A semiconductor material handling path optimization method based on big data analysis

Abstract

This invention discloses a semiconductor material handling path optimization method based on big data analysis, comprising the following steps: collecting operational data of the material handling system and aligning it with timestamps to form a rolling time window log dataset; constructing a heterogeneous graph based on the track or road network topology and generating feature vectors for nodes and edges; inputting the heterogeneous graph into an improved GAT model to output edge risk scores and node bottleneck heat; constructing a dynamic robust cost graph to generate a candidate path set and a candidate transfer action set; using an improved ALNS algorithm to solve for the output task and vehicle allocation, path sequence, and release time; performing trigger determination based on the rolling time window log dataset and executing local replanning upon triggering; and storing the results in a versioned manner and updating the improved GAT model. This invention improves the scheduling efficiency, path robustness, and dynamic response capability of the semiconductor material handling system, and is suitable for complex manufacturing scenarios.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing automation and intelligent scheduling technology, and in particular to a method for optimizing semiconductor material handling paths based on big data analysis. Background Technology

[0002] With the continuous expansion of semiconductor manufacturing scale and the ongoing improvement of wafer fab automation levels, automated material handling systems (AWDS) play an increasingly crucial role in the cross-process flow of wafer cassettes. Existing semiconductor material handling path planning and scheduling technologies mainly rely on shortest path algorithms based on static topology, rule-driven priority scheduling strategies, or deterministic optimization methods combined with simple historical statistics to achieve vehicle and task allocation and path planning. However, the following problems commonly exist in actual operation: On the one hand, the operating status of material handling systems exhibits significant time-varying and uncertainties. Track occupancy, switch mutual exclusion waiting, and station congestion change dynamically over time. Existing methods typically make decisions based on single-moment or short-term statistical information, which is insufficient to reflect the state evolution trend, resulting in lagging path planning results and inadequate response to sudden disturbances. On the other hand, semiconductor material handling networks consist of multiple types of entities, including track sections, switches, stations, and vehicles, with highly heterogeneous connections and constraints. Existing methods often use unified graph models or simplified abstractions, making it difficult to characterize the differences in conflict, congestion, and capacity limitations among different nodes and edges, thus affecting scheduling accuracy. Furthermore, traditional scheduling algorithms often rely on global replanning when congestion or anomalies occur, resulting in high computational overhead and insufficient real-time performance. They also lack the ability to specifically identify and locally optimize high-risk paths and bottleneck nodes, making it difficult to meet the stability and throughput efficiency requirements of large-scale wafer fabs.

[0003] Therefore, how to provide a semiconductor material handling path optimization method based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a semiconductor material handling path optimization method based on big data analysis. This invention performs rolling time window analysis on the operating data of the material handling system, and combines heterogeneous graph modeling, improved GAT model and improved ALNS algorithm to realize path planning, vehicle allocation and dynamic replanning for material handling tasks. It can effectively perceive path risks and system bottlenecks, and has the advantages of high scheduling efficiency, strong path robustness and strong response to system disturbances.

[0005] A semiconductor material handling path optimization method based on big data analysis according to an embodiment of the present invention includes the following steps: Step 1: Collect the operating data of the material handling system and align it according to a unified timestamp to form a rolling time window log dataset; Step 2: Construct a heterogeneous graph based on the track or road network topology, and generate feature vectors for the nodes and edges of the heterogeneous graph based on the rolling time window log dataset; Step 3: Input the heterogeneous graph into the improved GAT model. The improved GAT model consists of a heterogeneous graph encoding module, a time-varying state window encoding module, a heterogeneous relation attention message passing module, and a dual-output head scoring module, which outputs edge risk scores and node bottleneck heat. Step 4: Construct a dynamic robust cost graph based on edge risk scores and node bottleneck heat, and generate a set of candidate paths and a set of candidate transit actions under constraints; Step 5: Input the task set, candidate path set and candidate transfer action set into the improved ALNS algorithm, and solve and output the task and vehicle allocation, path sequence and release time based on the dynamic robust cost graph. The improved ALNS algorithm includes a feasibility mask insertion and filtering mechanism based on turnout mutual exclusion, capacity and time window constraints and a hot spot perception damage and repair mechanism based on edge risk scoring and node bottleneck heat. Step 6: Determine the trigger based on the rolling time window log dataset. When triggered, call the improved ALNS algorithm to perform local replanning on the affected tasks within the preset time budget and output the updated task and vehicle allocation, path sequence and release time. Step 7: Version the task and vehicle allocation, path sequence and release time, and update the improved GAT model.

[0006] Optionally, step one specifically includes: The operation data of the material handling system is collected. The operation data includes records of task generation, task modification and task completion in the task system, records of vehicle position, vehicle operation status and vehicle loading status in the vehicle control system, and records of turnout status, track section occupancy status and station entry and exit status in the line control system. A unified timestamp is assigned to each piece of operational data collected. Timestamp alignment is performed on operational data with different sampling periods and reporting frequencies. Status completion is performed on the timestamp-aligned operational data. When multiple state records of the same object exist in the same time bucket, they are sorted according to the order of the unified timestamp, and the state record with the latest timestamp is retained as the valid state in the current time bucket. Based on the time bucket sequence that has completed timestamp alignment and state completion, a rolling time window is constructed on the time axis according to the preset rolling window length and rolling step size, forming a rolling time window log dataset composed of multiple rolling time windows arranged in chronological order.

[0007] Optionally, step two specifically includes: A heterogeneous graph is constructed based on the track or road network topology data of the material handling system. The heterogeneous graph includes track segment nodes, turnout nodes, station nodes and vehicle nodes, as well as segment connection edges, turnout guiding edges, station access edges and vehicle occupancy edges. Based on the rolling time window log dataset, node feature vectors are generated for nodes and edge feature vectors are generated for edges. Missing values ​​are filled and numerical scaling is performed on the node feature vectors and edge feature vectors to obtain a heterogeneous graph.

