Self-adaptive visitor reservation management system and method based on multi-agent collaborative decision-making
By employing a multi-agent collaborative decision-making method, utilizing multi-agent heterogeneous graphs and an improved heterogeneous Transformer model, the problem of insufficient decision-making in existing visitor appointment management systems under dynamic environments is solved, achieving efficient and flexible visitor appointment management and improving the system's response speed and consistency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZERUI TECH GRP CO LTD
- Filing Date
- 2026-03-24
- Publication Date
- 2026-05-12
AI Technical Summary
The existing visitor appointment management system lacks the ability to dynamically model visitor behavior characteristics, visitor status, resource usage and security risks. It cannot make real-time adjustments during peak access times, resource conflicts, multi-role collaboration and temporary risk changes. Furthermore, the separation of process strategy generation and permission strategy generation leads to scheduling conflicts and permission delays.
A multi-agent collaborative decision-making method is adopted. Through multi-agent heterogeneous graphs and improved heterogeneous Transformer models, multiple relationships between visitors, visitors, meeting room resources, security units, access control paths and approval entities are constructed to achieve unified representation and reasoning. A dual-path decoding structure for process policy generation and permission policy generation is also constructed to achieve policy linkage and gating fusion.
It enables flexible decision-making in the face of dynamic access demands and resource conflicts, improves process consistency and decision automation, reduces scheduling conflicts and permission delays, and enhances the response speed and consistency of the access process.
Smart Images
Figure CN122024367A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent visitor management technology, and in particular to an adaptive visitor appointment management system and method based on multi-agent collaborative decision-making. Background Technology
[0002] Existing visitor appointment management systems typically rely on a rule-driven model, using pre-configured processes to handle visitor registration, approval workflows, meeting room occupancy registration, and access control. These systems, based on static processes and fixed logic, lack the ability to dynamically model visitor behavior characteristics, visitor status, resource usage, and security risks. Consequently, they struggle to adapt in real-time to peak visitor times, resource conflicts, multi-role collaboration, and unforeseen changes in risk.
[0003] Some solutions introduce basic graph structures or simple strategy models, but fail to establish complex dependencies between multiple types of entities that can simultaneously express visitor nodes, visited person nodes, meeting room resource nodes, security nodes, and access control path nodes. They also cannot combine access time, event triggering information, and approval links for unified reasoning.
[0004] Furthermore, existing systems typically separate process policy generation from permission policy generation, resulting in a lack of linkage between time coordination, resource scheduling, path planning, and access control. This can easily lead to scheduling conflicts, duplicate approvals, or permission delays. For scenarios involving cross-regional access, resource constraints, or security requirements, traditional methods cannot generate executable and implementable comprehensive decision-making strategies in real time based on the states of multiple parties.
[0005] Therefore, how to provide an adaptive visitor appointment management system and method based on multi-agent collaborative decision-making is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0006] One objective of this invention is to propose an adaptive visitor appointment management system and method based on multi-agent collaborative decision-making. This invention introduces a multi-agent heterogeneous graph modeling and improved heterogeneous Transformer model encoding mechanism to uniformly represent and reason about various relationships between visitors, visitors, meeting room resources, security units, access control paths, and approval entities. It also constructs a dual-path decoding structure for process policy generation and permission policy generation, ultimately achieving collaborative decision-making for access process scheduling, resource allocation, and permission control through policy-linked gating fusion. This invention can handle dynamic access demands, resource conflicts, and changes in security risks, and possesses advantages such as high process flexibility, good policy consistency, and a high degree of decision-making automation.
[0007] An adaptive visitor appointment management method based on multi-agent collaborative decision-making according to an embodiment of the present invention includes the following steps: Acquire visitor management-related data, perform structured processing on the visitor management-related data, and construct a multi-agent heterogeneous graph to form an initial heterogeneous graph representation set; Analyze visitor reservation requests, generate reservation request vectors, and associate the reservation request vectors with the initial heterogeneous graph representation set to obtain an extended heterogeneous graph structure; The extended heterogeneous graph structure is input into the input layer of the improved heterogeneous Transformer model, and correspondence-by-correspondence modulation processing is performed to obtain the event-modulated attention result set. Based on the event-modulated attention result set and the extended heterogeneous graph structure, message transformation and aggregation update are performed, and a global agent state representation set is obtained through stacked processing of the coding layer. The global agent state representation set is input into the process policy decoding path, filtered and aggregated for decoding, and a process policy vector is generated. The global agent state representation set is input into the permission policy decoding path, and features are extracted and decoded from a security perspective to generate a permission policy vector. The process strategy vector and the permission strategy vector are input into the policy linkage gating fusion module to obtain the joint policy output result and execute it.
[0008] Optionally, the steps of acquiring visitor management-related data, performing structured processing on the visitor management-related data, and constructing a multi-agent heterogeneous graph to form an initial heterogeneous graph representation set specifically include: Acquire visitor management-related data, including visitor identity information, visitor information, meeting room and workstation resource status information, security unit information, access control equipment information, and approval flow status information. Perform structured processing on the visitor management-related data, including standardization of original fields, parsing of entity relationships, verification of data validity, and integration of multi-source data, thereby forming an input data set for constructing a multi-agent heterogeneous graph. A set of nodes is constructed based on the input data set to form a multi-agent heterogeneous graph. The set of nodes includes visitor nodes, visited person nodes, conference room resource nodes, security nodes, access control path nodes, and approval nodes. An edge set is constructed based on the input data set to form access relationship edges, approval relationship edges, resource occupation relationship edges, passage path relationship edges, and event triggering relationship edges connecting each node. The node set and edge set are organized according to the node identifier and the reference relationship between the start and end nodes of the edge to form a multi-agent heterogeneous graph with a clear node connection structure; Obtain the corresponding node feature information for each type of node in the node set, generate node type embedding data based on the type of the node, and generate role embedding data based on the functional role of the node in the visitor appointment scenario. Combine the node feature information, node type embedding data and role embedding data to form node representation data. For each edge in the edge set, its edge type is determined, and corresponding edge type embedding data is generated based on the categories of access relationship edges, approval relationship edges, resource occupation relationship edges, travel path relationship edges, and event triggering relationship edges. For edges belonging to event triggering relationship edges, the event type associated with the event triggering relationship edge is further determined, and corresponding event type embedding data is generated. Based on the event occurrence time record associated with the event triggering relationship edge, the time difference information between the event occurrence time and the target reference time is determined, and time difference encoding processing is performed on the time difference information to obtain a combined structure of edge type embedding data, event type embedding data, and time difference encoded data. The node representation data, edge type embedding data, event type embedding data, and time difference encoded data are organized in a unified manner according to the correspondence between the node set and the edge set to form an initial heterogeneous graph representation set for representing the heterogeneous graph structure and associated attributes of multi-agent systems.
