A hierarchical multi-cluster resource and multi-scale task combination scheduling method based on hypergraph theory in intelligent manufacturing

Abstract

This invention provides a hierarchical multi-cluster resource and multi-scale task combination scheduling method based on hypergraph theory for intelligent manufacturing. Addressing the challenges of dispersed multi-cluster resources, coupled multi-scale tasks, and cross-level scheduling in the entire intelligent manufacturing production-monitoring-testing process, this method aims to achieve structured representation of resources and efficient task scheduling through hypergraph modeling, significantly improving the optimization efficiency of resource allocation and scheduling strategies. This method focuses on the hypergraph representation and scheduling optimization of resources and tasks in complex scenarios, constructing a hypergraph scheduling model integrating "intra-layer modeling, cross-layer association, feature extraction, and decision output," and proposing hypergraph construction and feature extraction algorithms. It covers intra-layer resource and cross-layer task modeling, adaptive attention fusion mechanisms, and spatial domain convolutional feature extraction strategies, and achieves accurate generation of resource allocation and scheduling schemes through a decision model. It is applicable to typical scenarios such as resource optimization throughout the intelligent manufacturing process, production monitoring and testing collaboration, and human-machine-material collaboration.

Description

Technical Field

[0001] This invention relates to the fields of intelligent manufacturing and artificial intelligence, specifically to a hierarchical multi-cluster resource and multi-scale task combination scheduling method based on hypergraph theory in intelligent manufacturing. Background Technology

[0002] In the "production-monitoring-testing" environment of the entire intelligent manufacturing process, the fragmentation of layered, multi-cluster resources and the scheduling silos of multi-scale tasks have long constrained the development of intelligent manufacturing. Intelligent manufacturing scenarios encompass multi-level clusters such as workshop level, production line level, workstation level, and equipment level. The resource status and task requirements at each level exist in the form of multimodal data, and their heterogeneity and unstructured characteristics make it difficult for traditional methods to achieve cross-domain knowledge association and reuse. On the one hand, resources and tasks in a single domain, lacking a cross-domain perspective, are difficult to adapt to the complex needs of new scenarios or new equipment; on the other hand, traditional knowledge fusion technologies, such as rule-based knowledge graphs and single-domain graph neural networks, mostly rely on manual annotation or simple similarity calculations, which cannot effectively capture high-order associations between cross-level features, resulting in low efficiency of experience reuse and long cold start cycles.

[0003] Against this backdrop, hypergraph technology, due to its ability to model high-order relationships and connect multiple nodes through hyperedges, has become a key tool for solving hierarchical multi-cluster resource scheduling modeling. However, existing hypergraph methods still face challenges in intelligent manufacturing scenarios: the structured representation of intra-layer resources and tasks needs to take into account both local and global features of multimodal data; the association between cross-layer resources and tasks needs to overcome hierarchical barriers to achieve semantic alignment; and feature fusion needs to dynamically balance the exclusivity of intra-layer features with the complementarity of cross-layer features.

[0004] To address the aforementioned issues, the "Hierarchical Multi-Cluster Resource and Multi-Scale Task Combination Scheduling Method Based on Hypergraph Theory in Intelligent Manufacturing" proposes a complete solution: It achieves unified structured representation of cross-level features through hypergraph modeling, extracts deep local and global features by combining spatial domain hypergraph convolution, and introduces an adaptive attention mechanism to dynamically fuse intra-domain and cross-level features. Finally, a scheduling decision model outputs reusable resource-task pairs. This method systematically solves the challenges of cross-domain association and feature convergence of resources and tasks in industrial environments, providing key technical support for balanced resource allocation, multi-scale task collaborative scheduling, and collaborative optimization of the entire production-monitoring-testing process in intelligent manufacturing. Summary of the Invention

[0005] Technical problem: At the hierarchical resource scheduling modeling level, resources and tasks in the entire intelligent manufacturing process are dispersed in multi-level clusters such as workshop level, production line level, and workstation level in a multimodal and heterogeneous manner. Traditional methods can only capture low-order relationships and it is difficult to explicitly model high-order implicit relationships between cross-level resources and tasks, such as the cross-level collaborative scheduling relationship between workstation-level equipment status and production line-level process parameter adjustment. It is urgent to realize the explicit expression of high-order relationships through hypergraph structure.

[0006] At the level of unstructured data representation, resource-task data exhibits heterogeneity (modal differences) and unstructuredness (lack of explicit graph structure). Traditional embedding methods (such as word embedding and simple vector projection) are difficult to preserve empirical semantic associations and local-global features. It is necessary to design multimodal embedding strategies to map discrete / unstructured data to a unified low-dimensional dense space, laying the foundation for the construction of hypergraph vertices.

[0007] At the level of intra-layer feature extraction, the local correlation of resources and tasks within a single-level cluster (such as common features of resources of similar devices) and global dependency (such as the whole process mode of production) need to be extracted by hypergraph convolution. However, traditional graph convolution only supports vertex-vertex information transmission and cannot handle the bidirectional information flow of hypergraph "vertex-hyperedge-vertex". A two-stage message passing mechanism (vertex → hyperedge → vertex) needs to be designed to realize end-to-end feature aggregation and optimization of hypergraph structure.