[0008] Optionally, the improved GAT model includes a heterogeneous graph encoding module, a time-varying state window encoding module, a heterogeneous relation attention message passing module, and a dual-output head scoring module: The heterogeneous graph encoding module sets a node type linear mapping layer for different types of nodes in the heterogeneous graph to obtain the initial embedding vector of the node, and sets an edge type linear mapping layer for different types of edges in the heterogeneous graph to obtain the initial embedding vector of the edge. The time-varying state window encoding module assembles the initial embedding vectors obtained by the same node or the same side in the heterogeneous graph into an embedding sequence in time order within multiple consecutive rolling time windows, and constructs a time window embedding vector based on the embedding sequence. The heterogeneous relation attention message passing module uses the edge type of the heterogeneous graph as the heterogeneous relation identifier. It sets relation attention parameters for segment connection edges, turnout guidance edges, station access edges, and vehicle occupancy edges respectively. For each incoming edge of the target node in the heterogeneous graph, the source node time window embedding vector and the target node time window embedding vector are linearly transformed to obtain the source node representation vector and the target node representation vector respectively. The attention weight is calculated based on the source node representation vector, the target node representation vector, and the edge time window embedding vector of the incoming edge. The source node representation vector under the same heterogeneous relation is multiplied by the corresponding attention weight and summed to obtain the relation aggregated message vector. The relation aggregated message vectors of different heterogeneous relations are linearly transformed and then added to obtain the fused message vector. The fused message vector is added to the target node time window embedding vector by residual addition and layer normalization is performed to obtain the updated embedding vector of the target node in the heterogeneous graph. The dual-output head scoring module is configured with an edge risk scoring output head and a node bottleneck heat output head. The edge risk scoring output head takes the updated edge embedding vector of each edge in the heterogeneous graph as input to obtain the edge risk score, and the node bottleneck heat output head takes the updated embedding vector of each node in the heterogeneous graph as input to obtain the node bottleneck heat.

[0009] Optionally, step four specifically includes: A dynamic robust cost graph is constructed based on edge risk scoring and node bottleneck heat. The dynamic robust cost graph uses the topology of the heterogeneous graph as the skeleton and the edges as the cost carriers, and generates an edge cost for each passable edge in the heterogeneous graph. For each edge in the heterogeneous graph, a basic passage cost is constructed. The basic passage cost is determined by the average passage delay and average queuing time obtained from the corresponding connection relationship of the edge in the rolling time window log dataset. For each edge in the heterogeneous graph, a risk penalty cost is constructed. The risk penalty cost is determined by the edge risk score of the edge. The edge risk score is mapped to a risk penalty coefficient, and the basic passage cost is multiplicatively weighted with the risk penalty coefficient to obtain the edge cost containing the risk term. For each edge in the heterogeneous graph, a bottleneck association penalty cost is constructed. The bottleneck association penalty cost is determined by the bottleneck heat of the adjacent nodes of the edge. The weighted sum of the bottleneck heat of the two ends of the edge is taken as the bottleneck penalty coefficient. The bottleneck penalty coefficient is used to multiplicatively weight the edge cost containing risk terms to obtain the dynamic robust edge cost. A dynamic robust cost graph is constructed based on the dynamic robust edge cost, and a candidate path set is generated under constraints, including turnout mutual exclusion constraints, section capacity constraints, and task time window constraints. Based on the generated candidate path set, a candidate transit action set is generated. The candidate transit action set refers to the station-level actions that occur during the path execution process. The candidate transit action set includes reversing actions, yielding and waiting actions, and cache dwelling actions.

[0010] Optionally, step five specifically includes: The input to the improved ALNS algorithm is constructed, which includes a task set, a candidate path set, a candidate transfer action set, and a dynamic robust cost graph; An initial solution is generated based on the task set. The initial solution includes task and vehicle allocation, path sequence and release time. When generating the initial solution, a path is selected from the candidate path set and a vehicle is allocated for each material handling task. Based on the edge sequence and corresponding dynamic robust edge cost of the selected path on the dynamic robust cost graph, the entry time and exit time of the current task on each edge are calculated, and the release time that satisfies the constraints of task release time and latest completion time is determined. Construct a discrete time slot occupancy table, which uses time buckets as time indexes and track sections, switches, and stations as resource indexes to record the time slots occupied by each task on the corresponding resources in the current solution. A feasibility mask insertion and filtering mechanism is implemented. A feasibility mask is generated for each candidate path and each candidate transfer action of the task to be inserted. The occupied time slot is compared with the discrete time slot occupancy table. When the same turnout is occupied by multiple tasks in the same discrete time slot, the number of occupied track sections or stations in the same discrete time slot exceeds the preset capacity, or the task execution time exceeds the interval between the task release time and the latest completion time, the corresponding candidate path or candidate transfer action is determined to be infeasible and removed. In each iteration, a destruction operator is executed to remove some tasks in the current solution, and a hotspot-aware destruction mechanism is executed. The hotspot-aware destruction mechanism constructs a hotspot set based on edge risk scores and node bottleneck heat. The hotspot set includes edges with edge risk scores higher than a preset first threshold and nodes with node bottleneck heat higher than a preset second threshold, and prioritizes removing tasks whose path sequences pass through the hotspot set. In each iteration, the repair operator will remove the task and reinsert it into the current solution, and execute the hot spot awareness repair mechanism. The hot spot awareness repair mechanism will prioritize the selection of candidate paths that avoid the hot spot set in the candidate path set, and prioritize the insertion of waiting or cached resident actions into the station where the node bottleneck heat is lower than the preset third threshold in the candidate transfer action set. For each insertion scheme, the feasibility mask insertion filtering mechanism will be called to filter infeasible schemes, and the scheme that minimizes the cumulative dynamic robust edge cost on the dynamic robust cost graph will be selected from the feasible schemes to update the task and vehicle allocation, path sequence and release time. When the preset number of iterations or the preset time budget is reached, the current optimal feasible solution is output.