[0009] Optionally, the step of parsing visitor reservation requests, generating a reservation request vector, and associating the reservation request vector with the initial heterogeneous graph representation set to obtain the extended heterogeneous graph structure specifically includes: Receive visitor reservation request data submitted by visitors. The visitor reservation request data includes at least the expected visit time, the area to be visited, the meeting room and workstation resources to be used, visitor identity information, and information on the number of accompanying persons. The visitor appointment request data is parsed and formatted. The expected visit time is converted into visit time window data, the area to be visited is converted into visit area data, the meeting room and workstation resources to be used are converted into resource demand data, the visitor identity information is converted into identity feature data, and the accompanying person information is converted into number feature data to obtain standard visitor appointment request data. The standard visitor reservation request data is encoded, which includes access time window data, access area data, resource demand data, identity feature data, and number of people feature data. The encoded features are combined in a preset order to generate a reservation demand vector corresponding to the visitor reservation request. In a multi-agent heterogeneous graph represented by an initial heterogeneous graph representation set, an appointment request node is created. The appointment demand vector is used as part of the node representation data of the appointment request node. Based on the access time window data, access area data, resource demand data, identity feature data, and number of people feature data contained in the appointment demand vector, corresponding appointment relationship edges are generated between the appointment request node and the visitor node, meeting room resource node, access control path node, and approval node. The appointment request node and the appointment relationship edges are added to the initial heterogeneous graph representation set to obtain an extended heterogeneous graph structure.
[0010] Optionally, the step of inputting the extended heterogeneous graph structure into the input layer of the improved heterogeneous Transformer model and performing correspondence-by-correspondence modulation processing to obtain the event-modulated attention result set specifically includes: The node representation data, edge type embedding data, event type embedding data, and time difference encoded data in the extended heterogeneous graph structure are input into the input layer of the improved heterogeneous Transformer model according to the correspondence between nodes and edges to obtain the input node sequence vector. The improved heterogeneous Transformer model includes an input layer, an encoding layer, a process policy decoding path, a permission policy decoding path, and a policy linkage gating fusion module. Based on the node type embedding vector contained in the node representation data and the edge type embedding vector contained in the edge type embedding data, basic attention weights are generated for any node pair. Based on the event type embedding vector corresponding to the edge in the event type embedding data and the time difference encoding vector corresponding to the edge in the time difference encoding data, an event gating factor for modulating the basic attention weight is generated. The basic attention weights and their corresponding event gating factors are modulated one-to-one to generate event-modulated attention results. Collect the event modulation attention results of all node pairs by node index to form an event modulation attention result set.
[0011] Optionally, the step of performing message transformation and aggregation updates based on the event-modulated attention result set and the extended heterogeneous graph structure, and obtaining the global agent state representation set through stacked coding layers, specifically includes: In the improved heterogeneous Transformer model, based on the event modulation attention result set and the node representation data in the extended heterogeneous graph structure, and based on the event modulation attention result and edge type embedding vector, relation-specific message transformation processing is performed on any node pair to generate relation-specific message vectors. Aggregate all relation-specific message vectors for the same target node to obtain the node's aggregated message vector; The node representation data of the target node is concatenated with the aggregated message vector of the target node, and the concatenated data is input into the node update processing structure. The concatenated data is linearly transformed according to the preset update weight to generate the update state representation of the target node. The updated state representations of all nodes are combined into a new set of node representations and input into the encoding layer of the next layer of the improved heterogeneous Transformer model for repeated message transformation, message aggregation and node update processing. Through multi-layer stacking, a global agent state representation set is generated.
[0012] Optionally, the step of inputting the global agent state representation set into the process policy decoding path, filtering and aggregating the decoding, and generating the process policy vector specifically includes: The global agent state representation set is input into the process strategy decoding path of the improved heterogeneous Transformer model. The updated state representations corresponding to visitor nodes, visited person nodes and meeting room resource nodes are selected from the global agent state representation set to form a process-related node state subset. Perform feature aggregation processing on the updated state representations in the process-related node state subset to obtain the process context vector; The process context vector is input into a multi-layer decoding network, and through fully connected transformation and nonlinear activation processing, a process policy hidden representation vector is generated. Through linear mapping, the implicit representation vector of the process strategy is decomposed into an access time window adjustment vector, a path planning vector, a resource allocation vector, and a process jump action probability vector. The access time window adjustment vector, path planning vector, resource allocation vector, and process jump action probability vector are combined to form the process strategy vector.
