Chip production logistics network optimization method and system based on graph neural network model

Abstract

The present invention provides a chip production logistics network optimization method and system based on a graph neural network model, which relates to the technical field of chip production logistics optimization, including collecting logistics data of multiple process nodes of a chip production line, calculating node criticality scores to determine node types, and constructing a multi-level logistics network graph including material transmission edges and priority-based resource competition edges; dividing logistics sub-networks, determining the correlation strength between sub-networks, constructing a sub-network state predictor and a collaborative scheduling module, and inputting the network topology structure and node features into the graph neural network for training to obtain an optimization model; predicting sub-network state changes based on the optimization model to generate resource allocation plans and node scheduling sequences, and triggering an adjustment mechanism when the predicted state exceeds a threshold, re-planning material allocation and transmission paths, and finally converting the plan into equipment control instructions and sending them to process nodes.

Description

Technical Field

[0001] The present invention relates to the technical field of chip production logistics optimization, and in particular to a chip production logistics network optimization method and system based on a graph neural network model. Background Art

[0002] Chip manufacturing is an extremely complex process involving hundreds of process steps, a large amount of materials, and complex equipment and resource scheduling. In order to ensure production efficiency and product quality, chip production logistics need to be efficiently managed and optimized. Traditional logistics management methods usually rely on experience and manual scheduling, which is difficult to adapt to the high complexity and dynamic changes of modern chip manufacturing. With the continuous advancement of chip manufacturing technology and the expansion of production scale, traditional logistics management methods are increasingly unable to meet the growing demand, resulting in problems such as extended production cycle, low resource utilization, and rising production costs.

[0003] There are still some shortcomings in the existing technology. The traditional optimization method ignores the coupling relationship between different types of networks in the logistics network, resulting in deviations between the optimization results and the actual production scenarios. The computational complexity of the centralized optimization strategy increases sharply with the increase of the network scale, and it is difficult to meet the real-time scheduling needs of large-scale chip production lines. There is a lack of systematic consideration of the priority of process nodes, and the material supply of key nodes cannot be guaranteed under resource competition, affecting production efficiency.

[0004] In summary, there is an urgent need for a chip production logistics network optimization method, which captures the coupling relationship between material transmission and resource competition by constructing a multi-level logistics network graph, uses sub-network division and collaborative scheduling mechanism to reduce the optimization complexity, introduces node priority constraints to ensure resource supply for key processes, and realizes accurate prediction and dynamic optimization of logistics network status based on a graph neural network model, so as to improve the logistics scheduling efficiency and robustness of chip production lines. The present invention can solve the problems in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a chip production logistics network optimization method and system based on a graph neural network model, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A chip production logistics network optimization method based on a graph neural network model is provided, comprising:

[0008] Collect logistics data of multiple process nodes on a chip production line; calculate the criticality score of the process node based on the logistics data, and determine the node type according to the criticality score; construct an initial logistics network diagram, establish material transmission edges between process nodes, and determine the material transmission network; establish priority-based resource competition edges based on node types, and determine the resource competition network; superimpose the material transmission network and the resource competition network to generate a multi-level logistics network diagram;

[0009] Divide the logistics subnetwork according to the material transmission network in the multi-level logistics network diagram; determine the correlation strength between the logistics subnetworks based on the resource competition network; extract the material flow characteristics of each logistics subnetwork and construct a subnetwork state predictor; establish a collaborative scheduling module using the correlation strength between adjacent subnetworks; input the topological structure and node characteristics of the logistics subnetwork into the graph neural network, perform model training with the output of the subnetwork state predictor and the information of the collaborative scheduling module as the optimization target, and obtain an optimized graph neural network model;

[0010] Based on the optimized graph neural network model, the state changes of the logistics sub-network are predicted, and the resource allocation plan is generated in combination with the connection relationship of the material transmission network; the node scheduling sequence is calculated according to the priority constraints of the resource competition network; when the predicted state exceeds the threshold, the adjustment mechanism is triggered according to the priority order of the node type, and the material allocation and transmission path of the relevant sub-network are re-planned; the resource allocation plan and node scheduling sequence are converted into equipment control instructions and sent to the corresponding process nodes.

[0011] In an optional embodiment,

[0012] Collecting logistics data of multiple process nodes on a chip production line; calculating the criticality score of the process node based on the logistics data, and determining the node type according to the criticality score includes:

[0013] Collect the material arrival time, material processing completion time and equipment utilization rate of each process node on the chip production line; calculate the time difference between the material arrival time and the material processing completion time to obtain the material processing cycle; count the material backlog per unit time, and divide the material backlog by the sampling time interval to obtain the material backlog change rate; construct a detection window, and set the length of the detection window to three times the material processing cycle calculated in the previous step; perform wavelet decomposition on the material backlog change rate within the detection window to obtain the high-frequency fluctuation coefficient and the low-frequency trend coefficient; calculate the ratio of the energy value of the high-frequency fluctuation coefficient to the total energy value, and mark the time point when the ratio exceeds the fluctuation threshold as a logistics abnormality point;

[0014] Calculate the number of occurrences of logistics anomalies within a unit time window to obtain the frequency of anomalies; extract the material backlog corresponding to the logistics anomalies, perform exponential smoothing on the material backlog, and obtain the anomaly intensity value; input the anomaly frequency and the anomaly intensity value into the recursive calculation unit to extract the anomaly evolution characteristics; perform weighted calculation on the anomaly evolution characteristics and the equipment utilization rate to obtain the node criticality score; sort the node criticality scores, mark the top one-third of the nodes as core nodes, and mark the remaining nodes as ordinary nodes to determine the node type.

[0015] In an optional embodiment,

[0016] Constructing an initial logistics network diagram, establishing material transmission edges between process nodes, and determining a material transmission network; establishing priority-based resource competition edges based on node types, and determining a resource competition network; superimposing the material transmission network and the resource competition network to generate a multi-level logistics network diagram includes:

[0017] Based on the material flow relationship between adjacent process nodes, a material transfer matrix is ​​created to construct an initial logistics network diagram; the number of material transfers between any two nodes in a unit time is counted, and the material transfer number is divided by the total material amount of the starting node to obtain the material flow ratio, and the material flow ratio is used as the weight of the material transfer edge to determine the material transfer network;

[0018] Based on the node type, the node criticality score product of the two core nodes is calculated to obtain a first competition coefficient, the node criticality score product of the core node and the common node is calculated to obtain a second competition coefficient, and the node criticality score product of the two common nodes is calculated to obtain a third competition coefficient, and the corresponding competition coefficient is used as the weight of the resource competition edge to determine the resource competition network; the material transmission network and the resource competition network are superimposed to generate a multi-level logistics network diagram;

[0019] Monitor the number of logistics anomaly points at each process node. When the number of logistics anomaly points exceeds the anomaly monitoring threshold, recalculate the anomaly frequency and anomaly intensity value of the corresponding process node; re-input the updated anomaly frequency and anomaly intensity value into the recursive calculation unit to obtain a new node criticality score; determine the node type according to the new node criticality score; based on the updated node type, recalculate the new first competition coefficient, the new second competition coefficient and the new third competition coefficient; collect the latest material transmission data to update the material flow ratio; based on the updated material flow ratio, the new first competition coefficient, the new second competition coefficient and the new third competition coefficient, regenerate a new resource competition network.

[0020] In an optional embodiment,

[0021] Divide the logistics subnetworks according to the material transmission network in the multi-level logistics network diagram; determine the correlation strength between the logistics subnetworks based on the resource competition network; extract the material flow characteristics of each logistics subnetwork and build a subnetwork state predictor; use the correlation strength between adjacent subnetworks to establish a collaborative scheduling module; input the topological structure and node characteristics of the logistics subnetwork into the graph neural network, and train the model with the output of the subnetwork state predictor and the information of the collaborative scheduling module as the optimization target, and the optimized graph neural network model includes:

[0022] Based on the material transmission network in the multi-level logistics network diagram, determine the material flow intensity between nodes and construct an affinity matrix; perform standardization on the affinity matrix to obtain a normalized Laplace matrix; analyze the eigenvalue distribution of the normalized Laplace matrix, determine the optimal number of sub-networks, and select corresponding eigenvectors to construct a feature matrix; perform clustering operations based on the feature matrix to obtain node attribution labels, and based on the node attribution labels, divide the material transmission network into a corresponding number of logistics sub-networks;

[0023] The ratio of the total material flow between adjacent logistics sub-networks to the maximum total material flow of each logistics sub-network is calculated to obtain the basic correlation strength between logistics sub-networks. Based on the pre-acquired dynamic change law of material flow, a time decay factor is introduced, and the time decay factor is multiplied by the basic correlation strength to obtain the time-varying correlation strength of the logistics sub-network.