[0008] At the level of cross-level fusion representation, the association between cross-level resources and tasks needs to break through the hierarchical barriers to achieve semantic alignment. Traditional cross-domain fusion methods (such as simple weighting based on similarity) rely on manually defined features, which are difficult to dynamically adapt to changes in industrial scenarios (such as new equipment and new processes). It is necessary to generate a joint representation that retains the core features of the current domain and integrates the related information of other domains through cross-level hyperedge construction (similarity calculation based on cross-level common features) and adaptive attention mechanism (dynamically allocating intra-layer / cross-level feature weights).

[0009] Technical Solution: This invention proposes a cross-level resource-task hypergraph representation and aggregation method for complex industrial environments. By constructing a hypergraph structure, it achieves unified representation and deep fusion of cross-level features, solving the core problems of resource-task fragmentation, difficulty in cross-domain association, and low reuse efficiency in industrial scenarios. Based on hypergraph modeling, this method combines spatial domain hypergraph convolution and adaptive attention mechanisms to realize a complete technical link from multi-source heterogeneous empirical data to accurately reusable resources-tasks. It mainly includes the following core steps and modules:

[0010] First, for cross-level resource-task data in industrial scenarios, an embedding model is used to map unstructured or weakly structured data to a low-dimensional dense semantic space, generating an initial embedding vector for each resource-task sample, thus completing a preliminary structured representation from discrete experience to continuous features. This embedding process takes into account the characteristics of multimodal data such as text, numerical, and time series data, employs a domain-adaptive encoder to extract local features, and aligns them to a unified space through cross-modal fusion technology, laying the foundation for subsequent hypergraph construction.

[0011] Secondly, a hypergraph structure is constructed based on the initial embedding vectors: using resource-task samples as the vertex set, highly correlated sample pairs are selected through attribute similarity or structural similarity, hyperedges are constructed, and a hypergraph adjacency matrix is ​​generated to achieve a localized structural representation of resources and tasks within the layer. This hypergraph structure can not only capture explicit resource-task associations, but also reveal potential collaborative relationships between experiences through higher-order hyperedges.

[0012] To address cross-level feature fusion, this invention designs a cross-level hypergraph construction and association mechanism: Based on common cross-domain features, the similarity of resources and tasks between different domains is calculated. Highly correlated sample pairs across domains are selected to construct cross-domain hyperedges, forming a cross-level hypergraph. Normalized hypergraph convolution operations combine features within the current domain with the cross-level hypergraph structure to generate cross-domain fusion experience, breaking through domain barriers to achieve semantic alignment. This mechanism, through cross-domain connections of hyperedges, enables resources and tasks from different domains to be interconnected and complementary in a shared semantic space.

[0013] At the feature fusion and enhancement level, an adaptive attention mechanism is introduced: Deep features within the domain and cross-level hypergraph features are taken as input. Attention weights are generated by calculating their similarity. After SoftMax normalization, the intra-domain and cross-domain features are weighted and summed to generate a joint representation that retains core features of the current domain while incorporating information from other domains. This mechanism can dynamically adjust the weights of intra-domain and cross-level features, ensuring that the fused empirical representation possesses both domain-specific expertise and the ability to absorb complementary knowledge from other domains.

[0014] Finally, the joint feature input scheduling decision model calculates the probability distribution of the correlation between features and candidate resource-tasks, and selects the top k high-probability resource-tasks as outputs. This directly guides the adjustment of equipment operating parameters, fault diagnosis, or optimization of operation strategies in industrial scenarios, achieving precise reuse of resources-tasks and rapid cold start in new scenarios. This method systematically solves the challenges of cross-domain correlation and feature convergence of resources-tasks in industrial environments through the high-order correlation modeling capabilities of hypergraphs, deep feature extraction from spatial domain convolution, and dynamic fusion of attention mechanisms. It provides end-to-end technical support for industrial intelligence, from experience representation to knowledge output.

[0015] The specific plan is as follows:

[0016] A hierarchical multi-cluster resource and multi-scale task combination scheduling method based on hypergraph theory in intelligent manufacturing comprises the following steps: First, the resource units and multi-scale task samples of the hierarchical multi-cluster are structurally organized as vertex sets of a hypergraph, and an initial embedding vector is generated for each vertex through an embedding model. Then, for a single-level cluster, based on aligned common features, a hyperedge construction method based on attribute similarity or structural similarity is used to screen highly correlated pairs between intra-layer resources and task samples, construct intra-layer hyperedges, and fuse them to form an intra-layer resource-task hypergraph, achieving local structural representation of intra-layer resources and tasks. Subsequently, spatial domain hypergraph convolution operations are applied to the constructed intra-layer resource-task hypergraph to extract deep local and global feature representations of intra-layer resources and tasks. Finally, to achieve cross-level association, cross-level common features are used... This paper calculates the cross-level similarity of resource and task samples among different level clusters, selects highly correlated cross-level pairs, constructs cross-level hyperedges and merges them to form a cross-level resource-task hypergraph, completes the structured representation of cross-level associations, and obtains the cross-level resource-task association features through hypergraph convolution. Next, an adaptive attention fusion mechanism is used to generate intra-level-cross-level joint feature representations of local features of intra-level resources and tasks and cross-level resource-task association features, realizing the complementary enhancement of cross-level resources and tasks. Finally, the intra-level-cross-level joint feature representations are input into the scheduling decision model, and the model inference outputs the optimal resource allocation and task scheduling scheme under a specific manufacturing scenario, completing the mapping from resource-task representation to scheduling decision, and systematically solving the problem of combined scheduling of hierarchical multi-cluster resources and multi-scale tasks in intelligent manufacturing environment.

