Wafer manufacturing scheduling trigger decision-making system considering multilevel relationship

Abstract

The invention relates to a wafer manufacturing scheduling trigger decision-making system considering a multilevel relationship. An information collection module receives multi-dimensional heterogeneous data from the system; the information processing module encodes and preprocesses the collected data; the multi-level relation sensing module is used for carrying out relation modeling on the processed data and constructing a multi-level relation model comprising a physical layer, a process layer and a scheduling layer; a scheduling triggering decision module performs scheduling decision and triggering on the multi-level relation model in combination with deep reinforcement learning to realize predictive triggering, event-driven triggering and multi-level collaborative triggering; and the GUI result output module visually displays the scheduling result and the system state. The problem that an existing wafer manufacturing scheduling trigger decision-making system is difficult to meet the requirement for high complexity of wafer manufacturing is solved, the problem of complex scheduling in the wafer manufacturing process can be effectively solved, the production efficiency and the equipment utilization rate are improved, the production cycle time is shortened, and the system adapts to the dynamically changing manufacturing environment.

Description

Technical Field

[0001] The present invention relates to semiconductor manufacturing, intelligent manufacturing and artificial intelligence technologies, and in particular to a wafer manufacturing scheduling triggering decision system that considers multi-level relationships. Background Art

[0002] Wafer manufacturing, a core link in the semiconductor industry chain, has complex workshops characterized by multiple processes (hundreds of steps), multiple devices (thousands of units), and multiple constraints (cleanliness, temperature, and timing). Frequent dynamic events such as equipment failures, process fluctuations, and order changes during production complicate wafer manufacturing operations and make it difficult to determine scheduling trigger timing. Traditional scheduling trigger decision-making methods face the following challenges: Single dimension of relational modeling: Existing methods mostly focus on the binary relationship between workpieces and machines, ignoring cross-layer interactions at the physical layer (equipment layout), process layer (process dependency), and scheduling layer (resource competition). For example, equipment location can affect workpiece flow time, thereby restricting process sequence decisions.

[0003] Insufficient reliability of scheduling trigger decisions: Traditional methods use scheduling strategies triggered by fixed time intervals or simple events, which lack predictive analysis of system status. In addition, traditional trigger decision models are difficult to embed in field knowledge such as process rules (such as photolithography alignment accuracy requirements) and equipment physical characteristics (such as diffusion furnace temperature curves), making scheduling decisions inefficient.

[0004] In recent years, manufacturing scheduling methods based on graph neural networks have gradually emerged. However, existing solutions have not yet solved the key issues of multi-level relationship modeling, intelligent triggering mechanisms, and physical constraint embedding, making it difficult to meet the high complexity requirements of wafer manufacturing. Summary of the Invention

[0005] In response to the problems of the existing wafer manufacturing scheduling trigger decision system in terms of single relationship modeling dimension, insufficient reliability of scheduling trigger decision, and difficulty in meeting the high complexity requirements of wafer manufacturing, the present invention provides a wafer manufacturing scheduling trigger decision system that considers multi-level relationships. Based on the consideration of the complex spatiotemporal coupling relationship of the wafer manufacturing workshop, the multi-level relationship between each processing area and equipment is analyzed, providing a decision support tool for the optimization control of the semiconductor manufacturing system.

[0006] The technical solution of the present invention is: A wafer manufacturing scheduling trigger decision system that considers multi-level relationships includes an information collection module, an information processing module, a multi-level relationship perception module, a scheduling trigger decision module and a GUI result output module; wherein the information collection module accepts multi-dimensional heterogeneous data from the system, including equipment status data, process flow data and material status data; the information processing module encodes and preprocesses the data collected by the information collection module, including process status coding, equipment status coding and material status coding; the multi-level relationship perception module performs relationship modeling on the data processed by the information processing module, and constructs a multi-level relationship model including a physical layer, a process layer and a scheduling layer; the scheduling trigger decision module combines the multi-level relationship model with deep reinforcement learning to make scheduling decisions and trigger, thereby realizing predictive triggering, event-driven triggering and multi-level collaborative triggering; the GUI result output module visualizes the scheduling results and system status generated by the scheduling trigger decision module. Further, specifically including: The information collection module includes a four-layer metadata architecture consisting of a meta-metamodel, a metamodel, metadata, and user data. First, the metamodel instances contained in the product, system, and equipment meta-metamodels are described based on ontology. The metadata involved is then described using the vocabulary, rules, and relationships in the metamodel instances. Finally, the user data is described using the metadata, thereby achieving a unified representation of multi-dimensional heterogeneous data in the wafer manufacturing system. The information processing module includes: a process status coding submodule, an equipment status coding submodule, and a material status coding submodule; wherein: the process status coding submodule encodes the status of each production process in time dimension; the equipment status coding submodule encodes the availability, performance and maintenance status of the equipment; and the material status coding submodule encodes the status and location of the material; The multi-level relationship perception module adopts a two-stage graph neural network architecture, including: a node feature extraction stage, which uses a graph convolutional network to extract feature representations of nodes at each layer; an inter-layer relationship modeling stage, which captures complex dependencies between different layers through a cross-layer message passing mechanism; The scheduling trigger decision module models the wafer manufacturing scheduling problem as a Markov decision process, where the state is represented by a multi-level system state encoded by a heterogeneous graph neural network; the action is the scheduling decision and trigger signal; and the reward function comprehensively considers multiple objectives such as production capacity, cycle time, and equipment utilization. The scheduling trigger decision module adopts the proximal policy optimization algorithm to train the scheduling policy network to achieve end-to-end scheduling decision learning; The scheduling trigger decision module adopts three trigger mechanisms: predictive triggering, active scheduling triggering based on state prediction; event-driven triggering, passive scheduling triggering based on key events; multi-level collaborative triggering, collaborative scheduling triggering based on inter-layer relationships; The system integrates knowledge learning technologies in the scheduling domain, including: process constraint embedding, encoding process flow constraints into the graph structure; equipment physical model, integrating equipment performance and physical characteristics into node features; and spatiotemporal coupling modeling, explicitly modeling the spatiotemporal coupling relationship of workpieces flowing between equipment.