[0024] Constructing a sub-network state predictor, the sub-network state predictor includes: collecting data on material backlog, turnover rate and load rate of nodes in each logistics sub-network, and constructing a time series feature sequence of node operation status; based on the time series feature sequence, performing a sliding window segmentation operation, extracting local time series features, and obtaining a low-dimensional feature representation through feature dimension reduction module mapping; calculating feature weights of the low-dimensional feature representation, and inputting the weighted feature into a pre-trained bidirectional long short-term memory network, outputting a node state prediction value, and calculating a state prediction deviation based on the actually collected operation state data and the node state prediction value;

[0025] Combine the node state prediction value of the current logistics sub-network, the node state prediction value of the adjacent logistics sub-network and the time-varying correlation strength to construct a cross-network feature vector, input it into the attention calculation unit, and generate the attention weight of the state interaction between the logistics sub-networks; multiply the attention weight with the node state prediction value of the adjacent logistics sub-network to obtain weighted state information; determine the state transition matrix by learning the mapping relationship between the current logistics sub-network and the adjacent logistics sub-network, map the weighted state information back to the current logistics sub-network, and generate collaborative scheduling information; compare the collaborative scheduling information with the preset scheduling target, and calculate the execution error of the collaborative scheduling information;

[0026] Extract the topological structure features and node operation status features of the logistics sub-network, and construct a node feature matrix based on the graph neural network; generate a query matrix, a key matrix and a value matrix through linear transformation; calculate the attention scores of the query matrix and the key matrix, and multiply the attention scores with the value matrix to obtain the fusion features; obtain the state prediction deviation of the sub-network state predictor and the execution error of the collaborative scheduling information, and construct an objective function containing network structure constraints; optimize the connection weights of the graph attention network based on the objective function, and output the optimized graph neural network model.

[0027] In an optional embodiment,

[0028] The time-varying correlation strength is formulated as follows:

[0029] ;

[0030] Among them, R(t) ij represents the time-varying association strength of two nodes i and i at time t, t represents time, i represents a logistics subnetwork, j represents another logistics subnetwork, k represents the time scale level, and λ k represents the weight parameter of scale k, represents the first wavelet basis function at scale k, x(·) represents the state sequence corresponding to the i logistics subnetwork, represents a time different from t, represents the second wavelet basis function at scale k, y(·) represents the state sequence corresponding to the j logistics subnetwork, represents the Laplace operator, S i represents the state distribution of the i logistics sub-network, S j represents the state distribution of the j logistics subnetwork, H(S i , S j ) represents the Wasserstein distance between two state distributions, α represents the time decay coefficient, t0 represents the initial time, and β represents the nonlinear adjustment factor;

[0031] In an optional embodiment,

[0032] Based on the optimized graph neural network model, the state changes of the logistics sub-network are predicted, and the resource allocation scheme generated by combining the connection relationship of the material transmission network includes:

[0033] The state prediction results and historical error data of the logistics sub-network are input into the graph neural network model, and the graph neural network model is used to construct the uncertainty interval of the state prediction based on the historical error data. According to the uncertainty interval, the process nodes in the logistics sub-network are divided into multiple levels of logistics nodes, including primary logistics nodes, secondary logistics nodes and buffer logistics nodes;

[0034] Based on the connection relationship of the logistics sub-network, a topological structure of the material transmission network is constructed, the time-varying transmission intensity between adjacent logistics nodes in the material transmission network is calculated, the material transmission bottleneck is determined based on the time-varying transmission intensity, and the transmission importance score of each logistics node is calculated according to the material transmission bottleneck;

[0035] According to the state prediction result and the uncertainty interval, the material demand interval of each logistics node is calculated, the material buffer capacity is configured for the first-level logistics node based on the transmission importance score, the material adjustment interval is set for the second-level logistics node, the material storage space is reserved for the buffer logistics node, and the material demand allocation model is generated;

[0036] Establish material flow constraints based on the topological structure of the material transmission network, set material priority guarantee constraints for the first-level logistics nodes, set material dynamic adjustment constraints for the second-level logistics nodes, configure material storage balance constraints for the buffer logistics nodes, and build transmission capacity constraints based on the material transmission bottleneck;

[0037] Material flow constraints, material priority constraints, material dynamic adjustment constraints, material storage balance constraints and transmission capacity constraints are constructed as resource allocation optimization objectives, and a deep reinforcement learning network is used to solve the resource allocation optimization objectives to obtain a material allocation strategy;

[0038] The robustness index of the material allocation strategy is calculated, the material allocation strategy is optimized online based on the robustness index, the material buffer capacity of the first-level logistics node is dynamically adjusted, and a material resource allocation plan is generated.

[0039] In an optional embodiment,

[0040] The scheduling sequence of computing nodes according to the priority constraints of the resource competition network includes:

[0041] Collect the connection relationship and competition relationship of the process nodes in the resource competition network, build a multi-dimensional heterogeneous graph structure, run the time-series attention network, extract the priority dependency features of the process nodes in the time dimension, run the space-series attention network, extract the resource competition features of the process nodes in the space dimension, and fuse the priority dependency features and the resource competition features to generate a node priority tensor;

[0042] Input the node priority tensor and the resource capacity information in the resource competition network into a hierarchical coding network for encoding to obtain a priority constraint vector and a resource constraint vector, and fuse the priority constraint vector and the resource constraint vector to generate a scheduling constraint graph;

[0043] Input the scheduling constraint graph into a probability coding network to obtain the probability distribution of process node execution time, predict the resource competition evolution trend between nodes based on the probability distribution, and generate a node execution time series according to the resource competition evolution trend;

[0044] An optimization function is constructed according to the scheduling constraint graph, the node execution time series is input into the optimization network to search the solution space, the scheduling strategy is extracted from the solution space search results, the scheduling strategy is input into the reinforcement learning network for optimization, an initial scheduling plan is generated, the initial scheduling plan is optimized online to obtain an optimized scheduling plan, the optimized scheduling plan is input into the graph neural network to detect priority constraint conflicts, the impact of the priority constraint conflicts is calculated, and the priority constraint conflicts are eliminated according to the impact; the scheduling plan after eliminating the priority constraint conflicts is iteratively optimized in combination with the scheduling constraint graph, and a node scheduling sequence is output.

[0045] According to a second aspect of the embodiments of the present invention,

[0046] A chip production logistics network optimization system based on a graph neural network model is provided, comprising:

[0047] The first unit is used to collect logistics data of multiple process nodes on a chip production line; calculate the criticality score of the process node based on the logistics data, and determine the node type according to the criticality score; construct an initial logistics network diagram, establish material transmission edges between process nodes, and determine the material transmission network; establish priority-based resource competition edges based on node types, and determine the resource competition network; superimpose the material transmission network and the resource competition network to generate a multi-level logistics network diagram;

[0048] The second unit is used to divide the logistics sub-network according to the material transmission network in the multi-level logistics network diagram; determine the correlation strength between the logistics sub-networks based on the resource competition network; extract the material flow characteristics of each logistics sub-network and build a sub-network state predictor; use the correlation strength between adjacent sub-networks to establish a collaborative scheduling module; input the topological structure and node characteristics of the logistics sub-network into the graph neural network, and train the model with the output of the sub-network state predictor and the information of the collaborative scheduling module as the optimization target to obtain an optimized graph neural network model;

[0049] The third unit is used to predict the state changes of the logistics sub-network based on the optimized graph neural network model, and generate a resource allocation plan based on the connection relationship of the material transmission network; calculate the node scheduling sequence according to the priority constraints of the resource competition network; when the predicted state exceeds the threshold, trigger the adjustment mechanism according to the priority order of the node type, and re-plan the material allocation and transmission path of the relevant sub-network; convert the resource allocation plan and node scheduling sequence into equipment control instructions, and send them to the corresponding process nodes.

[0050] According to a third aspect of the embodiments of the present invention,

[0051] An electronic device is provided, comprising:

[0052] processor;

[0053] a memory for storing processor-executable instructions;

[0054] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0055] A fourth aspect of the embodiments of the present invention is:

[0056] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0057] In an embodiment of the present invention, by constructing a multi-level logistics network diagram and combining it with a graph neural network model for prediction and optimization, resources can be effectively allocated and nodes can be scheduled, thereby reducing material flow time and equipment idle time, and ultimately improving chip production efficiency; based on priority resource competition edges and node types, more reasonable resource allocation can be achieved, resource conflicts and waste can be avoided, and resource utilization can be improved; through the sub-network state predictor and collaborative scheduling module, potential logistics bottlenecks and risks can be predicted, and the adjustment mechanism can be triggered in time to re-plan material distribution and transmission paths, thereby enhancing the system's ability to respond to changes and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a schematic diagram of the process of a chip production logistics network optimization method based on a graph neural network model according to an embodiment of the present invention;

[0059] Figure 2 It is a structural schematic diagram of a chip production logistics network optimization system based on a graph neural network model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0062] Figure 1 FIG. 1 is a flow chart of a chip production logistics network optimization method based on a graph neural network model according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0063] S101. Collect logistics data of multiple process nodes on the chip production line; calculate the criticality score of the process node based on the logistics data, and determine the node type according to the criticality score; construct an initial logistics network diagram, establish material transmission edges between process nodes, and determine the material transmission network; establish priority-based resource competition edges based on node types, and determine the resource competition network; superimpose the material transmission network and the resource competition network to generate a multi-level logistics network diagram;

[0064] Collect logistics data of each process node on the chip production line, including material receiving quantity, sending quantity, processing time, cache capacity, equipment utilization rate, and establish a logistics data collection table;

[0065] Calculate the criticality score of the process node based on the logistics data, calculate the node weight by analyzing the material flow frequency, processing time ratio, and equipment utilization level, and set the criticality threshold to divide the process nodes into three types: core nodes, key nodes, and ordinary nodes;

[0066] Construct an initial logistics network diagram, set process nodes as network vertices, and establish material transmission edges between adjacent process nodes based on the actual material flow path. Each transmission edge contains material transmission rate and transmission capacity attributes to form a material transmission network.

[0067] Construct resource competition relationships based on node types. Core nodes have the highest priority, followed by key nodes, and common nodes have the lowest priority. Resource competition edges are established between process nodes that share the same processing equipment or transmission equipment. The weight of the competition edge is determined by the difference in node priority.

[0068] The material transmission network and resource competition network are superimposed and integrated to generate a multi-level logistics network diagram that includes material transmission relationships and resource competition relationships, thereby realizing global modeling of the chip production line logistics system.

