated payoff vector; The hysteresis-free efficiency model is updated based on the target node context vector, actual processing efficiency, and corresponding forgetting coefficient to obtain the updated efficiency model. The pure lag output model is updated based on the actual feedback duration and historical sample sequence to obtain the updated lag model. Based on the task type and the target edge node identifier, the update efficiency model and the update lag model are combined to obtain the update model group.

[0014] Optionally, the output of the cooperative scheduling result specifically includes: Based on the task type and target edge node identifier, the updated parameter matrix, updated profit vector, updated model group and actual feedback time are written into the same parameter record to obtain the updated parameter record; Establish a parameter record sequence based on the updated parameter records and existing parameter records in the historical task execution records; Upon entering the next scheduling cycle, update parameter records are extracted from the parameter record sequence based on the task type of the task to be distributed and the identifier of each edge node to obtain the cycle matching record; The timeliness coefficient is calculated based on the scheduling time of the next scheduling cycle, the actual feedback time in the cycle matching record, and the allowable delay of the task. Calculate the periodic parameter set based on the timeliness coefficient, periodic matching records, and the improved LinUCB parameter matrix and improved LinUCB benefit vector corresponding to the task type. The periodic update records are obtained by updating and improving the LinUCB algorithm and the prediction model group based on the periodic parameter group and the periodic matching records. The node context set for the next scheduling cycle is processed based on the periodic update records to obtain the collaborative scheduling result.

[0015] The beneficial effects of this invention are: (1) This invention performs dimensionless mapping of task attributes, resource requirements, task constraint parameters and node state parameters, and uses the diagonal elements of the improved LinUCB parameter matrix to scale each data index to form a node context set, so that the input features can be adjusted with the historical parameter state, reducing the impact of sudden changes in edge node state and differences in index values ​​on the node scoring process.

[0016] (2) This invention generates the estimated processing efficiency and the estimated feedback time through the lag-free efficiency model and the pure lag output model, and uses the deviation between the actual processing efficiency and the delayed efficiency prediction value to correct the current efficiency prediction value, so that the task processing efficiency prediction and feedback delay are separated from each other, providing a continuous efficiency basis for node selection before the actual processing feedback returns.

[0017] (3) The improved LinUCB algorithm in this invention corrects the exploration coefficient based on the estimated feedback time, corrects the node estimated return based on the estimated processing efficiency, and uses the estimated processing efficiency to temporarily update the benefit vector and the actual processing efficiency deviation to correct the temporary benefit vector. At the same time, it integrates the historical update parameters and the basic parameters based on the parameter timeliness coefficient, so that the contribution of each indicator can be dynamically adjusted according to the task type, node running status and feedback timeliness. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a collaborative task distribution and computing power scheduling method for supply chain edge nodes proposed in this invention; Figure 2The flowchart of the improved LinUCB algorithm for collaborative task distribution and computing power scheduling method of supply chain edge nodes proposed in this invention is shown. Figure 3 A comparison chart of task completion times for different scheduling methods. Detailed Implementation

[0019] Example 1: Reference Figures 1-3 A collaborative task distribution and computing power scheduling method for supply chain edge nodes, comprising: Collect the task type, task attributes, resource requirements, task constraint parameters, and node status parameters of each edge node for the tasks to be distributed, and establish scheduling status data. Read historical task execution records, filter edge nodes that meet resource requirements based on scheduling status data, and combine the task attributes of the task to be distributed with the node status of the filtered edge nodes to obtain a set of node contexts; Establish a predictive model group based on historical task execution records; Input the node context set into the prediction model set to obtain the prediction efficiency set and the prediction duration set; The improved LinUCB algorithm is executed based on the node context set, the estimated efficiency set, and the estimated duration set to obtain the node score set. The target edge node is determined based on the node score set, the task distribution instruction is sent to the target edge node, and the LinUCB algorithm is updated and improved based on the target node context vector and the target node estimated efficiency to obtain the task execution record, temporary parameter matrix and temporary profit vector. Receive the actual processing feedback returned by the target edge node, calculate the actual processing efficiency and processing efficiency deviation based on the task execution record, and update the temporary revenue vector and the prediction model group based on the temporary parameter matrix, the actual processing efficiency and the processing efficiency deviation to obtain the updated parameter matrix, the updated revenue vector and the updated model group. Store the updated parameter matrix, updated benefit vector, and updated model group according to the task type and target edge node identifier. In the next scheduling cycle, call the updated parameter matrix, updated benefit vector, and updated model group to output the collaborative scheduling result.

