Iot service exception positioning model deployment method based on dynamic subnet segmentation

Abstract

This invention provides a method for deploying an IoT service anomaly localization model based on dynamic subnet segmentation. The method includes: constructing a dynamic graph sequence of IoT service interactions; generating an initial set of IoT subnets based on the dynamic graph sequence; constructing a hybrid expert model for each initial subnet using a hybrid expert framework; dynamically adjusting the boundaries of the initial subnets using expert model inference loss and vertex migration metric functions to obtain subnets with refined boundaries; obtaining subnet structure and attribute embeddings using the Weisfeiler-Lehman algorithm; constructing graph kernel similarity based on optimal transmission distance; and classifying all subnets; and performing training, parameter transfer, lightweight fine-tuning, and online deployment of the service anomaly localization model for subnets of the same type. This method can dynamically segment subnets in large-scale IoT service interaction networks, identify functionally similar subnets, effectively transfer the service anomaly localization model between subnets of the same type, and reduce model training resource consumption.

Description

Technical Field

[0001] This invention relates to the fields of IoT service operation and maintenance, graph data analysis and model migration technology, and in particular to an IoT service anomaly location model migration method based on dynamic subnetting. Background Technology

[0002] With the development of the Internet of Things (IoT), edge computing, and intelligent operation and maintenance technologies, the scale of terminal devices, network nodes, business services, and sensed data in IoT systems continues to grow. The IoT has formed a large-scale network environment with multi-level topologies, dynamic service call relationships, and complex operational metrics.

[0003] In complex IoT application environments, service networks typically have the following characteristics: First, the scale is large, with diverse node types and service functions, making it difficult for a single global model to balance training cost and model accuracy. Second, the network topology and indicator status change dynamically, and fixed partitions cannot accurately reflect the true affiliation of nodes. Third, there are subnets with similar functions in different regions or different business scenarios, which provides a basis for model reuse and migration deployment; Fourth, subnet similarity is reflected not only in the topology, but also in multi-dimensional information such as node attributes, edge weight relationships, service calls, and operational metrics.

[0004] Currently, existing methods for anomaly localization in IoT services include: training models uniformly on a global network, or training models separately within manually pre-defined fixed subnets. The drawbacks of these existing methods include: the former requires processing all network nodes, resulting in high training and inference overhead, easily consuming IoT computing resources and affecting the normal operation of business systems; the latter relies on static topology or manual experience for segmentation, making it difficult to adapt to the dynamic network structure of IoT. When network boundary division is unreasonable, nodes within the same subnet have significant functional differences, while different subnets may have highly similar structures and operating modes, leading to redundant model training, low transfer efficiency, and unstable anomaly localization results. Summary of the Invention

[0005] This invention provides a deployment method for an IoT service anomaly location model based on dynamic subnetting, so as to effectively improve the operating efficiency of IoT systems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution.

[0007] A method for deploying an IoT service anomaly localization model based on dynamic subnetting, comprising: Construct a dynamic graph sequence of service interactions for the Internet of Things (IoT), and generate an initial set of subnets for the IoT based on the dynamic graph sequence. For each initial subnet, a hybrid expert model oriented towards a hybrid expert framework is constructed. The boundary of the initial subnet is dynamically adjusted using the expert model inference loss and vertex migration metric function to obtain a subnet with refined boundary. The subnet structure and attribute embeddings are obtained using a graph feature encoding algorithm. Graph kernel similarity is constructed based on the optimal transmission distance, and all subnets are classified. Training, parameter transfer, lightweight fine-tuning, and online deployment of service anomaly localization models for similar subnets.