[0069] In this embodiment, by collecting logistics data of multiple process nodes and calculating the criticality scores of the nodes, it is possible to more accurately identify key nodes in the production process, optimize resource allocation and logistics routes, and reduce unnecessary resource waste and logistics delays; by superimposing the material transmission network and the resource competition network, a multi-level logistics network diagram is generated, which can fully reflect the multi-level relationships and competition between nodes in the logistics network, and help solve coordination problems in complex logistics systems; priority resource competition edges are established based on criticality scores and node types, which can achieve priority sorting of resource scheduling, avoid low-priority nodes occupying key resources, and improve the overall scheduling efficiency of the system; by constructing an initial logistics network diagram and comprehensively considering material transmission and resource competition between nodes, the performance of the logistics network in a dynamic environment can be better analyzed and optimized, and the stability and anti-interference ability of the production system can be improved.

[0070] S102. Divide the logistics subnetwork according to the material transmission network in the multi-level logistics network diagram; determine the correlation strength between the logistics subnetworks based on the resource competition network; extract the material flow characteristics of each logistics subnetwork and construct a subnetwork state predictor; establish a collaborative scheduling module using the correlation strength between adjacent subnetworks; input the topological structure and node characteristics of the logistics subnetwork into the graph neural network, perform model training with the output of the subnetwork state predictor and the information of the collaborative scheduling module as the optimization target, and obtain an optimized graph neural network model;

[0071] In this embodiment, by dividing the logistics sub-networks and extracting features for each sub-network, the analysis and optimization process of complex logistics networks can be simplified, and the scalability and computational efficiency of the model can be improved; by using the resource competition network to determine the correlation strength between sub-networks, and coordinating the interactions between sub-networks through the collaborative scheduling module, the global resource allocation can be optimized, and resource conflicts and scheduling delays between sub-networks can be reduced; based on the logistics sub-network state predictor, combined with the powerful modeling capabilities of the graph neural network, the state changes of the sub-network can be more accurately predicted, providing reliable data support for subsequent optimization; through the training of the graph neural network model, with the sub-network state and collaborative scheduling as the optimization goals, real-time adjustment and optimization can be achieved in a dynamically changing production environment, thereby improving the overall efficiency and stability of the logistics network; the topological structure and node characteristics of the logistics sub-network are incorporated into the optimization process, and the network structure information is fully utilized, so that the model can better capture the complex relationships in the logistics system and improve the accuracy and reliability of the optimization effect.

[0072] S103. Based on the optimized graph neural network model, the state changes of the logistics sub-network are predicted, and the resource allocation plan is generated in combination with the connection relationship of the material transmission network; the node scheduling sequence is calculated according to the priority constraints of the resource competition network; when the predicted state exceeds the threshold, the adjustment mechanism is triggered according to the priority order of the node type, and the material allocation and transmission path of the relevant sub-network are re-planned; the resource allocation plan and node scheduling sequence are converted into equipment control instructions, and sent to the corresponding process nodes.

[0073] The optimized graph neural network model is applied to the logistics sub-network status prediction. Based on the material cache, equipment status and processing progress of the process nodes in the sub-network, the material demand and output changes of each process node in the future time window are predicted.

[0074] According to the connection relationship of the material transmission network, the material transmission volume and transmission timing between adjacent process nodes are calculated, and combined with the node material demand prediction results, a resource allocation plan that meets the material balance constraints is generated; according to the priority constraints of the nodes in the resource competition network, the process nodes competing for the same resources are sorted in time, and a node scheduling sequence is generated to ensure that the resource requirements of core nodes and key nodes are met first;

[0075] Monitor the status changes of the logistics sub-network in real time. When the predicted status exceeds the preset threshold, trigger the network adjustment mechanism according to the priority order of the node type, give priority to ensuring the material supply of high-priority nodes, and re-plan the material allocation strategy and transmission path of the affected sub-network;

[0076] The resource allocation plan and node scheduling sequence are converted into standard equipment control instructions, including material transfer instructions, processing start and stop instructions, and equipment scheduling instructions, which are sent to the corresponding process nodes for execution through the control system.

[0077] In this embodiment, by predicting the state changes of the logistics sub-network based on the optimized graph neural network model, potential anomalies or bottlenecks can be perceived in advance in the production process, providing a basis for dynamic adjustment; a resource allocation plan is generated in combination with the connection relationship of the material transmission network, and a scheduling sequence is calculated based on the priority constraints of the resource competition network, so that efficient resource allocation and priority scheduling of nodes can be achieved, reducing conflicts and delays; when the predicted state exceeds the threshold, an adjustment mechanism can be triggered to dynamically plan material allocation and transmission paths to ensure the continued efficient operation of the logistics network under abnormal circumstances; the resource allocation plan and scheduling sequence are converted into equipment control instructions and sent to process nodes to directly guide the operation of production equipment, realizing full-process intelligence from optimization decision-making to actual execution; by combining prediction, adjustment and control, the state of the sub-network and resource allocation can be fully coordinated to maximize the efficiency and stability of the overall logistics network.

[0078] In an optional implementation, collecting logistics data of multiple process nodes on a chip production line; calculating the criticality score of the process node based on the logistics data, and determining the node type according to the criticality score includes:

[0079] Collect the material arrival time, material processing completion time and equipment utilization rate of each process node on the chip production line; calculate the time difference between the material arrival time and the material processing completion time to obtain the material processing cycle; count the material backlog per unit time, and divide the material backlog by the sampling time interval to obtain the material backlog change rate; construct a detection window, and set the length of the detection window to three times the material processing cycle calculated in the previous step; perform wavelet decomposition on the material backlog change rate within the detection window to obtain the high-frequency fluctuation coefficient and the low-frequency trend coefficient; calculate the ratio of the energy value of the high-frequency fluctuation coefficient to the total energy value, and mark the time point when the ratio exceeds the fluctuation threshold as a logistics abnormality point;

[0080] Calculate the number of occurrences of logistics anomalies within a unit time window to obtain the frequency of anomalies; extract the material backlog corresponding to the logistics anomalies, perform exponential smoothing on the material backlog, and obtain the anomaly intensity value; input the anomaly frequency and the anomaly intensity value into the recursive calculation unit to extract the anomaly evolution characteristics; perform weighted calculation on the anomaly evolution characteristics and the equipment utilization rate to obtain the node criticality score; sort the node criticality scores, mark the top one-third of the nodes as core nodes, and mark the remaining nodes as ordinary nodes to determine the node type.

[0081] In an optional implementation, first, the material arrival time, material processing completion time, and equipment utilization data of each process node on the chip production line are collected. For example, at a certain time point, the material arrival time of node A is 8 am, the processing completion time is 10 am, and the equipment utilization rate is 80%.

[0082] Next, the time difference between the material arrival time and the material completion time is calculated to obtain the material processing cycle. For example, the material processing cycle of node A is 2 hours.

[0083] Then, count the amount of material backlog per unit time. For example, within 1 hour, node A has a backlog of 50 materials. Divide the amount of material backlog by the sampling time interval to get the material backlog change rate. For example, if the sampling time interval is 1 hour, the material backlog change rate of node A is 50 per hour.

[0084] Construct a detection window with a length set to three times the material processing cycle. For example, if the material processing cycle of node A is 2 hours, the detection window length is 6 hours.

[0085] The material backlog change rate in the detection window is subjected to wavelet decomposition to obtain the high-frequency fluctuation coefficient and the low-frequency trend coefficient. Wavelet decomposition can be understood as decomposing the backlog change rate into a fast-changing part (high frequency) and a slow-changing part (low frequency).

[0086] Calculate the ratio of the energy value of the high-frequency fluctuation coefficient to the total energy value. If the ratio exceeds the preset fluctuation threshold (for example, 0.5), the time point is marked as a logistics anomaly point. This indicates that the backlog change rate at this time point fluctuates greatly, and there may be logistics anomalies.

[0087] Calculate the number of times a logistics anomaly occurs within a unit time window (for example, 1 day) to get the frequency of anomaly occurrence. For example, if node A has 3 logistics anomalies in one day, the frequency of anomaly occurrence is 3 times / day.

[0088] Extract the material backlog corresponding to the logistics anomaly point. Perform exponential smoothing on these material backlogs to obtain the anomaly intensity value. Exponential smoothing can be understood as a weighted average of the backlog over a period of time to reduce the impact of short-term fluctuations.

[0089] The anomaly frequency and anomaly intensity values ​​are input into the recursive calculation unit to extract the anomaly evolution characteristics. The recursive calculation unit can be understood as a cyclic calculation process. Through continuous iterative calculation, the law of the change of anomaly frequency and anomaly intensity values ​​over time is extracted.

[0090] The abnormal evolution characteristics and equipment utilization are weighted to obtain the node criticality score. For example, the weight of the abnormal evolution characteristics is 0.7, and the weight of the equipment utilization is 0.3.

[0091] Sort the criticality scores of all nodes. Mark the top one-third of the nodes as core nodes and the rest as ordinary nodes to determine the node type.

[0092] In this embodiment, by identifying and resolving logistics anomalies, material backlogs and production delays can be reduced, thereby improving chip production efficiency; by optimizing logistics processes, material waste and equipment idle time can be reduced, thereby reducing production costs; and by focusing on monitoring and managing key nodes, the stability and reliability of the production line can be improved.