[0020] In this embodiment, establishing the scheduling status data specifically involves: Collect the task type, task attributes, resource requirements, task processing volume, and allowed task latency of the tasks to be distributed. Write the task processing volume and allowed task latency into the task constraint parameters and associate them with the tasks to be distributed to obtain the task collection record. In Example 1, a corresponding task collection record is established for each task to be distributed, and the task type, task attributes, resource requirements, task processing volume, and allowable task latency are written sequentially into the same task collection record. The task processing volume and allowable task latency are synchronously written into the task constraint parameter field, so that the task constraint parameters are in the same task collection record as the task type, task attributes, and resource requirements of the same task to be distributed. After writing is completed, the contents of the task collection record are checked item by item according to the task to be distributed. If the contents are complete, the task collection record is saved. If there are missing contents, the corresponding contents are added and the records are checked again. Collect the node status parameters of each edge node according to the edge node identifier, and associate the task collection record with each node status parameter to obtain the task node record. The node status parameters include node status, available computing power, task queue length, communication bandwidth and communication latency. In Example 1, node status records are established according to edge node identifiers, and node status, available computing power, task queue length, communication bandwidth, and communication latency are written into the corresponding node status records. The task acquisition records corresponding to the same task to be distributed are copied and appended to each node status record one by one according to the edge node identifiers, so that each edge node identifier corresponds to a set of task acquisition records and node status parameters. The correspondence between the edge node identifiers and node status parameters in each set of records is checked. After the matching is completed, the records are saved according to the edge node identifiers to obtain the task node records. Generate computing power constraint markers based on resource requirements and available computing power in the task node records; In Example 1, resource requirements are compared with the corresponding available computing power item by item. When all resource requirements are not greater than the corresponding available computing power, the corresponding edge node is recorded as having sufficient computing power. When any resource requirement is greater than the corresponding available computing power, the corresponding edge node is recorded as having insufficient computing power. Generate a delay constraint flag based on communication bandwidth, communication latency, allowable task latency, and the amount of task data in the task attributes; In Example 1, the task data volume in the task attribute is divided by the communication bandwidth to obtain the task transmission duration. The task transmission duration is added to the communication delay to obtain the expected communication duration. When the expected communication duration is not greater than the allowed delay of the task, the corresponding edge node is recorded as the delay satisfaction state. When the expected communication duration is greater than the allowed delay of the task, the corresponding edge node is recorded as the delay non-satisfaction state. The task node record, computing power constraint mark, and latency constraint mark are associated according to the edge node identifier to obtain scheduling status data. The association is to retain the edge node record where both the computing power constraint mark and the latency constraint mark are satisfied, and the retained edge node record is used as the basis for screening edge nodes that meet resource requirements.

[0021] In this embodiment, obtaining the node context set specifically refers to: Extract edge node records that meet resource requirements from the scheduling status data, and associate the task attributes of the tasks to be distributed with the node status of the edge nodes to obtain candidate combination records; In Example 1, candidate combination records are established according to edge node identifiers. The task type, task processing volume, resource requirements and allowable latency of the same task to be distributed are written into each candidate combination record. The available computing power, task queue length, communication bandwidth and communication latency of each edge node are written into the candidate combination record with the same edge node identifier, so that each candidate combination record corresponds to one edge node. Based on the task type, extract the historical values ​​of each data indicator from the historical task execution records, and combine them with candidate combination records to determine the value range of each data indicator, thus obtaining the indicator range set; In Example 1, when there are historical values ​​for the corresponding task type, the minimum and maximum values ​​in the historical values ​​are determined as the value range of the corresponding data indicator; when there are no historical values ​​for the corresponding task type, the minimum and maximum values ​​of each data indicator in the candidate combination record are determined as the value range of the corresponding data indicator. Each data indicator in the candidate combination record is matched with the indicator range set, and the value range of the corresponding data indicator is adjusted according to the matching result to obtain the updated range set. In Example 1, the current data indicator in the candidate combination record is compared with the minimum and maximum values ​​of similar data indicators in the indicator range set. When the current data indicator is less than the minimum value of similar data indicators, the current data indicator is determined as the updated minimum value while the maximum value of similar data indicators is maintained. When the current data indicator is greater than the maximum value of similar data indicators, the current data indicator is determined as the updated maximum value while the minimum value of similar data indicators is maintained. When the current data indicator is between the minimum and maximum values ​​of similar data indicators, the minimum and maximum values ​​of similar data indicators are maintained. Based on the update range set, perform dimensionless mapping on the task processing volume, resource requirements, allowed task latency, available computing power, task queue length, communication bandwidth and communication latency respectively to obtain the initial mapping set; In Example 1, for each type of data indicator, the difference between the current data indicator and the minimum updated value of the same type of data indicator in the update range set is calculated; when the maximum updated value of the same type of data indicator is greater than the minimum updated value, the difference is divided by the difference between the maximum updated value and the minimum updated value of the same type of data indicator; when the maximum updated value of the same type of data indicator is equal to the minimum updated value, the mapping value of the current data indicator is determined to be 0. Extract the diagonal elements from the improved LinUCB parameter matrix corresponding to the task type, excluding the task type encoding dimension. Scale the data indicators in the initial mapping set according to each diagonal element to obtain a matrix scaling set. The scaling involves matching each diagonal element with each data indicator in a fixed splicing order, adding 1 to each diagonal element and taking the square root, and then dividing the corresponding data indicator by the square root result. The task type codes are set according to the order of task types, and the task type codes and matrix scaling sets are concatenated to obtain the node context vector. The concatenation is performed in a fixed order of task type code, task processing volume, resource requirements, allowable task latency, available computing power, task queue length, communication bandwidth and communication latency. The data position and vector dimension of each node's context vector are checked according to the edge node identifier, and the node context vectors are summarized to obtain a node context set. The summarization is to merge the node context vectors with the same data position and vector dimension.