[0011] Optionally, step six specifically includes: Trigger determination is performed based on the rolling time window log dataset. The trigger determination takes the alignment record of the current rolling time window and the previous rolling time window as input, and calculates the change in track section occupancy rate, the change in the number of turnout mutual exclusion waiting times, and the change in task timeout risk, respectively. A trigger condition is determined when any of the following conditions are met: the change in track occupancy rate of any track section exceeds the preset occupancy threshold; the change in the number of turnout mutual exclusion waiting times of any turnout exceeds the preset change threshold; the change in task timeout risk exceeds the preset risk threshold; or the rolling time window log dataset contains records of abnormal vehicle stagnation or unpassable track sections. After triggering, the set of affected tasks is determined. The set of affected tasks consists of tasks that meet the following conditions: the path sequence contains an edge with an edge risk score higher than a preset edge risk score threshold, or the path sequence passes through a node with a bottleneck heat higher than a preset bottleneck heat threshold, or the path sequence contains a track segment corresponding to a track segment that is impassable, or the expected arrival time is later than the latest completion time. After being triggered, the preset time budget is started, and within the preset time budget, the improved ALNS algorithm is called to perform local replanning on the affected task set. Only the task and vehicle allocation, path sequence and release time of the affected tasks are updated. During the update, the dynamic robust cost map is used as the cost basis and the feasibility mask insertion and filtering mechanism is called to ensure that the turnout mutual exclusion constraint, capacity constraint and time window constraint are satisfied. When the preset time budget is reached or all affected task sets have a feasible update plan, output the updated task and vehicle allocation, path sequence and release time.

[0012] Optionally, step seven specifically includes: Task and vehicle allocation, path sequence and release time are stored in a versioned manner according to a rolling time window. The actual execution results of the corresponding version are written into the rolling time window log dataset. When the preset update cycle arrives, the improved GAT model is updated based on the versioned storage results and the actual execution results.

[0013] The beneficial effects of this invention are: This invention provides a unified modeling of the time-varying operating state of semiconductor material handling systems based on a rolling time window log dataset. It uses a heterogeneous graph to characterize various types of entities and their constraints among track sections, switches, stations, and vehicles. An improved graph attention network outputs edge risk scores and node bottleneck heat, enabling refined quantification of path conflict risks and system bottlenecks. Based on this, a dynamic robust cost graph is constructed, combined with an improved adaptive large neighborhood search algorithm featuring feasibility mask insertion filtering and hotspot perception and repair mechanisms. This achieves collaborative optimization of task and vehicle allocation, path sequences, and release times. Furthermore, triggered local replanning avoids global recalculation, significantly improving the system's response speed and operational stability to congestion, disturbances, and anomalies. This invention effectively reduces the frequency of switch mutual exclusion conflicts and section congestion, improving the robustness and reliability of material handling path planning. It has significant engineering application value for enhancing the throughput, on-time delivery capability, and intelligent operation level of automated material handling systems in large-scale wafer fabs. Attached Figure Description

[0014] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a semiconductor material handling path optimization method based on big data analysis proposed in this invention; Figure 2 This is a schematic diagram of a semiconductor material handling path optimization method based on big data analysis proposed in this invention; Figure 3This is a framework diagram of the improved GAT model in the semiconductor material handling path optimization method based on big data analysis proposed in this invention. Detailed Implementation

[0015] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0016] refer to Figures 1-3 A semiconductor material handling path optimization method based on big data analysis includes the following steps: Step 1: Collect the operating data of the material handling system and align it according to a unified timestamp to form a rolling time window log dataset; Step 2: Construct a heterogeneous graph based on the track or road network topology, and generate feature vectors for the nodes and edges of the heterogeneous graph based on the rolling time window log dataset; Step 3: Input the heterogeneous graph into the improved GAT model. The improved GAT model consists of a heterogeneous graph encoding module, a time-varying state window encoding module, a heterogeneous relation attention message passing module, and a dual-output head scoring module, which outputs edge risk scores and node bottleneck heat. Step 4: Construct a dynamic robust cost graph based on edge risk scores and node bottleneck heat, and generate a set of candidate paths and a set of candidate transit actions under constraints; Step 5: Input the task set, candidate path set, and candidate transfer action set into the improved ALNS algorithm. Based on the dynamic robust cost graph, the algorithm outputs the task and vehicle allocation, path sequence, and release time. The improved ALNS algorithm includes a feasibility mask insertion and filtering mechanism based on turnout mutual exclusion, capacity, and time window constraints, and a hotspot perception, damage, and repair mechanism based on edge risk scoring and node bottleneck heat. Step 6: Determine the trigger based on the rolling time window log dataset. When triggered, call the improved ALNS algorithm to perform local replanning on the affected tasks within the preset time budget and output the updated task and vehicle allocation, path sequence and release time. Step 7: Version the task and vehicle allocation, path sequence and release time, and update the improved GAT model.

[0017] In this embodiment, step one specifically includes: Collect operational data from the material handling system. The operational data includes records of task generation, task modification and task completion in the task system, records of vehicle position, vehicle operation status and vehicle loading status in the vehicle control system, and records of turnout status, track section occupancy status and station entry and exit status in the track control system. A unified timestamp is assigned to each piece of operational data collected. The unified timestamp uses the master clock of the material handling system as the time base and converts the local time fields from different subsystems into a unified timestamp. For running data with different sampling periods and reporting frequencies, perform timestamp alignment processing, divide the time axis into continuous time buckets according to the preset time granularity, and map each piece of running data to the time bucket corresponding to the same timestamp; The runtime data after timestamp alignment is processed to complete the status. For vehicle status, section occupancy status or turnout status that is missing in a certain time bucket, the most recently recorded valid status of the object on the time axis is used to fill it. When multiple state records of the same object exist in the same time bucket, they are sorted according to the order of the unified timestamp, and the state record with the latest timestamp is retained as the valid state in the current time bucket. Based on the time bucket sequence that has completed timestamp alignment and status completion, a rolling time window is constructed on the time axis according to the preset rolling window length and rolling step size, forming a rolling time window log dataset composed of multiple rolling time windows arranged in chronological order. Each rolling time window contains the operation status records of tasks, vehicles, switches, sections and stations within the covered time interval.