[0013] Optionally, the step of inputting the global agent state representation set into the permission policy decoding path, performing feature extraction and decoding from a security perspective, and generating a permission policy vector specifically includes: The global agent state representation set is input into the permission policy decoding path of the improved heterogeneous Transformer model. The updated state representations corresponding to security nodes, access control path nodes and area nodes are selected from the global agent state representation set to form a subset of security-related node states. Perform feature aggregation processing on the subset of security-related node states to obtain a security context vector; The permission context vector is input into the permission matrix generation structure. A linear mapping process is performed on the permission context vector, and the access region time matrix is generated by transforming the permission context vector element by element using the access region time weight matrix. The permission context vector is input into the authorization instruction generation structure. A linear mapping process is performed on the permission context vector, and the authorization instruction generation weight matrix is used to transform the permission context vector element by element to generate the authorization instruction set vector. The permission context vector is input into the permission revocation instruction generation structure. A linear mapping process is performed on the permission context vector, and the permission context vector is transformed element by element using the weight matrix generated by the permission revocation instruction to generate a set vector of permission revocation instructions. Input the permission context vector into the risk constraint generation structure, perform linear mapping on the permission context vector, and use the risk constraint generation weight matrix to transform the permission context vector element by element to generate the risk constraint vector. The access area time matrix, authorization instruction set vector, revocation instruction set vector, and risk constraint vector are combined to form the permission policy vector.
[0014] Optionally, the step of inputting the process policy vector and the permission policy vector into the policy linkage gating fusion module, obtaining the joint policy output result, and executing it specifically includes: The process strategy vector and the permission strategy vector are input into the policy linkage gating fusion module, and the process strategy vector and the permission strategy vector are concatenated to form a policy combination vector; The policy combination vector is input into the gating coefficient generation structure, and the gating coefficient vector is generated through linear mapping and activation processing. The process policy vector and the permission policy vector are weighted and fused based on the gating coefficient vector to obtain the joint policy output result. The visitor appointment process is scheduled based on the output of the joint strategy, and the visitor access time window is adjusted according to the time-related content in the output of the joint strategy. Based on the output of the joint strategy, perform resource status update processing, and update the occupancy status of meeting rooms and workstations according to the resource-related content in the output of the joint strategy. Based on the output of the joint strategy, perform visitor and visitee notification distribution processing, and convert the action content contained in the output of the joint strategy into corresponding notification instructions; Access control processing is performed based on the output of the federated policy, mapping the authorization or revocation content in the federated policy output to the access control adjustment result.
[0015] The adaptive visitor appointment management based on multi-agent collaborative decision-making according to embodiments of the present invention includes the following modules: The data processing module is used to acquire visitor management-related data, perform structured processing on the visitor management-related data, and form an input data set; The heterogeneous graph construction module is used to construct a multi-agent heterogeneous graph based on the input data set, forming an initial set of heterogeneous graph representations. The appointment request parsing module is used to parse visitor appointment requests, generate appointment request vectors, and associate appointment request vectors with the initial heterogeneous graph representation set to obtain an extended heterogeneous graph structure. The event-gated attention module is used to input the extended heterogeneous graph structure into the input layer of the improved heterogeneous Transformer model, perform correspondence-by-correspondence modulation processing, and obtain the event-modulated attention result set. The heterogeneous relation update module is used to perform message transformation and aggregation update based on the event modulation attention result set and the extended heterogeneous graph structure, and obtains the global agent state representation set through stacked processing of the coding layer. The process strategy generation module is used to input the global agent state representation set into the process strategy decoding path, perform filtering and aggregation decoding, and generate process strategy vectors. The permission policy generation module is used to input the global agent state representation set into the permission policy decoding path, perform feature extraction and decoding from a security perspective, and generate a permission policy vector. The strategy linkage gating fusion module is used to process the process strategy vector and the permission strategy vector to obtain the joint strategy output result and execute it.
[0016] The beneficial effects of this invention are: This invention constructs a multi-agent heterogeneous graph comprising visitor nodes, visited person nodes, meeting room resource nodes, security nodes, access control path nodes, and approval nodes. This enables a unified representation of multiple types of entities and their relationships within the appointment process, allowing the system to simultaneously process access time, resource usage, approval links, and security path information within the same structure. By introducing node type embedding, edge type embedding, event type embedding, and time difference encoding, the model can characterize cross-entity dependencies under dynamic changes in access demands, improving the accuracy of state representation in complex business scenarios.
[0017] This invention encodes an extended heterogeneous graph structure using an improved heterogeneous Transformer model. By introducing an event gating factor on top of the basic attention weights, it generates event-modulated attention results, enabling the model to differentiate its focus based on access conflicts, approval delays, and changes in path risks. Through this event-sensitive graph attention structure, the system can obtain a more business-logical global agent state representation when resources are scarce, access is concentrated, or security status changes, providing a reliable basis for subsequent policy generation.
[0018] This invention constructs a process policy generation path and a permission policy generation path, and in the policy linkage gating fusion module, it achieves weighted fusion of process policy vectors and permission policy vectors through gating coefficient vectors, enabling access scheduling, resource allocation, path planning, and permission control to be executed under a unified joint policy. This structure can reduce the conflicts caused by the separation of process and security policies, improve the coordination of access process adjustment, resource status update, notification distribution, and permission authorization, and enable the system to achieve higher response speed and consistency in dynamic scenarios. Attached Figure Description
[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 The flowchart shows the adaptive visitor appointment management method based on multi-agent collaborative decision-making proposed in this invention. Figure 2 This is a schematic diagram of the improved heterogeneous Transformer model structure in the adaptive visitor reservation management system and method based on multi-agent collaborative decision-making proposed in this invention. Figure 3 This is a schematic diagram of the adaptive visitor appointment management system based on multi-agent collaborative decision-making proposed in this invention. Detailed Implementation
[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0021] refer to Figure 1 and Figure 2 An adaptive visitor appointment management method based on multi-agent collaborative decision-making includes the following steps: Acquire visitor management-related data, perform structured processing on the visitor management-related data, and construct a multi-agent heterogeneous graph to form an initial heterogeneous graph representation set; Analyze visitor reservation requests, generate reservation request vectors, and associate the reservation request vectors with the initial heterogeneous graph representation set to obtain an extended heterogeneous graph structure; The extended heterogeneous graph structure is input into the input layer of the improved heterogeneous Transformer model, and correspondence-by-correspondence modulation processing is performed to obtain the event-modulated attention result set. Based on the event-modulated attention result set and the extended heterogeneous graph structure, message transformation and aggregation update are performed, and a global agent state representation set is obtained through stacked processing of the coding layer. The global agent state representation set is input into the process policy decoding path, filtered and aggregated for decoding, and a process policy vector is generated. The global agent state representation set is input into the permission policy decoding path, and features are extracted and decoded from a security perspective to generate a permission policy vector. The process strategy vector and the permission strategy vector are input into the policy linkage gating fusion module to obtain the joint policy output result and execute it.