[0022] In this embodiment, the step of establishing a prediction model group based on historical task execution records specifically includes: The historical task execution records are divided according to task type and edge node identifier, and then arranged according to the chronological order of the completion time of the historical tasks to obtain the historical sample sequence. In Example 1, record groups are established according to the task type and edge node identifier in the historical task execution records. Historical task execution records with the same task type and the same edge node identifier are grouped into the same record group, and historical task execution records with different task types or different edge node identifiers are grouped into different record groups. The historical task execution records in each record group are arranged in chronological order of the completion time of the historical tasks to obtain the historical sample sequence. Historical node context vectors and historical actual processing efficiency are extracted from historical sample sequences, and a correspondence between historical node context vectors and historical actual processing efficiency is established to obtain the recognition sample set. In Example 1, historical node context vectors and historical actual processing efficiency are extracted from the historical task execution records corresponding to each historical sample sequence. A one-to-one correspondence is established according to the same historical task execution record. Each historical node context vector and the historical actual processing efficiency in the same historical task execution record are combined to form an identification sample. The identification samples are arranged in chronological order of the completion time of the historical tasks to obtain the identification sample set. The forgetting coefficients are determined by the order of the identified sample set and the context vectors of adjacent historical nodes, thus obtaining the forgetting coefficient sequence. In Example 1, the dimensions corresponding to communication bandwidth and task queue length are extracted from the context vectors of adjacent historical nodes. The difference between adjacent mapping values ​​is calculated, and the larger of the two differences is limited to the range of 0 to 1. The difference between 1 and the limited result is determined as the forgetting coefficient. Based on the forgetting coefficient sequence, recursive least squares identification is performed on the identification sample set to obtain the efficiency parameter set; In Example 1, the previous parameter estimation result is adjusted according to the current forgetting coefficient according to the order of the identified sample set. The current efficiency prediction value is calculated using the adjusted parameter estimation result and the current historical node context vector. The difference between the current historical actual processing efficiency and the current efficiency prediction value is determined as the efficiency identification bias. The parameter estimation result is updated according to the efficiency identification bias and the current historical node context vector. The efficiency parameter set is associated with the corresponding task type and edge node identifier. The historical node context vector is processed according to the efficiency parameter set, and the processing results are stored in correspondence with the historical actual processing efficiency to obtain the lag-free efficiency model. In Example 1, the estimation result of the last parameter in the efficiency parameter group is determined as the efficiency parameter vector corresponding to the task type and edge node identifier. The efficiency parameter vector is then multiplied with the context vector of each historical node in the corresponding historical sample sequence to obtain the lag-free efficiency output sequence. The historical node context vector, the lag-free efficiency output, and the historical actual processing efficiency are stored in the order of the historical task completion time to obtain the lag-free efficiency model. The feedback time length is determined based on the historical task completion time and the historical actual feedback time in the historical sample sequence, and the sequence offset is determined based on the feedback time length. In Example 1, the first historical task completion time that is no earlier than the historical actual feedback time is found in the historical sample sequence, and the sequence position corresponding to the found historical task completion time is subtracted from the sequence position corresponding to the current historical task completion time. The position of the historical output of the no-hysteresis efficiency model is adjusted according to the sequence offset and correlated with the corresponding historical actual processing efficiency to obtain the pure-hysteresis output model. The position adjustment is to move the position of the historical output of the corresponding historical task execution record backward according to the sequence offset. Based on task type and edge node identifier, the lag-free efficiency model and the pure lag output model are combined to obtain the prediction model group.

[0023] In Example 1, model records are established according to task type and edge node identifier. The lag-free efficiency model and the pure lag output model with the same task type and edge node identifier are written into the same model record. The output end of the lag-free efficiency model is connected to the input end of the pure lag output model, so that the lag-free efficiency value output by the lag-free efficiency model enters the pure lag output model for sequence offset processing. The records of each model are then summarized.

[0024] In this embodiment, inputting the node context set into the prediction model group specifically means: Based on the task type and edge node identifier, the context vector of each node in the node context set is matched with the predicted model group to obtain the model matching record; In Example 1, a no-hysteresis efficiency model and a pure-hysteresis output model with the same task type and edge node identifier are found, and the node context vector, the no-hysteresis efficiency model, and the pure-hysteresis output model are written into the same record. Based on the model matching record processing node context vector, the current efficiency prediction value is obtained; In Example 1, the node context vector in the model matching record is input into the corresponding lag-free efficiency model. The efficiency parameter vector corresponding to the task type and edge node identifier in the lag-free efficiency model is extracted. The efficiency parameter vector and the node context vector are inner producted according to the same dimension position to obtain the current efficiency prediction value. Based on the pure lag output model in the model matching record, process the current efficiency prediction value to obtain the delayed efficiency prediction value and the estimated feedback time; In Example 1, the current efficiency prediction value is written into the output position corresponding to the current task in the corresponding pure delay output model. The sequence offset and feedback time length corresponding to the task type and edge node identifier in the pure delay output model are extracted. The current efficiency prediction value is moved from the output position corresponding to the current task to the target output position according to the sequence offset. The current efficiency prediction value corresponding to the target output position is determined as the delayed efficiency prediction value, and the feedback time length corresponding to the sequence offset is determined as the estimated feedback time. Based on the task type and edge node identifier, the most recent feedback task is found from the historical task execution record, and the most recent feedback task is matched with the pure delay output model to obtain the feedback matching record; In Example 1, task records that have returned actual processing feedback are filtered from historical task execution records according to task type and edge node identifier. The filtered task records are arranged in chronological order of the actual feedback time in history, and the task record with the latest actual feedback time in history is determined as the most recent feedback task. A pure delay output model with the same task type and edge node identifier as the most recent feedback task is found, and the target output position in the pure delay output model is located according to the sequence offset corresponding to the most recent feedback task. The delay efficiency prediction value corresponding to the target output position is extracted, and the actual processing efficiency, sequence offset, and delay efficiency prediction value of the most recent feedback task are written into the same record to obtain the feedback matching record. The feedback correction value is calculated based on the feedback matching record. The calculation of the feedback correction value is to subtract the corresponding delay efficiency prediction value from the actual processing efficiency of the most recent feedback task. When there is no most recent feedback task, the feedback correction value is determined to be 0. The estimated processing efficiency is calculated based on the current efficiency prediction value and the feedback correction value, and the estimated processing efficiency is summarized according to the edge node identifier to obtain the estimated efficiency set. The calculation of the estimated processing efficiency is to add the current efficiency prediction value and the feedback correction value. The estimated feedback time is summarized according to the edge node identifier to obtain the estimated time set. The summary is written according to the arrangement position of each estimated processing efficiency in the estimated efficiency set and the corresponding estimated feedback time.