[0017] Furthermore, in the structured organization of the hypergraph, the set of intelligent manufacturing resources and tasks is defined as follows: This collection is distributed across multiple levels of clusters, including workshop-level clusters. Production line-level clusters Workstation-level clusters Three typical levels, with resources and tasks at each level represented as follows: , , Each level of resources and tasks is described by multimodal data features of the corresponding manufacturing scenario; a hypergraph is used. A structured representation of resources and tasks is provided, in which the vertex set... For the sample set of resources and tasks, the edge set This is used to describe the relationship between resources and tasks, and its construction is based on the similarity characteristics of resource and task data; at the same time, the adjacency matrix of the hypergraph is defined. Represents the relationship between vertices and hyperedges in a hypergraph. Represents the current vertex Belongs to superedge This paper proposes a formal representation of the relationship between vertices and hyperedges through binary numerical encoding. Simultaneously, it employs an embedding layer or feature encoder to perform preliminary vector representation processing on the heterogeneous raw data. This process maps discrete or unstructured resource-task data into a low-dimensional dense semantic space, generating an initial embedding vector for each vertex. .

[0018] Furthermore, the method for constructing the intra-layer resource-task hypergraph employs a hypergraph. The resources and tasks within a single-level cluster are represented in a structured manner, where the vertex set... Defined as the sample set of each resource-task within this domain, edge set Edge sets are used to describe the relationships between different resource units and task samples within the same level of cluster. The construction of hyperedges needs to be based on the similarity features of resource-task data, and is achieved by mining the potential relationships between data. Specifically, hyperedges are constructed in two ways: one is based on common attribute feature extraction, that is, extracting key attribute features from the multimodal data of resource-task samples. If two samples are highly consistent in core attributes, a hyperedge connection is established; the other is based on similarity measurement or vector distance measurement. By calculating the cosine similarity and Euclidean distance index between sample feature vectors, a threshold is set to filter highly similar sample pairs, and then a hyperedge is generated.

[0019] Furthermore, the core purpose of the spatial domain hypergraph convolution operation of the intra-layer resource-task hypergraph is to directly extract features of the intra-layer resource-task hypergraph through hypergraph structure manipulation. This process does not rely on spectral decomposition methods, but is implemented through a two-stage message passing mechanism: the input is the constructed intra-layer resource-task hypergraph and initial vertex features; firstly, vertex-to-hyperedge feature aggregation is performed, passing the feature information of each vertex to its corresponding hyperedge, generating hyperedge features through the aggregation operation of vertex features within the hyperedge, thereby capturing the local association information and common features of vertices within the hyperedge; subsequently, hyperedge-to-vertex feature propagation is performed, passing the hyperedge features back to each associated vertex, updating vertex features through the reception and integration of hyperedge features by the vertex, so that the vertex retains its original information while incorporating the association information of its corresponding hyperedge, realizing the deep optimization of feature extraction and vertex representation of the intra-layer resource-task hypergraph.

[0020] Furthermore, defined in a matrix with an association matrix Hypergraph Above, interactive neighborhood relationships Defined as: Define hyperedges based on interactive neighborhood relationships Vertex interaction neighborhood set and vertex superedge Hyper-edge interactive neighborhood set Define a hyperedge Vertex interaction neighborhood set: Define vertices Hyperedge interaction neighborhood set: Following the above definition, a message-passing mechanism for hypergraph convolution in the hypergraph space domain is introduced; given a hypergraph... one of the vertices ,in It is the weight of the hyperedge, and the goal is to aggregate the set of interaction neighborhoods from its hyperedge. The message; in order to obtain each hyperedge in this set Hyperedge messages need to be aggregated from the set of interaction neighborhoods of its vertices. The message; then, the two stages of hypergraph convolution constitute a sequence from the vertex feature set. arrive The closed-loop message passing process, the first A convolution of a layer with the same spatial hypergraph is defined as:

[0021]

[0022] in It is the first vertex in layer The input feature vector, It is the updated vertex Features; It is a super-edge The news, Is with super-edge The weight of the association; Represents vertices The message; It is a super-edge The hyperedge feature is the first An element of the hyperedge feature set in the layer; These are the vertex message function, hyperedge update message function, hyperedge message function, and vertex update function in layer t, respectively.