[0008] Preferably, the step of constructing the service interaction dynamic graph sequence for the Internet of Things (IoT), and generating an initial set of subnets for the IoT based on the service interaction dynamic graph sequence, includes: Multi-source data is collected from the IoT system, including device ledgers, network topology, service registration information, service call logs, operational metrics, fault alarms, and business context data. This multi-source data is mapped into a unified service interaction sequence diagram. The nodes in this diagram include device nodes, service nodes, business nodes, and network resource nodes, while the edges include physical connection relationships, service call relationships, dependency relationships, control relationships, and data flow relationships. For each time slice... t By combining node operation metrics, edge weights, service call frequency, exception tags, and business context information, a corresponding dynamic graph of service interaction is generated. ,in Represents a set of nodes. Denotes the set of edges. Represents the node feature matrix, This represents the edge weight or service interaction strength matrix. Indicates a label indicating an abnormal node or service; The service interaction dynamic graph is partitioned using a graph partitioning algorithm to obtain an initial set of subnets. , Indicates the first Individual Network.

[0009] Preferably, the step of constructing a hybrid expert model for each initial subnet using a hybrid expert framework, and dynamically adjusting the boundaries of the initial subnets using expert model inference loss and vertex migration metric functions to obtain subnets with refined boundaries, includes: Configure a graph neural network-based expert model for each initial subnet. k Each initial subnet is trained separately. k An expert model based on a graph neural network { Each expert model is trained only on the corresponding subnet. The training tasks include node function classification and service status classification. The expert model takes the node features, adjacency matrix, edge weight matrix and historical index sequence in the subnet as input and outputs the node function category and service operation status classification. For subnets Define training loss ( | ) represents the expert model In subnet The overall convergence loss over the subnet; for any node within the subnet ,definition ( | This indicates that the node is defined by the expert model. The node-level loss generated during inference is applied to the nodes within each subnet after each training cycle. ( | Sort the results and select the one with the largest loss. Q Each node forms a candidate migration set. ; For candidate nodes Design a vertex migration metric function Score ( → ), Score ( → )=α·Spinner( → )-β· Loss ( | )-γ·Balance( ), among which, Spinner ( → ) represents a node Migrate to subnet Subsequent changes in weights or zonal balancing returns Loss ( | ) represents the target expert model For nodes The inference loss, Balance( ) represents the size or load penalty term of the target subnet, where α, β, and γ are adjustable weights for candidate nodes. ,choose Score ( → The highest-ranking subnet will be designated as the new home subnet; Repeat the expert model training, candidate node identification, migration metric calculation, and node reassignment process until the node-level loss generated during inference within each subnet is lower than a preset threshold, or the maximum number of iterations is reached, resulting in a dynamically refined set of subnets. .

[0010] Preferably, the step of using a graph feature encoding algorithm to obtain subnet structure and attribute embeddings, constructing graph kernel similarity based on optimal transmission distance, and classifying all subnets includes: For any two subnets after boundary refinement, construct non-attribute network representation and attribute network representation respectively. The non-attribute network retains only nodes, edges, node types, and edge types for calculating structural similarity; the attribute network retains node indicators, edge weights, call frequency, anomaly history, and knowledge semantic attributes for calculating attribute similarity. For any two subnets, let the embedding matrices of the graph feature encoding algorithm be respectively... and Calculate the optimal transmission distance between the two embedding matrices. , It is to solve and The distance matrix between each vertex feature in the matrix. Represents random variables and The joint distribution It is a transfer matrix. It uses the Frobenius dot product. For non-attribute networks, the node distance is calculated using Hamming distance; for attribute networks, the node distance is calculated using Euclidean distance or weighted distance, thus determining the optimal transmission distance. Convert to kernel ; For the non-attribute network representations of any two subnets, a non-attribute structure graph kernel is calculated according to the above processing procedure. For the attribute network representations of any two subnets, an attribute graph kernel is calculated according to the above processing procedure. The non-attribute structure graph kernel and the attribute graph kernel are multiplied to obtain the comprehensive similarity of the two subnets. Based on the comprehensive similarity of the two subnets, a threshold judgment or clustering algorithm is used to identify functionally similar subnet clusters, and all subnets are classified.