[0025] In this embodiment, the improved LinUCB algorithm is specifically as follows: According to the edge node identifier, the node context vector in the node context set is matched with the estimated processing efficiency in the estimated efficiency set and the estimated feedback time in the estimated time set to obtain the node input record. The matching of the position means that the node context vector, estimated processing efficiency and estimated feedback time with the same edge node identifier are written into the same node input record. Extract the improved LinUCB parameter matrix, improved LinUCB reward vector, and basic exploration coefficient according to the task type, and determine the reward parameter vector based on the improved LinUCB parameter matrix and improved LinUCB reward vector. In Example 1, the improved LinUCB parameter matrix is ​​obtained by multiplying the inverse matrix of the improved LinUCB parameter matrix with the improved LinUCB benefit vector. The improved LinUCB parameter matrix is ​​generated by feeding back each diagonal element (excluding the task type encoding dimension) to the node context vector, and then accumulating and updating it using the outer product matrix corresponding to the node context vector after scaling each diagonal element. The improved LinUCB benefit vector is temporarily updated by accumulating the product of the target node context vector and the estimated efficiency of the target node to the LinUCB benefit vector, and then corrected by accumulating the product of the target node context vector and the processing efficiency deviation to the temporary benefit vector after the actual processing feedback is returned. The base exploration coefficient is adjusted based on the estimated feedback duration and allowable task delay recorded in the input records of each node to obtain the adjusted exploration coefficient. In Example 1, the estimated feedback time is divided by the allowable task delay, and the resulting ratio is limited to the range of 0 to 1 to obtain the feedback delay ratio. The feedback delay ratio is added to 1 and then multiplied by the basic exploration coefficient to obtain the modified exploration coefficient. Calculate the efficiency correction coefficient based on each estimated processing efficiency in the estimated efficiency set; In Example 1, the maximum estimated processing efficiency is extracted from the estimated efficiency set. When the maximum estimated processing efficiency is greater than 0, each estimated processing efficiency is divided by the maximum estimated processing efficiency. When the maximum estimated processing efficiency is equal to 0, the efficiency correction coefficient is determined to be 0. Calculate the estimated node return based on the node context vector, return parameter vector, and efficiency correction coefficient; In Example 1, the node context vector and the reward parameter vector are multiplied by an inner product to obtain the basic estimated reward, and then the basic estimated reward is multiplied by the corresponding efficiency correction coefficient. Calculate the node confidence increment based on the node context vector, the improved LinUCB parameter matrix, and the corrected exploration coefficients; In Example 1, the transpose of the node context vector, the inverse of the improved LinUCB parameter matrix, and the node context vector are multiplied sequentially, the square root of the multiplication result is taken, and the square root result is multiplied with the corresponding modified exploration coefficient. The estimated return for each node is added to the corresponding node confidence increment to obtain the node selection score. The node selection scores are then aggregated according to the edge node identifier to obtain the node score set.

[0026] In this embodiment, obtaining the task execution record, temporary parameter matrix, and temporary profit vector specifically involves: The target edge nodes are determined based on the node score set, and the target node records are obtained. In Example 1, edge nodes are arranged in descending order of node selection score. When the maximum node selection score corresponds to one edge node, the corresponding edge node is determined as the target edge node. When the maximum node selection score corresponds to multiple edge nodes, the corresponding estimated feedback time is extracted from the estimated time set, and the edge node with the smallest estimated feedback time is determined as the target edge node. Extract the target node context vector and target node prediction efficiency from the target node record, and establish the target input record; In Example 1, according to the edge node identifier in the target node record, the target node context vector and the target node estimated efficiency are extracted from the node context set and the estimated efficiency set, respectively, and the target node context vector, the target node estimated efficiency, the task type and the target edge node are written into the same record; Based on the target input record, a task distribution instruction is sent to the target edge node, and a task execution record is established. The establishment of the task execution record involves recording the task sending time and associating the target input record, the task sending time, and the task distribution instruction. Calculate the temporary parameter matrix based on the target node context vector and the improved LinUCB parameter matrix corresponding to the task type; In Example 1, the target node context vector is multiplied by the transpose of the target node context vector to obtain the target outer product matrix, and the target outer product matrix is ​​accumulated into the improved LinUCB parameter matrix corresponding to the task type. Calculate the temporary benefit vector based on the target node context vector, the target node estimated efficiency, and the improved LinUCB benefit vector for the corresponding task type. In Example 1, the target node context vector is multiplied by the target node estimated efficiency to obtain the estimated revenue vector, and the estimated revenue vector is accumulated into the improved LinUCB revenue vector of the corresponding task type.