[0022] In this embodiment, the steps of acquiring visitor management-related data, performing structured processing on the visitor management-related data, and constructing a multi-agent heterogeneous graph to form an initial heterogeneous graph representation set specifically include: Acquire visitor management-related data, including visitor identity information, visitor information, meeting room and workstation resource status information, security unit information, access control equipment information, and approval flow status information. Perform structured processing on the visitor management-related data, including standardization of original fields, parsing of entity relationships, verification of data validity, and integration of multi-source data, thereby forming an input data set for constructing a multi-agent heterogeneous graph. A set of nodes is constructed based on the input data set to form a multi-agent heterogeneous graph. The set of nodes includes visitor nodes, visited person nodes, conference room resource nodes, security nodes, access control path nodes, and approval nodes. An edge set is constructed based on the input data set to form access relationship edges, approval relationship edges, resource occupation relationship edges, passage path relationship edges, and event triggering relationship edges connecting each node. The node set and edge set are organized according to the node identifier and the reference relationship between the start and end nodes of the edge to form a multi-agent heterogeneous graph with a clear node connection structure; Obtain the corresponding node feature information for each type of node in the node set, generate node type embedding data based on the type of the node, and generate role embedding data based on the functional role of the node in the visitor appointment scenario. Combine the node feature information, node type embedding data and role embedding data to form node representation data. For each edge in the edge set, its edge type is determined, and corresponding edge type embedding data is generated based on the categories of access relationship edges, approval relationship edges, resource occupation relationship edges, travel path relationship edges, and event triggering relationship edges. For edges belonging to event triggering relationship edges, the event type associated with the event triggering relationship edge is further determined, and corresponding event type embedding data is generated. Based on the event occurrence time record associated with the event triggering relationship edge, the time difference information between the event occurrence time and the target reference time is determined, and time difference encoding processing is performed on the time difference information to obtain a combined structure of edge type embedding data, event type embedding data, and time difference encoded data. The node representation data, edge type embedding data, event type embedding data, and time difference encoded data are organized in a unified manner according to the correspondence between the node set and the edge set to form an initial heterogeneous graph representation set for representing the heterogeneous graph structure and associated attributes of multi-agent systems.
[0023] In this embodiment, the step of parsing visitor reservation requests, generating a reservation request vector, and associating the reservation request vector with the initial heterogeneous graph representation set to obtain the extended heterogeneous graph structure specifically includes: Receive visitor reservation request data submitted by visitors. The visitor reservation request data includes at least the expected visit time, the area to be visited, the meeting room and workstation resources to be used, visitor identity information, and information on the number of accompanying persons. The visitor appointment request data is parsed and formatted. The expected visit time is converted into visit time window data, the area to be visited is converted into visit area data, the meeting room and workstation resources to be used are converted into resource demand data, the visitor identity information is converted into identity feature data, and the accompanying person information is converted into number feature data to obtain standard visitor appointment request data. The standard visitor reservation request data is encoded, which includes access time window data, access area data, resource demand data, identity feature data, and number of people feature data. The encoded features are combined in a preset order to generate a reservation demand vector corresponding to the visitor reservation request. In a multi-agent heterogeneous graph represented by an initial heterogeneous graph representation set, an appointment request node is created. The appointment demand vector is used as part of the node representation data of the appointment request node. Based on the access time window data, access area data, resource demand data, identity feature data, and number of people feature data contained in the appointment demand vector, corresponding appointment relationship edges are generated between the appointment request node and the visitor node, meeting room resource node, access control path node, and approval node. The appointment request node and the appointment relationship edges are added to the initial heterogeneous graph representation set to obtain an extended heterogeneous graph structure.
[0024] In this embodiment, the step of inputting the extended heterogeneous graph structure into the input layer of the improved heterogeneous Transformer model and performing correspondence-by-correspondence modulation processing to obtain the event modulation attention result set specifically includes: The node representation data, edge type embedding data, event type embedding data, and time difference encoded data in the extended heterogeneous graph structure are input into the input layer of the improved heterogeneous Transformer model according to the correspondence between nodes and edges to obtain the input node sequence vector. The improved heterogeneous Transformer model includes an input layer, an encoding layer, a process policy decoding path, a permission policy decoding path, and a policy linkage gating fusion module. Based on the node type embedding vector contained in the node representation data and the edge type embedding vector contained in the edge type embedding data, basic attention weights are generated for any node pair. The specific steps of generating basic attention weights for any pair of nodes include: for any pair of nodes, assigning the node... The node type embedding vector is used to input the query projection matrix to generate the query vector, which will then embed the node type into the vector. The node type embedding vector is input into the key projection matrix to generate the key vector. The edge type embedding data is input into the edge type projection matrix to generate the edge type projection vector. The edge type projection vector and the key vector are additively combined to obtain the relation modulation key vector. Attention scoring is calculated based on the query vector and the relation modulation key vector. The attention scores of all outgoing edges are normalized to generate basic attention weights. Based on the event type embedding vector corresponding to the edge in the event type embedding data and the time difference encoding vector corresponding to the edge in the time difference encoding data, an event gating factor for modulating the basic attention weight is generated. The generation of the event gating factor specifically includes combining the event type embedding vector and the time difference encoding vector, and sequentially performing weighted transformation and gating activation processing to obtain the event gating factor used to modulate the basic attention weights; The basic attention weights and their corresponding event gating factors are modulated one-to-one to generate event-modulated attention results. The terminal modulation process refers to applying the event gating factor multiplicatively to the basic attention weight that is consistent with its index position, thereby generating an event modulation attention result that is structurally corresponding to each item in the set of basic attention weights. Collect the event modulation attention results of all node pairs by node index to form an event modulation attention result set.