[0023] Furthermore, the method for constructing the cross-level resource-task hypergraph employs a hypergraph. Structured representation of cross-level resources and tasks, including vertex sets. Edge set is defined as the set of resource units and task samples within different levels of clusters. Used to describe the relationship between resources and task samples between different levels of clusters; edge set The construction of the system needs to be based on the similarity characteristics of resource-task data in each domain, and is achieved by mining the potential relationships between data. The specific steps are as follows: First, core attribute features are extracted from the multimodal data of cross-level resource-task samples through attribute feature extraction; then, cross-domain sample pairs with the same or highly similar core attributes are identified and merged into a hyperedge, thereby establishing a structured relationship between cross-domain samples in the hypergraph, and finally achieving effective fusion and representation of cross-level resources-tasks.

[0024] Furthermore, the hypergraph convolution obtains cross-level fusion features, and the cross-level fusion features are calculated through normalized hypergraph convolution operations. Based on the cross-domain expert hypergraph, a corresponding adjacency matrix is ​​constructed, and it is set in the workshop-level cluster. and production line-level clusters Construct a cross-level hypergraph between them, and define the cross-level hypergraph adjacency matrix as follows: The degree of the vertex is The degree of the hyperedge is Production line-level clusters The domain characteristics are ,pass: Obtain the corresponding workshop-level cluster and production line-level clusters Cross-domain integration experience, among which This represents a nonlinear activation function; this operation normalizes the product of the adjacency matrix and the degree matrix, thus transforming the intra-layer features of the production line cluster B into... With cross-level hypergraph structure By combining these elements, cross-domain association and feature fusion of resources and tasks between workshop-level cluster A and production line-level cluster B can be achieved, thereby obtaining the fused cross-level feature representation.

[0025] Furthermore, the adaptive attention fusion mechanism is implemented as follows: First, the local features of intra-layer resource-task and the associated features of cross-layer resource-task are used as inputs to the attention mechanism. Then, the attention module calculates the similarity between intra-domain and cross-domain features to generate attention weights that reflect their correlation. After being normalized by SoftMax, these weights are weighted and summed for intra-domain and cross-domain features respectively, so that intra-domain features absorb the associated information of other domains, and cross-domain features enhance the adaptability of the current domain. Finally, the weighted intra-domain features and cross-domain features are fused through residual connection or splicing operations to generate a weighted fusion feature representation that retains the core scheduling features of the current layer and integrates the associated information of other layers.

[0026] Furthermore, the scheduling decision model takes the joint feature vector obtained by fusion of hypergraph convolution and attention as input, and outputs the fit distribution of each candidate resource configuration and task scheduling scheme through a multi-objective optimization decision head. Specifically, the model calculates the probability value based on the correlation between the feature vector and the candidate resource-task, forming a probability distribution reflecting the applicability of the resource-task. Subsequently, the top k high-fit scheduling schemes are selected as output based on the descending order of the probability distribution. Finally, these Top-k resources-tasks provide direct decision-making basis for the balanced allocation of multi-cluster resources, multi-scale task collaborative scheduling and full-process optimization in intelligent manufacturing scenarios by parsing the equipment operation specifications, fault handling strategies or process parameter suggestions contained therein, thereby achieving accurate matching of resources and tasks and scenario adaptation of scheduling strategies.

[0027] Beneficial effects:

[0028] (1) Breaking the scheduling silo and improving the efficiency of cross-level resource allocation: In traditional industrial scenarios, resources and tasks are stored in a scattered manner due to differences in domain, equipment, and process, forming "data silos". Cross-level feature reuse requires manual sorting or repeated labeling, which is inefficient. This invention explicitly models the high-order association relationship of cross-level features through a hypergraph structure, mapping the scattered multi-level resources and tasks to a unified hypergraph space, so that cross-level resources and tasks can be directly associated through hyperedges, realizing the global balanced allocation of hierarchical multi-cluster resources.

[0029] (2) Dynamic fusion enhances global coordination and supports multi-scale task scheduling decisions: Industrial scenarios often face multiple objective conflicts (such as efficiency and energy consumption, accuracy and cost), which are difficult to balance dynamically using traditional methods. This invention uses an adaptive attention mechanism to dynamically allocate the weights of intra-layer resource features (local accuracy) and cross-layer fusion features (global coordination) according to scheduling requirements, generating a joint representation that retains the core capabilities of the domain while integrating knowledge from other domains.

[0030] (3) Empowering full-process collaborative scheduling and promoting the upgrading of intelligent manufacturing: The joint experience representation generated by this method can be directly input into the scheduling decision model and output multi-label probability distribution, providing interpretable and traceable scheduling decision basis for multi-cluster resource balance allocation, multi-scale task collaborative scheduling, and full-process optimization of production-monitoring-testing, supporting the full-process collaborative optimization of intelligent manufacturing for ASEAN cooperation. Attached Figure Description

[0031] Figure 1 It is a spatial domain hypergraph convolution flowchart that demonstrates a hierarchical multi-cluster resource and multi-scale task combination scheduling method based on hypergraph theory in intelligent manufacturing.