[0011] Preferably, the training, parameter transfer, lightweight fine-tuning, and online deployment of the service anomaly localization model for similar subnets include: For each category of subnet clusters, select subnets with sufficient sample size, high label quality, or strong business representativeness as source subnets. Train the service anomaly localization model on the source subnets. After the service anomaly localization model of the source subnet is trained, transfer the parameters, feature encoder, or graph convolutional layer of the service anomaly localization model of the source subnet to the same target subnet. If the target subnet has labeled samples, update the parameters of the anomaly localization model by fine-tuning the samples; if the target subnet lacks labeled samples, use pseudo-labels, self-training, or feature distribution alignment to perform unsupervised adaptation of the anomaly localization model. The transfer training objective function of the service anomaly localization model is expressed as: ,in This indicates a loss in the operation and maintenance tasks. This represents the feature distribution alignment loss between the source subnet and the target subnet. The parameter regularization term is represented by μ and ν, which are weight coefficients. After the migration of the service anomaly localization model is completed, the same or similar anomaly localization models are deployed to the same type of subnet and online inference is performed. When the service anomaly localization model detects a shift in subnet structure, service call pattern, or indicator distribution during online operation, it triggers the re-processing of subnet segmentation, subnet similarity recalculation, and service anomaly localization model migration.

[0012] As can be seen from the technical solution provided by the present invention, the method of the present invention can dynamically segment subnets and identify subnets with similar functions in a large-scale Internet of Things service interaction network, and effectively migrate the service anomaly localization model between subnets of the same type, thereby reducing the consumption of model training resources and improving intelligent operation and maintenance capabilities.

[0013] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of the invention. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 A flowchart illustrating the deployment method of an IoT service anomaly location model based on dynamic subnetting, provided in an embodiment of the present invention; Figure 2 This is a structural diagram of a wire-laying unit provided in an embodiment of the present invention.

[0016] Figure 3This is a schematic diagram illustrating the migration of a similar subnet mining and service anomaly localization model provided in an embodiment of the present invention. Detailed Implementation

[0017] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0018] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or couplings. The term “and / or” as used herein includes any and all combinations of one or more of the associated listed items.

[0019] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0020] To facilitate understanding of the embodiments of the present invention, the following will provide further explanation and description with reference to the accompanying drawings and several specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0021] This invention proposes a deployment method for IoT service anomaly localization models based on dynamic subnet segmentation and similarity modeling. First, a dynamic graph representation is generated based on IoT service interaction time-series data, and initial subnets are obtained using a conventional graph partitioning algorithm. Then, a MOE (Mixture of Experts)-GNN (Graph Neural Network) boundary refinement model oriented towards a hybrid expert framework is constructed. A corresponding graph neural network expert model is trained for each initial subnet, and the subnet boundaries are dynamically adjusted using node inference loss and vertex migration metric functions. After obtaining stable subnets, the Weisfeiler-Lehman algorithm is used to obtain the subnet structure and attribute embeddings, and graph kernel similarity is constructed based on the optimal transmission distance to identify functionally similar subnets. Finally, the service anomaly localization model is trained, parameter transferred, lightweight fine-tuned, and deployed online for similar subnets, thereby reducing the overall network training cost while ensuring anomaly localization effectiveness.

[0022] The present invention may include the following four main parts: a dynamic graph sequence construction scheme, a dynamic subnet segmentation scheme based on boundary refinement, a similar subnet modeling scheme based on WL-OT (Weisfeiler–Lehman Optimal Transport) graph kernel, and a model migration and deployment scheme for similar subnets.

[0023] 1. Construct a dynamic graph sequence of service interactions.

[0024] First, this invention unifies the representation of devices, services, subnets, gateways, sensors, links, call relationships, and operational metrics in an IoT network as a service interaction sequence diagram. Nodes in the service interaction sequence diagram include device nodes, service nodes, business nodes, and network resource nodes, while edges include physical connection relationships, service call relationships, dependency relationships, control relationships, and data flow relationships. For each time slice... t By combining node operation metrics, edge weights, service call frequency, exception tags, and business context information, a corresponding dynamic graph of service interaction is generated. ,in Represents a set of nodes. Denotes the set of edges. Represents the node feature matrix, This represents the edge weight or service interaction strength matrix. Indicates a node or service exception label. For consecutive time slices, a dynamic graph sequence of service interactions can be obtained. , ,..., This is used to reflect the historical evolution of IoT services and operational status.