[0018] In this embodiment, step two specifically includes: Heterogeneous graphs are constructed based on the track or road network topology data of the material handling system. The heterogeneous graphs include track segment nodes, turnout nodes, station nodes and vehicle nodes, as well as segment connecting edges, turnout guiding edges, station access edges and vehicle occupancy edges. The segment connecting edges are set according to the travel direction allowed by the topology, the turnout guiding edges contain branch identifiers, the station access edges contain entry direction identifiers, and the vehicle occupancy edges are updated with the rolling time window. Based on the rolling time window log dataset, node feature vectors are generated for each node. The node feature vectors of track section nodes include occupancy duration, occupancy count, passage count, and average passage delay. The node feature vectors of turnout nodes include switching count, mutual exclusion waiting count, and average mutual exclusion waiting duration. The node feature vectors of station nodes include entry count, exit count, average dwell time, and queue length. The node feature vectors of vehicle nodes include empty time, loading time, stationary time, and cumulative travel time. Based on the rolling time window log dataset, edge feature vectors are generated for each edge. The edge feature vectors of the segment connectivity edge include the number of passes, average queuing time, and number of conflict triggers. The edge feature vectors of the turnout guidance edge include the number of branch selections, average switching delay, and number of mutual exclusion blocking. The edge feature vectors of the station access edge include the number of inbound blocking, average inbound waiting time, and buffer occupancy ratio. The edge feature vectors of the vehicle occupancy edge include the binding time, number of state switching, and number of abnormal stops. Missing value imputation and numerical scaling are performed on the node feature vector and edge feature vector. When imputing missing values, the frequency of events that have not occurred is set to zero, and the missing state features are filled with the encoding value corresponding to the most recent valid state in the rolling time window log data. When scaling the numerical values, each feature dimension is scaled according to the minimum and maximum values ​​of the preset statistical interval to obtain a heterogeneous graph.

[0019] In this embodiment, the improved GAT model includes a heterogeneous graph encoding module, a time-varying state window encoding module, a heterogeneous relational attention message passing module, and a dual-output head scoring module: The heterogeneous graph encoding module sets up a node type linear mapping layer for different types of nodes in the heterogeneous graph. The node feature vectors of track section nodes, turnout nodes, station nodes, and vehicle nodes are input into the corresponding node type linear mapping layer to obtain the initial embedding vector of the node. The module also sets up an edge type linear mapping layer for different types of edges in the heterogeneous graph. The edge feature vectors of section connecting edges, turnout guiding edges, station access edges, and vehicle occupancy edges are input into the corresponding edge type linear mapping layer to obtain the initial embedding vector of the edge. The time-varying state window encoding module forms an embedding sequence by arranging the initial embedding vectors of the same node or the same edge in the heterogeneous graph within multiple consecutive rolling time windows. The embedding vectors at each time step in the embedding sequence are superimposed with the position encoding vector and input into the gated recurrent unit network to obtain the hidden state sequence. Each hidden state in the hidden state sequence is linearly transformed to obtain a scalar weight, and the scalar weight is normalized by softmax to obtain the convergence weight. Each hidden state is multiplied by the corresponding convergence weight and summed to obtain the time window embedding vector of the corresponding node or edge in the heterogeneous graph in the current rolling time window. The heterogeneous relation attention message passing module uses the edge type of the heterogeneous graph as the heterogeneous relation identifier. Relationship attention parameters are set for segment connection edges, turnout guidance edges, station access edges, and vehicle occupancy edges respectively. For each incoming edge of the target node in the heterogeneous graph, the source node time window embedding vector and the target node time window embedding vector are linearly transformed to obtain the source node representation vector and the target node representation vector respectively. The source node representation vector, the target node representation vector and the edge time window embedding vector of the incoming edge are concatenated. The concatenation result is inner productd with the attention vector of the corresponding heterogeneous relation and activated by LeakyReLU. Softmax normalization is performed to obtain the attention weight. The source node representation vectors under the same heterogeneous relation are multiplied with the corresponding attention weight and summed to obtain the relational aggregated message vector. The relational aggregated message vectors of different heterogeneous relations are linearly transformed and then added to obtain the fused message vector. The fused message vector is residually added with the time window embedding vector of the target node and layer normalization is performed to obtain the updated embedding vector of the target node in the heterogeneous graph. The dual-output scoring module is configured with an edge risk scoring output head and a node bottleneck heat output head. The edge risk scoring output head takes the updated edge embedding vector of each edge in the heterogeneous graph as input. The updated edge embedding vector is obtained by concatenating the corresponding edge time window embedding vector and the updated embedding vectors of the two end nodes, followed by a linear transformation and ReLU activation. The edge risk scoring output head performs a linear transformation, ReLU activation, and linear transformation on the updated edge embedding vector in sequence to obtain the edge risk score, and then applies the sigmoid function to the edge risk score to obtain an edge risk score with a value between zero and one. The node bottleneck heat output head takes the updated embedding vector of each node in the heterogeneous graph as input. The node bottleneck heat output head performs a linear transformation, ReLU activation, and linear transformation on the updated embedding vector in sequence to obtain the node bottleneck score, and then applies the sigmoid function to the node bottleneck score to obtain a node bottleneck heat with a value between zero and one.

[0020] The improved GAT model used in this embodiment maintains the same basic computational paradigm as the original GAT model. Both models use linear mapping of adjacent node features, calculation of attention weights, and weighted aggregation to update node representations to complete information transmission, thus inheriting the advantages of the original GAT model in local information fusion and non-uniform neighborhood modeling. The difference lies in that this invention extends the original model at the structural level to address the characteristics of semiconductor material handling scenarios, such as diverse entity types, significant state evolution over time, and risks mainly concentrated in edge and constraint relationships. It introduces heterogeneous graph encoding, time-varying state window encoding, heterogeneous relationship attention message passing, and dual-output head scoring.

[0021] The benefits of these improvements include the ability of the model to not only distinguish between different types of nodes and edges in terms of structure and semantics, but also to integrate state change information across multiple rolling time windows, enabling early detection of risks such as track segment conflicts, switch mutual exclusion waiting, and station congestion. Furthermore, by outputting edge risk scores and node bottleneck heat values ​​respectively, the model results are directly transformed into quantifiable indicators usable for subsequent path cost construction and scheduling optimization. This effectively enhances the adaptability of path planning and scheduling decisions to dynamic disturbances and its ability to specifically address system bottlenecks, thereby improving the robustness and overall efficiency of the material handling system under complex operating conditions.