[0032] Figure 2 This is a schematic diagram of the main principle of the method of the present invention. Detailed Implementation

[0033] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0034] like Figure 2 As shown, this invention provides a hypergraph representation and aggregation method based on cross-level resource-task in an industrial environment.

[0035] (1) Constructing a hierarchical multi-cluster resource-task hypergraph in the context of intelligent manufacturing: In the process of industrial intelligence, the integration and reuse of cross-domain resources and tasks is the key to breaking through the limitations of single-domain knowledge and improving decision-making capabilities in complex scenarios. However, resources and tasks in industrial scenarios often exhibit significant characteristics of multi-domain distribution, unstructured nature, and fragmentation. To systematically solve this problem, this invention first proposes a structured representation method based on hypergraphs: defining the resource-task set as This set is distributed across multiple levels of clusters; in this embodiment, it is a workshop-level cluster. Production line-level clusters Workstation-level clusters Taking three typical industrial sectors as examples, the resource-task structure at each level is represented as follows: , , Each level of resources and tasks is described by multimodal data features of the corresponding manufacturing scenario. Hypergraphs are further employed. A structured representation of resources and tasks is provided, in which the vertex set... For the sample set of resources and tasks, the edge set This is used to describe the relationship between resources and tasks, and its construction is based on the similarity characteristics of resource-task data. It also defines the adjacency matrix of the hypergraph. Represents the relationship between vertices and hyperedges in a hypergraph. Represents the current vertex Belongs to superedge A formal representation of the relationship between vertices and hyperedges is achieved through binary numerical encoding. Simultaneously, an embedding layer or feature encoder performs preliminary vector representation processing on the heterogeneous raw data. This process maps discrete or unstructured resource-task data into a low-dimensional dense semantic space, generating an initial embedding vector for each vertex. .

[0036] (2) Intra-layer hypergraph modeling and representation stage: In order to represent the relationship between resources and tasks within the layer and obtain the main feature representation of resources and tasks within the layer, a hypergraph is first used. A structured representation of resources and tasks within a single industrial sector, where the vertex set... Defined as the sample set of each resource-task within this domain, edge set This is used to describe the relationships between different resource-task samples within the same domain. Since resources and tasks are inherently unstructured or weakly structured data, edge sets... The construction of hyperedges needs to be based on the similarity features of resource-task data, and is achieved by mining the potential relationships between the data. Specifically, hyperedges can be constructed in two ways: one is based on common attribute feature extraction, that is, extracting key attribute features from the multimodal data of resource-task samples. If two samples are highly consistent in core attributes, a hyperedge connection is established; the other is based on similarity metrics or vector distance metrics. By calculating indicators such as cosine similarity and Euclidean distance between sample feature vectors, a threshold is set to filter highly similar sample pairs, thereby generating hyperedges. The above methods formally model the relationships between resource-task samples, aiming to accurately mine the common features and intrinsic connections of resources and tasks within the same level, providing a structured input foundation for subsequent feature aggregation and deep representation of intra-layer hypergraphs.

[0037] like Figure 1 As shown, the core purpose of the intra-layer hypergraph spatial domain convolution operation is to directly extract features of the intra-layer resource-task hypergraph through hypergraph structure manipulation. This process does not rely on spectral decomposition methods, but is implemented through a two-stage message passing mechanism: the input is the constructed intra-layer resource-task hypergraph and initial vertex features; firstly, vertex-to-hyperedge feature aggregation is performed, passing the feature information of each vertex to its corresponding hyperedge, generating hyperedge features through the aggregation operation of vertex features within the hyperedge, thereby capturing the local association information and common features of vertices within the hyperedge; then, hyperedge-to-vertex feature propagation is performed, passing the hyperedge features back to each associated vertex, updating vertex features through the reception and integration of hyperedge features by the vertex, so that the vertex retains its original information while incorporating the association information of its corresponding hyperedge, ultimately achieving deep optimization of feature extraction and vertex representation of the intra-layer resource-task hypergraph.

[0038] Defined in a matrix of association Hypergraph Above, interactive neighborhood relationships Defined as: Define hyperedges based on interactive neighborhood relationships Vertex interaction neighborhood set and vertex superedge Hyper-edge interactive neighborhood set Define a hyperedge Vertex interaction neighborhood set: Define vertices Hyperedge interaction neighborhood set: Following the above definition, a message-passing mechanism for hypergraph convolution in the hypergraph space domain is introduced. Given a hypergraph... one of the vertices ,in It is the weight of the hyperedge, and the goal is to aggregate the set of interaction neighborhoods from its hyperedge. The message. In order to obtain each hyperedge in this set. Hyperedge messages need to be aggregated from the set of interaction neighborhoods of its vertices. The message. Then, the two stages of hypergraph convolution constitute a feature set from the vertices. arrive The closed-loop message passing process, the first A convolution of a layer with the same spatial hypergraph can be defined as:

[0039]

[0040] in It is the first vertex in layer The input feature vector, It is the updated vertex Its characteristics. It is a super-edge The news, Is with super-edge The weight of the association. Represents vertices The news. It is a super-edge The hyperedge feature is the first An element of the hyperedge feature set in the layer. These are the vertex message function, hyperedge update message function, hyperedge message function, and vertex update function in layer t, respectively. Through the above steps, deep optimization of feature extraction and vertex representation of the resource-task hypergraph within the layer can be achieved.

