An edge computing method for electronic component processing data

Abstract

The application discloses an edge computing method for electronic component processing data, and relates to the technical field of industrial edge computing.The method comprises the following steps: an edge gateway collects multi-source heterogeneous process data of a processing device and performs preprocessing, a causal semantic graph of process parameters and quality indexes is constructed by a dynamic semantic dictionary and a graph embedding method, the comprehensive resource margin of an edge node and the emergency degree score of a detection task are calculated, a state vector is combined as an input of a reinforcement learning scheduling agent, an action is adaptively selected among local processing, edge-edge collaborative processing, cloud offloading and degradation to a pruning model processing, and an emergency degradation processing is triggered when the resource margin is lower than a threshold value and the task emergency degree is higher than a threshold value.The application improves the edge-side causal decision information density, realizes a dynamic balance among detection accuracy, response real-time performance and heterogeneous computing power, and is suitable for quality detection and intelligent production process management of an electronic component production line.

Description

Technical Field

[0001] This invention relates to the field of industrial edge computing technology, specifically to an edge computing method for electronic component processing data. Background Technology

[0002] The manufacturing process of electronic components typically involves multiple steps, such as surface mount technology (SMT), reflow soldering, automated optical inspection, and automated X-ray inspection. Each step is equipped with data acquisition units such as programmable logic controllers (PLCs), temperature sensors, pressure sensors, and industrial cameras. In the field of industrial internet, unified communication protocols are widely used for on-site data acquisition and transmission, and edge gateways are used to perform local preliminary data processing to reduce cloud transmission bandwidth pressure and response latency. Meanwhile, graph embedding representation learning technology, which maps graph structure nodes to low-dimensional vector spaces, has been applied in the fields of recommendation systems and knowledge graph construction. Deep reinforcement learning has also been used in decision optimization scenarios such as resource scheduling and task allocation.

[0003] Currently, edge computing implementation schemes in the electronic component processing industry that are similar to this invention typically operate as follows: each process device uploads raw data such as mounting pressure, temperature zone curves, and detection images to the edge gateway through a unified communication protocol. The edge gateway performs standardization processing on the data, such as format cleaning, unit conversion, and timestamp alignment, to form a data set with a unified structure. The edge gateway allocates processing tasks to local computing units or uploads them to the cloud server for inference calculations according to pre-set fixed rules or a fixed cloud-edge collaboration ratio. The detection results are fed back to the manufacturing execution system for production decisions. Structurally, this scheme consists of a field device layer, an edge gateway layer, and a cloud processing layer, with the edge gateway acting as a transit hub between the field devices and the cloud.

[0004] The above-mentioned scheme has the following shortcomings: First, since the scheme only performs data format cleaning and aggregation at the edge gateway side, it does not establish a deep semantic association model between cross-process parameters and final quality defects at the data level. As a result, the edge nodes cannot identify the causal relationship between process parameter fluctuations and quality defects in real time. This leads to a large amount of raw data being transmitted and stored indiscriminately, and the causal information with decision-making value is not effectively extracted. This results in a large amount of data transmission but a low density of effective decision-making information, causing a lag in the root cause location and real-time interception of defective products. Second, since the scheme adopts a fixed rule or fixed ratio cloud-edge collaborative scheduling method, it does not make adaptive decisions based on the dynamic resource reserves of the edge nodes and the urgency of specific tasks. During periods of intensive detection tasks, the edge nodes are prone to computing power overload, causing high-priority detection tasks to be treated equally with low-priority tasks, resulting in high processing latency. It is difficult to complete the interception of defective products within the time window allowed by the process, and it is difficult to simultaneously meet the requirements of detection accuracy, heterogeneous hardware compatibility, and real-time response. Therefore, we propose an edge computing method for electronic component processing data. Summary of the Invention

[0005] The purpose of this invention is to provide an edge computing method for electronic component processing data.

[0006] To address the shortcomings of the existing technologies, the technical problems to be solved by this invention include two aspects: First, how to perform real-time deep semantic alignment on multi-source heterogeneous process data generated by multiple processes of electronic components at the edge, thereby directly establishing a causal mapping relationship between process parameters and quality defects at the edge, avoiding the problem of low effective decision information density caused by merely staying at the level of data format uniformity. Second, how to achieve adaptive task scheduling based on the urgency of tasks and the dynamic resource reserve of nodes under the conditions of highly heterogeneous computing power and memory and limited resources at edge nodes, thereby ensuring detection accuracy while avoiding detection delays exceeding the time window allowed by the process due to computing power overload.