[0007] Furthermore, the information collection module is as follows: The information collection module is responsible for collecting multi-dimensional heterogeneous data from the wafer manufacturing system, including equipment status data, process flow data and material status data; this module adopts a four-layer metadata architecture including: meta-metamodel, namely ontology: describing basic concepts and defining the ontology of the wafer manufacturing field; metamodel: defining entity types and relationships in the wafer manufacturing field based on ontology; metadata: describing the attributes and relationships of specific entities; user data: dynamic data in the actual production process; through a hierarchical structure, the system can uniformly express the multi-dimensional heterogeneous data of the wafer manufacturing system.

[0008] Furthermore, the information processing module is as follows: Process status coding submodule: encodes the status of each production process in the order of time dimension; the process status coding can be expressed as:

[0009] in, E process ( i , t ) indicates the process i In time t The encoding vector of P i Indicates process parameters, S i ( t ) indicates the process status, T i ( t ) indicates process time information, f process is the encoding function; Equipment status coding submodule: encodes the availability, performance, and maintenance status of the equipment; the equipment status coding can be expressed as:

[0010] in, E equipment ( j ,t ) indicates a device j In time t The encoding vector of A j ( t ) indicates device availability, P j ( t ) indicates the performance of the device, M j ( t ) indicates the equipment maintenance status, f equipment is the encoding function; Material status coding submodule: encodes the status and location of the material; the material status coding can be expressed as:

[0011] in, E material ( k , t ) indicates material k In time t The encoding vector of S k ( t ) indicates the material status, L k ( t ) indicates the material location, Q k ( t ) indicates the material yield information, f material is the encoding function; The encoding results form a unified state representation as the input of the graph neural network.

[0012] Furthermore, the multi-level relationship perception module is as follows: The multi-level relationship perception module builds a multi-level relationship model based on a heterogeneous graph neural network, including: Physical layer: represents physical entities and their relationships, which can be represented as graphs G physical =( V physical , E physical ),in, V physical is a collection of physical entity nodes, E physical is the set of physical relationship edges; Process layer: represents process steps, process parameters and their dependencies, which can be represented as a graph G process =( Vprocess , E process ),in, V process is a collection of process nodes, E process is the set of process dependency edges; Scheduling layer: represents scheduling decisions, resource allocation and their constraints, which can be represented as a graph G scheduling =( V scheduling , E scheduling ), V scheduling is the set of scheduling nodes, E scheduling is the set of scheduling constraint edges; The system adopts a two-stage graph neural network architecture: Node feature extraction stage: Use graph convolutional network to extract the feature representation of nodes at each layer; for nodes v , its feature update can be expressed as:

[0013] in, Representation node v In the l The feature representation of the layer, N ( v ) represents a node v The neighbor set of AGGREGATE is an aggregate function, W (l) and b (l) is a learnable parameter, s is the activation function, u Representation node v adjacent nodes of ; Inter-layer relationship modeling stage: capture the complex dependencies between different layers through the cross-layer message passing mechanism; for nodes from different layers and , its inter-layer message passing can be expressed as:

[0014]

[0015] in, m o,p Represents a slave node v p Pass to node v o News, eo,p is the edge feature between nodes, N cross ( v o ) is a node v o The cross-layer neighbor set of f and P is a learnable function, is a node v o The feature representation updated by inter-layer message passing, is a node v o The original feature representation of is a node v p feature representation.