[0027] In this embodiment, obtaining the updated parameter matrix, the updated profit vector, and the updated model group specifically involves: Receive the actual processing feedback returned by the target edge node, and associate the actual processing feedback with the task execution record to obtain the feedback association record. The association is to find the corresponding task execution record according to the task type and the target edge node identifier, and write the actual processing feedback into the found task execution record. The actual processing efficiency and actual feedback time are calculated by linking feedback records. In Example 1, the task processing volume, task start time, and task completion time are extracted from the feedback association record. The task processing volume is divided by the time length from the task start time to the task completion time. The calculation of the actual feedback time is to extract the task completion time and the actual feedback time from the feedback association record and calculate the time length from the task completion time to the actual feedback time. Calculate the processing efficiency deviation based on the actual processing efficiency and the estimated efficiency of the target node; In Example 1, the estimated efficiency of the target node is extracted from the task execution record, and the actual processing efficiency is subtracted from the estimated efficiency of the target node. The updated profit vector is calculated based on the target node context vector, processing efficiency deviation, and temporary profit vector. In Example 1, the target node context vector is multiplied by the processing efficiency deviation to obtain the revenue correction vector, and the revenue correction vector is accumulated into the temporary revenue vector; A parameter update record is established based on the temporary parameter matrix and the updated revenue vector. The parameter update record is established by determining the temporary parameter matrix as the updated parameter matrix and writing the updated parameter matrix and the updated revenue vector into the same record according to the task type. The hysteresis-free efficiency model is updated based on the target node context vector, actual processing efficiency, and corresponding forgetting coefficient to obtain the updated efficiency model. In Example 1, the target node context vector and the actual processing efficiency are written into the corresponding identification sample set, recursive least squares identification is performed according to the corresponding forgetting coefficient, the efficiency parameter set is updated, and the historical node context vector is processed according to the updated efficiency parameter set. The pure lag output model is updated based on the actual feedback duration and historical sample sequence to obtain the updated lag model. In Example 1, the corresponding sequence offset is determined based on the actual feedback duration and historical sample sequence, and the historical output position of the update efficiency model is adjusted according to the corresponding sequence offset; According to the task type and the target edge node identifier, the update efficiency model and the update lag model are combined to obtain an update model group. The combination is to write the update efficiency model and the update lag model with the same task type and target edge node identifier into the same model record, and connect the output of the update efficiency model to the input of the update lag model.

[0028] In this embodiment, the output of the cooperative scheduling result specifically refers to: Based on the task type and target edge node identifier, the updated parameter matrix, updated profit vector, updated model group and actual feedback time are written into the same parameter record to obtain the updated parameter record; A parameter record sequence is established based on the updated parameter record and the existing parameter record in the historical task execution record. The establishment of the parameter record sequence involves finding existing parameter records with the same task type and target edge node identifier, and arranging the updated parameter record and the found existing parameter record in the order of the actual feedback time. After entering the next scheduling cycle, update parameter records are extracted from the parameter record sequence according to the task type of the task to be distributed and the identifier of each edge node to obtain the cycle matching record. The extraction of update parameter records involves finding parameter records with the same task type and edge node identifier, and extracting the update parameter record with the latest actual feedback time from the found parameter records. The timeliness coefficient is calculated based on the scheduling time of the next scheduling cycle, the actual feedback time in the cycle matching record, and the allowable delay of the task. In Example 1, the time length between the scheduling time and the actual feedback time of the next scheduling cycle is calculated, the time length is divided by the allowable delay of the task, the resulting ratio is limited to the range of 0 to 1, and the difference between 1 and the limited ratio is calculated. Calculate the periodic parameter set based on the timeliness coefficient, periodic matching records, and the improved LinUCB parameter matrix and improved LinUCB benefit vector corresponding to the task type. In Example 1, the updated parameter matrix and the updated revenue vector in the periodic matching record are multiplied by the time factor, the improved LinUCB parameter matrix and the improved LinUCB revenue vector are multiplied by the difference between 1 and the time factor, and the products of the same type are added together. The periodic update records are obtained by updating and improving the LinUCB algorithm and the prediction model group based on the periodic parameter group and the periodic matching records. In Example 1, the periodic parameter group is written into the improved LinUCB algorithm, and the updated model group in the periodic matching record is written into the prediction model group corresponding to the task type and edge node identifier. When no periodic matching record is found, the improved LinUCB parameter matrix, improved LinUCB revenue vector and prediction model group corresponding to the task type are written into the periodic update record. The node context set for the next scheduling cycle is processed based on the periodic update records to obtain the collaborative scheduling result.

[0029] In Example 1, according to the task type and edge node identifier, the periodic parameter set and update model set are extracted from the periodic update record. The node context set is input into the update model set to obtain the estimated efficiency set and estimated duration set. The improved LinUCB algorithm is executed according to the periodic parameter set, node context set, estimated efficiency set and estimated duration set to obtain the node score set. The target edge node is determined according to the node score set. The task type of the task to be distributed, the target edge node and the task distribution instruction are associated to obtain the collaborative scheduling result.

[0030] Example 2: To verify the feasibility of this invention in practice, it was applied to a supply chain collaborative processing scenario, performing edge task distribution and computing power scheduling for order verification, inventory matching, transportation status analysis, and abnormal task identification. The implementation scenario sets up six edge nodes, each connected to a supply chain business terminal and a data acquisition device, with a scheduling cycle of 30 seconds. Tasks to be distributed are categorized into order verification, inventory matching, and transportation status analysis tasks, each with task attributes, resource requirements, processing volume, and allowable latency. Each edge node experiences changes in available computing power, task queue accumulation, communication bandwidth fluctuations, and communication latency variations across different scheduling cycles. When using fixed indicator weights for node scoring, the same weight configuration is difficult to adapt to different task types simultaneously. Even after node status changes, the original indicator contribution relationship is still used. When actual processing feedback is delayed, the current scheduling cycle cannot promptly obtain the corresponding actual processing efficiency, leading to a deviation between node scoring and task execution conditions.

[0031] During implementation, the task type, task attributes, resource requirements, task processing volume, and allowable latency of the tasks to be distributed are collected. The task processing volume and allowable latency are written into the task constraint parameters to establish a task collection record. Node status, available computing power, task queue length, communication bandwidth, and communication latency are collected according to the edge node identifier. The task collection record is written into the node record corresponding to each node status parameter to form a task node record. Resource requirements are compared with available computing power item by item to generate a computing power constraint flag. The task data volume in the task attributes is divided by the communication bandwidth to obtain the task transmission duration. The task transmission duration is added to the communication latency to obtain the estimated communication duration, which is then compared with the allowable latency to generate a latency constraint flag. Edge node records where both computing power constraint flags and latency constraint flags meet the requirements are retained to establish scheduling status data.