[0007] To address the aforementioned technical problems, this invention provides an edge computing method for electronic component processing data, comprising: an edge gateway collecting multi-source heterogeneous process data generated by electronic component processing equipment via a communication protocol; preprocessing the collected multi-source heterogeneous process data to obtain standardized multi-source feature vectors; for each type of preprocessed process parameter or quality indicator, searching for the existence of a corresponding semantic node in a dynamic semantic dictionary maintained by the edge gateway; if it exists, updating the value of the semantic node; if it does not exist, adding a corresponding semantic node in the dynamic semantic dictionary and assigning it an initial semantic vector; using each semantic node in the dynamic semantic dictionary as a vertex, and using the process parameter nodes as the basis for the calculation. The associations between points and quality indicator nodes, established based on historical statistical correlations, are used as initial edges to construct a process relationship graph. Random walks are performed on this graph to obtain node sequences, which are then trained to map each semantic node to an embedding vector in a unified semantic vector space. Based on historical sequence data within a sliding time window, conditional mutual information between established process parameter nodes and quality indicator nodes is calculated to quantify the causal contribution of corresponding process parameter nodes to quality indicator nodes. The weights of corresponding edges in the process relationship graph are updated based on the calculated conditional mutual information to obtain a causal semantic graph. The embedding vectors corresponding to the currently collected process parameters are then compared with... The system performs nearest neighbor retrieval on established causal relationship patterns in the semantic vector space. When the retrieved causal strength exceeds a preset threshold and the corresponding quality indicator shows a deteriorating trend, it outputs a process anomaly warning and a corresponding defect propagation risk assessment value. It also calculates the comprehensive resource reserve of edge nodes in real time. This comprehensive resource reserve is obtained by weighting and summing the ratios of available computing power to total computing power, available memory to total memory, and available bandwidth to total bandwidth, according to their respective computing power weights, memory weights, and bandwidth weights. For the detection task to be processed, the system considers the defect propagation risk assessment value, the reciprocal of the remaining processable time window length for the task, and the processing batch to which the task belongs. The urgency score of the task is calculated by weighted summation of the criticality weights. The overall resource availability, urgency score, and real-time network bandwidth status are combined into an environment state vector, which is input into the policy network of the reinforcement learning scheduling agent. The agent selects one of four candidate actions to execute: local processing, edge-to-edge collaborative processing, cloud offloading, and downgrading to a pruning model. Before the selection, it is determined whether the overall resource availability is lower than a preset resource depletion threshold and whether the urgency score is higher than a preset urgency threshold. If both conditions are met, the decision-making process of the policy network is skipped, and the action of downgrading to a pruning model is directly triggered, calling a pre-trained pruned lightweight detection model to process the task.

[0008] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0009] 1. This invention differs from the existing combination of conventional graph embedding representation learning and conventional reinforcement learning scheduling, which are applied independently. It directly uses the defect propagation risk assessment value output by the causal semantic graph as the core input of the urgency score of the reinforcement learning scheduling agent. This enables the causal inference results of process anomalies to drive resource scheduling decisions in real time. It overcomes the technical limitations of existing edge computing solutions where data semantic analysis and task scheduling are separated and cannot be linked to respond, forming a non-obvious technical effect of semantic perception and resource scheduling collaborative decision-making.

[0010] 2. This invention employs an adaptive offloading strategy that dually senses resource availability and task urgency, enabling edge systems to complete high-priority tasks in a relatively short time even when computing power is limited. This achieves a dynamic balance between detection accuracy, real-time response, and heterogeneous hardware computing power, alleviating the problem of delays in high-priority task processing caused by fixed-rule scheduling. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the overall process in an embodiment of the present invention;

[0012] Figure 2 This is a schematic diagram of the causal semantic graph construction sub-process in an embodiment of the present invention;

[0013] Figure 3 This is a schematic diagram of the reinforcement learning adaptive scheduling subprocess in an embodiment of the present invention. Detailed Implementation

[0014] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0015] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0016] I. System Overall Architecture:

[0017] The edge computing method for electronic component processing data provided in this embodiment is applied to a system that, from bottom to top, includes a field device layer, an edge gateway layer, an edge scheduling decision layer, and a multi-level processing resource layer.