[0025] 2. Divide the dynamic subnet

[0026] Figure 2 A flowchart illustrating dynamic subnetting and subnet boundary refinement provided in an embodiment of the present invention. Figure 2 As shown, for large-scale service interaction dynamic graph sequences, this invention first uses the conventional graph partitioning algorithm Kernighan-Lin to generate multiple initial subnet sets. The initial segmentation is used to reduce the global network size and provide a base partition for subsequent expert model training.

[0027] Based on the initial subnet, this invention constructs a boundary refinement model oriented towards a hybrid expert framework. For k Each initial subnet is trained separately. k Expert models of graph neural networks { Each expert model is trained only on its corresponding subnet. Training tasks can include node function classification and service status classification. The expert model learns a consistent representation of nodes within the subnet by aggregating the topological neighborhood features, temporal index features, and knowledge semantic features of nodes within the subnet.

[0028] For subnets and its expert model Define training loss ( | This indicates that the expert model is in the subnet. The overall convergence loss over the subnet; for any node within the subnet ,definition ( | This represents the node-level loss generated by the current expert model during inference. If the node... If a node exhibits a high inference loss within the current subnet, it indicates that the node's fit with the current subnet's topology, attribute characteristics, or functional pattern is low, and its subnet affiliation needs to be reassessed.

[0029] The key to boundary refinement lies in identifying poorly performing candidate nodes and recalculating their affiliation. Specifically, after each training cycle, this invention refines the boundary values ​​of nodes within each subnet. ( | Sort the results and select the one with the largest loss. Q Each node forms a candidate migration set. For candidate nodes This invention designs a vertex migration metric function. Score ( → The algorithm comprehensively considers the degree of topology matching between the node after migration and the target subnet, the inference loss of the target expert model for the node, and the load balancing constraints of the target subnet.

[0030] The vertex migration metric function can be expressed as: Score ( → )=α·Spinner( → )-β· Loss ( | )-γ·Balance( Among them, Spinner ( → ) represents a node Migrate to subnet Subsequent changes in weights or zonal balancing returns Loss ( | ) represents the target expert model For nodes The inference loss, Balance( ) represents the size or load penalty term of the target subnet, where α, β, and γ are adjustable weights. For candidate ,choose Score ( → The highest-ranking subnet is designated as the new home subnet.

[0031] The process of expert model training, candidate node identification, migration metric calculation, and node reassignment is repeated until the inference loss of nodes within each subnet is lower than a preset threshold, or the maximum number of iterations is reached. This yields a set of dynamically refined subnets. This dynamic subnet set can more accurately reflect the real business boundaries and functional aggregation relationships in the Internet of Things (IoT) network.

[0032] 3. Similar Subnet Modeling

[0033] After obtaining the set of subnets with refined boundaries, this invention further models the similarity between subnets. Similarity modeling includes two aspects: structural similarity and attribute similarity. Structural similarity mainly reflects the degree of similarity in node types, edge types, topological patterns, and service call paths within the subnets; attribute similarity mainly reflects the degree of similarity in node performance indicators, edge weight strength, service call frequency, historical anomaly patterns, and knowledge semantic attributes.

[0034] This invention first employs a graph feature encoding algorithm to embed each subnet. For non-attribute subnets, the WL algorithm iteratively aggregates the discrete labels of nodes and their neighbors to generate hash-compressed node structure features. For attribute subnets, the WL algorithm further introduces node attributes, edge weights, and index features during the neighbor aggregation process, thereby obtaining a subnet embedding matrix containing topological structure and operational attributes. The aforementioned graph feature encoding algorithm needs to satisfy the following: generating node-level feature representations or node embedding matrices based on node neighborhood, connectivity, node type, edge type, and / or attribute information. The aforementioned graph feature encoding algorithm can be the Weisfeiler-Lehman algorithm, GCN, GraphSAGE, GAT, GIN, spectral graph feature extraction, etc. The Weisfeiler-Lehman algorithm is essentially a method that forms graph features by iteratively aggregating neighborhood labels, and is not the only means of obtaining subnet structure embeddings. GCN, GraphSAGE, GAT, and GIN can also generate node representations based on graph structure and node attributes.