[0025] In this embodiment, the step of performing message transformation and aggregation update based on the event modulation attention result set and the extended heterogeneous graph structure, and obtaining the global agent state representation set through stacked coding layers, specifically includes: In the improved heterogeneous Transformer model, based on the event-modulated attention result set and the node representation data in the extended heterogeneous graph structure, and based on the event-modulated attention result and edge type embedding vector, relation-specific message transformation processing is performed on any node pair to generate relation-specific message vectors: ; in, Indicates from node Pointing to node Relationship-specific message vectors Represents a node With nodes Event modulation attention results between Representation based on edges The relationship-specific transformation matrix determined by the edge-type embedding vector. Represents an edge-type embedding vector. Represents nodes in an extended heterogeneous graph structure The nodes represent data; Aggregate all relation-specific message vectors for the same target node to obtain the node's aggregated message vector; The node representation data of the target node is concatenated with the aggregated message vector of the target node, and the concatenated data is input into the node update processing structure. The concatenated data is linearly transformed according to the preset update weight to generate the update state representation of the target node. The updated state representations of all nodes are combined into a new set of node representations and input into the encoding layer of the next layer of the improved heterogeneous Transformer model for repeated message transformation, message aggregation and node update processing. Through multi-layer stacking, a global agent state representation set is generated.
[0026] In this embodiment, the step of inputting the global agent state representation set into the process policy decoding path, filtering and aggregating the decoding to generate the process policy vector specifically includes: The global agent state representation set is input into the process strategy decoding path of the improved heterogeneous Transformer model. The updated state representations corresponding to visitor nodes, visited person nodes and meeting room resource nodes are selected from the global agent state representation set to form a process-related node state subset. Perform feature aggregation processing on the updated state representations in the process-related node state subset to obtain the process context vector; The feature aggregation process includes collecting each updated state representation in the process-related node state subset one by one according to a preset aggregation order, and performing at least one aggregation method among weighted summation, maximum value selection, or average processing on the updated state representation, and using the aggregated result as the process context vector. The process context vector is input into a multi-layer decoding network, and through fully connected transformation and nonlinear activation processing, a process policy hidden representation vector is generated. Through linear mapping, the implicit representation vector of the process strategy is decomposed into an access time window adjustment vector, a path planning vector, a resource allocation vector, and a process jump action probability vector. The linear mapping includes performing linear transformation processing on the implicit representation vector of the process strategy according to a preset set of mapping matrices. The set of mapping matrices includes a mapping matrix for generating an access time window adjustment vector, a mapping matrix for generating a path planning vector, a mapping matrix for generating a resource allocation vector, and a mapping matrix for generating a process jump action probability vector. The access time window adjustment vector, path planning vector, resource allocation vector, and process jump action probability vector are combined to form the process strategy vector.
[0027] In this embodiment, the step of inputting the global agent state representation set into the permission policy decoding path, performing feature extraction and decoding from a security perspective, and generating a permission policy vector specifically includes: The global agent state representation set is input into the permission policy decoding path of the improved heterogeneous Transformer model. The updated state representations corresponding to security nodes, access control path nodes and area nodes are selected from the global agent state representation set to form a subset of security-related node states. Perform feature aggregation processing on the subset of security-related node states to obtain a security context vector; The feature aggregation process includes reading the updated state representation of each node in the security-related node state subset item by item according to a preset aggregation rule, performing a vector-level aggregation operation on the updated state representation, the aggregation operation including at least one of weighted summation processing, maximum value selection processing, and average value calculation processing, and outputting the aggregated result as a permission context vector. The permission context vector is input into the permission matrix generation structure. A linear mapping process is performed on the permission context vector, and the access region time matrix is generated by transforming the permission context vector element by element using the access region time weight matrix. The permission context vector is input into the authorization instruction generation structure. A linear mapping process is performed on the permission context vector, and the authorization instruction generation weight matrix is used to transform the permission context vector element by element to generate the authorization instruction set vector. The permission context vector is input into the permission revocation instruction generation structure. A linear mapping process is performed on the permission context vector, and the permission context vector is transformed element by element using the weight matrix generated by the permission revocation instruction to generate a set vector of permission revocation instructions. Input the permission context vector into the risk constraint generation structure, perform linear mapping on the permission context vector, and use the risk constraint generation weight matrix to transform the permission context vector element by element to generate the risk constraint vector. The access area time matrix, authorization instruction set vector, revocation instruction set vector, and risk constraint vector are combined to form the permission policy vector.
[0028] In this embodiment, the step of linking the process strategy vector and the permission strategy vector into a gating fusion module to obtain the joint strategy output result and execute it specifically includes: The process strategy vector and the permission strategy vector are input into the policy linkage gating fusion module. The process strategy vector and the permission strategy vector are concatenated to form a policy combination vector. The concatenation process includes reading all vector components of the process strategy vector and the permission strategy vector one by one according to the corresponding index order according to the vector arrangement rules, and arranging and combining each component of the process strategy vector and each component of the permission strategy vector in sequence to form a policy combination vector that is structurally continuous and dimensionally expanded. The vector arrangement rule includes placing all components in the process strategy vector in their original component order, and then appending all components in the permission strategy vector in their original component order after all components in the process strategy vector, so that the concatenated strategy combination vector maintains a linear arrangement structure with the process strategy vector first and the permission strategy vector last in the component arrangement. The policy combination vector is input into the gating coefficient generation structure, and the gating coefficient vector is generated through linear mapping and activation processing. The process policy vector and the permission policy vector are weighted and fused based on the gating coefficient vector to obtain the joint policy output result. The weighted fusion process includes performing element-wise weighted processing on the value of each item in the gating coefficient vector and the value of the process policy vector at the corresponding position, performing element-wise weighted processing on the complement value of the gating coefficient vector and the value of the permission policy vector at the corresponding position, and adding the two sets of weighted results item by item at the corresponding position to generate a joint policy output result. The visitor appointment process is scheduled based on the output of the joint strategy, and the visitor access time window is adjusted according to the time-related content in the output of the joint strategy. Based on the output of the joint strategy, perform resource status update processing, and update the occupancy status of meeting rooms and workstations according to the resource-related content in the output of the joint strategy. Based on the output of the joint strategy, perform visitor and visitee notification distribution processing, and convert the action content contained in the output of the joint strategy into corresponding notification instructions; Access control processing is performed based on the output of the federated policy, mapping the authorization or revocation content in the federated policy output to the access control adjustment result.