[0041] (3) Cross-level hypergraph modeling and representation stage: In order to integrate the representation of cross-level resources and tasks, a hypergraph is first used. Structured representation of cross-level resources and tasks, where vertex sets Defined as a set of resource-task samples within different industrial sectors, edge set This is used to describe the relationships between resource-task samples from different domains. Since resources and tasks are inherently unstructured or weakly structured data, edge sets... The construction of this system needs to be based on the similarity characteristics of resource-task data from different domains, and is achieved by mining the potential relationships between the data. The specific steps are as follows: First, core attribute features are extracted from the multimodal data of cross-level resource-task samples through attribute feature extraction; then, cross-domain sample pairs with the same or highly similar core attributes are identified and merged into a hyperedge, thereby establishing a structured relationship between cross-domain samples in the hypergraph, and finally achieving effective fusion and representation of cross-level resources-tasks.

[0042] The cross-level resource-task fusion process computes cross-domain fusion experience through normalized hypergraph convolution operations. Based on the cross-domain expert hypergraph, the corresponding adjacency matrix is ​​constructed. Here, we assume a workshop-level cluster. and production line-level clusters Construct a cross-level hypergraph between them, and define the cross-level hypergraph adjacency matrix as follows: The degree of the vertex is The degree of the hyperedge is ,domain The domain characteristics are ,pass: Obtain the corresponding workshop-level cluster and production line-level clusters Cross-domain integration experience, among which This represents a nonlinear activation function. This operation normalizes the product of the adjacency matrix and the degree matrix, thus representing the intra-domain features of the production line cluster B. With cross-level hypergraph structure By combining these elements, cross-domain association and feature fusion of resources and tasks between workshop-level cluster A and production line-level cluster B can be achieved, thereby obtaining the fused cross-level feature representation.

[0043] (4) Intra-layer and cross-layer feature fusion stage: In complex decision-making in industrial scenarios, cross-domain fusion of resources and tasks is a key link to improve the generalization and robustness of experience. To achieve this goal, this invention proposes a resource-task fusion mechanism with dynamic weight allocation: taking the feature representations of intra-layer resources and tasks and cross-layer resources and tasks as inputs, feature fusion is achieved by dynamically allocating the weights of the two. The specific process is as follows: First, the local features of intra-layer resources-tasks and the associated features of cross-layer resources-tasks are used as inputs to the attention mechanism. Then, the attention module generates attention weights that reflect the correlation between intra-domain and cross-domain features by calculating the similarity between the two. After SoftMax normalization, these weights are weighted and summed for intra-domain and cross-domain features respectively, so that intra-domain features absorb the associated information of other domains, and cross-domain features strengthen the adaptability of the current domain. Finally, the weighted intra-domain features and cross-domain features are fused through residual connection or splicing operations to generate a weighted fusion experience representation that retains the core features of the current domain resources-tasks and integrates the associated experience of other domains, thereby improving the generalization and robustness of the current domain resources-tasks.

[0044] (5) Scheduling Decision Stage: The fusion classification stage is a key link in the process of resource-task transition from feature representation to specific decision implementation. This stage takes the resource-task feature vector obtained by fusion of hypergraph convolution and attention as input, and outputs the probability distribution of each candidate resource-task through a multi-label classification head. The specific process is as follows: The scheduling decision model takes the resource-task feature vector obtained by fusion of hypergraph convolution and attention as input, and outputs the probability distribution of each candidate resource-task through a multi-label classification head. Specifically, the model calculates the probability value based on the correlation between the feature vector and the candidate resource-task, forming a probability distribution that reflects the applicability of the resource-task. Subsequently, the top k high-probability resource-tasks are selected as outputs based on the descending order of the probability distribution. Finally, these Top-k degree schemes provide direct decision-making basis for the balanced allocation of multi-cluster resources, multi-scale task collaborative scheduling, and full-process production optimization in intelligent manufacturing scenarios by analyzing the information such as resource configuration schemes, task allocation strategies, and process parameter optimization suggestions contained therein. This enables accurate matching of resources and tasks and dynamic adaptation of scheduling strategies, and can be packaged into intelligent components and integrated into the intelligent manufacturing full-process collaborative optimization platform for ASEAN cooperation.

[0045] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered within the scope of protection of this invention.