[0016] Furthermore, the scheduling trigger decision module is as follows: The scheduling trigger decision module combines deep reinforcement learning to make scheduling decisions and triggers. The system models the wafer manufacturing scheduling problem as a Markov decision process: State: The multi-level system state encoded by the heterogeneous graph neural network can be expressed as s t = { , , }; in, yes t The diagram of the physical layer at this moment is represented as follows: yes t The diagram of the process layer at the moment is represented. yes t Graph representation of the time scheduling layer; Action: Scheduling decision and trigger signal, which can be expressed as a t ={ d t , t t},in d t is a scheduling decision, t t is the trigger signal; Reward: Taking into account the multiple objectives of production capacity, cycle time, and equipment utilization, it can be expressed as:

[0017] in, r throughput , r cycle_time , r utilizationRepresents the rewards related to production capacity, cycle time and equipment utilization, respectively. r trigger Indicates triggering of related rewards. w 1, w 2, w 3. w 4 is the weight coefficient; The system uses a proximal policy optimization algorithm to train the scheduling policy network, and its objective function is:

[0018] in, r t ( i ) represents the probability ratio of the new and old strategies, represents the advantage function estimate, e is the cropping parameter, is the time step t Expected calculation of clip is the clipping function.

[0019] Furthermore, the mechanism for triggering the scheduling trigger decision module is as follows: Predictive triggering: proactive scheduling triggering based on state prediction; the system predicts future states s t+Δt , when the predicted status meets the triggering conditions, the scheduling decision is triggered in advance:

[0020] in, f trigger To trigger the evaluation function, i predictive is the predictive trigger threshold, I is the indicator function; Event-driven triggering: passive scheduling triggering based on key events; when the system detects a key event When , the scheduling decision is triggered:

[0021] in, E critical It is a collection of key events; multi-level collaborative triggering: collaborative scheduling triggering based on inter-layer relationships; the system comprehensively considers the status and relationships of different levels and performs collaborative triggering:

[0022] in, f collaborative is the collaborative trigger evaluation function, i collaborative is the collaborative triggering threshold.

[0023] Furthermore, the system integrates scheduling domain knowledge learning technology to embed wafer manufacturing domain knowledge and physical constraints into the model: Process constraint embedding: Encode the process flow constraints into the graph structure; for the process constraint c, its embedding can be expressed as:

[0024]

[0025] in, f constraint is the constraint encoding function, C is the set of process constraints, g embed is the graph embedding function; It is the updated process layer graph structure after embedding process constraints. e c It is a process constraint c Embedding vector after processing by the numbering function; Device physical model: Integrate device performance and physical characteristics into node characteristics; for device nodes v equipment , its feature enhancement can be expressed as:

[0026] in, P equipment Indicates the physical characteristics of the device. f physics is the physical property encoding function, is a node v equipment The original feature representation of is a node v equipment New feature representation after feature enhancement; Spatiotemporal coupling modeling: Explicitly modeling the spatiotemporal coupling relationship of artifacts flowing between devices; w Slave device I To device J The flow of , its spatiotemporal coupling relationship can be expressed as:

[0027] in, L I and L J Representation device I and J location, Represents workpiece w Slave device I To device J The flow time,f spatiotemporal is the spatiotemporal coupling encoding function.

[0028] Furthermore, the system workflow is as follows: Data collection and preprocessing: Collect multi-source heterogeneous data from MES, ERP devices, and sensors, and represent them uniformly through a multi-dimensional state coding mechanism; Multi-level relationship perception: Construct heterogeneous graph representations for the physical, process, and scheduling layers, and extract multi-level features and relationships through a two-stage graph neural network. Scheduling status assessment: Evaluate scheduling requirements based on current system status and historical data, calculate key performance indicators and bottleneck analysis; Scheduling decisions and triggering: Scheduling decisions are generated through a deep reinforcement learning policy network, and the timing of scheduling execution is determined based on the trigger mechanism; Result feedback and optimization: Record scheduling execution results and system responses, and continuously optimize scheduling strategies through online learning.

[0029] Furthermore, the basic concepts described in the meta-metamodel include products, systems, and equipment; the entity types and relationships in the metamodel include equipment type, process type, and material type; the attributes and relationships of the entities in the metadata include parameters of specific equipment and requirements of specific processes; and the dynamic data in the user data includes equipment status, workpiece location, and processing progress.

[0030] The beneficial effects of the present invention are: It can effectively handle complex scheduling issues in the wafer manufacturing process, improve production efficiency and equipment utilization, reduce production cycle time, and adapt to the dynamically changing manufacturing environment. The details are as follows: 1. Through multi-level heterogeneous graph representation, the complex relationships in the wafer manufacturing system are fully captured to improve the accuracy and effectiveness of scheduling decisions.