[0029] refer to Figure 3 An adaptive visitor appointment management system based on multi-agent collaborative decision-making includes the following modules: The data processing module is used to acquire visitor identity information, visitor information, meeting room and workstation resource status information, security unit information, access control equipment information and approval flow status information, and perform field parsing and structured processing on the data to form an input data set for constructing a multi-agent heterogeneous graph; The heterogeneous graph construction module is used to generate visitor nodes, visited person nodes, meeting room resource nodes, security nodes, access control path nodes and approval nodes based on the input data set, and to construct access relationship edges, approval relationship edges, resource occupation relationship edges, passage path relationship edges and event trigger relationship edges to form an initial heterogeneous graph representation set; The appointment request parsing module is used to parse visitor appointment requests, extract access time windows, access areas, resource requirements, identity characteristics and number of people information, generate appointment request vectors, and associate appointment request vectors with the initial heterogeneous graph representation set to obtain an extended heterogeneous graph structure. The event-gated attention module is used to input the extended heterogeneous graph structure into the improved heterogeneous Transformer model, generate a basic attention weight set based on node type embedding and edge type embedding, and generate an event gating factor set based on event type embedding and temporal difference encoding, and perform correspondence-by-correspondence modulation processing to obtain the event-modulated attention result set. The heterogeneous relation update module is used to perform relation-specific message transformation, message aggregation and node state update processing based on the event modulation attention result set, and generates a global agent state representation set through the encoding layer of a multi-layer improved heterogeneous Transformer model; The process strategy generation module is used to select the updated state representations corresponding to visitor nodes, visited person nodes and meeting room resource nodes from the global agent state representation set, perform feature aggregation and multi-layer decoding processing on the updated state representations, and generate access time window adjustment vector, path planning vector, resource allocation vector and process jump action set to form process strategy vector. The permission policy generation module is used to select the updated state representations corresponding to security nodes, access control path nodes and area nodes from the global intelligent agent state representation set, perform security perspective feature aggregation and multi-way instruction mapping on the updated state representations, and generate access area time matrix, authorization instruction set, revocation instruction set and risk constraint information to form permission policy vector. The strategy linkage gating fusion module is used to input the process strategy vector and the permission strategy vector into the strategy linkage gating fusion structure, generate the gating coefficient vector, perform weighted fusion processing on the process strategy vector and the permission strategy vector to obtain the joint strategy output result, and complete visitor appointment process scheduling, resource status update, notification distribution and access control based on the joint strategy output result.
[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to a comprehensive office scenario. This scenario includes multiple access areas, varying numbers of meeting rooms and workstations, multiple security nodes, and cross-floor access control paths, while experiencing a high daily visitor demand. Traditional visitor appointment systems in such scenarios are prone to problems such as long visitor queuing times, uneven meeting room occupancy, delayed access authorization, and cross-area access failures due to static processes, frequent resource conflicts, and complex approval chains, thus affecting overall management efficiency. This invention utilizes multi-agent heterogeneous graph modeling and an improved heterogeneous Transformer model encoding method to uniformly process information on visitors, those being visited, meeting room resources, security units, and access control paths, enabling the system to dynamically analyze complex dependencies and generate strategies.
[0031] In this scenario, the data processing module collects visitor identity data, visitor scheduling information, real-time occupancy data of meeting rooms and workstations, security patrol status data, and access control path status data to form a structured data set. The heterogeneous graph construction module generates an initial heterogeneous graph representation containing multiple types of nodes and relationships. Subsequently, the system analyzes the visitor's access time requirements, access areas, number of visitors, and resource type requirements, and adds this information to the extended heterogeneous graph. The event gating attention module introduces event type and time difference encoding into the model, enabling the system to identify access peaks, temporary scheduling changes, and risk event triggers. The heterogeneous relationship update module forms a global agent state representation based on the encoding results, allowing the system to simultaneously see resource status, approval links, and security risks in a single inference. The process strategy generation module generates strategies for access time adjustment, path planning, and resource allocation based on the global state. The permission strategy generation module generates an access area time matrix, authorization instructions, and revocation instructions. The strategy fusion module finally provides a joint strategy, enabling the visitor appointment process to achieve consistent scheduling under time, resource, and security constraints.
[0032] By applying the method of this invention and comparing the core indicators of the traditional system and the system of this invention within a continuous observation period, the beneficial effects of this invention can be objectively reflected. Data collection covers scenarios with different workdays and different access densities. The results show that this invention can reduce appointment conflict rates, improve the efficiency of the approval process, and reduce access control authorization delays. When meeting resources are scarce, the resource allocation strategy generated by this invention can more effectively avoid the duplication of similar resources and significantly reduce visitor waiting time in multi-area access scenarios. At high access densities, the system can automatically avoid congested access control points when scheduling visitor paths, thereby improving the success rate of passage.