[0022] In this embodiment, step four specifically includes: A dynamic robust cost graph is constructed based on edge risk scoring and node bottleneck heat. The dynamic robust cost graph uses the topology of the heterogeneous graph as the skeleton and the edges as the cost carriers, generating edge costs for each passable edge in the heterogeneous graph. For each edge in the heterogeneous graph, a basic passage cost is constructed. The basic passage cost is determined by the average passage delay and average queuing time obtained from the corresponding connection relationship of the edge in the rolling time window log dataset. For each edge in the heterogeneous graph, a risk penalty cost is constructed. The risk penalty cost is determined by the edge risk score of the edge. The edge risk score is mapped to a risk penalty coefficient, and the risk penalty coefficient is multiplicatively weighted with the basic passage cost to obtain the edge cost containing the risk term. For each edge in the heterogeneous graph, a bottleneck association penalty cost is constructed. The bottleneck association penalty cost is determined by the bottleneck heat of the adjacent nodes of the edge. The weighted sum of the bottleneck heat of the two ends of the edge is taken as the bottleneck penalty coefficient. The edge cost containing risk terms is multiplicatively weighted with the bottleneck penalty coefficient to obtain the dynamic robust edge cost. A dynamic robust cost graph is constructed based on the dynamic robust edge cost, and a candidate path set is generated under constraints. The constraints include turnout mutual exclusion constraints, section capacity constraints, and task time window constraints. The candidate path set consists of multiple paths from the starting station to the ending station that satisfy the constraints, and each path is represented by an edge sequence arranged in the order of travel. Based on the generated candidate path set, a candidate transfer action set is generated. The candidate transfer action set is a station-level action that occurs during the execution of the task. The candidate transfer action set includes reversing action, yielding and waiting action, and buffering and dwelling action. The reversing action is the action of the vehicle completing the change of travel direction at a designated station or a designated reversing section. The yielding and waiting action is the action of the vehicle staying at a designated station or a designated waiting section to satisfy the turnout mutual exclusion constraint. The buffering and dwelling action is the action of the vehicle staying at a station with buffering capacity to satisfy the section capacity constraint.

[0023] In this embodiment, step five specifically includes: The input to the improved ALNS algorithm is constructed. The input includes a task set, a candidate path set, a candidate transfer action set, and a dynamic robust cost graph. The task set consists of multiple material handling tasks. Each material handling task includes a task identifier, a starting station, an ending station, a task release time, a latest completion time, and a task priority. The task release time is the earliest time when the vehicle is allowed to start executing the task, and the latest completion time is the latest time when the task is allowed to be completed. An initial solution is generated based on the task set. The initial solution includes task and vehicle allocation, path sequence and release time. When generating the initial solution, a path is selected from the candidate path set and a vehicle is allocated for each material handling task. Based on the edge sequence and corresponding dynamic robust edge cost of the selected path on the dynamic robust cost graph, the entry time and exit time of the current task on each edge are calculated, and the release time that satisfies the constraints of task release time and latest completion time is determined. Construct a discrete time slot occupancy table. The discrete time slot occupancy table uses time buckets as time indexes and track sections, switches and stations as resource indexes to record the time slots occupied by each task on the corresponding resources in the current solution. A feasibility mask insertion and filtering mechanism is implemented. For each candidate path and each candidate transfer action of the task to be inserted, a feasibility mask is generated. The feasibility mask consists of the edge sequence of the candidate path and the occupied time slot of the candidate transfer action in the discrete time slot occupancy table. The occupied time slot is compared with the discrete time slot occupancy table one time slot at a time. When the same turnout is occupied by multiple tasks in the same discrete time slot, the number of occupied track sections or stations in the same discrete time slot exceeds the preset capacity, or the task execution time exceeds the interval between the task release time and the latest completion time, the corresponding candidate path or candidate transfer action is determined to be infeasible and removed. In each iteration, a destruction operator is executed to remove some tasks in the current solution, and a hotspot-aware destruction mechanism is executed. The hotspot-aware destruction mechanism constructs a hotspot set based on edge risk scores and node bottleneck heat. The hotspot set includes edges with edge risk scores higher than a preset first threshold and nodes with node bottleneck heat higher than a preset second threshold, and prioritizes removing tasks whose path sequences pass through the hotspot set. In each iteration, the repair operator removes the task and re-inserts it into the current solution, and executes the hotspot awareness repair mechanism. The hotspot awareness repair mechanism prioritizes candidate paths that avoid the hotspot set from the candidate path set, and prioritizes inserting waiting or cached resident actions into stations where the node bottleneck heat is lower than the preset third threshold from the candidate transfer action set. For each insertion scheme, the feasibility mask insertion filtering mechanism is called to filter infeasible schemes, and the scheme that minimizes the cumulative dynamic robust edge cost on the dynamic robust cost graph is selected from the feasible schemes to update the task and vehicle allocation, path sequence and release time. When the preset number of iterations or the preset time budget is reached, the current optimal feasible solution is output. The current optimal feasible solution includes task and vehicle allocation, path sequence and release time.

[0024] In this embodiment, the improved ALNS algorithm maintains the same overall solution process as the original ALNS algorithm, including iterative processes of initial solution generation, removal of some tasks by the destruction operator, and re-insertion of tasks by the repair operator. The difference lies in the introduction of a feasibility mask insertion and filtering mechanism to address the tightly intertwined constraints of turnout mutual exclusion, section capacity, and task time windows. This allows candidate paths and transfer actions to undergo time-slot-by-time constraint verification during the insertion phase. Simultaneously, a hotspot-aware destruction and repair mechanism based on edge risk scoring and node bottleneck heat is introduced to guide the search focus towards high-risk and high-congestion areas.

[0025] The improvements effectively reduced the generation of infeasible solutions, increased search efficiency and solution stability, and enabled scheduling results to more accurately avoid congested and conflict areas, thereby enhancing the real-time response capability and overall operating efficiency of the semiconductor material handling system under complex operating conditions.