[0093] In an optional implementation, constructing an initial logistics network diagram, establishing material transmission edges between process nodes, and determining a material transmission network; establishing priority-based resource competition edges based on node types, and determining a resource competition network; and superimposing the material transmission network and the resource competition network to generate a multi-level logistics network diagram includes:

[0094] Based on the material flow relationship between adjacent process nodes, a material transfer matrix is ​​created to construct an initial logistics network diagram; the number of material transfers between any two nodes in a unit time is counted, and the material transfer number is divided by the total material amount of the starting node to obtain the material flow ratio, and the material flow ratio is used as the weight of the material transfer edge to determine the material transfer network;

[0095] Based on the node type, the node criticality score product of the two core nodes is calculated to obtain a first competition coefficient, the node criticality score product of the core node and the common node is calculated to obtain a second competition coefficient, and the node criticality score product of the two common nodes is calculated to obtain a third competition coefficient, and the corresponding competition coefficient is used as the weight of the resource competition edge to determine the resource competition network; the material transmission network and the resource competition network are superimposed to generate a multi-level logistics network diagram;

[0096] Monitor the number of logistics anomaly points at each process node. When the number of logistics anomaly points exceeds the anomaly monitoring threshold, recalculate the anomaly frequency and anomaly intensity value of the corresponding process node; re-input the updated anomaly frequency and anomaly intensity value into the recursive calculation unit to obtain a new node criticality score; determine the node type according to the new node criticality score; based on the updated node type, recalculate the new first competition coefficient, the new second competition coefficient and the new third competition coefficient; collect the latest material transmission data to update the material flow ratio; based on the updated material flow ratio, the new first competition coefficient, the new second competition coefficient and the new third competition coefficient, regenerate a new resource competition network.

[0097] In an optional implementation, first, construct an initial logistics network diagram and a material transfer network. Treat each process node as a node in the network, and create a material transfer matrix based on the material flow relationship between adjacent process nodes. For example, assuming that there are three process nodes A, B, and C, the material transfer volume from A to B is 100, the material transfer volume from A to C is 50, and the material transfer volume from B to C is 200, then a 3x3 material transfer matrix can be constructed. Then, count the number of material transfers between any two nodes in unit time, for example, 100 units of material are transferred from A to B per hour. Divide the number of material transfers by the total amount of material at the starting node, for example, node A outputs a total of 150 units of material per hour, then the material flow ratio from A to B is 100 / 150=0.67. The calculated material flow ratio is used as the weight of the material transfer edge to determine the material transfer network.

[0098] Next, construct a resource competition network. According to the node type (such as core node and ordinary node), calculate the competition coefficient between nodes. First, it is necessary to determine the criticality score of the node. The criticality score of the node can be calculated based on factors such as the node's throughput and the number of connections. For example, if the throughput of node A is 1000 and the number of connections is 5, then the criticality score of node A can be calculated as 10 according to the preset rules. Then, calculate the product of the criticality scores of the two core nodes as the first competition coefficient, the product of the criticality scores of the core node and the ordinary node as the second competition coefficient, and the product of the criticality scores of the two ordinary nodes as the third competition coefficient. Use these competition coefficients as the weights of the resource competition edges to determine the resource competition network. For example, node A and node B are both core nodes, and their criticality scores are 10 and 8 respectively, then the first competition coefficient between them is 10×8=80.

[0099] Then, the material transmission network and resource competition network are superimposed to generate a multi-level logistics network diagram. The multi-level logistics network diagram comprehensively considers the two aspects of material transmission and resource competition, and can more comprehensively reflect the operating status of the logistics system.

[0100] Finally, dynamic adjustments are made. Continuously monitor the number of logistics anomalies at each process node, such as delayed delivery, material damage, etc. When the number of logistics anomalies exceeds the preset anomaly monitoring threshold, recalculate the anomaly frequency and anomaly intensity value of the corresponding process node. For example, if the number of anomalies at node A exceeds the threshold, the number of anomalies at node A and the loss caused by each anomaly are recounted, and the new anomaly frequency and anomaly intensity values ​​are calculated according to the preset rules. The updated anomaly frequency and anomaly intensity values ​​are re-entered into the criticality score calculation unit to obtain a new node criticality score. Re-determine the node type based on the new node criticality score, and recalculate the competition coefficient based on the updated node type. At the same time, collect the latest material transfer data to update the material flow ratio. Finally, based on the updated material flow ratio and competition coefficient, regenerate a new resource competition network and multi-level logistics network diagram.

[0101] In this embodiment, by analyzing the material transmission network, logistics bottlenecks can be identified, material transmission paths can be optimized, and material residence time can be reduced, thereby improving logistics efficiency; by analyzing the resource competition network, resource competition hotspots can be identified, resource allocation strategies can be optimized, resource conflicts can be avoided, and resource utilization can be improved; by dynamically adjusting the network structure, logistics strategies can be adjusted in a timely manner according to abnormal situations, thereby avoiding abnormal situations from having a significant impact on the logistics system and enhancing the robustness of the system.

[0102] In an optional implementation, the logistics subnetworks are divided according to the material transmission network in the multi-level logistics network diagram; the association strength between the logistics subnetworks is determined based on the resource competition network; the material flow characteristics are extracted for each logistics subnetwork, and a subnetwork state predictor is constructed; the association strength between adjacent subnetworks is used to establish a collaborative scheduling module; the topological structure and node characteristics of the logistics subnetwork are input into the graph neural network, and the model is trained with the output of the subnetwork state predictor and the information of the collaborative scheduling module as the optimization target, and the optimized graph neural network model includes:

[0103] Based on the material transmission network in the multi-level logistics network diagram, determine the material flow intensity between nodes and construct an affinity matrix; perform standardization on the affinity matrix to obtain a normalized Laplace matrix; analyze the eigenvalue distribution of the normalized Laplace matrix, determine the optimal number of sub-networks, and select corresponding eigenvectors to construct a feature matrix; perform clustering operations based on the feature matrix to obtain node attribution labels, and based on the node attribution labels, divide the material transmission network into a corresponding number of logistics sub-networks;

[0104] The ratio of the total material flow between adjacent logistics sub-networks to the maximum total material flow of each logistics sub-network is calculated to obtain the basic correlation strength between logistics sub-networks. Based on the pre-acquired dynamic change law of material flow, a time decay factor is introduced, and the time decay factor is multiplied by the basic correlation strength to obtain the time-varying correlation strength of the logistics sub-network.

[0105] Constructing a sub-network state predictor, the sub-network state predictor includes: collecting data on material backlog, turnover rate and load rate of nodes in each logistics sub-network, and constructing a time series feature sequence of node operation status; based on the time series feature sequence, performing a sliding window segmentation operation, extracting local time series features, and obtaining a low-dimensional feature representation through feature dimension reduction module mapping; calculating feature weights of the low-dimensional feature representation, and inputting the weighted feature into a pre-trained bidirectional long short-term memory network, outputting a node state prediction value, and calculating a state prediction deviation based on the actually collected operation state data and the node state prediction value;

[0106] Combine the node state prediction value of the current logistics sub-network, the node state prediction value of the adjacent logistics sub-network and the time-varying correlation strength to construct a cross-network feature vector, input it into the attention calculation unit, and generate the attention weight of the state interaction between the logistics sub-networks; multiply the attention weight with the node state prediction value of the adjacent logistics sub-network to obtain weighted state information; determine the state transition matrix by learning the mapping relationship between the current logistics sub-network and the adjacent logistics sub-network, map the weighted state information back to the current logistics sub-network, and generate collaborative scheduling information; compare the collaborative scheduling information with the preset scheduling target, and calculate the execution error of the collaborative scheduling information;

[0107] Extract the topological structure features and node operation status features of the logistics sub-network, and construct a node feature matrix based on the graph neural network; generate a query matrix, a key matrix and a value matrix through linear transformation; calculate the attention scores of the query matrix and the key matrix, and multiply the attention scores with the value matrix to obtain the fusion features; obtain the state prediction deviation of the sub-network state predictor and the execution error of the collaborative scheduling information, and construct an objective function containing network structure constraints; optimize the connection weights of the graph attention network based on the objective function, and output the optimized graph neural network model.

[0108] The affinity matrix is ​​specifically used to quantify the similarity or degree of association between nodes in the system. It is a symmetric matrix in which each element represents the strength of association or similarity between two nodes. In a logistics network, the affinity matrix is ​​usually calculated based on the material flow intensity, reflecting the logistics relationship between nodes in the network.

[0109] The intelligent collaborative scheduling method of multi-level logistics networks can effectively improve the overall operating efficiency and resource utilization of logistics networks. The core lies in decomposing complex logistics networks into multiple sub-networks, considering the mutual influence between sub-networks, and conducting collaborative scheduling.

[0110] First, according to the material transmission path in the multi-level logistics network, the material flow intensity between nodes is determined. For example, if there are 1,000 pieces of goods transported between nodes A and B every day, their material flow intensity is 1,000. The material flow intensity between all nodes is sorted into a matrix, called the affinity matrix. Then, the affinity matrix is ​​standardized to obtain the normalized Laplace matrix. By analyzing the eigenvalue distribution of the normalized Laplace matrix, it can be determined how many subnetworks are most appropriate to divide the logistics network into. Assuming that the analysis results show that it is best to divide the network into 3 subnetworks, the corresponding 3 eigenvectors are selected to form a feature matrix. Finally, a clustering operation is performed based on the feature matrix to obtain the subnetwork label to which each node belongs. For example, the label of node A is 1, indicating that it belongs to subnetwork 1; the label of node B is 2, indicating that it belongs to subnetwork 2. In this way, the subnetwork division of the material transmission network is completed.