[0031] 2. Through a two-stage graph neural network architecture, it effectively handles complex relationships within and between layers, enhancing the model's expressiveness and learning capabilities.

[0032] 3. Through the intelligent trigger mechanism, precise scheduling intervention can be achieved to reduce unnecessary scheduling calculations and system interference.

[0033] 4. Improve model interpretability and reliability by scheduling domain knowledge learning, integrating domain knowledge and physical constraints. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 System structure diagram of the present invention; Figure 2 This is a metadata model framework diagram of the present invention; Figure 3This is a schematic diagram of information coding of the present invention; Figure 4 Schematic diagram of the multi-level heterogeneous graph structure of the present invention; Figure 5 This is a scheduling trigger decision flow chart of the present invention. DETAILED DESCRIPTION

[0035] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0036] A wafer manufacturing scheduling trigger decision system that considers multi-level relationships includes an information collection module, an information processing module, a multi-level relationship perception module, a scheduling trigger decision module and a GUI result output module, wherein the information collection module accepts multi-dimensional heterogeneous data from the system, including equipment status data, process flow data and material status data; the information processing module encodes and preprocesses the data collected by the information collection module, including process status coding, equipment status coding and material status coding; the multi-level relationship perception module performs relationship modeling on the data processed by the information processing module, and constructs a multi-level relationship model including a physical layer, a process layer and a scheduling layer; the scheduling trigger decision module combines the multi-level relationship model constructed by the multi-level relationship perception module with deep reinforcement learning to make scheduling decisions and trigger, thereby realizing predictive triggering, event-driven triggering and multi-level collaborative triggering; the GUI result output module visualizes the scheduling results and system status generated by the scheduling trigger decision module.

[0037] The information collection module includes: a four-layer metadata architecture of meta-metamodel (i.e., ontology), metamodel, metadata and user data. First, based on ontology, the metamodel instances contained in the product, system and equipment meta-metamodel are described, and the vocabulary, rules and relationships in the metamodel instance are used to describe the metadata involved. Finally, the user data is described through metadata, thereby realizing a unified expression of multi-dimensional heterogeneous data of the wafer manufacturing system.

[0038] The information processing module includes: a process status coding submodule, an equipment status coding submodule and a material status coding submodule, wherein: the process status coding submodule encodes the status of each production process in order of time dimension; the equipment status coding submodule encodes the availability, performance and maintenance status of the equipment; the material status coding submodule encodes the status and position of materials such as wafers and masks.

[0039] The multi-level relationship perception module adopts a two-stage graph neural network architecture, including: a node feature extraction stage, using a graph convolutional network to extract the feature representation of nodes in each layer; an inter-layer relationship modeling stage, capturing the complex dependencies between different layers through a cross-layer message passing mechanism.

[0040] The scheduling trigger decision module models the wafer manufacturing scheduling problem as a Markov decision process, in which: the state is represented by a multi-level system state encoded by a heterogeneous graph neural network; the action is the scheduling decision and trigger signal; and the reward function comprehensively considers multiple objectives such as production capacity, cycle time, and equipment utilization.

[0041] The scheduling trigger decision module adopts the proximal policy optimization (PPO) algorithm to train the scheduling policy network to achieve end-to-end scheduling decision learning.

[0042] The scheduling trigger decision module is designed with three trigger mechanisms: predictive triggering, which is an active scheduling trigger based on state prediction; event-driven triggering, which is a passive scheduling trigger based on key events; and multi-level collaborative triggering, which is a collaborative scheduling trigger based on inter-layer relationships.

[0043] The system integrates scheduling domain knowledge learning technologies, including: process constraint embedding, encoding process flow constraints into the graph structure; equipment physical model, integrating equipment performance and physical characteristics into node features; and spatiotemporal coupling modeling, explicitly modeling the spatiotemporal coupling relationship of workpieces flowing between equipment.

[0044] In order to realize the intelligent scheduling of wafer manufacturing system, the present invention proposes a wafer manufacturing scheduling trigger decision system considering multi-level relationships. The system is divided into five parts: information collection module, information processing module, multi-level relationship perception module, scheduling trigger decision module and GUI result output module. Figure 1 shown.

[0045] The information collection module is responsible for collecting multi-dimensional heterogeneous data from the wafer manufacturing system, including equipment status data, process data and material status data. This module adopts a four-layer metadata architecture, such as Figure 2 As shown in the figure, it includes: (1) Meta-metamodel (ontology): describes basic concepts such as products, systems, and equipment, and defines the ontology of the wafer manufacturing field. (2) Metamodel: defines the entity types and relationships in the wafer manufacturing field based on the ontology, such as equipment type, process type, material type, etc. (3) Metadata: describes the attributes and relationships of specific entities, such as the parameters of specific equipment, the requirements of specific processes, etc. (4) User data: dynamic data in the actual production process, such as equipment status, workpiece location, processing progress, etc. Through the hierarchical structure, the system can uniformly express the multi-dimensional heterogeneous data of the wafer manufacturing system, providing a basis for subsequent graph neural network modeling.