[0018] The field equipment layer includes multiple heterogeneous production equipment such as pick-and-place machines, reflow ovens, AOI inspection machines, and AXI inspection machines. Each device sends raw data such as placement pressure parameters, temperature zone temperature curve characteristics, and inspection image characteristics to the edge gateway layer via OPC-UA or MQTT protocols.

[0019] The edge gateway layer is deployed on the production line and contains a dynamic semantic dictionary module and a causal semantic graph embedding module. The dynamic semantic dictionary module is used to maintain a semantic node mapping table that can be dynamically updated according to the types of process parameters and quality indicators. The causal semantic graph embedding module is used to perform graph embedding calculations on the nodes in the semantic dictionary and dynamically update the graph edge weights based on causal strength. The two modules achieve real-time support of the semantic dictionary for the graph embedding process through bidirectional data interaction.

[0020] The edge scheduling decision layer is equipped with a resource reserve perception module, a task urgency assessment module, and a reinforcement learning scheduling agent. The outputs of the three modules are used as components of the environment state vector and input into the policy network of the scheduling agent, which then outputs scheduling actions.

[0021] The multi-level processing resource layer includes edge local processing units, edge-to-edge collaborative processing units (i.e., adjacent edge nodes), cloud processing center, and pruned lightweight detection model. The scheduling agent allocates tasks to the corresponding processing resources based on the decision results, and directly calls the pruned lightweight detection model to complete the processing when the emergency degradation path is triggered.

[0022] II. Real-time Causal Semantic Graph Embedding Method for Cross-Process Multi-Source Heterogeneous Process Data:

[0023] In this embodiment, the edge gateway first collects the timing data of the placement pressure of the chip mounter, the temperature curve characteristics of each temperature zone of the reflow soldering, and the feature vectors of the AOI and AXI detection images in real time through the OPC-UA or MQTT protocol. The feature vectors include the defect location, defect size and grayscale features. The collected data is then filtered and normalized to form a standardized set of multi-source feature vectors.

[0024] For each type of process parameter or quality indicator after preprocessing, the corresponding semantic node is searched in the dynamic semantic dictionary to see if it already exists. If it already exists, the latest value of the node is updated. If it does not exist, for example, if a new equipment type or a new detection indicator appears, the corresponding node is added to the semantic dictionary and an initial semantic vector is assigned to it.

[0025] Using each node in the dynamic semantic dictionary as the vertex of the graph, and the association between the process parameter node and the quality index node based on historical statistical correlation or prior process knowledge as the initial edge of the graph, a directed process relationship graph is constructed. The initial weight of the edge is set according to the Pearson correlation coefficient of historical batch data or expert prior rules.

[0026] A depth-constrained random walk sampling is performed on the process relationship graph to generate a node sequence. The node sequence is then trained using a Skip-gram model with negative sampling, mapping each semantic node to a low-dimensional dense vector in a unified high-dimensional semantic vector space. The optimization objective function for this process is as follows:

[0027] ;

[0028] in, This represents the loss function for training graph embeddings. This represents the set of positive sample node pairs obtained by random walk sampling. This represents pairs of nodes that co-occur within the same walk window. This represents the set of negative sample node pairs obtained through negative sampling. Represents a node Non-co-occurring nodes with random sampling The negative sample pairs formed and Representing nodes respectively With nodes Embedded vectors in semantic vector space This represents the Sigmoid activation function;

[0029] For the process parameter nodes and quality index nodes that have established connections in the semantic vector space, conditional mutual information is calculated based on historical sequence data within a sliding time window to quantify the causal contribution of process parameter changes to quality indices. The calculation formula is as follows:

[0030] ;

[0031] in, Indicates process parameter nodes Quality indicator nodes The causal strength, This indicates conditional mutual information operations. Indicates process parameter nodes At any moment The value of , and These represent the quality indicator nodes. At any moment and the next moment The value of is used to calculate the causal strength value and update the weight of the corresponding edge in the process relationship diagram. The larger the weight, the more significant the causal influence of the process parameter on the quality index.