[0111] Next, determine the strength of association between subnetworks. For example, 5,000 pieces of goods are exchanged between subnetwork 1 and subnetwork 2 every day, subnetwork 1 itself handles 8,000 pieces of goods every day, subnetwork 2 itself handles 7,000 pieces of goods every day, and subnetwork 3 itself handles 9,000 pieces of goods every day. Then the basic association strength between subnetwork 1 and subnetwork 2 is the ratio of 5,000 to the largest cargo handling volume of 9,000 among the three subnetworks. Considering the dynamic changes in material flow, a time decay factor is introduced. For example, if the material flow volume in the current time period increases by 20% compared with the previous time period, the time decay factor can be set to 1.2. Multiply the basic association strength by the time decay factor to obtain the time-varying association strength between subnetworks.

[0112] Then, a subnetwork state predictor is constructed. For example, the material backlog, turnover rate, and load rate data of each node in subnetwork 1 are collected to construct a time series feature sequence of the node's operating status. The time series feature sequence is segmented by a sliding window to extract local time series features. For example, a sliding window of length 10 is used to take the feature data of the past 10 time points as a local time series feature. The high-dimensional local time series features are mapped into low-dimensional feature representations through a feature dimensionality reduction module. The feature weights of the low-dimensional feature representation are calculated, and the weighted features are input into a pre-trained bidirectional long short-term memory network to output the node state prediction value. The predicted value is compared with the actual collected operating status data to calculate the state prediction deviation.

[0113] In order to achieve collaborative scheduling between sub-networks, it is necessary to build a collaborative scheduling module. The node state prediction value of the current sub-network, the node state prediction value of the adjacent sub-network, and the time-varying correlation strength between them are combined into a cross-network feature vector, which is input into the attention calculation unit to generate the attention weight of the state interaction between sub-networks. The attention weight is multiplied by the node state prediction value of the adjacent sub-network to obtain the weighted state information. By learning the mapping relationship between the current sub-network and the adjacent sub-network, the state transition matrix is ​​determined, and the weighted state information is mapped back to the current sub-network to generate collaborative scheduling information. The collaborative scheduling information is compared with the preset scheduling target, and the execution error of the collaborative scheduling information is calculated.

[0114] Finally, extract the topological structure features and node operation status features of the logistics sub-network, and build a node feature matrix based on the graph neural network. Generate the query matrix, key matrix, and value matrix through linear transformation. Calculate the attention score of the query matrix and the key matrix, and multiply the attention score with the value matrix to obtain the fusion feature. Obtain the state prediction deviation of the sub-network state predictor and the execution error of the collaborative scheduling information, and build an objective function containing network structure constraints. Optimize the connection weights of the graph attention network based on the objective function, and output the optimized graph neural network model.

[0115] In this embodiment, by constructing a sub-network state predictor and a collaborative scheduling module, the operating state of the sub-network can be predicted more accurately, providing a more reliable basis for scheduling decisions; by considering the correlation strength and state interaction between sub-networks, more effective collaborative scheduling can be achieved, resource allocation can be optimized, and the overall operating efficiency of the logistics network can be improved; through the learning and optimization of graph neural networks, the model can adapt to complex network structures and dynamically changing material flows, thereby enhancing the robustness of the scheduling solution.

[0116] In an optional implementation manner, the time-varying correlation strength is expressed as follows:

[0117] ;

[0118] Among them, R(t) ij represents the time-varying association strength of two nodes i and i at time t, t represents time, i represents a logistics subnetwork, j represents another logistics subnetwork, k represents the time scale level, and λ k represents the weight parameter of scale k, represents the first wavelet basis function at scale k, x(·) represents the state sequence corresponding to the i logistics subnetwork, represents a time different from t, represents the second wavelet basis function at scale k, y(·) represents the state sequence corresponding to the j logistics subnetwork, represents the Laplace operator, S i represents the state distribution of the i logistics sub-network, S j represents the state distribution of the j logistics subnetwork, H(S i , S j ) represents the Wasserstein distance between two state distributions, α represents the time decay coefficient, t0 represents the initial time, and β represents the nonlinear adjustment factor;

[0119] The formula is used to calculate the time-varying association strength between two logistics sub-networks, which can more accurately capture the dynamic relationship between them.

[0120] First, we need to obtain the state sequence data of each logistics subnetwork. For example, the state sequence of a logistics subnetwork can be the number of packages it handles daily, or the average driving speed of its vehicles, etc. Suppose we have two logistics subnetworks A and B, and we obtain the number of packages they handled daily in the past year as state sequence data. The state sequence of subnetwork A is [1000, 1200, 1100, ..., 3000], and the state sequence of subnetwork B is [500, 600, 550, ..., 1500].

[0121] Next, multi-scale wavelet decomposition is performed on the state sequence of each logistics sub-network. Wavelet decomposition can decompose a time series into fluctuation components at different time scales. For example, we can decompose the state sequences of A and B into short-term (daily), medium-term (weekly) and long-term (monthly) fluctuation components respectively.

[0122] Then, the correlation between the two sub-network state sequences at each time scale is calculated. The wavelet correlation coefficient is used here, which can measure the degree of correlation between two sequences at different time scales. For example, we can calculate the wavelet correlation coefficient between A and B at the daily, weekly, and monthly scales. Assume that the wavelet correlation coefficient between A and B at the daily scale is 0.8, the wavelet correlation coefficient at the weekly scale is 0.6, and the wavelet correlation coefficient at the monthly scale is 0.4.

[0123] Next, we weight the correlations at different time scales and get a comprehensive correlation strength. The weight parameter reflects the contribution of different time scales to the overall correlation strength. For example, we can set the weight of the daily scale to 0.5, the weight of the weekly scale to 0.3, and the weight of the monthly scale to 0.2. Then the comprehensive correlation strength between A and B is 0.8×0.5+0.6×0.3+0.4×0.2=0.66.

[0124] At the same time, the distance between the state distributions of the two sub-networks is calculated. The state distribution can reflect the overall operation status of the sub-network. Here, the Wasserstein distance is used to measure the degree of difference between the two distributions. For example, the Wasserstein distance between the package quantity distributions of A and B can be calculated. Assume that the calculated distance is 0.2.

[0125] Finally, the final time-varying correlation strength is calculated by combining the comprehensive correlation strength, state distribution distance and time decay factor. The time decay factor reflects the influence of historical data on the current correlation strength. For example, assuming that the current time is t, the initial time is t0, the time decay coefficient is 0.9, and the nonlinear adjustment factor is 1.2, the final time-varying correlation strength can be expressed as 0.66×(1-0.2)×0.9 (t-t0) ×1.2=0.47×0.9 (t-t0) .

[0126] In this embodiment, the dynamic correlation between logistics sub-networks can be more accurately characterized because it takes into account the correlation at different time scales and the differences in state distribution; it can provide a more refined correlation strength analysis because it can capture the changing trend of correlation strength over time; it can support more effective logistics network management and optimization because it can help identify key logistics sub-networks and the mutual influence relationships between them.

[0127] In an optional implementation, predicting the state change of the logistics sub-network based on the optimized graph neural network model and generating a resource allocation plan in combination with the connection relationship of the material transmission network includes:

[0128] The state prediction results and historical error data of the logistics sub-network are input into the graph neural network model, and the graph neural network model is used to construct the uncertainty interval of the state prediction based on the historical error data. According to the uncertainty interval, the process nodes in the logistics sub-network are divided into multiple levels of logistics nodes, including primary logistics nodes, secondary logistics nodes and buffer logistics nodes;

[0129] Based on the connection relationship of the logistics sub-network, a topological structure of the material transmission network is constructed, the time-varying transmission intensity between adjacent logistics nodes in the material transmission network is calculated, the material transmission bottleneck is determined based on the time-varying transmission intensity, and the transmission importance score of each logistics node is calculated according to the material transmission bottleneck;

[0130] According to the state prediction result and the uncertainty interval, the material demand interval of each logistics node is calculated, the material buffer capacity is configured for the first-level logistics node based on the transmission importance score, the material adjustment interval is set for the second-level logistics node, the material storage space is reserved for the buffer logistics node, and the material demand allocation model is generated;

[0131] Establish material flow constraints based on the topological structure of the material transmission network, set material priority guarantee constraints for the first-level logistics nodes, set material dynamic adjustment constraints for the second-level logistics nodes, configure material storage balance constraints for the buffer logistics nodes, and build transmission capacity constraints based on the material transmission bottleneck;

[0132] Material flow constraints, material priority constraints, material dynamic adjustment constraints, material storage balance constraints and transmission capacity constraints are constructed as resource allocation optimization objectives, and a deep reinforcement learning network is used to solve the resource allocation optimization objectives to obtain a material allocation strategy;

[0133] The robustness index of the material allocation strategy is calculated, the material allocation strategy is optimized online based on the robustness index, the material buffer capacity of the first-level logistics node is dynamically adjusted, and a material resource allocation plan is generated.

[0134] The uncertainty interval specifically refers to a mathematical representation used to characterize the possible range of values ​​of a variable, reflecting the magnitude of changes in the prediction results or parameter estimates due to uncertainty factors. In the prediction of the state of the logistics subnetwork, the uncertainty interval is constructed based on the historical error data and the output of the prediction model to describe the upper and lower limits of the true value of the state of the logistics subnetwork. In this way, the credibility and stability of the prediction results can be quantified. The upper limit of the uncertainty interval represents the maximum value of the prediction result under the current conditions, which usually corresponds to a more optimistic estimate; the lower limit represents the possible minimum value, corresponding to a more conservative estimate. The width of the interval reflects the size of the uncertainty, that is, the credibility level of the prediction result. The smaller the width, the higher the reliability of the prediction result; the larger the width, the higher the uncertainty. The main role of the uncertainty interval is reflected in many aspects. First, it provides a basis for resource planning and node priority division of the logistics network. By referring to the uncertainty interval, the process nodes can be divided into logistics nodes of different levels, so that resource allocation and scheduling are more targeted. Secondly, the uncertainty interval can be used as an important basis for dynamic optimization, helping to adjust the material demand configuration in real time to ensure that the logistics system has sufficient robustness and adaptability when the environment changes.