[0046] The information processing module encodes and preprocesses the collected data, such as Figure 3 Shown, including: (1) Process status coding submodule: encodes the status of each production process in the order of time dimension. The process status coding can be expressed as:

[0047] in, E process ( i , t ) indicates the process i In time t The encoding vector of P i Indicates process parameters, S i ( t ) indicates the process status, T i ( t ) indicates process time information, f process is the encoding function.

[0048] (2) Equipment status coding submodule: Encodes the availability, performance, and maintenance status of the equipment. The equipment status coding can be expressed as:

[0049] in, E equipment ( j , t ) indicates a device j In time t The encoding vector of A j ( t ) indicates device availability, P j ( t ) indicates the performance of the device, M j ( t ) indicates the equipment maintenance status, f equipment is the encoding function.

[0050] (3) Material status coding submodule: encodes the status and position of materials such as wafers and photomasks. The material status coding can be expressed as:

[0051] in, E material ( k ,t ) indicates material k In time t The encoding vector of S k ( t ) indicates the material status, L k ( t ) indicates the material location, Q k ( t ) indicates the material yield information, f material is the encoding function.

[0052] The encoding results form a unified state representation as the input of the graph neural network.

[0053] The multi-level relationship perception module builds a multi-level relationship model based on heterogeneous graph neural network, such as Figure 4 Shown, including: (1) Physical layer: represents physical entities such as equipment and materials and their relationships, which can be represented as graphs. G physical =( V physical , E physical ),in, V physical is a collection of physical entity nodes, E physical is the set of physical relationship edges.

[0054] (2) Process layer: represents the process steps, process parameters and their dependencies, which can be represented as a graph G process =( V process , E process ),in, V process is a collection of process nodes, E process is the set of process dependency edges.

[0055] (3) Scheduling layer: represents scheduling decisions, resource allocation and their constraints, which can be represented as a graph G scheduling =( V scheduling , E scheduling ), V scheduling is the set of scheduling nodes, E scheduling is the set of scheduling constraint edges.

[0056] The system adopts a two-stage graph neural network architecture: (1) Node feature extraction stage: Use graph convolutional networks to extract the feature representation of nodes at each layer. v , its feature update can be expressed as:

[0057] in, Representation node v In the l The feature representation of the layer, N ( v ) represents a node v The neighbor set of AGGREGATE is an aggregate function, W (l) and b (l) is a learnable parameter, s is the activation function, u Representation node v adjacent nodes.

[0058] (2) Inter-layer relationship modeling stage: Capture the complex dependencies between different layers through the cross-layer message passing mechanism; for nodes from different layers and , its inter-layer message passing can be expressed as:

[0059]

[0060] in, m o,p Represents a slave node v p Pass to node v o News, e o,p is the edge feature between nodes, N cross ( v o ) is a node v o The cross-layer neighbor set of f and P is a learnable function, is a node v o The feature representation updated by inter-layer message passing, is a node v o The original feature representation of is a node v p feature representation.

[0061] The scheduling trigger decision module combines deep reinforcement learning to make scheduling decisions and triggers, such as Figure 5 As shown in Figure 2, the system models the wafer manufacturing scheduling problem as a Markov decision process: (1) State: The multi-level system state encoded by the heterogeneous graph neural network can be expressed as s t = { , , }; in, yes t The diagram of the physical layer at this moment is represented as follows: yes t The diagram of the process layer at the moment is represented. yes t Graph representation of the time scheduling layer.

[0062] (2) Action: Scheduling decision and trigger signal, which can be expressed as a t ={ d t , t t},in d t is a scheduling decision, t t is the trigger signal.

[0063] (3) Rewards: Taking into account multiple objectives such as production capacity, cycle time, and equipment utilization, it can be expressed as:

[0064] in, r throughput , r cycle_time , r utilization Represents the rewards related to production capacity, cycle time and equipment utilization, respectively. r trigger Indicates triggering of related rewards. w 1, w 2, w 3. w 4 is the weight coefficient.

[0065] The system uses the proximal policy optimization (PPO) algorithm to train the scheduling policy network, and its objective function is:

[0066] in, r t ( i) represents the probability ratio of the new and old strategies, represents the advantage function estimate, e is the cropping parameter, is the time step t Expected calculation of clip is the clipping function.

[0067] The system is designed with three trigger mechanisms: (1) Predictive triggering: proactive scheduling triggering based on state prediction. The system predicts the future state. s t+Δt , when the predicted status meets the triggering conditions, the scheduling decision is triggered in advance:

[0068] in, f trigger To trigger the evaluation function, i predictive is the predictive trigger threshold, and I is the indicator function.