[0035] For any two subnets, their embedding matrices are respectively and This invention calculates the optimal transmission distance between the two. Among them, among them, It is to solve and The distance matrix between each vertex feature. Represents random variables and The joint distribution of . It is a transfer matrix. It's the Frobenius dot product; the distance formula measures how far the data is distributed from the given distribution. "Move into distribution" The minimum average distance required for movement is determined. For non-attribute networks, Hamming distance can be used for node distance; for attribute networks, Euclidean distance or weighted distance can be used. Next, the optimal transmission distance is converted into a graph kernel. Furthermore, this invention calculates the non-attribute structure graph kernel and the attribute graph kernel separately, and multiplies them to obtain a comprehensive similarity. By calculating the comprehensive similarity of subnets pairwise, a subnet similarity matrix can be constructed, and functionally similar subnet clusters can be identified based on threshold determination or clustering algorithms.

[0036] 4. Migration and Deployment of Service Anomaly Localization Model for Similar Subnets

[0037] For each similar subnet cluster identified, this invention selects a subnet with sufficient sample size, high label quality, or strong business representativeness as the source subnet, and trains the service anomaly localization model on the source subnet.

[0038] In the target subnet, this invention transfers the parameters of the service anomaly localization model trained in the source subnet to the target subnet and performs lightweight adaptation based on the data conditions of the target subnet. If the target subnet has a small number of labeled samples, a few-sample fine-tuning method is used to update the classification layer or some high-level parameters; if the target subnet lacks labeled samples, unsupervised adaptation is performed using pseudo-labels, self-training, or feature distribution alignment. The transfer training objective can be represented as... ,in This indicates a loss in the operation and maintenance tasks. This represents the feature distribution alignment loss between the source subnet and the target subnet. This represents the parameter regularization term, where μ and ν are weight coefficients.

[0039] The processing flow of the IoT service anomaly location model deployment method based on dynamic subnet segmentation and similarity modeling provided in this embodiment of the invention is as follows: Figure 1 As shown, it includes the following steps: Step S1: Construct a dynamic graph sequence of service interactions The system collects device ledgers, network topology, service registration information, service call logs, operational metrics, fault alarms, and business context data from the IoT system, mapping this multi-source data into a unified service interaction sequence diagram. Nodes include devices, services, subnets, links, business tasks, and runtime events; relationships include connections, calls, deployments, dependencies, control, and triggers. The service interaction sequence diagram is sliced ​​according to fixed time windows, resulting in multiple dynamic service interaction diagrams. Each dynamic diagram includes node characteristics, edge weight characteristics, service call characteristics, and anomaly label information.

[0040] Step S2: Generate the initial subnet set

[0041] Based on the dynamic graph of service interactions, a graph partitioning algorithm is used for initial partitioning. The initial partitioning comprehensively considers node connectivity, service call intensity, node type, metric similarity, and business affiliation constraints to obtain an initial set of subnets. During initial partitioning, partition size constraints and edge cut constraints can be set to avoid subnets being too large or too small, and to reduce the unreasonable splitting of strongly dependent nodes.

[0042] Step S3: Train the MOE-GNN expert model

[0043] like Figure 2As shown, a corresponding graph neural network expert model is configured for each initial subnet. For any subnet, the expert model takes the node features, adjacency matrix, edge weight matrix, and historical index sequence within the subnet as input, and outputs the node function category and service operation status classification. The training objective of the expert model is to minimize the task loss of the nodes within the subnet, so that the nodes within the same subnet maintain a high degree of consistency in topology, attribute features, and anomaly patterns.