[0135] In an optional implementation, an optimized graph neural network model is used to predict the future state of each node in the logistics subnetwork. The model can not only predict the node state, but also combine historical error data to evaluate the uncertainty of the prediction results and generate a range of state changes rather than a single point value. For example, the material demand for a node in the next hour is not predicted to be 100 pieces, but a demand range of 90 to 110 pieces.

[0136] Then, the nodes in the logistics subnetwork are graded according to the predicted uncertainty interval. Nodes with high uncertainty are classified as first-level logistics nodes, which need to be configured with a higher material buffer capacity to cope with fluctuations; nodes with medium uncertainty are classified as second-level logistics nodes, which allow their material demand to be adjusted within a certain range; nodes with low uncertainty are classified as buffer logistics nodes, which only require a certain amount of material storage space to be reserved. For example, a raw material supply node has a high uncertainty in material demand forecast due to large fluctuations in the upstream supply chain, so it is classified as a first-level logistics node; a production line node has relatively stable material demand, so it is classified as a second-level logistics node; the finished product warehouse node has the smallest demand fluctuation, so it is classified as a buffer logistics node.

[0137] Next, the connection relationship of the logistics sub-network is analyzed to construct the topological structure of the material transmission network. The time-varying transmission intensity between adjacent nodes is calculated. For example, the transmission capacity from node A to node B may be different in different time periods. In the morning, there are more transport vehicles and the transmission intensity is high, while in the afternoon, there are fewer vehicles and the transmission intensity is low. By analyzing the transmission intensity, the material transmission bottleneck is identified. For example, the transmission capacity of some connecting lines is insufficient, resulting in material backlogs. According to the material transmission bottleneck, the transmission importance score of each node is calculated. The upstream nodes of the bottleneck route have higher scores because the material supply of these nodes has a greater impact on the entire network.

[0138] According to the state prediction results and uncertainty interval of the node, the material demand interval of each node is calculated. Different material strategies are configured for nodes of different levels in combination with the transmission importance score of the node. A higher material buffer capacity is configured for the first-level logistics node. For example, if the predicted demand interval is 90 to 110 pieces, a buffer capacity of 120 pieces is configured. A material adjustment interval is set for the second-level logistics node. For example, its material demand is allowed to fluctuate between 80 and 120 pieces. Material storage space is reserved for the buffer logistics node. For example, if the predicted demand is 100 pieces, storage space for 110 pieces is reserved. Finally, a material demand allocation model is generated.

[0139] According to the topological structure of the material transmission network and the material demand of the nodes, a series of constraints are established, including material flow constraints, for example, the output of node A must be equal to the sum of the inputs of the connected nodes; material priority guarantee constraints, giving priority to meeting the needs of the first-level logistics nodes; material dynamic adjustment constraints, allowing the needs of the second-level logistics nodes to be adjusted within a certain range; material storage balance constraints, the storage capacity of the buffer logistics node is maintained at a reasonable level; transmission capacity constraints, the transmission volume of any connecting line cannot exceed its capacity limit.

[0140] The above constraints are constructed as resource allocation optimization objectives, and the optimization objectives are solved using a deep reinforcement learning network. The deep reinforcement learning network learns the best material allocation strategy by continuously simulating the material flow and allocation process.

[0141] Finally, the robustness index of the material allocation strategy is calculated. For example, the material flow in different scenarios is simulated to evaluate the stability of the strategy. Based on the robustness index, the strategy is optimized online. For example, if the buffer capacity of a first-level logistics node is not enough to cope with fluctuations, its buffer capacity is dynamically increased. Finally, a robust material resource allocation plan is generated.

[0142] In this embodiment, by combining historical error data and uncertainty intervals, the prediction results are closer to the actual situation, reducing the risks brought by prediction errors; material resources are configured differently according to the importance of the nodes and demand fluctuations, avoiding waste and shortage of resources and improving resource utilization efficiency; by considering various constraints and online optimization mechanisms, the material allocation plan can adapt to various emergencies and changes, ensuring the stable operation of the logistics network.

[0143] In an optional implementation, computing a node scheduling sequence according to a priority constraint of a resource contention network includes:

[0144] Collect the connection relationship and competition relationship of the process nodes in the resource competition network, build a multi-dimensional heterogeneous graph structure, run the time-series attention network, extract the priority dependency features of the process nodes in the time dimension, run the space-series attention network, extract the resource competition features of the process nodes in the space dimension, and fuse the priority dependency features and the resource competition features to generate a node priority tensor;

[0145] Input the node priority tensor and the resource capacity information in the resource competition network into a hierarchical coding network for encoding to obtain a priority constraint vector and a resource constraint vector, and fuse the priority constraint vector and the resource constraint vector to generate a scheduling constraint graph;

[0146] Input the scheduling constraint graph into a probability coding network to obtain the probability distribution of process node execution time, predict the resource competition evolution trend between nodes based on the probability distribution, and generate a node execution time series according to the resource competition evolution trend;

[0147] An optimization function is constructed according to the scheduling constraint graph, the node execution time series is input into the optimization network to search the solution space, the scheduling strategy is extracted from the solution space search results, the scheduling strategy is input into the reinforcement learning network for optimization, an initial scheduling plan is generated, the initial scheduling plan is optimized online to obtain an optimized scheduling plan, the optimized scheduling plan is input into the graph neural network to detect priority constraint conflicts, the impact of the priority constraint conflicts is calculated, and the priority constraint conflicts are eliminated according to the impact; the scheduling plan after eliminating the priority constraint conflicts is iteratively optimized in combination with the scheduling constraint graph, and a node scheduling sequence is output.

[0148] Collect resource competition network information and construct a multi-dimensional heterogeneous graph. First, collect the connection relationship of process nodes, such as node A is connected to node B, node B is connected to node C, and so on. Then, collect the competition relationship of process nodes, such as node A and node B compete for resource X, node C and node D compete for resource Y, and so on. Resources can be of various types, and the collected connection relationship and competition relationship data are used to construct a multi-dimensional heterogeneous graph structure. The nodes of this graph represent process nodes, and the edges represent the connection or competition relationship between nodes. The type of edge can be used to distinguish between connection relationships and competition relationships, and can also be used to represent different types of resource competition. For example, a solid line represents a connection relationship, a dotted line represents a competition relationship, and different colors represent different types of resource competition. Suppose there are four nodes A, B, C, and D. A is connected to B, B is connected to C, A and B compete for resource X, and C and D compete for resource Y. In the constructed graph, there is a solid line edge between A and B, a solid line edge between B and C, a dotted line edge between A and B to represent competition for resource X, and a dotted line edge between C and D to represent competition for resource Y.

[0149] Extract node priority features. Based on the constructed multi-dimensional heterogeneous graph, the temporal attention network is used to extract the priority dependency features of the process nodes in the time dimension. For example, if node A must be executed before node B, node A takes precedence over node B in the time dimension. Then use the spatial attention network to extract the resource competition features of the process nodes in the spatial dimension. For example, if nodes A and B compete for the same resources, they have a competitive relationship in the spatial dimension. The time dimension priority dependency features and the spatial dimension resource competition features are fused to generate a node priority tensor. This tensor contains the priority information of each node and can be used for subsequent scheduling decisions. For example, if node A has a higher time priority and less resource competition with other nodes, its priority tensor value will be higher. Assuming that the time priority of node A is 0.8 and the spatial competition degree is 0.2, the fused priority tensor value is [0.8, 0.2].

[0150] Generate a scheduling constraint graph. Input the node priority tensor and the resource capacity information in the resource competition network into the hierarchical coding network for encoding to obtain a priority constraint vector and a resource constraint vector. Resource capacity information, for example, the capacity of resource X is 10, the capacity of resource Y is 5, and so on. Fuse the priority constraint vector and the resource constraint vector to generate a scheduling constraint graph. The scheduling constraint graph contains the priority constraints and resource constraint information of the node, which can be used to guide the scheduling of the node. For example, if node A must be executed before node B, and both nodes A and B compete for resource X, the scheduling constraint graph will reflect this constraint relationship. Assuming that the priority constraint vector of node A is [0.9] and the resource constraint vector is [0.5], the fused scheduling constraint vector is [0.9, 0.5].

[0151] Predict node execution time series. Input the scheduling constraint graph into the probability coding network to obtain the probability distribution of process node execution time. For example, the execution time of node A may be 1 to 3 time units, with probabilities of 0.2, 0.5, and 0.3, respectively. Predict the resource competition evolution trend between nodes based on the probability distribution. For example, if the probability of nodes A and B competing for resource X in the same time period is high, there is a potential resource conflict between them. Generate node execution time series based on the resource competition evolution trend. For example, if it is predicted that nodes A and B will compete for resource X in the same time period, their execution order can be adjusted to avoid conflicts. Assume that the execution time series of node A is [1, 2, 3], and the corresponding probabilities are [0.2, 0.5, 0.3], and the execution time series of node B is [2, 3, 4], and the corresponding probabilities are [0.3, 0.4, 0.3], respectively.