[0026] In this embodiment, step six specifically includes: Trigger determination is performed based on the rolling time window log dataset. The trigger determination takes the aligned records of the current rolling time window and the previous rolling time window as input, and calculates the changes in track section occupancy rate, the changes in the number of turnout mutual exclusion waits, and the changes in task timeout risk. The changes in track section occupancy rate are the difference between the ratio of the track section occupancy time in the current rolling time window to the length of the rolling time window and the corresponding ratio in the previous rolling time window. The changes in the number of turnout mutual exclusion waits are the difference between the number of turnout mutual exclusion waits in the current rolling time window and the corresponding number in the previous rolling time window. The changes in task timeout risk are the difference between the number of tasks in the current rolling time window whose expected arrival time is later than the latest completion time and the corresponding number in the previous rolling time window. A trigger condition is determined when any of the following conditions are met: the change in track occupancy rate of any track section exceeds the preset occupancy threshold; the change in the number of turnout mutual exclusion waiting times of any turnout exceeds the preset change threshold; the change in task timeout risk exceeds the preset risk threshold; or the rolling time window log dataset contains records of abnormal vehicle stagnation or unpassable track sections. After triggering, the set of affected tasks is determined. The set of affected tasks consists of tasks that meet the following conditions: the path sequence contains an edge with an edge risk score higher than the preset edge risk score threshold, or the path sequence passes through a node with a bottleneck heat higher than the preset bottleneck heat threshold, or the path sequence contains a track segment corresponding to a track segment that is impassable, or the expected arrival time is later than the latest completion time. After being triggered, the preset time budget is started, and within the preset time budget, the improved ALNS algorithm is called to perform local replanning on the affected task set. Only the task and vehicle allocation, path sequence and release time of the affected tasks are updated. During the update, the dynamic robust cost map is used as the cost basis and the feasibility mask insertion and filtering mechanism is called to ensure that the turnout mutual exclusion constraint, capacity constraint and time window constraint are satisfied. When the preset time budget is reached or all affected task sets have a feasible update plan, output the updated task and vehicle allocation, path sequence and release time.

[0027] In this embodiment, step seven specifically includes: Task and vehicle allocation, path sequence and release time are stored in a versioned manner according to a rolling time window. The actual execution results of the corresponding version are written into the rolling time window log dataset. The actual execution results include the actual release time, actual arrival time and actual path sequence of the task. The improved GAT model is updated based on the versioned storage results and the actual execution results when the preset update cycle arrives.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to an automated material handling system in a 12-inch semiconductor wafer manufacturing plant. This wafer fab uses a track-based automated material handling system responsible for the automatic transfer of wafer cassettes between multiple process equipment, including photolithography, etching, deposition, and cleaning. The entire material handling network consists of 312 track sections, 68 switches, 54 stations, and 86 vehicles. The tracks have a mixed structure of multiple loops and branches, with switches densely distributed near critical process areas. During peak periods, the system experiences high concurrent task loads, leading to significant congestion and mutual exclusion waiting issues.

[0029] In this embodiment, the operational data of the material handling system is collected with a minimum sampling granularity of 1 second. The rolling time window length is set to 300 seconds, and the rolling step size is set to 60 seconds, meaning the rolling time window log dataset is updated every minute. The collected data includes task generation and completion records, vehicle positions and statuses, track section occupancy statuses, switch switching and mutual exclusion waiting records, and station entry and exit statuses. The collected data is uniformly mapped to the system master clock timestamp, then time-aligned and status-completed to construct a continuous and stable rolling time window log dataset.

[0030] A heterogeneous graph is constructed based on the existing track topology of the wafer fab. Track sections, switches, stations, and vehicles are represented as different types of nodes, while track connectivity, switch branching, station access, and vehicle occupancy are represented as different types of edges. Node and edge features are generated from statistics within a rolling time window, such as occupancy duration, number of passes, and average passage delay for track section nodes; mutual exclusion waiting times and average waiting times for switch nodes; and conflict trigger counts and queuing times for edge features. All features are uniformly normalized before being input into the model.

[0031] Regarding model parameter settings, the node and edge embedding dimensions of the improved GAT model are set to 64 dimensions. The most recent five rolling time windows are selected as the input sequence in the time-varying state window encoding module, and the hidden layer dimension of the gated recurrent unit is set to 64. The edge risk score and node bottleneck heat output by the model are both mapped to the interval between 0 and 1.

[0032] During the scheduling phase, the dynamic robust cost graph uses the basic passage cost as its core and incorporates edge risk scoring and node bottleneck heat for multiplicative weighting. Each task in the task set includes a defined release time and latest completion time, with the time window width set to vary from 15 to 45 minutes depending on process requirements. The improved ALNS algorithm has a single-step solution time budget of 3 seconds and a maximum number of iterations of 200. In the local replanning trigger thresholds, the track section occupancy rate change threshold is set to 0.15, the turnout mutual exclusion waiting count change threshold is set to 5, and the task timeout risk change threshold is set to 3 tasks.

[0033] To verify the effectiveness of this invention, the method of this invention was compared with two other methods: a traditional rule-based scheduling method based on static shortest paths, and a standard ALNS scheduling method that does not introduce risk scoring and bottleneck heat. The three methods were subjected to continuous 24-hour simulation tests under the same operating data and task load conditions. The statistical results are shown in Table 1 below: Table 1. Performance Comparison of Different Material Handling Path Planning and Scheduling Methods

[0034] As shown in Table 1, the traditional static shortest path method, due to its inability to perceive changes in system state, tends to concentrate tasks on local shortest paths under high load conditions, leading to severe congestion at switches and sections and the highest task timeout rate. The standard ALNS method improves scheduling quality through global search, but due to the lack of explicit modeling of high-risk edges and bottleneck nodes, recurring congestion still exists in some key areas.