Claims

1. A hierarchical multi-cluster resource and multi-scale task combination scheduling method based on hypergraph theory in intelligent manufacturing, characterized by the following steps: The following steps are taken: First, the hierarchical multi-cluster resource units and multi-scale task samples are structured and organized as the vertex set of the hypergraph. At the same time, the initial embedding vector corresponding to each vertex is generated through the embedding model. Subsequently, for single-level clusters, based on aligned common features, a hyperedge construction method based on attribute similarity or structural similarity is adopted to screen highly correlated pairs between intra-layer resources and task samples, construct intra-layer hyperedges, and fuse them to form an intra-layer resource-task hypergraph, realizing the local structured representation of intra-layer resources and tasks. Then, spatial domain hypergraph convolution operation is applied to the constructed intra-layer resource-task hypergraph to extract deep local and global feature representations of intra-layer resources and tasks. Furthermore, to achieve cross-level association, cross-level similarity of resource and task samples between different level clusters is calculated based on cross-level common features, highly correlated pairs across levels are screened, cross-level hyperedges are constructed, and fused to form a cross-level resource-task hypergraph. A source-task hypergraph is used to complete the structured representation of cross-level associations and obtain cross-level resource-task association features through hypergraph convolution. Next, an adaptive attention fusion mechanism is used to generate intra-level-cross-level joint feature representations of local features of intra-level resources-tasks and cross-level resource-task association features, achieving complementary enhancement of cross-level resources and tasks. Finally, the intra-level-cross-level joint feature representations are input into the scheduling decision model, and the model inference outputs the optimal resource allocation and task scheduling scheme under specific manufacturing scenarios, completing the mapping from resource-task representations to scheduling decisions, and systematically solving the problem of combined scheduling of hierarchical multi-cluster resources and multi-scale tasks in intelligent manufacturing environment.

2. The hierarchical multi-cluster resource and multi-scale task combination scheduling method based on hypergraph theory in intelligent manufacturing according to claim 1, characterized in that: In the structured organization of the hypergraph, the set of intelligent manufacturing resources and tasks is defined as follows: This collection is distributed across multiple levels of clusters, including workshop-level clusters. Production line-level clusters Workstation-level clusters Three typical levels, with resources and tasks at each level represented as follows: , , Each level of resources and tasks is described by multimodal data features of the corresponding manufacturing scenario; a hypergraph is used. A structured representation of resources and tasks is provided, in which the vertex set... For the sample set of resources and tasks, the edge set This is used to describe the relationship between resources and tasks, and its construction is based on the similarity characteristics of resource and task data; at the same time, the adjacency matrix of the hypergraph is defined. Represents the relationship between vertices and hyperedges in a hypergraph. Represents the current vertex Belongs to superedge This paper proposes a formal representation of the relationship between vertices and hyperedges through binary numerical encoding. Simultaneously, it employs an embedding layer or feature encoder to perform preliminary vector representation processing on the heterogeneous raw data. This process maps discrete or unstructured resource-task data into a low-dimensional dense semantic space, generating an initial embedding vector for each vertex. .

3. The hierarchical multi-cluster resource and multi-scale task combination scheduling method based on hypergraph theory in intelligent manufacturing according to claim 2, characterized in that: The method for constructing the intra-layer resource-task hypergraph uses a hypergraph. The resources and tasks within a single-level cluster are represented in a structured manner, where the vertex set... Defined as the sample set of each resource-task within this domain, edge set Edge sets are used to describe the relationships between different resource units and task samples within the same level of cluster. The construction of hyperedges needs to be based on the similarity features of resource-task data, and is achieved by mining the potential relationships between data. Specifically, hyperedges are constructed in two ways: one is based on common attribute feature extraction, that is, extracting key attribute features from the multimodal data of resource-task samples. If two samples are highly consistent in core attributes, a hyperedge connection is established; the other is based on similarity measurement or vector distance measurement. By calculating the cosine similarity and Euclidean distance index between sample feature vectors, a threshold is set to filter highly similar sample pairs, and then a hyperedge is generated.

4. The hierarchical multi-cluster resource and multi-scale task combination scheduling method based on hypergraph theory in intelligent manufacturing according to claim 3, characterized in that: The core purpose of the spatial domain hypergraph convolution operation of the intra-layer resource-task hypergraph is to directly extract features of the intra-layer resource-task hypergraph through hypergraph structure. This process does not rely on spectral decomposition methods, but is implemented through a two-stage message passing mechanism: the input is the constructed intra-layer resource-task hypergraph and initial vertex features; firstly, vertex-to-hyperedge feature aggregation is performed, passing the feature information of each vertex to its corresponding hyperedge, and generating hyperedge features through the aggregation operation of vertex features within the hyperedge, thereby capturing the local association information and common features of vertices within the hyperedge; Subsequently, feature propagation from hyperedge to vertex is performed, and the hyperedge features are passed back to each associated vertex. Vertex features are updated by receiving and integrating the hyperedge features, so that vertices can integrate the associated information of their hyperedges while retaining the original information, thereby realizing the deep optimization of feature extraction and vertex representation of the resource-task hypergraph within the layer.