[0033] Table 1: Performance Comparison Data in Visitor Appointment Management Scenarios
[0034] As can be seen from the table above, this invention demonstrates significant improvements over traditional visitor management methods in several core performance indicators for visitor appointment management scenarios. Firstly, regarding process stability, the multi-agent heterogeneous graph structure and event-modulated attention mechanism constructed in this invention can more comprehensively express the complex relationships between visitor nodes, resource nodes, approval nodes, and security nodes, enabling the system to identify potential conflicts in advance when analyzing appointment requests. The data shows that the visitor appointment conflict rate of this invention is 4.8%, significantly lower than the 17.4% of traditional methods, a reduction of over 72%. This indicates that unified state modeling based on multi-agents can effectively reduce appointment failures caused by resource contention or process inconsistencies.
[0035] Secondly, regarding resource scheduling efficiency, this invention achieves dynamic allocation of meeting room and workstation resources through a process strategy generation module. The data shows that the resource occupancy balance of meeting rooms improved from 0.63 in the traditional method to 0.89, an improvement of 41.3%. This result demonstrates that this invention can utilize global agent state representation to adopt more reasonable scheduling strategies during periods of high resource competition, reducing resource idleness and local congestion, and making resource allocation more balanced in both spatial and temporal dimensions.
[0036] Regarding access control and authorization, the authorization policy generation module and policy-linked gating fusion module of this invention can simultaneously handle access area time constraints, authorization rules, and risk information. The cross-regional access failure rate has decreased from 9.6% in traditional methods to 2.1%, a reduction of over 78%, demonstrating that this invention provides a more robust access policy in complex environments with multiple access control paths and multiple security nodes. Simultaneously, the average access control authorization latency has decreased from 14.2 seconds to 3.7 seconds, a reduction of 73.9%, indicating that this invention utilizes the authorization policy aggregated from event gating attention and security node features, enabling the system to generate authorization and revocation actions more quickly, reducing on-site delays.
[0037] Regarding process response speed, the average visitor waiting time decreased from 11.8 minutes using traditional methods to 4.3 minutes, and the average approval process time was shortened from 23.4 minutes to 9.6 minutes, representing improvements of 63.5% and 59%, respectively. Because this invention can simultaneously generate process policies and permission policies, and further integrates them through linked gating, enabling collaborative output of both policies within the same framework, it reduces waiting time caused by process breakpoints and permission delays.
[0038] Finally, in dynamically changing scenarios, the success rate of automatic adjustment of abnormal processes in this invention is increased from 54.7% in the traditional method to 92.3%, an improvement of approximately 68.7%. This indicates that in the event of sudden resource changes, approval node delays, or abnormal access control paths, the strategy generation method of this invention based on global state and event-modulated attention can achieve a higher proportion of automatic adjustment, ensuring the stability and continuity of the visitor process under high load and complex conditions.
[0039] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. An adaptive visitor appointment management method based on multi-agent collaborative decision-making, characterized in that, Includes the following steps: Acquire visitor management-related data, perform structured processing on the visitor management-related data, and construct a multi-agent heterogeneous graph to form an initial heterogeneous graph representation set; Analyze visitor reservation requests, generate reservation request vectors, and associate the reservation request vectors with the initial heterogeneous graph representation set to obtain an extended heterogeneous graph structure; The extended heterogeneous graph structure is input into the input layer of the improved heterogeneous Transformer model, and correspondence-by-correspondence modulation processing is performed to obtain the event-modulated attention result set. Based on the event-modulated attention result set and the extended heterogeneous graph structure, message transformation and aggregation update are performed, and a global agent state representation set is obtained through stacked processing of the coding layer. The global agent state representation set is input into the process policy decoding path, filtered and aggregated for decoding, and a process policy vector is generated. The global agent state representation set is input into the permission policy decoding path to perform feature extraction and decoding from a security perspective, and generate a permission policy vector. The process strategy vector and the permission strategy vector are input into the policy linkage gating fusion module to obtain the joint policy output result and execute it.
2. The adaptive visitor appointment management method based on multi-agent collaborative decision-making according to claim 1, characterized in that, The process of acquiring visitor management-related data, structuring the visitor management-related data, and constructing a multi-agent heterogeneous graph to form an initial heterogeneous graph representation set specifically includes: Acquire visitor management-related data and perform structured processing on the visitor management-related data to form an input data set; Construct a set of nodes and a set of edges connecting the nodes of a multi-agent heterogeneous graph based on the input data set; The node set and edge set are organized according to the node identifier and the start and end node reference relationship of the edge to form a multi-agent heterogeneous graph; Obtain node feature information, node type embedding data, and role embedding data, and combine them to form node representation data; Determine the edge type of each edge in the edge set, and generate the corresponding edge type embedding data, event type embedding data, and time difference encoding data; The node representation data, edge type embedded data, event type embedded data, and time difference encoded data are organized to form an initial heterogeneous graph representation set.
3. The adaptive visitor appointment management method based on multi-agent collaborative decision-making according to claim 1, characterized in that, The process of parsing visitor reservation requests, generating reservation request vectors, and associating these vectors with the initial heterogeneous graph representation set to obtain the extended heterogeneous graph structure specifically includes: Receive visitor appointment request data, perform field parsing and format specification processing on the visitor appointment request data to obtain standard visitor appointment request data; The standard visitor appointment request data is encoded, and the encoded features are combined in a preset order to generate an appointment demand vector corresponding to the visitor appointment request. Create an appointment request node in a multi-agent heterogeneous graph, and generate corresponding appointment relationship edges between the appointment request node and the visitor node, meeting room resource node, access control path node, and approval node. Add the reservation request node and reservation relationship edge to the initial heterogeneous graph representation set to obtain the extended heterogeneous graph structure.
4. The adaptive visitor appointment management method based on multi-agent collaborative decision-making according to claim 1, characterized in that, The step of inputting the extended heterogeneous graph structure into the input layer of the improved heterogeneous Transformer model and performing correspondence-by-correspondence modulation processing to obtain the event modulation attention result set specifically includes: The extended heterogeneous graph structure is input into the input layer of the improved heterogeneous Transformer model to obtain the input node sequence vector; Based on the node type embedding vector contained in the node representation data and the edge type embedding vector contained in the edge type embedding data, basic attention weights are generated for any node pair. Based on the event type embedding vector corresponding to the edge in the event type embedding data and the time difference encoding vector corresponding to the edge in the time difference encoding data, an event gating factor for modulating the basic attention weight is generated. The basic attention weights and their corresponding event gating factors are modulated one-to-one to generate event-modulated attention results. Collect the event modulation attention results of all node pairs by node index to form an event modulation attention result set.