[0035] In comparison, the method of this invention demonstrates significant advantages in key indicators such as average task completion time, task timeout rate, and turnout mutual exclusion waiting. Particularly in terms of the number of track section congestion occurrences, the method of this invention reduces congestion by approximately 49.6% compared to the standard ALNS method, indicating that the dynamic robust cost graph constructed based on edge risk scoring and node bottleneck heat can effectively guide paths to avoid high-risk areas. Furthermore, in terms of local replanning success rate, the method of this invention achieves 93.5%, significantly higher than the standard ALNS method, verifying the effectiveness of the triggered local replanning mechanism under dynamic disturbance scenarios.

[0036] As can be seen from the above embodiments, this invention, by introducing a rolling time window, heterogeneous graph modeling, an improved graph attention network, and a risk-aware adaptive large neighborhood search algorithm, achieves accurate perception and targeted optimization of operational risks and bottlenecks in semiconductor material handling systems. Without significantly increasing computational overhead, it effectively reduces the probability of congestion and conflicts, improves task on-time completion rates and overall system throughput, and possesses good engineering feasibility and practical application value.

[0037] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for optimizing semiconductor material handling paths based on big data analysis, characterized in that, Includes the following steps: Step 1: Collect the operating data of the material handling system and align it according to a unified timestamp to form a rolling time window log dataset; Step 2: Construct a heterogeneous graph based on the track or road network topology, and generate feature vectors for the nodes and edges of the heterogeneous graph based on the rolling time window log dataset; Step 3: Input the heterogeneous graph into the improved GAT model. The improved GAT model consists of a heterogeneous graph encoding module, a time-varying state window encoding module, a heterogeneous relation attention message passing module, and a dual-output head scoring module, which outputs edge risk scores and node bottleneck heat. Step 4: Construct a dynamic robust cost graph based on edge risk scores and node bottleneck heat, and generate a set of candidate paths and a set of candidate transit actions under constraints; Step 5: Input the task set, candidate path set and candidate transfer action set into the improved ALNS algorithm, and solve and output the task and vehicle allocation, path sequence and release time based on the dynamic robust cost graph. The improved ALNS algorithm includes a feasibility mask insertion and filtering mechanism based on turnout mutual exclusion, capacity and time window constraints and a hot spot perception damage and repair mechanism based on edge risk scoring and node bottleneck heat. Step 6: Determine the trigger based on the rolling time window log dataset. When triggered, call the improved ALNS algorithm to perform local replanning on the affected tasks within the preset time budget and output the updated task and vehicle allocation, path sequence and release time. Step 7: Version the task and vehicle allocation, path sequence and release time, and update the improved GAT model.

2. The semiconductor material handling path optimization method based on big data analysis according to claim 1, characterized in that, Step one specifically includes: The operation data of the material handling system is collected. The operation data includes records of task generation, task modification and task completion in the task system, records of vehicle position, vehicle operation status and vehicle loading status in the vehicle control system, and records of turnout status, track section occupancy status and station entry and exit status in the line control system. A unified timestamp is assigned to each piece of operational data collected. Timestamp alignment is performed on operational data with different sampling periods and reporting frequencies. Status completion is performed on the timestamp-aligned operational data. When multiple state records of the same object exist in the same time bucket, they are sorted according to the order of the unified timestamp, and the state record with the latest timestamp is retained as the valid state in the current time bucket. Based on the time bucket sequence that has completed timestamp alignment and state completion, a rolling time window is constructed on the time axis according to the preset rolling window length and rolling step size, forming a rolling time window log dataset composed of multiple rolling time windows arranged in chronological order.

3. The semiconductor material handling path optimization method based on big data analysis according to claim 1, characterized in that, Step two specifically includes: A heterogeneous graph is constructed based on the track or road network topology data of the material handling system. The heterogeneous graph includes track segment nodes, turnout nodes, station nodes and vehicle nodes, as well as segment connection edges, turnout guiding edges, station access edges and vehicle occupancy edges. Based on the rolling time window log dataset, node feature vectors are generated for nodes and edge feature vectors are generated for edges. Missing values ​​are filled and numerical scaling is performed on the node feature vectors and edge feature vectors to obtain a heterogeneous graph.

4. The semiconductor material handling path optimization method based on big data analysis according to claim 1, characterized in that, The improved GAT model includes a heterogeneous graph encoding module, a time-varying state window encoding module, a heterogeneous relational attention message passing module, and a dual-output head scoring module. The heterogeneous graph encoding module sets a node type linear mapping layer for different types of nodes in the heterogeneous graph to obtain the initial embedding vector of the node, and sets an edge type linear mapping layer for different types of edges in the heterogeneous graph to obtain the initial embedding vector of the edge. The time-varying state window encoding module assembles the initial embedding vectors obtained by the same node or the same side in the heterogeneous graph into an embedding sequence in time order within multiple consecutive rolling time windows, and constructs a time window embedding vector based on the embedding sequence. The heterogeneous relation attention message passing module uses the edge type of the heterogeneous graph as the heterogeneous relation identifier. It sets relation attention parameters for segment connection edges, turnout guidance edges, station access edges, and vehicle occupancy edges respectively. For each incoming edge of the target node in the heterogeneous graph, the source node time window embedding vector and the target node time window embedding vector are linearly transformed to obtain the source node representation vector and the target node representation vector respectively. The attention weight is calculated based on the source node representation vector, the target node representation vector, and the edge time window embedding vector of the incoming edge. The source node representation vector under the same heterogeneous relation is multiplied by the corresponding attention weight and summed to obtain the relation aggregated message vector. The relation aggregated message vectors of different heterogeneous relations are linearly transformed and then added to obtain the fused message vector. The fused message vector is added to the target node time window embedding vector by residual addition and layer normalization is performed to obtain the updated embedding vector of the target node in the heterogeneous graph. The dual-output head scoring module is configured with an edge risk scoring output head and a node bottleneck heat output head. The edge risk scoring output head takes the updated edge embedding vector of each edge in the heterogeneous graph as input to obtain the edge risk score, and the node bottleneck heat output head takes the updated embedding vector of each node in the heterogeneous graph as input to obtain the node bottleneck heat.