[0069] (2) Event-driven triggering: Passive scheduling triggering based on key events (such as urgent orders, equipment abnormalities, etc.). When the system detects a key event When , the scheduling decision is triggered:

[0070] in, E critical is a set of key events.

[0071] (3) Multi-level collaborative triggering: Collaborative scheduling triggering based on inter-layer relationships. The system comprehensively considers the status and relationships of different levels and performs collaborative triggering:

[0072] in, f collaborative is the collaborative trigger evaluation function, i collaborative is the collaborative triggering threshold.

[0073] The system integrates scheduling domain knowledge learning technology to embed wafer manufacturing domain knowledge and physical constraints into the model: (1) Process constraint embedding: Encode the process flow constraints into the graph structure. For the process constraint c, its embedding can be expressed as:

[0074]

[0075] in, f constraintis the constraint encoding function, C is the set of process constraints, g embed is the graph embedding function. It is the updated process layer graph structure after embedding process constraints. e c It is a process constraint c The embedding vector after being processed by the numbering function.

[0076] (2) Device physical model: Integrate device performance and physical characteristics into node characteristics. v equipment , its feature enhancement can be expressed as:

[0077] in, P equipment Indicates the physical characteristics of the device. f physics is the physical property encoding function, is a node v equipment The original feature representation of is a node v equipment New feature representation after feature enhancement.

[0078] (3) Spatiotemporal coupling modeling: Explicitly modeling the spatiotemporal coupling relationship of artifacts flowing between devices; w Slave device I To device J The flow of , its spatiotemporal coupling relationship can be expressed as:

[0079] in, L I and L J Representation device I and J location, Represents workpiece w Slave device I To device J The flow time, f spatiotemporal is the spatiotemporal coupling encoding function.

[0080] The system's workflow is as follows: (1) Data collection and preprocessing: Collect multi-source heterogeneous data from MES (manufacturing information management system), ERP (enterprise resource planning system) equipment and sensors, and represent them uniformly through a multi-dimensional state coding mechanism.

[0081] (2) Multi-level relationship perception: Construct heterogeneous graph representations of the physical layer, process layer, and scheduling layer, and extract multi-level features and relationships through a two-stage graph neural network.

[0082] (3) Scheduling status evaluation: Evaluate scheduling requirements based on current system status and historical data, calculate key performance indicators and bottleneck analysis.

[0083] (4) Scheduling decision and triggering: Scheduling decisions are generated through a deep reinforcement learning policy network, and the timing of scheduling execution is determined based on the trigger mechanism.

[0084] (5) Result feedback and optimization: Record scheduling execution results and system responses, and continuously optimize scheduling strategies through online learning.

[0085] The above-described embodiment merely represents one embodiment of the present invention. While the description is relatively specific and detailed, it should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art would be able to make various modifications and improvements without departing from the spirit of the present invention, and these modifications and improvements fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be determined by the appended claims.

Claims

1. A wafer manufacturing scheduling trigger decision system considering multi-level relationships, characterized in that: It includes an information collection module, an information processing module, a multi-level relationship perception module, a scheduling trigger decision module and a GUI result output module; among them, the information collection module accepts multi-dimensional heterogeneous data from the system, including equipment status data, process flow data and material status data; the information processing module encodes and preprocesses the data collected by the information collection module, including process status coding, equipment status coding and material status coding; the multi-level relationship perception module performs relationship modeling on the data processed by the information processing module, and constructs a multi-level relationship model including the physical layer, process layer and scheduling layer; the scheduling trigger decision module combines the multi-level relationship model with deep reinforcement learning to make scheduling decisions and trigger, realizing predictive triggering, event-driven triggering and multi-level collaborative triggering; the GUI result output module visualizes the scheduling results and system status generated by the scheduling trigger decision module.