[0032] Extract edge node records that meet resource requirements from the scheduling status data. Establish candidate combination records according to the edge node identifiers. Write the task type, task processing volume, resource requirements, and allowable latency of the same task to be distributed into each candidate combination record, along with the available computing power, task queue length, communication bandwidth, and communication latency of the corresponding edge node. During implementation, call the most recent 1000 historical task execution records and extract the historical minimum and maximum values ​​of each data indicator according to the task type. When the current data indicator exceeds the historical value range, update the corresponding boundary using the current data indicator. When the updated maximum value is greater than the updated minimum value, divide the difference between the current data indicator and the updated minimum value by the difference between the updated maximum value and the updated minimum value; when they are equal, set the mapping value to 0 to obtain the initial mapping set. Extract each diagonal element from the improved LinUCB parameter matrix corresponding to the task type, excluding the task type encoding dimension. Add 1 to each diagonal element, take the square root, and use the square root result to scale the corresponding data indicator to obtain a matrix scaling set. The nodes are concatenated in a fixed order according to task type encoding, task processing volume, resource requirements, allowable task latency, available computing power, task queue length, communication bandwidth, and communication latency to form a node context vector, which is then aggregated to obtain a node context set.

[0033] Historical task execution records are categorized based on task type and edge node identifier. Records with the same task type and edge node identifier are grouped together and arranged according to their historical task completion time to obtain a historical sample sequence. Historical node context vectors and historical actual processing efficiencies are extracted from each historical task execution record to form an identification sample set. The forgetting coefficient is determined based on the difference in mapping values ​​between the communication bandwidth dimension and the task queue length dimension in the context vectors of adjacent historical nodes; the forgetting coefficient is limited to the range of 0 to 1. Recursive least squares identification is performed on the identification sample set based on the forgetting coefficient sequence to generate an efficiency parameter set, and a lag-free efficiency model is established accordingly. The feedback time length is determined based on the historical task completion time and the historical actual feedback time. The first historical task completion time no earlier than the historical actual feedback time is found in the historical sample sequence, the sequence position difference is calculated, and the sequence offset is obtained, thus establishing a pure lag output model. The lag-free efficiency model and the pure lag output model with the same task type and edge node identifier are written into the same model record to form a prediction model group.

[0034] In the current scheduling cycle, the node context vector is input into the corresponding lag-free efficiency model. The current efficiency prediction value is obtained by performing an inner product operation between the efficiency parameter vector and the node context vector. The current efficiency prediction value is written into the pure lag output model, and the output position is adjusted according to the sequence offset to obtain the delayed efficiency prediction value. The corresponding feedback time length is determined as the estimated feedback duration. The most recently fed-back task is found according to the task type and edge node identifier. The actual processing efficiency is subtracted from the corresponding delayed efficiency prediction value to obtain the feedback correction value. The current efficiency prediction value and the feedback correction value are then added to obtain the estimated processing efficiency. These are summarized to form an estimated efficiency set and an estimated duration set.

[0035] When executing the improved LinUCB algorithm, node input records are established according to edge node identifiers, and the improved LinUCB parameter matrix, improved LinUCB reward vector, and basic exploration coefficients are extracted. The estimated feedback duration is divided by the allowable task latency, and the resulting ratio is limited to the range of 0 to 1. This ratio is then added to 1 and multiplied by the basic exploration coefficient to obtain the corrected exploration coefficient. Each estimated processing efficiency is divided by the maximum estimated processing efficiency to obtain the efficiency correction coefficient. The estimated node reward is calculated based on the node context vector, reward parameter vector, and efficiency correction coefficient. The node confidence increment is calculated based on the node context vector, improved LinUCB parameter matrix, and corrected exploration coefficient. The two are added together to obtain the node selection score. When node selection scores are the same, the edge node with the smaller estimated feedback duration is selected as the target edge node.

[0036] After sending the task distribution instruction to the target edge node, the target node's context vector is multiplied by its transpose to form the target outer product matrix, which is then added to the improved LinUCB parameter matrix to obtain a temporary parameter matrix. The target node's context vector is multiplied by the target node's estimated efficiency, and this multiplication is added to the improved LinUCB reward vector to obtain a temporary reward vector. After the actual processing feedback is returned, the task processing volume is divided by the time from task start to task completion to obtain the actual processing efficiency. The time from task completion to the actual feedback time is determined as the actual feedback duration. The actual processing efficiency is subtracted from the target node's estimated efficiency to obtain the processing efficiency deviation. The product of the target node's context vector and the processing efficiency deviation is added to the temporary reward vector to obtain the updated reward vector, and the lag-free efficiency model and the pure lag output model are updated accordingly.

[0037] The scheduling process ran for 20 consecutive cycles, handling a total of 300 tasks, including 120 order verification tasks, 90 inventory matching tasks, and 90 transportation status analysis tasks. The scheduling results of five representative tasks are listed in Table 1.

[0038] Table 1. Supply Chain Edge Task Scheduling Results The estimated and actual processing efficiencies of the five tasks in Table 1 differ by 1.5, 1.3, 0.8, 1.5, and 0.7 items per second, respectively, indicating that the estimated results are close to the actual processing feedback. The completion time of each task is lower than the allowed latency and lower than the completion time corresponding to the fixed-weight processing method. For the transportation status analysis task T063, the allowed latency is 650 milliseconds, and the estimated feedback time is 214 milliseconds. After adjusting the exploration coefficient based on the estimated feedback time, the system selects the target edge node that matches the communication conditions and node load, reducing the completion time from 621 milliseconds in the fixed-weight processing method to 548 milliseconds. The estimated processing efficiency of the order verification task T118 is 91.3 items per second, and the actual processing efficiency is 89.8 items per second. The processing efficiency deviation formed after the actual processing feedback is returned is written into the updated revenue vector and used for node score calculation in the next scheduling cycle.