[0044] Step S4: Identify high-loss nodes and refine their boundaries

[0045] After each round of expert model training, the inference loss of each node within a subnet is calculated, and the Q nodes with the highest losses from each subnet are selected as candidate migration nodes. For any candidate node, inference is performed using the expert model corresponding to different target subnets, and the migration metric Score is calculated by combining the Spinner score, target inference loss, and subnet balancing penalty. The target subnet with the highest migration metric is selected as the node's new home, and the subnet partitioning results are updated.

[0046] Repeat the above training, identification, and transfer operations until the average training loss of all subnets stabilizes, or the inference loss of candidate nodes falls below a preset threshold. After boundary refinement, the original initial subnets are adjusted into a dynamic set of subnets. This process can migrate nodes with low fit to the current subnet to more suitable subnets, thereby improving the functional consistency and model adaptability within each subnet.

[0047] Step S5: Discover similar subnets

[0048] Figure 3 This is a schematic diagram illustrating the migration of a similar subnet mining and service anomaly localization model provided in an embodiment of the present invention, as shown below. Figure 3 As shown, for any two subnets after boundary refinement, non-attribute network representations and attribute network representations are constructed respectively. The non-attribute network retains only nodes, edges, node types, and edge types for calculating structural similarity; the attribute network further retains node indices, edge weights, call frequencies, anomaly history, and knowledge semantic attributes for calculating attribute similarity.

[0049] WL embedding is performed on each subnet to obtain a subnet node-level embedding matrix. The optimal transmission distance between the embedding matrices of two subnets is then calculated and converted into a graph kernel similarity. For the non-attribute network representations of any two subnets, a non-attribute structure graph kernel is calculated according to the above process; for the attribute network representations of any two subnets, an attribute graph kernel is calculated according to the above process. The combined similarity between the non-attribute structure graph kernel and the attribute graph kernel is obtained.

[0050] Based on the comprehensive similarity of two subnets, threshold judgment or clustering algorithms are used to identify functionally similar subnet clusters, resulting in several functionally similar subnet clusters. All subnets are then classified.

[0051] S6: Service Anomaly Localization Model Migration and Deployment

[0052] Within each similar subnet cluster, subnets with sufficient data, high label reliability, and stable historical operation are selected as source subnets to train service anomaly localization models. The transfer training objective function for the service anomaly localization model is expressed as follows: ,in This indicates a loss in the operation and maintenance tasks. This represents the feature distribution alignment loss between the source subnet and the target subnet. This represents the parameter regularization term, where μ and ν are weight coefficients. The service anomaly localization model includes a feature encoding module, a relationship feature propagation module, a feature fusion module, and an anomaly localization module.

[0053] The feature encoding module is used to perform feature encoding on the input subnet data. The input subnet data includes service nodes, service call edges, node types, edge types, node operation indicators, etc. in the subnet, which are converted into corresponding feature vectors through discrete feature encoding or vector mapping.

[0054] The relational feature propagation module is used to perform feature aggregation and information propagation between adjacent service nodes based on the calling relationships between them. This ensures that the feature representation of each service node simultaneously includes the service's own running state and the state information of its neighboring services with which it has a calling relationship. This can be implemented using graph-based information propagation methods such as graph convolution, neighborhood aggregation, and graph attention.

[0055] The feature fusion module is used to fuse the topological features, operating status features, and temporal change features of service nodes to obtain the fused feature representation of each service node.

[0056] The anomaly localization module is used to calculate the anomaly probability or anomaly score corresponding to each service node based on the fusion feature representation of each service node, and obtain the anomaly score vector corresponding to the service node in the subnet; service nodes with anomaly probability greater than a preset anomaly threshold are identified as candidate anomaly services, and the anomaly service localization result is determined based on the anomaly score, service call relationship and anomaly propagation direction of the candidate anomaly services.