5. The semiconductor material handling path optimization method based on big data analysis according to claim 1, characterized in that, Step four specifically includes: A dynamic robust cost graph is constructed based on edge risk scoring and node bottleneck heat. The dynamic robust cost graph uses the topology of the heterogeneous graph as the skeleton and the edges as the cost carriers, and generates an edge cost for each passable edge in the heterogeneous graph. For each edge in the heterogeneous graph, a basic passage cost is constructed. The basic passage cost is determined by the average passage delay and average queuing time obtained from the corresponding connection relationship of the edge in the rolling time window log dataset. For each edge in the heterogeneous graph, a risk penalty cost is constructed. The risk penalty cost is determined by the edge risk score of the edge. The edge risk score is mapped to a risk penalty coefficient, and the basic passage cost is multiplicatively weighted with the risk penalty coefficient to obtain the edge cost containing the risk term. For each edge in the heterogeneous graph, a bottleneck association penalty cost is constructed. The bottleneck association penalty cost is determined by the bottleneck heat of the adjacent nodes of the edge. The weighted sum of the bottleneck heat of the two ends of the edge is taken as the bottleneck penalty coefficient. The bottleneck penalty coefficient is used to multiplicatively weight the edge cost containing risk terms to obtain the dynamic robust edge cost. A dynamic robust cost graph is constructed based on the dynamic robust edge cost, and a candidate path set is generated under constraints, including turnout mutual exclusion constraints, section capacity constraints, and task time window constraints. Based on the generated candidate path set, a candidate transit action set is generated. The candidate transit action set refers to the station-level actions that occur during the path execution process. The candidate transit action set includes reversing actions, yielding and waiting actions, and cache dwelling actions.

6. The semiconductor material handling path optimization method based on big data analysis according to claim 1, characterized in that, Step five specifically includes: The input to the improved ALNS algorithm is constructed, which includes a task set, a candidate path set, a candidate transfer action set, and a dynamic robust cost graph; An initial solution is generated based on the task set. The initial solution includes task and vehicle allocation, path sequence and release time. When generating the initial solution, a path is selected from the candidate path set and a vehicle is allocated for each material handling task. Based on the edge sequence and corresponding dynamic robust edge cost of the selected path on the dynamic robust cost graph, the entry time and exit time of the current task on each edge are calculated, and the release time that satisfies the constraints of task release time and latest completion time is determined. Construct a discrete time slot occupancy table, which uses time buckets as time indexes and track sections, switches, and stations as resource indexes to record the time slots occupied by each task on the corresponding resources in the current solution. A feasibility mask insertion and filtering mechanism is implemented. A feasibility mask is generated for each candidate path and each candidate transfer action of the task to be inserted. The occupied time slot is compared with the discrete time slot occupancy table. When the same turnout is occupied by multiple tasks in the same discrete time slot, the number of occupied track sections or stations in the same discrete time slot exceeds the preset capacity, or the task execution time exceeds the interval between the task release time and the latest completion time, the corresponding candidate path or candidate transfer action is determined to be infeasible and removed. In each iteration, a destruction operator is executed to remove some tasks in the current solution, and a hotspot-aware destruction mechanism is executed. The hotspot-aware destruction mechanism constructs a hotspot set based on edge risk scores and node bottleneck heat. The hotspot set includes edges with edge risk scores higher than a preset first threshold and nodes with node bottleneck heat higher than a preset second threshold, and prioritizes removing tasks whose path sequences pass through the hotspot set. In each iteration, the repair operator will remove the task and reinsert it into the current solution, and execute the hot spot awareness repair mechanism. The hot spot awareness repair mechanism will prioritize the selection of candidate paths that avoid the hot spot set in the candidate path set, and prioritize the insertion of waiting or cached resident actions into the station where the node bottleneck heat is lower than the preset third threshold in the candidate transfer action set. For each insertion scheme, the feasibility mask insertion filtering mechanism will be called to filter infeasible schemes, and the scheme that minimizes the cumulative dynamic robust edge cost on the dynamic robust cost graph will be selected from the feasible schemes to update the task and vehicle allocation, path sequence and release time. When the preset number of iterations or the preset time budget is reached, the current optimal feasible solution is output.

7. The semiconductor material handling path optimization method based on big data analysis according to claim 1, characterized in that, Step six specifically includes: Trigger determination is performed based on the rolling time window log dataset. The trigger determination takes the alignment record of the current rolling time window and the previous rolling time window as input, and calculates the change in track section occupancy rate, the change in the number of turnout mutual exclusion waiting times, and the change in task timeout risk, respectively. A trigger condition is determined when any of the following conditions are met: the change in track occupancy rate of any track section exceeds the preset occupancy threshold; the change in the number of turnout mutual exclusion waiting times of any turnout exceeds the preset change threshold; the change in task timeout risk exceeds the preset risk threshold; or the rolling time window log dataset contains records of abnormal vehicle stagnation or unpassable track sections. After triggering, the set of affected tasks is determined. The set of affected tasks consists of tasks that meet the following conditions: the path sequence contains an edge with an edge risk score higher than a preset edge risk score threshold, or the path sequence passes through a node with a bottleneck heat higher than a preset bottleneck heat threshold, or the path sequence contains a track segment corresponding to a track segment that is impassable, or the expected arrival time is later than the latest completion time. After being triggered, the preset time budget is started, and within the preset time budget, the improved ALNS algorithm is called to perform local replanning on the affected task set. Only the task and vehicle allocation, path sequence and release time of the affected tasks are updated. During the update, the dynamic robust cost map is used as the cost basis and the feasibility mask insertion and filtering mechanism is called to ensure that the turnout mutual exclusion constraint, capacity constraint and time window constraint are satisfied. When the preset time budget is reached or all affected task sets have a feasible update plan, output the updated task and vehicle allocation, path sequence and release time.

8. The semiconductor material handling path optimization method based on big data analysis according to claim 1, characterized in that, Step seven specifically includes: Task and vehicle allocation, path sequence and release time are stored in a versioned manner according to a rolling time window. The actual execution results of the corresponding version are written into the rolling time window log dataset. When the preset update cycle arrives, the improved GAT model is updated based on the versioned storage results and the actual execution results.

Images