[0032] The real-time process parameters collected in the current batch are embedded into a vector and the known historical parameters and quality causal relationship patterns in the semantic vector space are searched for nearest neighbors. If the causal strength of the retrieved data exceeds a preset threshold and the corresponding quality index shows a deteriorating trend, cross-batch process anomaly warning information is output. The defect propagation risk assessment value corresponding to the warning information is used as an important input for the subsequent task scheduling process. The above graph construction, embedding calculation and edge weight update process form a closed loop with the continuous arrival of new data, realizing the dynamic update of the causal semantic graph.

[0033] III. Adaptive offloading strategy for edge tasks based on dynamic awareness of resource availability and task urgency:

[0034] In this embodiment, the comprehensive resource reserve of each edge node is calculated in real time, and the calculation formula is as follows:

[0035] ;

[0036] in, Represents edge nodes The overall resource surplus, and These represent the node's currently available computing power and total computing power, respectively. and These represent the currently available memory and total memory of the node, respectively. and These represent the currently available network bandwidth and total bandwidth of the node, respectively. , , These are the weighting coefficients corresponding to the three indicators: computing power, memory, and bandwidth, respectively, satisfying... ;

[0037] For each pending detection task, an urgency score is calculated by considering its potential defect propagation risk and the remaining processing time window. The calculation formula is as follows:

[0038] ;

[0039] in, Indicates task Urgency rating, This represents the defect propagation risk assessment value for the corresponding batch of the task, output by the aforementioned causal semantic graph embedding module. This indicates the remaining processable time window length for the task. This indicates the criticality weight of the processing batch to which the task belongs. , , These are the corresponding weighting coefficients;

[0040] The resource availability of each edge node, the urgency score of the task to be processed, and the real-time network bandwidth status are combined into the current environmental state vector, which is used as the input to the reinforcement learning scheduling agent policy network. The scheduling agent's policy network outputs the action probability distribution in the action space (local processing, edge-to-edge collaborative processing, cloud offloading, and degrading to pruning model processing) according to the current state, and selects the action to be executed according to the distribution. The policy network adopts a lightweight actor-critic structure to ensure that the memory occupied on the edge nodes is kept within a limited range.

[0041] Before executing the regular decision, it is first determined whether the current node's resource balance is lower than the preset resource depletion threshold and whether the task urgency is higher than the preset urgency threshold. If both conditions are met, the regular decision-making process of the policy network is skipped, and the action of downgrading to the pruning model is directly triggered. The pre-trained pruned version of the lightweight detection model is called to process the task to ensure that the high-priority alarm task is completed in a short time.

[0042] After the action is completed, the actual latency, detection accuracy loss, and detection accuracy of this task are collected, and the instant reward value is calculated using the following formula:

[0043] ;

[0044] in, Indicates time The immediate reward obtained after performing an action. This indicates the actual processing delay of this task. This indicates the loss of detection accuracy due to model downgrading or unloading. This indicates the detection accuracy rate of this task. , , These are the corresponding weighting coefficients;

[0045] Store the experience tuple consisting of the current state, the action performed, the reward obtained, and the next state into the experience replay buffer, and iteratively update the value function as follows:

[0046] ;

[0047] in, Indicates the state Next action The value estimate, Indicates the learning rate. This represents a discount factor used to weigh the importance of future rewards. Indicates the next state The strategy network is deployed by taking the largest value among all the optional actions and combining offline pre-training with online fine-tuning. That is, the initial training is completed offline on the historical dataset, and then the parameters are continuously fine-tuned online in small steps during the actual production line operation to adapt to the dynamic changes in the production line conditions.

[0048] IV. Overall Implementation Process:

[0049] The overall implementation process of this embodiment spans two major modules: causal semantic graph embedding and adaptive task offloading. After the raw data from multiple processes is accessed by the edge gateway, it first undergoes causal semantic graph embedding to obtain abnormal warning information and defect propagation risk assessment values. Then, it enters the state awareness stage to calculate resource reserves and task urgency. Finally, the reinforcement learning scheduling agent completes the adaptive offloading decision. The processing results are ultimately fed back to the manufacturing execution system or enterprise resource planning system. The processing results are also used as new historical data to feed back into the dynamic update process of the semantic graph, thus forming a complete closed loop of data collection, semantic extraction, task scheduling, and result feedback.