[0152] Generate and optimize the scheduling plan. Construct an optimization function based on the scheduling constraint graph, input the node execution time series into the optimization network to search the solution space, and extract the scheduling strategy from the search results. Input the scheduling strategy into the reinforcement learning network for optimization to generate an initial scheduling plan. Perform online optimization on the initial scheduling plan to obtain the optimized scheduling plan. For example, dynamically adjust the execution order of nodes according to resource competition and node priority. Input the optimized scheduling plan into the graph neural network to detect priority constraint conflicts and calculate the impact of the conflict. For example, if node A must be executed before node B, but node B is executed before node A in the scheduling plan, there is a priority constraint conflict. Eliminate the priority constraint conflict according to the impact. For example, adjust the execution order of node B and node A. Combine the scheduling plan after eliminating the conflict with the scheduling constraint graph for iterative optimization, and finally output the node scheduling sequence. For example, the final output scheduling sequence is [A, C, B, D].

[0153] In this embodiment, by considering the priority constraints and resource competition relationships between nodes, resource conflicts and deadlocks can be effectively avoided, thereby improving scheduling efficiency; by predicting the evolution trend of resource competition between nodes, resource allocation can be planned in advance to avoid resource waste, thereby optimizing resource utilization; through reinforcement learning and online optimization, the scheduling plan can be continuously adjusted to adapt to the dynamically changing environment, thereby enhancing the robustness of the scheduling plan.

[0154] The present invention provides another optional embodiment, including:

[0155] Collect production line logistics data: Collect historical records of wafer group transportation between the loading port and buffer node on the machine, including the start time, completion time, actual transportation path length, task queue waiting time, and number of transfers during transportation; record the status data of automatic transport vehicles within and between areas, including the current location coordinates of the equipment, the current task queue length, the equipment idle rate, and the equipment working time;

[0156] Calculate node criticality: Based on historical handling data, calculate the average processing capacity indicators of the loading port of the machine, including processing batches per unit time, processing time distribution, and equipment availability; analyze the transfer service indicators of the buffer, including material turnover rate, buffer occupancy rate, and peak processing capacity; combine the node dependency relationship of the product process route to establish a node criticality scoring model; set key process machines such as lithography machines and etching machines with criticality scores in the top 20% as core nodes, set the buffers with the top 30% transfer frequency as key nodes, and set the rest as ordinary nodes;

[0157] Construct a material transmission network: Take the loading port and buffer on the machine as network nodes, establish material transmission edges according to process flow specifications and actual material flow data, and record the actual path distance, historical average transportation time, and maximum channel capacity for each transmission edge; divide the network into two levels: intra-regional transmission layer and inter-regional transmission layer, and establish a connection relationship matrix between the levels through shared nodes;

[0158] Establish a resource competition network: Identify the node groups that share the same automatic handling equipment and build a resource competition relationship based on priority; set the handling task priority of the core node to level 1, the key node to level 2, and the ordinary node to level 3; the weight of the competition edge is calculated based on the priority difference of the connected nodes, and the larger the weight, the more significant the priority difference;

[0159] Divide logistics subnetworks: Divide independent logistics subnetworks according to regional boundaries. Each subnetwork contains all loading port nodes and buffer nodes in the region. Calculate the correlation strength index between adjacent subnetworks, which is determined by the number of shared boundary nodes, the frequency of cross-regional material flow, and the degree of equipment sharing.

[0160] Predict network status: Use the graph neural network model to predict the transport demand and equipment load level of each sub-network in the next 2 hours; Combine the connection relationship of the material transmission network to calculate the feasibility of the direct transport plan between computers, including path conditions, time window matching, and equipment availability; Evaluate the benefits of skipping buffer transfers, including saved waiting time and transport distance;

[0161] Generate scheduling plans: Based on demand forecast results, calculate the number of automated transport vehicles required for each area to achieve balanced equipment load distribution; prioritize direct transportation routes for core nodes to ensure timely supply of materials for key processes; monitor network status in real time, and when the predicted load exceeds the threshold, trigger the adjustment mechanism according to the node priority order to dynamically optimize the equipment allocation plan and transportation route planning within the region.

[0162] Figure 2 Schematic diagram of the structure of a chip production logistics network optimization system based on a graph neural network model according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0163] The first unit is used to collect logistics data of multiple process nodes on a chip production line; calculate the criticality score of the process node based on the logistics data, and determine the node type according to the criticality score; construct an initial logistics network diagram, establish material transmission edges between process nodes, and determine the material transmission network; establish priority-based resource competition edges based on node types, and determine the resource competition network; superimpose the material transmission network and the resource competition network to generate a multi-level logistics network diagram;

[0164] The second unit is used to divide the logistics sub-network according to the material transmission network in the multi-level logistics network diagram; determine the correlation strength between the logistics sub-networks based on the resource competition network; extract the material flow characteristics of each logistics sub-network and build a sub-network state predictor; use the correlation strength between adjacent sub-networks to establish a collaborative scheduling module; input the topological structure and node characteristics of the logistics sub-network into the graph neural network, and train the model with the output of the sub-network state predictor and the information of the collaborative scheduling module as the optimization target to obtain an optimized graph neural network model;

[0165] The third unit is used to predict the state changes of the logistics sub-network based on the optimized graph neural network model, and generate a resource allocation plan based on the connection relationship of the material transmission network; calculate the node scheduling sequence according to the priority constraints of the resource competition network; when the predicted state exceeds the threshold, trigger the adjustment mechanism according to the priority order of the node type, and re-plan the material allocation and transmission path of the relevant sub-network; convert the resource allocation plan and node scheduling sequence into equipment control instructions, and send them to the corresponding process nodes.

[0166] According to a third aspect of the embodiments of the present invention,

[0167] An electronic device is provided, comprising:

[0168] processor;

[0169] a memory for storing processor-executable instructions;

[0170] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0171] According to a fourth aspect of the embodiments of the present invention,

[0172] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0173] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0174] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A chip production logistics network optimization method based on a graph neural network model, characterized in that: include: Collect logistics data of multiple process nodes on chip production lines; Calculate the criticality score of the process node based on the logistics data, and determine the node type according to the criticality score; Construct the initial logistics network diagram, establish the material transmission edges between process nodes, and determine the material transmission network; establish the priority-based resource competition edges based on the node type, and determine the resource competition network; Superimposing the material transmission network and the resource competition network to generate a multi-level logistics network diagram; Dividing the logistics sub-network according to the material transmission network in the multi-level logistics network diagram; Determine the association strength between logistics sub-networks based on the resource competition network; extract material flow characteristics for each logistics sub-network and build a sub-network state predictor; and establish a collaborative scheduling module using the association strength between adjacent sub-networks; The topological structure and node features of the logistics sub-network are input into the graph neural network, and the output of the sub-network state predictor and the information of the collaborative scheduling module are used as optimization targets for model training to obtain an optimized graph neural network model; Based on the optimized graph neural network model, the state changes of the logistics sub-network are predicted, and the resource allocation plan is generated in combination with the connection relationship of the material transmission network; the node scheduling sequence is calculated according to the priority constraints of the resource competition network; when the predicted state exceeds the threshold, the adjustment mechanism is triggered according to the priority order of the node type, and the material allocation and transmission path of the relevant sub-network are re-planned; the resource allocation plan and node scheduling sequence are converted into equipment control instructions and sent to the corresponding process nodes.

2. The method according to claim 1, characterized in that Collect logistics data of multiple process nodes on chip production lines; Calculating the criticality score of the process node based on the logistics data, and determining the node type according to the criticality score includes: Collect the material arrival time, material processing completion time and equipment utilization rate of each process node on the chip production line; calculate the time difference between the material arrival time and the material processing completion time to obtain the material processing cycle; count the material backlog per unit time, and divide the material backlog by the sampling time interval to obtain the material backlog change rate; construct a detection window, and set the length of the detection window to three times the material processing cycle calculated in the previous step; perform wavelet decomposition on the material backlog change rate within the detection window to obtain the high-frequency fluctuation coefficient and the low-frequency trend coefficient; calculate the ratio of the energy value of the high-frequency fluctuation coefficient to the total energy value, and mark the time point when the ratio exceeds the fluctuation threshold as a logistics abnormality point; Calculate the number of occurrences of logistics anomalies within a unit time window to obtain the frequency of anomalies; extract the material backlog corresponding to the logistics anomalies, perform exponential smoothing on the material backlog, and obtain the anomaly intensity value; input the anomaly frequency and the anomaly intensity value into the recursive calculation unit to extract the anomaly evolution characteristics; perform weighted calculation on the anomaly evolution characteristics and the equipment utilization rate to obtain the node criticality score; sort the node criticality scores, mark the top one-third of the nodes as core nodes, and mark the remaining nodes as ordinary nodes to determine the node type.

3. The method according to claim 2, characterized in that Construct the initial logistics network diagram, establish the material transmission edges between process nodes, and determine the material transmission network; establish the priority-based resource competition edges based on the node type, and determine the resource competition network; The material transmission network and resource competition network are superimposed to generate a multi-level logistics network diagram including: Based on the material flow relationship between adjacent process nodes, a material transfer matrix is ​​created to construct an initial logistics network diagram; the number of material transfers between any two nodes in a unit time is counted, and the material transfer number is divided by the total material amount of the starting node to obtain the material flow ratio, and the material flow ratio is used as the weight of the material transfer edge to determine the material transfer network; Based on the node type, the node criticality score product of the two core nodes is calculated to obtain a first competition coefficient, the node criticality score product of the core node and the common node is calculated to obtain a second competition coefficient, and the node criticality score product of the two common nodes is calculated to obtain a third competition coefficient, and the corresponding competition coefficient is used as the weight of the resource competition edge to determine the resource competition network; the material transmission network and the resource competition network are superimposed to generate a multi-level logistics network diagram; Monitor the number of logistics anomaly points at each process node. When the number of logistics anomaly points exceeds the anomaly monitoring threshold, recalculate the anomaly frequency and anomaly intensity value of the corresponding process node; re-input the updated anomaly frequency and anomaly intensity value into the recursive calculation unit to obtain a new node criticality score; determine the node type according to the new node criticality score; based on the updated node type, recalculate the new first competition coefficient, the new second competition coefficient and the new third competition coefficient; collect the latest material transmission data to update the material flow ratio; based on the updated material flow ratio, the new first competition coefficient, the new second competition coefficient and the new third competition coefficient, regenerate a new resource competition network.