[0039] During cross-cycle invocation, the system stores the update parameter matrix, update benefit vector, update model group, and actual feedback time according to task type and target edge node identifier. At the start of the next scheduling cycle, the update parameter record with the latest actual feedback time is extracted, the time length between the scheduling time and the actual feedback time is calculated, and compared with the task's allowed latency to generate a timeliness coefficient. The timeliness coefficient is used to fuse the update parameters with the improved LinUCB parameters corresponding to the task type, so that the update results with shorter intervals maintain a high proportion, and the proportion of historical update parameters is reduced when the interval is close to the task's allowed latency. After 20 scheduling cycles, 286 out of 300 tasks were completed within the task's allowed latency, while 267 were completed using the fixed-weight processing method; the average estimated processing efficiency deviation of this invention was 2.1%, while the deviation based on the historical average efficiency under the fixed-weight processing method was 6.8%; the difference between the maximum and minimum number of node task allocations decreased from 47 to 29.

[0040] This embodiment demonstrates that the present invention, through the collaborative processing of node context set, prediction model group and improved LinUCB algorithm, completes node selection by using predicted processing efficiency and predicted feedback duration before actual feedback is returned. After the actual processing feedback is returned, the benefit vector and prediction model group are corrected, and parameter calls across scheduling cycles are completed through parameter timeliness coefficient. This allows task type, node running status and feedback timeliness to jointly participate in node score calculation, forming a continuous collaborative task distribution and computing power scheduling process.

[0041] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for collaborative task distribution and computing power scheduling at supply chain edge nodes, characterized in that, include: Collect the task type, task attributes, resource requirements, task constraint parameters, and node status parameters of each edge node for the tasks to be distributed, and establish scheduling status data. Read historical task execution records, filter edge nodes that meet resource requirements based on scheduling status data, and combine the task attributes of the task to be distributed with the node status of the filtered edge nodes to obtain a set of node contexts; Establish a predictive model group based on historical task execution records; Input the node context set into the prediction model set to obtain the prediction efficiency set and the prediction duration set; The improved LinUCB algorithm is executed based on the node context set, the estimated efficiency set, and the estimated duration set to obtain the node score set. The target edge node is determined based on the node score set, the task distribution instruction is sent to the target edge node, and the LinUCB algorithm is updated and improved based on the target node context vector and the target node estimated efficiency to obtain the task execution record, temporary parameter matrix and temporary profit vector. Receive the actual processing feedback returned by the target edge node, calculate the actual processing efficiency and processing efficiency deviation based on the task execution record, and update the temporary revenue vector and the prediction model group based on the temporary parameter matrix, the actual processing efficiency and the processing efficiency deviation to obtain the updated parameter matrix, the updated revenue vector and the updated model group. Store the updated parameter matrix, updated benefit vector, and updated model group according to the task type and target edge node identifier. In the next scheduling cycle, call the updated parameter matrix, updated benefit vector, and updated model group to output the collaborative scheduling result.

2. The collaborative task distribution and computing power scheduling method for supply chain edge nodes according to claim 1, characterized in that, The establishment of scheduling status data specifically involves: Collect the task type, task attributes, resource requirements, task processing volume, and allowed task latency of the tasks to be distributed. Write the task processing volume and allowed task latency into the task constraint parameters and associate them with the tasks to be distributed to obtain the task collection record. Collect the node status parameters of each edge node according to the edge node identifier, and associate the task collection record with each node status parameter to obtain the task node record. The node status parameters include node status, available computing power, task queue length, communication bandwidth and communication latency. Generate computing power constraint markers based on resource requirements and available computing power in the task node records; Generate a delay constraint flag based on communication bandwidth, communication latency, allowable task latency, and the amount of task data in the task attributes; By associating task node records, computing power constraint tags, and latency constraint tags according to edge node identifiers, scheduling status data is obtained.

3. The collaborative task distribution and computing power scheduling method for supply chain edge nodes according to claim 1, characterized in that, The obtained node context set is specifically as follows: Extract edge node records that meet resource requirements from the scheduling status data, and associate the task attributes of the tasks to be distributed with the node status of the edge nodes to obtain candidate combination records; Based on the task type, extract the historical values ​​of each data indicator from the historical task execution records, and combine them with candidate combination records to determine the value range of each data indicator, thus obtaining the indicator range set; Each data indicator in the candidate combination record is matched with the indicator range set, and the value range of the corresponding data indicator is adjusted according to the matching result to obtain the updated range set. Based on the update range set, perform dimensionless mapping on the task processing volume, resource requirements, allowed task latency, available computing power, task queue length, communication bandwidth and communication latency respectively to obtain the initial mapping set; Extract the diagonal elements of the improved LinUCB parameter matrix corresponding to the task type, excluding the task type encoding dimension, and scale the data indicators in the initial mapping set according to the diagonal elements to obtain the matrix scaling set. Set the task type codes according to the order of task types, and concatenate the task type codes with the matrix scaling set to obtain the node context vector; The data location and vector dimension of each node's context vector are checked according to the edge node identifier, and the node context vectors are summarized to obtain the node context set.