[0050] V. Alternative Implementation Methods:

[0051] In another embodiment of the present invention, the semantic alignment of multi-source heterogeneous process data across processes can be achieved by using a time-series feature encoding model based on a self-attention mechanism instead of the graph embedding method. This model directly encodes the multi-source time-series feature sequences and outputs a unified semantic vector. The attention weight matrix indirectly reflects the degree of influence of different process parameters on quality indicators, thus adapting to application scenarios where process parameters have strong time-series dependence characteristics.

[0052] In another embodiment of the present invention, the edge task adaptive unloading decision-making process can use a heuristic scheduling algorithm based on fuzzy logic rules to replace the reinforcement learning scheduling agent. That is, the fuzzy membership function and inference rule base between resource reserve, task urgency and scheduling action are set in advance according to historical experience. The scheduling action is directly output through fuzzy inference. Thus, even in scenarios where the computing power of edge nodes is limited and the conditions for deploying reinforcement learning models are not available, the adaptive task allocation effect with dual perception of resource reserve and task urgency can still be achieved.

[0053] VI. System and Storage Media Implementation:

[0054] Corresponding to the above method, this embodiment also provides an edge computing system for electronic component processing data, including an edge gateway and an edge scheduling decision module communicatively connected to the edge gateway. The edge gateway is equipped with a dynamic semantic dictionary module and a causal semantic graph embedding module, which are used to perform various processes in the aforementioned causal semantic graph embedding method. The edge scheduling decision module is equipped with a resource reserve perception module, a task urgency assessment module, and a reinforcement learning scheduling agent, which are used to perform various processes in the aforementioned edge task adaptive offloading strategy, and establish data connection relationships with the edge local processing unit, the edge-to-edge collaborative processing unit, the cloud processing center, and the pruned lightweight detection model, respectively.

[0055] This embodiment also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the various steps in the aforementioned edge computing method for electronic component processing data.

[0056] This embodiment also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements each step in the aforementioned edge computing method for processing data of electronic components.

[0057] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. An edge computing method for electronic component processing data, comprising the step of an edge gateway collecting multi-source heterogeneous process data generated by electronic component processing equipment via a communication protocol, characterized in that, Also includes: The collected multi-source heterogeneous process data are preprocessed to obtain standardized multi-source feature vectors; For each type of process parameter or quality indicator after preprocessing, the existence of a corresponding semantic node is searched in the dynamic semantic dictionary maintained by the edge gateway. If it exists, the value of the semantic node is updated. If it does not exist, a corresponding semantic node is added in the dynamic semantic dictionary and an initial semantic vector is assigned to it. Using each semantic node in the dynamic semantic dictionary as a vertex and the association between process parameter nodes and quality index nodes based on historical statistical correlation as the initial edge, a process relationship graph is constructed. Random walk sampling is performed on the process relationship graph to obtain a node sequence, and the node sequence is trained to map each semantic node to an embedding vector in a unified semantic vector space; Based on historical sequence data within a sliding time window, the conditional mutual information between established process parameter nodes and quality indicator nodes is calculated to quantify the causal contribution of the corresponding process parameter nodes to the quality indicator nodes. The weights of the corresponding edges in the process relationship graph are updated according to the calculated conditional mutual information to obtain a causal semantic graph. The embedded vector corresponding to the currently collected process parameters is searched for nearest neighbors with the established causal association patterns in the semantic vector space. When the causal strength of the retrieved data exceeds a preset threshold and the corresponding quality indicators show a deterioration trend, the process anomaly warning information and the corresponding defect propagation risk assessment value are output. The comprehensive resource reserve of the edge node is calculated in real time. The comprehensive resource reserve is obtained by weighting and summing the ratio of the available computing power of the edge node to the total computing power, the ratio of available memory to the total memory, and the ratio of available bandwidth to the total bandwidth according to the corresponding computing power weight, memory weight, and bandwidth weight, respectively. The sum of the computing power weight, memory weight, and bandwidth weight is 1. For the inspection task to be processed, the urgency score of the task is calculated by weighted summation based on the defect propagation risk assessment value, the reciprocal of the remaining processable time window length of the task, and the criticality weight of the processing batch to which the task belongs. The comprehensive resource balance, the urgency score, and the real-time network bandwidth status are combined into an environment state vector, which is then input into the policy network of the reinforcement learning scheduling agent. One of the four candidate actions is selected for execution: local processing, edge-to-edge collaborative processing, cloud offloading, and downgrading to pruning model processing. Before making the selection, it is determined whether the overall resource balance is lower than the preset resource depletion threshold and whether the urgency score is higher than the preset urgency threshold. If both conditions are met, the decision-making process of the policy network is skipped, and the action of downgrading to the pruning model is directly triggered. The pre-trained pruned lightweight detection model is then called to process the task.