[0057] After the service anomaly localization model of the source subnet is trained, its parameters, feature encoder, or graph convolutional layer are transferred to a similar target subnet. The parameters of the service anomaly localization model are then fine-tuned using a small number of samples from the target subnet. For target subnets lacking labeled samples, lightweight adaptation can be achieved through pseudo-labels, self-supervised consistency constraints, or feature distribution alignment.

[0058] After the service anomaly localization model in the target subnet is migrated, similar or related anomaly localization models are deployed to similar subnets and online inference is performed. When the service anomaly localization model detects significant shifts in subnet structure, service call patterns, or indicator distribution during online operation, this invention triggers a re-segmentation, similarity recalculation, or model re-migration process. The system re-executes steps S2 to S6, forming a closed loop of "segmentation—similarity modeling—migration deployment—online feedback." Therefore, this invention can reduce redundant model training and computational resource consumption in large-scale IoT environments while maintaining good anomaly localization accuracy, achieving continuous adaptation to dynamic IoT environments.

[0059] In summary, this invention significantly reduces training and inference resource consumption in large-scale IoT scenarios by replacing global model training with dynamic subnetting; it improves subnet functional consistency by refining the MOE-GNN boundary, allowing subnetting to consider topology, dynamic metrics, and model inference loss simultaneously; it achieves a unified measurement of subnet structure and attribute similarity through WL embedding, optimal transmission distance, and graph kernel similarity fusion; it reduces redundant training and improves the reusability and deployment efficiency of the service anomaly localization model across different business areas by migrating similar subnet models; and it enhances the method's adaptability to dynamic network changes and business model shifts through an online feedback update mechanism.

[0060] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0061] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for apparatus or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The apparatus and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0062] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for deploying an IoT service anomaly location model based on dynamic subnetting, characterized in that, include: Construct a dynamic graph sequence of service interactions for the Internet of Things (IoT), and generate an initial set of subnets for the IoT based on the dynamic graph sequence. For each initial subnet, a hybrid expert model oriented towards a hybrid expert framework is constructed. The boundary of the initial subnet is dynamically adjusted using the expert model inference loss and vertex migration metric function to obtain a subnet with refined boundary. The subnet structure and attribute embeddings are obtained using a graph feature encoding algorithm. Graph kernel similarity is constructed based on the optimal transmission distance, and all subnets are classified. Training, parameter transfer, lightweight fine-tuning, and online deployment of service anomaly localization models for similar subnets.

2. The method according to claim 1, characterized in that, The aforementioned construction of the service interaction dynamic graph sequence for the Internet of Things (IoT), and the generation of an initial set of subnets for the IoT based on the service interaction dynamic graph sequence, includes: Multi-source data is collected from the IoT system, including device ledgers, network topology, service registration information, service call logs, operational metrics, fault alarms, and business context data. This multi-source data is mapped into a unified service interaction sequence diagram. The nodes in this diagram include device nodes, service nodes, business nodes, and network resource nodes, while the edges include physical connection relationships, service call relationships, dependency relationships, control relationships, and data flow relationships. For each time slice... t By combining node operation metrics, edge weights, service call frequency, exception tags, and business context information, a corresponding dynamic graph of service interaction is generated. ,in Represents a set of nodes. Denotes the set of edges. Represents the node feature matrix, This represents the edge weight or service interaction strength matrix. Indicates a label indicating an abnormal node or service; The service interaction dynamic graph is partitioned using a graph partitioning algorithm to obtain an initial set of subnets. , Indicates the first Individual Network.