4. The method according to claim 1, characterized in that: Divide the logistics subnetworks according to the material transmission network in the multi-level logistics network diagram; determine the association strength between the logistics subnetworks based on the resource competition network; extract the material flow characteristics of each logistics subnetwork and build a subnetwork state predictor; and establish a collaborative scheduling module using the association strength between adjacent subnetworks; The topological structure and node features of the logistics sub-network are input into the graph neural network, and the model training is performed with the output of the sub-network state predictor and the information of the collaborative scheduling module as the optimization target, and the optimized graph neural network model is obtained, including: Based on the material transmission network in the multi-level logistics network diagram, determine the material flow intensity between nodes and construct an affinity matrix; perform standardization on the affinity matrix to obtain a normalized Laplace matrix; analyze the eigenvalue distribution of the normalized Laplace matrix, determine the optimal number of sub-networks, and select corresponding eigenvectors to construct a feature matrix; perform clustering operations based on the feature matrix to obtain node attribution labels, and based on the node attribution labels, divide the material transmission network into a corresponding number of logistics sub-networks; The ratio of the total material flow between adjacent logistics sub-networks to the maximum total material flow of each logistics sub-network is calculated to obtain the basic correlation strength between logistics sub-networks. Based on the pre-acquired dynamic change law of material flow, a time decay factor is introduced, and the time decay factor is multiplied by the basic correlation strength to obtain the time-varying correlation strength of the logistics sub-network. Constructing a sub-network state predictor, the sub-network state predictor includes: collecting data on material backlog, turnover rate and load rate of nodes in each logistics sub-network, and constructing a time series feature sequence of node operation status; based on the time series feature sequence, performing a sliding window segmentation operation, extracting local time series features, and obtaining a low-dimensional feature representation through feature dimension reduction module mapping; calculating feature weights of the low-dimensional feature representation, and inputting the weighted feature into a pre-trained bidirectional long short-term memory network, outputting a node state prediction value, and calculating a state prediction deviation based on the actually collected operation state data and the node state prediction value; Combine the node state prediction value of the current logistics sub-network, the node state prediction value of the adjacent logistics sub-network and the time-varying correlation strength to construct a cross-network feature vector, input it into the attention calculation unit, and generate the attention weight of the state interaction between the logistics sub-networks; multiply the attention weight with the node state prediction value of the adjacent logistics sub-network to obtain weighted state information; determine the state transition matrix by learning the mapping relationship between the current logistics sub-network and the adjacent logistics sub-network, map the weighted state information back to the current logistics sub-network, and generate collaborative scheduling information; compare the collaborative scheduling information with the preset scheduling target, and calculate the execution error of the collaborative scheduling information; Extract the topological structure features and node operation status features of the logistics sub-network, and construct a node feature matrix based on the graph neural network; generate a query matrix, a key matrix and a value matrix through linear transformation; calculate the attention scores of the query matrix and the key matrix, and multiply the attention scores with the value matrix to obtain the fusion features; obtain the state prediction deviation of the sub-network state predictor and the execution error of the collaborative scheduling information, and construct an objective function containing network structure constraints; optimize the connection weights of the graph attention network based on the objective function, and output the optimized graph neural network model.

5. The method according to claim 4, characterized in that The time-varying correlation strength is formulated as follows: ； Among them, R(t) ij represents the time-varying correlation strength of two nodes i and i at time t, t represents time, i represents a logistics subnetwork, j represents another logistics subnetwork, k represents the time scale level, and λ k represents the weight parameter of scale k, represents the first wavelet basis function at scale k, x(·) represents the state sequence corresponding to the i logistics subnetwork, represents a time different from t, represents the second wavelet basis function at scale k, y(·) represents the state sequence corresponding to the j logistics subnetwork, represents the Laplace operator, S i represents the state distribution of the i logistics sub-network, S j represents the state distribution of the j logistics subnetwork, H(S i , S j ) represents the Wasserstein distance between two state distributions, α represents the time decay coefficient, t0 represents the initial time, and β represents the nonlinear adjustment factor.

6. The method according to claim 1, characterized in that Based on the optimized graph neural network model, the state changes of the logistics sub-network are predicted, and the resource allocation scheme generated by combining the connection relationship of the material transmission network includes: The state prediction results and historical error data of the logistics sub-network are input into the graph neural network model, and the graph neural network model is used to construct the uncertainty interval of the state prediction based on the historical error data. According to the uncertainty interval, the process nodes in the logistics sub-network are divided into multiple levels of logistics nodes, including primary logistics nodes, secondary logistics nodes and buffer logistics nodes; Based on the connection relationship of the logistics sub-network, a topological structure of the material transmission network is constructed, the time-varying transmission intensity between adjacent logistics nodes in the material transmission network is calculated, the material transmission bottleneck is determined based on the time-varying transmission intensity, and the transmission importance score of each logistics node is calculated according to the material transmission bottleneck; According to the state prediction result and the uncertainty interval, the material demand interval of each logistics node is calculated, the material buffer capacity is configured for the first-level logistics node based on the transmission importance score, the material adjustment interval is set for the second-level logistics node, the material storage space is reserved for the buffer logistics node, and the material demand allocation model is generated; Establish material flow constraints based on the topological structure of the material transmission network, set material priority guarantee constraints for the first-level logistics nodes, set material dynamic adjustment constraints for the second-level logistics nodes, configure material storage balance constraints for the buffer logistics nodes, and build transmission capacity constraints based on the material transmission bottleneck; Material flow constraints, material priority constraints, material dynamic adjustment constraints, material storage balance constraints and transmission capacity constraints are constructed as resource allocation optimization objectives, and a deep reinforcement learning network is used to solve the resource allocation optimization objectives to obtain a material allocation strategy; The robustness index of the material allocation strategy is calculated, the material allocation strategy is optimized online based on the robustness index, the material buffer capacity of the first-level logistics node is dynamically adjusted, and a material resource allocation plan is generated.

7. The method according to claim 1, characterized in that The scheduling sequence of computing nodes according to the priority constraints of the resource competition network includes: Collect the connection relationship and competition relationship of the process nodes in the resource competition network, build a multi-dimensional heterogeneous graph structure, run the time-series attention network, extract the priority dependency features of the process nodes in the time dimension, run the space-series attention network, extract the resource competition features of the process nodes in the space dimension, and fuse the priority dependency features and the resource competition features to generate a node priority tensor; Input the node priority tensor and the resource capacity information in the resource competition network into a hierarchical coding network for encoding to obtain a priority constraint vector and a resource constraint vector, and fuse the priority constraint vector and the resource constraint vector to generate a scheduling constraint graph; Input the scheduling constraint graph into a probability coding network to obtain the probability distribution of process node execution time, predict the resource competition evolution trend between nodes based on the probability distribution, and generate a node execution time series according to the resource competition evolution trend; An optimization function is constructed according to the scheduling constraint graph, the node execution time series is input into the optimization network to search the solution space, the scheduling strategy is extracted from the solution space search results, the scheduling strategy is input into the reinforcement learning network for optimization, an initial scheduling plan is generated, the initial scheduling plan is optimized online to obtain an optimized scheduling plan, the optimized scheduling plan is input into the graph neural network to detect priority constraint conflicts, the impact of the priority constraint conflicts is calculated, and the priority constraint conflicts are eliminated according to the impact; the scheduling plan after eliminating the priority constraint conflicts is iteratively optimized in combination with the scheduling constraint graph, and a node scheduling sequence is output.

8. A chip production logistics network optimization system based on a graph neural network model, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to collect logistics data of multiple process nodes on the chip production line; Calculate the criticality score of the process node based on the logistics data, and determine the node type according to the criticality score; Construct the initial logistics network diagram, establish the material transmission edges between process nodes, and determine the material transmission network; establish the priority-based resource competition edges based on the node type, and determine the resource competition network; Superimposing the material transmission network and the resource competition network to generate a multi-level logistics network diagram; The second unit is used to divide the logistics sub-network according to the material transmission network in the multi-level logistics network diagram; Determine the association strength between logistics sub-networks based on the resource competition network; extract material flow characteristics for each logistics sub-network and build a sub-network state predictor; and establish a collaborative scheduling module using the association strength between adjacent sub-networks; The topological structure and node features of the logistics sub-network are input into the graph neural network, and the output of the sub-network state predictor and the information of the collaborative scheduling module are used as optimization targets for model training to obtain an optimized graph neural network model; The third unit is used to predict the state changes of the logistics sub-network based on the optimized graph neural network model, and generate a resource allocation plan based on the connection relationship of the material transmission network; calculate the node scheduling sequence according to the priority constraints of the resource competition network; when the predicted state exceeds the threshold, trigger the adjustment mechanism according to the priority order of the node type, and re-plan the material allocation and transmission path of the relevant sub-network; convert the resource allocation plan and node scheduling sequence into equipment control instructions, and send them to the corresponding process nodes.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images