2. The edge computing method for electronic component processing data according to claim 1, characterized in that, The step of mapping each semantic node to an embedding vector in a unified semantic vector space includes: training the node sequence using a Skip-gram model with negative sampling, so that the similarity of co-occurring node pairs within the same walk window in the semantic vector space is higher than that of randomly sampled non-co-occurring node pairs, thereby obtaining a low-dimensional dense embedding vector corresponding to each semantic node.

3. The edge computing method for electronic component processing data according to claim 1, characterized in that, The steps for determining that the corresponding quality indicator has a deteriorating trend include: comparing the value change trend of the quality indicator node within the current sliding time window with the historical value distribution of the quality indicator node; when the degree to which the value change trend deviates from the historical value distribution exceeds a preset deviation threshold, it is determined that the quality indicator node has a deteriorating trend.

4. The edge computing method for electronic component processing data according to claim 1, characterized in that, Also includes: The experience tuple, consisting of the state, action, reward, and next state obtained after the action is performed, is stored in the experience replay buffer. The immediate reward value is calculated by weighting the actual latency, detection accuracy loss, and detection accuracy of this task. Based on the experience tuple in the experience replay buffer, the value function of the policy network is iteratively updated. The magnitude of the iterative update is determined by the learning rate, the immediate reward value, and the maximum value among the value estimates corresponding to each candidate action in the next state.

5. The edge computing method for electronic component processing data according to claim 4, characterized in that: The policy network is deployed using an offline pre-training combined with online fine-tuning approach. This involves first performing initial offline training of the policy network on a historical dataset, and then continuously fine-tuning the parameters of the policy network in small steps during the actual operation of the electronic component processing production line.

6. The edge computing method for electronic component processing data according to claim 1, characterized in that: The multi-source heterogeneous process data includes the placement pressure timing data of the chip mounting equipment, the temperature curve characteristics of the reflow soldering equipment, and the detection image feature vectors output by the optical inspection equipment and the X-ray inspection equipment. The detection image feature vectors include defect location features, defect size features, and grayscale features.

7. An edge computing system for electronic component processing data, characterized in that, It includes an edge gateway and an edge scheduling decision module that is communicatively connected to the edge gateway; The edge gateway is equipped with a dynamic semantic dictionary module and a causal semantic graph embedding module. The dynamic semantic dictionary module is used to maintain a semantic node mapping table corresponding to process parameters and quality indicators. The causal semantic graph embedding module is used to perform the steps of constructing a process relationship graph, calculating embedding vectors, calculating conditional mutual information, and updating edge weights in claim 1. The edge scheduling decision module includes a resource availability perception module, a task urgency assessment module, a reinforcement learning scheduling agent, and a resource depletion threshold determination unit. The resource depletion threshold determination unit is used to determine whether the overall resource availability is lower than a preset resource depletion threshold and whether the urgency score is higher than a preset urgency threshold before the reinforcement learning scheduling agent executes a regular decision. When both conditions are met, a trigger signal is sent to the pruned lightweight detection model. The resource availability perception module is used to calculate the overall resource availability of the edge nodes. The task urgency assessment module is used to calculate the urgency score of the detection task to be processed. The reinforcement learning scheduling agent is used to execute the steps of selecting candidate actions based on the environmental state vector and determining whether to trigger a downgrade to the pruned model for processing. The edge scheduling decision module also establishes data connection relationships with the edge local processing unit, the edge-to-edge collaborative processing unit, the cloud processing center, and the pruned lightweight detection model.

8. The edge computing system for electronic component processing data according to claim 7, characterized in that: The edge scheduling decision module is also equipped with a resource depletion threshold determination unit. The resource depletion threshold determination unit is used to determine whether the comprehensive resource reserve is lower than the preset resource depletion threshold and whether the urgency score is higher than the preset urgency threshold before the reinforcement learning scheduling agent executes the regular decision. When the determination result is that both conditions are met at the same time, a trigger signal is sent to the pruned lightweight detection model.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 6.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 6.

Images