2. The wafer manufacturing scheduling trigger decision system considering multi-level relationships according to claim 1 is characterized in that: Specifically include: The information collection module includes a four-layer metadata architecture consisting of a meta-metamodel, a metamodel, metadata, and user data. First, the metamodel instances contained in the product, system, and equipment meta-metamodels are described based on ontology. The metadata involved is then described using the vocabulary, rules, and relationships in the metamodel instances. Finally, the user data is described using the metadata, thereby achieving a unified representation of multi-dimensional heterogeneous data in the wafer manufacturing system. The information processing module includes: a process status coding submodule, an equipment status coding submodule, and a material status coding submodule; wherein: the process status coding submodule encodes the status of each production process in time dimension; the equipment status coding submodule encodes the availability, performance and maintenance status of the equipment; and the material status coding submodule encodes the status and location of the material; The multi-level relationship perception module adopts a two-stage graph neural network architecture, including: a node feature extraction stage, which uses a graph convolutional network to extract feature representations of nodes at each layer; an inter-layer relationship modeling stage, which captures complex dependencies between different layers through a cross-layer message passing mechanism; The scheduling trigger decision module models the wafer manufacturing scheduling problem as a Markov decision process, where the state is represented by a multi-level system state encoded by a heterogeneous graph neural network; the action is the scheduling decision and trigger signal; and the reward function comprehensively considers multiple objectives such as production capacity, cycle time, and equipment utilization. The scheduling trigger decision module adopts the proximal policy optimization algorithm to train the scheduling policy network to achieve end-to-end scheduling decision learning; The scheduling trigger decision module adopts three trigger mechanisms: predictive triggering, active scheduling triggering based on state prediction; event-driven triggering, passive scheduling triggering based on key events; multi-level collaborative triggering, collaborative scheduling triggering based on inter-layer relationships; The system integrates knowledge learning technologies in the scheduling domain, including: process constraint embedding, encoding process flow constraints into the graph structure; equipment physical model, integrating equipment performance and physical characteristics into node features; and spatiotemporal coupling modeling, explicitly modeling the spatiotemporal coupling relationship of workpieces flowing between equipment.

3. The wafer manufacturing scheduling trigger decision system considering multi-level relationships according to claim 1 is characterized in that: The information collection module is as follows: The information collection module is responsible for collecting multi-dimensional heterogeneous data from the wafer manufacturing system, including equipment status data, process flow data and material status data; this module adopts a four-layer metadata architecture including: meta-metamodel, namely ontology: describing basic concepts and defining the ontology of the wafer manufacturing field; metamodel: defining entity types and relationships in the wafer manufacturing field based on ontology; metadata: describing the attributes and relationships of specific entities; user data: dynamic data in the actual production process; through a hierarchical structure, the system can uniformly express the multi-dimensional heterogeneous data of the wafer manufacturing system.

4. The wafer manufacturing scheduling trigger decision system considering multi-level relationships according to claim 1 is characterized in that: The information processing modules are as follows: Process status coding submodule: encodes the status of each production process in the order of time dimension; the process status coding can be expressed as: in, E process ( i , t ) indicates the process i In time t The encoding vector of P i Indicates process parameters, S i ( t ) indicates the process status, T i ( t ) indicates process time information, f process is the encoding function; Equipment status coding submodule: encodes the availability, performance, and maintenance status of the equipment; the equipment status coding can be expressed as: in, E equipment ( j , t ) indicates a device j In time t The encoding vector of A j ( t ) indicates device availability, P j ( t ) indicates the performance of the device, M j ( t ) indicates the equipment maintenance status, f equipment is the encoding function; Material status coding submodule: encodes the status and location of the material; the material status coding can be expressed as: in, E material ( k , t ) indicates material k In time t The encoding vector of S k ( t ) indicates the material status, L k ( t ) indicates the material location, Q k ( t ) indicates the material yield information, f material is the encoding function; The encoding results form a unified state representation as the input of the graph neural network.

5. The wafer manufacturing scheduling trigger decision system considering multi-level relationships according to claim 1, characterized in that: The multi-level relationship perception module is as follows: The multi-level relationship perception module builds a multi-level relationship model based on a heterogeneous graph neural network, including: Physical layer: represents physical entities and their relationships, which can be represented as graphs G physical =( V physical , E physical ),in, V physical is a collection of physical entity nodes, E physical is the set of physical relationship edges; Process layer: represents process steps, process parameters and their dependencies, which can be represented as a graph G process =( V process , E process ),in, V process is a collection of process nodes, E process is the set of process dependency edges; Scheduling layer: represents scheduling decisions, resource allocation and their constraints, which can be represented as a graph G scheduling =( V scheduling , E scheduling ), V scheduling is the set of scheduling nodes, E scheduling is the set of scheduling constraint edges; The system adopts a two-stage graph neural network architecture: Node feature extraction stage: Use graph convolutional network to extract the feature representation of nodes at each layer; for nodes v , its feature update can be expressed as: in, Representation node v In the l The feature representation of the layer, N ( v ) represents a node v The neighbor set of AGGREGATE is an aggregate function, W (l) and b (l) is a learnable parameter, σ is the activation function, u Representation node v adjacent nodes of ; Inter-layer relationship modeling stage: capture the complex dependencies between different layers through the cross-layer message passing mechanism; for nodes from different layers and , its inter-layer message passing can be expressed as: in, m o,p Represents a slave node v p Pass to node v o News, e o,p is the edge feature between nodes, N cross ( v o ) is a node v o The cross-layer neighbor set of φ and Ψ is a learnable function, is a node v o The feature representation updated by inter-layer message passing, is a node v o The original feature representation of is a node v p feature representation.