3. The method according to claim 1, characterized in that, The process of constructing a hybrid expert model for each initial subnet using a hybrid expert framework, and dynamically adjusting the boundaries of the initial subnets using expert model inference loss and vertex migration metric functions to obtain subnets with refined boundaries, includes: Configure a graph neural network-based expert model for each initial subnet. k Each initial subnet is trained separately. k An expert model based on a graph neural network { Each expert model is trained only on the corresponding subnet. The training tasks include node function classification and service status classification. The expert model takes the node features, adjacency matrix, edge weight matrix and historical index sequence in the subnet as input and outputs the node function category and service operation status classification. For subnets Define training loss ( | ) represents the expert model In subnet The overall convergence loss over the subnet; for any node within the subnet ,definition ( | This indicates that the node is defined by the expert model. The node-level loss generated during inference is applied to the nodes within each subnet after each training cycle. ( | Sort the results and select the one with the largest loss. Q Each node forms a candidate migration set. ; For candidate nodes Design a vertex migration metric function Score ( → ), Score ( → )=α·Spinner( → )-β· Loss ( | )-γ·Balance( ), among which, Spinner ( → ) represents a node Migrate to subnet Subsequent changes in weights or zonal balancing returns Loss ( | ) represents the target expert model For nodes The inference loss, Balance( ) represents the size or load penalty term of the target subnet, where α, β, and γ are adjustable weights for candidate nodes. ,choose Score ( → The highest-ranking subnet will be designated as the new home subnet; Repeat the expert model training, candidate node identification, migration metric calculation, and node reassignment process until the node-level loss generated during inference within each subnet is lower than a preset threshold, or the maximum number of iterations is reached, resulting in a dynamically refined set of subnets. .

4. The method according to claim 3, characterized in that, The method described above utilizes graph feature encoding algorithms to obtain subnet structure and attribute embeddings, constructs graph kernel similarity based on optimal transmission distance, and classifies all subnets into categories, including: For any two subnets after boundary refinement, construct non-attribute network representation and attribute network representation respectively. The non-attribute network retains only nodes, edges, node types, and edge types for calculating structural similarity; the attribute network retains node indicators, edge weights, call frequency, anomaly history, and knowledge semantic attributes for calculating attribute similarity. For any two subnets, let the embedding matrices of the graph feature encoding algorithm be respectively... and Calculate the optimal transmission distance between the two embedding matrices. , It is to solve and The distance matrix between each vertex feature in the matrix. Represents random variables and The joint distribution It is a transfer matrix. It uses the Frobenius dot product. For non-attribute networks, the node distance is calculated using Hamming distance; for attribute networks, the node distance is calculated using Euclidean distance or weighted distance, thus determining the optimal transmission distance. Convert to kernel ; For the non-attribute network representations of any two subnets, a non-attribute structure graph kernel is calculated according to the above processing procedure. For the attribute network representations of any two subnets, an attribute graph kernel is calculated according to the above processing procedure. The non-attribute structure graph kernel and the attribute graph kernel are multiplied to obtain the comprehensive similarity of the two subnets. Based on the comprehensive similarity of the two subnets, a threshold judgment or clustering algorithm is used to identify functionally similar subnet clusters, and all subnets are classified.

5. The method according to any one of claims 1 to 4, characterized in that, The training, parameter transfer, lightweight fine-tuning, and online deployment of the service anomaly localization model for similar subnets include: For each category of subnet clusters, select subnets with sufficient sample size, high label quality, or strong business representativeness as source subnets. Train the service anomaly localization model on the source subnets. After the service anomaly localization model of the source subnet is trained, transfer the parameters, feature encoder, or graph convolutional layer of the service anomaly localization model of the source subnet to the same target subnet. If the target subnet has labeled samples, update the parameters of the anomaly localization model by fine-tuning the samples; if the target subnet lacks labeled samples, use pseudo-labels, self-training, or feature distribution alignment to perform unsupervised adaptation of the anomaly localization model. The transfer training objective function of the service anomaly localization model is expressed as: ,in This indicates a loss in the operation and maintenance tasks. This represents the feature distribution alignment loss between the source subnet and the target subnet. The parameter regularization term is represented by μ and ν, which are weight coefficients. After the migration of the service anomaly localization model is completed, the same or similar anomaly localization models are deployed to the same type of subnet and online inference is performed. When the service anomaly localization model detects a shift in subnet structure, service call pattern, or indicator distribution during online operation, it triggers the re-processing of subnet segmentation, subnet similarity recalculation, and service anomaly localization model migration.

Images