5. A hierarchical multi-cluster resource and multi-scale task combination scheduling method based on hypergraph theory in intelligent manufacturing, as described in claim 4, is characterized in that: Defined in a matrix of association Hypergraph Above, interactive neighborhood relationships Defined as: Define hyperedges based on interactive neighborhood relationships Vertex interaction neighborhood set and vertex superedge Hyper-edge interactive neighborhood set Define a hyperedge Vertex interaction neighborhood set: Define vertices Hyperedge interaction neighborhood set: Following the above definition, a message-passing mechanism for hypergraph convolution in the hypergraph space domain is introduced; given a hypergraph... one of the vertices ,in It is the weight of the hyperedge, and the goal is to aggregate the set of interaction neighborhoods from its hyperedge. The message; in order to obtain each hyperedge in this set Hyperedge messages need to be aggregated from the set of interaction neighborhoods of its vertices. The message; then, the two stages of hypergraph convolution constitute a sequence from the vertex feature set. arrive The closed-loop message passing process, the first A convolution of a layer with the same spatial hypergraph is defined as: in It is the first vertex in layer The input feature vector, It is the updated vertex Features; It is a super-edge The news, Is with super-edge The weight of the association; Represents vertices The message; It is a super-edge The hyperedge feature is the first An element of the hyperedge feature set in the layer; These are the vertex message function, hyperedge update message function, hyperedge message function, and vertex update function in layer t, respectively.

6. A hierarchical multi-cluster resource and multi-scale task combination scheduling method based on hypergraph theory in intelligent manufacturing, as described in claim 5, is characterized in that: The method for constructing the cross-level resource-task hypergraph uses a hypergraph. Structured representation of cross-level resources and tasks, including vertex sets. Edge set is defined as the set of resource units and task samples within different levels of clusters. Used to describe the relationship between resources and task samples between different levels of clusters; edge set The construction of the system needs to be based on the similarity characteristics of resource-task data in each domain, and is achieved by mining the potential relationships between data. The specific steps are as follows: First, core attribute features are extracted from the multimodal data of cross-level resource-task samples through attribute feature extraction; then, cross-domain sample pairs with the same or highly similar core attributes are identified and merged into a hyperedge, thereby establishing a structured relationship between cross-domain samples in the hypergraph, and finally achieving effective fusion and representation of cross-level resources-tasks.

7. A hierarchical multi-cluster resource and multi-scale task combination scheduling method based on hypergraph theory in intelligent manufacturing, as described in claim 6, is characterized in that: The hypergraph convolution obtains cross-level fusion features, and the cross-level fusion features are calculated through normalized hypergraph convolution operations. Based on the cross-domain expert hypergraph, a corresponding adjacency matrix is ​​constructed, and it is set in the workshop-level cluster. and production line-level clusters Construct a cross-level hypergraph between them, and define the cross-level hypergraph adjacency matrix as follows: The degree of the vertex is The degree of the hyperedge is Production line-level clusters The domain characteristics are ,pass: Obtain the corresponding workshop-level cluster and production line-level clusters Cross-domain integration experience, among which This represents a nonlinear activation function; this operation normalizes the product of the adjacency matrix and the degree matrix, thus transforming the intra-layer features of the production line cluster B into... With cross-level hypergraph structure By combining these elements, cross-domain association and feature fusion of resources and tasks between workshop-level cluster A and production line-level cluster B can be achieved, thereby obtaining the fused cross-level feature representation.

8. A hierarchical multi-cluster resource and multi-scale task combination scheduling method based on hypergraph theory in intelligent manufacturing, as described in claim 7, is characterized in that: The adaptive attention fusion mechanism operates as follows: First, the local features of intra-layer resources-tasks and the associated features of cross-layer resources-tasks are used as inputs to the attention mechanism. Then, the attention module calculates the similarity between intra-domain and cross-domain features to generate attention weights that reflect their correlation. After SoftMax normalization, these weights are weighted and summed for intra-domain and cross-domain features respectively, so that intra-domain features absorb the associated information from other domains, and cross-domain features enhance the adaptability of the current domain. Finally, the weighted intra-domain features and cross-domain features are fused through residual connection or concatenation operations to generate a weighted fusion feature representation that retains the core scheduling features of the current layer and incorporates the associated information from other layers.

9. A hierarchical multi-cluster resource and multi-scale task combination scheduling method based on hypergraph theory in intelligent manufacturing, as described in claim 8, is characterized in that: The scheduling decision model takes the joint feature vector obtained by fusion of hypergraph convolution and attention as input, and outputs the fit distribution of each candidate resource configuration and task scheduling scheme through a multi-objective optimization decision head. Specifically, the model calculates the probability value based on the correlation between the feature vector and the candidate resource-task, forming a probability distribution reflecting the applicability of the resource-task. Then, the top k high-fit scheduling schemes are selected as output based on the descending order of the probability distribution. Finally, these Top-k resources-tasks provide direct decision-making basis for the balanced allocation of multi-cluster resources, multi-scale task collaborative scheduling and full-process optimization in intelligent manufacturing scenarios by parsing the equipment operation specifications, fault handling strategies or process parameter suggestions contained therein, thus realizing accurate matching of resources and tasks and scenario adaptation of scheduling strategies.

Images