4. The collaborative task distribution and computing power scheduling method for supply chain edge nodes according to claim 1, characterized in that, The establishment of the prediction model group based on historical task execution records specifically includes: Historical task execution records are divided according to task type and edge node identifier, and then arranged to obtain a historical sample sequence. Historical node context vectors and historical actual processing efficiency are extracted from historical sample sequences, and a correspondence is established to obtain the recognition sample set; The forgetting coefficients are determined by the order of the identified sample set and the context vectors of adjacent historical nodes, thus obtaining the forgetting coefficient sequence. Based on the forgetting coefficient sequence, recursive least squares identification is performed on the identification sample set to obtain the efficiency parameter set; The efficiency parameter set is associated with the corresponding task type and edge node identifier. The historical node context vector is processed according to the efficiency parameter set, and the processing results are stored in correspondence with the historical actual processing efficiency to obtain the lag-free efficiency model. The feedback time length is determined based on the historical task completion time and the historical actual feedback time in the historical sample sequence, and the sequence offset is determined based on the feedback time length. The position of the hysteresis-free efficiency model is adjusted based on the sequence offset and correlated with the corresponding historical actual processing efficiency to obtain the pure hysteresis output model. Based on task type and edge node identifier, the lag-free efficiency model and the pure lag output model are combined to obtain the prediction model group.

5. The collaborative task distribution and computing power scheduling method for supply chain edge nodes according to claim 1, characterized in that, The step of inputting the node context set into the prediction model group specifically involves: Based on the task type and edge node identifier, the context vector of each node in the node context set is matched with the predicted model group to obtain the model matching record; Based on the model matching record processing node context vector, the current efficiency prediction value is obtained; Based on the pure lag output model in the model matching record, process the current efficiency prediction value to obtain the delayed efficiency prediction value and the estimated feedback time; Based on the task type and edge node identifier, the most recent task that has been responded to is found from the historical task execution records and matched with the pure lag output model to obtain the feedback matching record; Calculate the feedback correction value based on the feedback matching record; The estimated processing efficiency is calculated based on the current efficiency prediction value and the feedback correction value, and the estimated processing efficiency is summarized according to the edge node identifier to obtain the estimated efficiency set. The estimated feedback duration is obtained by summarizing the edge node identifiers.

6. The collaborative task distribution and computing power scheduling method for supply chain edge nodes according to claim 1, characterized in that, The improved LinUCB algorithm is as follows: According to the edge node identifier, the node context vector in the node context set is matched with the estimated processing efficiency in the estimated efficiency set and the estimated feedback time in the estimated time set to obtain the node input record. Extract the improved LinUCB parameter matrix, improved LinUCB reward vector, and basic exploration coefficient according to the task type, and determine the reward parameter vector; The base exploration coefficient is adjusted based on the estimated feedback duration and allowable task delay recorded in the input records of each node to obtain the adjusted exploration coefficient. Calculate the efficiency correction coefficient based on each estimated processing efficiency in the estimated efficiency set; Calculate the estimated node return based on the node context vector, return parameter vector, and efficiency correction coefficient; Calculate the node confidence increment based on the node context vector, the improved LinUCB parameter matrix, and the corrected exploration coefficients; The estimated return for each node is added to the corresponding node confidence increment to obtain the node selection score. The node selection scores are then aggregated according to the edge node identifier to obtain the node score set.

7. The collaborative task distribution and computing power scheduling method for supply chain edge nodes according to claim 1, characterized in that, The process of obtaining the task execution record, temporary parameter matrix, and temporary profit vector is as follows: The target edge nodes are determined based on the node score set, and the target node records are obtained. Extract the target node context vector and target node prediction efficiency from the target node record, and establish the target input record; Based on the target input record, send task distribution instructions to the target edge node and establish a task execution record; Calculate the temporary parameter matrix based on the target node context vector and the improved LinUCB parameter matrix corresponding to the task type; Calculate the temporary revenue vector based on the target node context vector, the target node estimated efficiency, and the improved LinUCB revenue vector for the corresponding task type.

8. The collaborative task distribution and computing power scheduling method for supply chain edge nodes according to claim 1, characterized in that, The process of obtaining the updated parameter matrix, updated profit vector, and updated model set specifically involves: Receive the actual processing feedback returned by the target edge node, and associate the actual processing feedback with the task execution record to obtain the feedback association record; The actual processing efficiency and actual feedback time are calculated by linking feedback records. Calculate the processing efficiency deviation based on the actual processing efficiency and the estimated efficiency of the target node; The updated profit vector is calculated based on the target node context vector, processing efficiency deviation, and temporary profit vector. Establish parameter update records based on the temporary parameter matrix and the updated payoff vector; The hysteresis-free efficiency model is updated based on the target node context vector, actual processing efficiency, and corresponding forgetting coefficient to obtain the updated efficiency model. The pure lag output model is updated based on the actual feedback duration and historical sample sequence to obtain the updated lag model. Based on the task type and the target edge node identifier, the update efficiency model and the update lag model are combined to obtain the update model group.

9. The collaborative task distribution and computing power scheduling method for supply chain edge nodes according to claim 1, characterized in that, The output of the collaborative scheduling result is specifically as follows: Based on the task type and target edge node identifier, the updated parameter matrix, updated profit vector, updated model group and actual feedback time are written into the same parameter record to obtain the updated parameter record; Establish a parameter record sequence based on the updated parameter records and existing parameter records in the historical task execution records; Upon entering the next scheduling cycle, update parameter records are extracted from the parameter record sequence based on the task type of the task to be distributed and the identifier of each edge node to obtain the cycle matching record; The timeliness coefficient is calculated based on the scheduling time of the next scheduling cycle, the actual feedback time in the cycle matching record, and the allowable delay of the task. Calculate the periodic parameter set based on the timeliness coefficient, periodic matching records, and the improved LinUCB parameter matrix and improved LinUCB benefit vector corresponding to the task type. The periodic update records are obtained by updating and improving the LinUCB algorithm and the prediction model group based on the periodic parameter group and the periodic matching records. The node context set for the next scheduling cycle is processed based on the periodic update records to obtain the collaborative scheduling result.

Images