5. The adaptive visitor appointment management method based on multi-agent collaborative decision-making according to claim 1, characterized in that, The process of performing message transformation and aggregation updates based on the event-modulated attention result set and the extended heterogeneous graph structure, and obtaining the global agent state representation set through stacked coding layers, specifically includes: In the improved heterogeneous Transformer model, based on the event-modulated attention results and edge type embedding vectors, relation-specific message transformation processing is performed on any node pair to generate relation-specific message vectors; Aggregate all relation-specific message vectors for the same target node to obtain the node's aggregated message vector; The node representation data of the target node is concatenated with the aggregated message vector of the target node, and then a linear transformation is performed to generate the updated state representation of the target node. The updated state representations of all nodes are combined into a new set of node representations and input into the encoding layer of the next layer of the improved heterogeneous Transformer model for repeated message transformation, message aggregation and node update processing. Through multi-layer stacking, a global agent state representation set is generated.
6. The adaptive visitor appointment management method based on multi-agent collaborative decision-making according to claim 1, characterized in that, The step of inputting the global agent state representation set into the process policy decoding path, filtering and aggregating the decoding, and generating the process policy vector specifically includes: The global agent state representation set is input into the process strategy decoding path of the improved heterogeneous Transformer model. The updated state representations corresponding to visitor nodes, visited person nodes and meeting room resource nodes are selected from the global agent state representation set to form a process-related node state subset. Perform feature aggregation processing on the updated state representations in the process-related node state subset to obtain the process context vector; The process context vector is input into a multi-layer decoding network, and through fully connected transformation and nonlinear activation processing, a process policy hidden representation vector is generated. Through linear mapping, the implicit representation vector of the process strategy is decomposed into an access time window adjustment vector, a path planning vector, a resource allocation vector, and a process jump action probability vector. The access time window adjustment vector, path planning vector, resource allocation vector, and process jump action probability vector are combined to form the process strategy vector.
7. The adaptive visitor appointment management method based on multi-agent collaborative decision-making according to claim 1, characterized in that, The step of inputting the global agent state representation set into the permission policy decoding path, performing feature extraction and decoding from a security perspective, and generating a permission policy vector specifically includes: The global agent state representation set is input into the permission policy decoding path of the improved heterogeneous Transformer model. The updated state representations corresponding to security nodes, access control path nodes and area nodes are selected from the global agent state representation set to form a subset of security-related node states. Perform feature aggregation processing on the security-related node state subset to obtain the security context vector. Input the permission context vector into the permission matrix generation structure to generate the access area time matrix. Input the permission context vector into the authorization instruction generation structure to generate an authorization instruction set vector; input the permission context vector into the revocation instruction generation structure to generate a revocation instruction set vector; input the permission context vector into the risk constraint generation structure to generate a risk constraint vector. The access area time matrix, authorization instruction set vector, revocation instruction set vector, and risk constraint vector are combined to form the permission policy vector.
8. The adaptive visitor appointment management method based on multi-agent collaborative decision-making according to claim 1, characterized in that, The module that integrates the process strategy vector and the permission strategy vector into a gating system to obtain and execute the joint strategy output includes: The process strategy vector and the permission strategy vector are input into the policy linkage gating fusion module, and the process strategy vector and the permission strategy vector are concatenated to form a policy combination vector; The policy combination vector is input into the gating coefficient generation structure, and the gating coefficient vector is generated through linear mapping and activation processing. The process policy vector and the permission policy vector are weighted and fused based on the gating coefficient vector to obtain the joint policy output result. The visitor appointment process is scheduled based on the output of the joint strategy, and the visitor access time window is adjusted according to the time-related content in the output of the joint strategy. Based on the output of the joint strategy, perform resource status update processing, and update the occupancy status of meeting rooms and workstations according to the resource-related content in the output of the joint strategy. Based on the output of the joint strategy, perform visitor and visitee notification distribution processing, and convert the action content contained in the output of the joint strategy into corresponding notification instructions; Access control processing is performed based on the output of the federated policy, mapping the authorization or revocation content in the federated policy output to the access control adjustment result.
9. The adaptive visitor reservation management system based on multi-agent collaborative decision-making according to claim 1, executing the adaptive visitor reservation management method based on multi-agent collaborative decision-making according to any one of claims 1 to 8, characterized in that, Includes the following modules: The data processing module is used to acquire visitor management-related data, perform structured processing on the visitor management-related data, and form an input data set; The heterogeneous graph construction module is used to construct a multi-agent heterogeneous graph based on the input data set, forming an initial set of heterogeneous graph representations. The appointment request parsing module is used to parse visitor appointment requests, generate appointment request vectors, and associate appointment request vectors with the initial heterogeneous graph representation set to obtain an extended heterogeneous graph structure. The event-gated attention module is used to input the extended heterogeneous graph structure into the input layer of the improved heterogeneous Transformer model, perform correspondence-by-correspondence modulation processing, and obtain the event-modulated attention result set. The heterogeneous relation update module is used to perform message transformation and aggregation update based on the event modulation attention result set and the extended heterogeneous graph structure, and obtains the global agent state representation set through stacked processing of the coding layer. The process strategy generation module is used to input the global agent state representation set into the process strategy decoding path, perform filtering and aggregation decoding, and generate process strategy vectors. The permission policy generation module is used to input the global agent state representation set into the permission policy decoding path, perform feature extraction and decoding from a security perspective, and generate a permission policy vector. The strategy linkage gating fusion module is used to process the process strategy vector and the permission strategy vector to obtain the joint strategy output result and execute it.