6. The wafer manufacturing scheduling trigger decision system considering multi-level relationships according to claim 1, characterized in that: The scheduling trigger decision module is as follows: The scheduling trigger decision module combines deep reinforcement learning to make scheduling decisions and triggers. The system models the wafer manufacturing scheduling problem as a Markov decision process: State: The multi-level system state encoded by the heterogeneous graph neural network can be expressed as s t = { , , }; in, yes t The diagram of the physical layer at this moment is represented as follows: yes t The diagram of the process layer at the moment is represented. yes t Graph representation of the time scheduling layer; Action: Scheduling decision and trigger signal, which can be expressed as a t ={ d t , τ t },in d t is a scheduling decision, τ t is the trigger signal; Reward: Taking into account the multiple objectives of production capacity, cycle time, and equipment utilization, it can be expressed as: in, r throughput , r cycle_time , r utilization Represents the rewards related to production capacity, cycle time and equipment utilization, respectively. r trigger Indicates triggering of related rewards. w 1, w 2, w 3. w 4 is the weight coefficient; The system uses a proximal policy optimization algorithm to train the scheduling policy network, and its objective function is: in, r t ( θ ) represents the probability ratio of the new and old strategies, represents the advantage function estimate, ε is the cropping parameter, is the time step t Expected calculation of clip is the clipping function.

7. The wafer manufacturing scheduling trigger decision system considering multi-level relationships according to claim 1, characterized in that: The mechanism for triggering the scheduling trigger decision module is as follows: Predictive triggering: proactive scheduling triggering based on state prediction; the system predicts future states s t+Δt , when the predicted status meets the triggering conditions, the scheduling decision is triggered in advance: in, f trigger To trigger the evaluation function, θ predictive is the predictive trigger threshold, I is the indicator function; Event-driven triggering: passive scheduling triggering based on key events; when the system detects a key event When , the scheduling decision is triggered: in, E critical is a set of key events; Multi-level collaborative triggering: Collaborative scheduling triggering based on inter-layer relationships; the system comprehensively considers the status and relationships of different levels to perform collaborative triggering: in, f collaborative is the collaborative trigger evaluation function, θ collaborative is the collaborative triggering threshold.

8. The wafer manufacturing scheduling trigger decision system considering multi-level relationships according to claim 1, characterized in that: The system integrates scheduling domain knowledge learning technology to embed wafer manufacturing domain knowledge and physical constraints into the model: Process constraint embedding: encoding process flow constraints into the graph structure; For the process constraint c, its embedding can be expressed as: in, f constraint is the constraint encoding function, C is the set of process constraints, g embed is the graph embedding function; It is the updated process layer graph structure after embedding process constraints. e c It is a process constraint c Embedding vector after processing by the numbering function; Device physical model: Integrate device performance and physical characteristics into node characteristics; for device nodes v equipment , its feature enhancement can be expressed as: in, P equipment Indicates the physical characteristics of the device. f physics is the physical property encoding function, is a node v equipment The original feature representation of is a node v equipment New feature representation after feature enhancement; Spatiotemporal coupling modeling: Explicitly modeling the spatiotemporal coupling relationship of artifacts flowing between devices; w Slave device I To device J The spatiotemporal coupling relationship of the flow can be expressed as: in, L I and L J Representation device I and J location, Indicates workpiece w Slave device I To device J The flow time, f spatiotemporal is the spatiotemporal coupling encoding function.

9. The wafer manufacturing scheduling trigger decision system considering multi-level relationships according to claim 1, characterized in that: The system's workflow is as follows: Data collection and preprocessing: Collect multi-source heterogeneous data from MES, ERP devices, and sensors, and represent them uniformly through a multi-dimensional state coding mechanism; Multi-level relationship perception: Construct heterogeneous graph representations for the physical, process, and scheduling layers, and extract multi-level features and relationships through a two-stage graph neural network. Scheduling status assessment: Evaluate scheduling requirements based on current system status and historical data, calculate key performance indicators and bottleneck analysis; Scheduling decisions and triggering: Scheduling decisions are generated through a deep reinforcement learning policy network, and the timing of scheduling execution is determined based on the trigger mechanism; Result feedback and optimization: Record scheduling execution results and system responses, and continuously optimize scheduling strategies through online learning.

10. The wafer manufacturing scheduling trigger decision system considering multi-level relationships according to claim 3, characterized in that: The basic concepts described in the meta-metamodel include products, systems, and equipment; the entity types and relationships in the metamodel include equipment type, process type, and material type; the attributes and relationships of the entities in the metadata include parameters of specific equipment and requirements of specific processes; the dynamic data in the user data includes equipment status, workpiece location, and processing progress.

Images