Industrial Internet of Things equipment detection method and system based on federal map neural network

Abstract

The invention relates to the technical field of industrial Internet of Things detection, in particular to an industrial Internet of Things equipment detection method and system based on a federal map neural network. The method comprises the steps of obtaining real-time operation state data of each device and a measurement value of a sensor, constructing a topological graph by using input data as device node features, introducing a knowledge graph based on the topological graph, fusing a client device state and knowledge graph information, and obtaining a client device state information fusion result. Introducing an attention mechanism to adaptively adjust the weight of the local model of the client; the method comprises the following steps: fusing node features, performing feature enhancement based on the fused node features, establishing a federated graph neural network client dynamic scheduling strategy based on reinforcement learning, and performing a node classification task of graph structure data by using a global model. The method improves the performance of the model in the absence of annotation data, enables the model to learn more robust features, and improves the judgment accuracy of equipment faults.

Description

Technical Field

[0001] The present invention relates to the field of industrial Internet of Things detection technology, and in particular to an industrial Internet of Things equipment detection method and system based on a federated graph neural network. Background Art

[0002] With the rapid development of information technology and communication technology, the Industrial Internet of Things, as an important branch of the Internet of Things technology, has been widely used in many fields such as manufacturing, energy, and transportation. The Industrial Internet of Things connects various devices, sensors, machines, and control systems to the Internet of Things to achieve information exchange and collaborative work between devices. In the Industrial Internet of Things, equipment fault detection is a key link to ensure the stable operation of equipment. There are complex dependencies between devices in the Industrial Internet of Things. Traditional fault detection methods rely on manual inspections and regular inspections, and often cannot effectively utilize the interactive information between devices, limiting the accuracy of fault detection.

[0003] Federated graph neural networks combine the advantages of federated learning and graph neural networks, and are particularly suitable for distributed equipment fault detection tasks in the Industrial Internet of Things. However, the application of federated graph neural networks in the field of Industrial Internet of Things faces two key challenges. First, the cost of data annotation in the Industrial Internet of Things is high. Due to the huge amount of data, complex scenarios and high professional knowledge requirements, it is very difficult to obtain a large amount of labeled data. This makes traditional supervised learning methods based on a large amount of labeled data not well applicable to the Industrial Internet of Things. Secondly, the performance and data resources of client devices in the Industrial Internet of Things vary significantly. The computing power, storage space and network bandwidth of different devices vary. This may lead to uneven resource utilization or performance bottlenecks under a fixed client selection strategy. At this stage, a method and system for industrial Internet of Things device detection based on federated graph neural networks are needed. Summary of the invention

[0004] In order to solve the problem of low accuracy of equipment fault detection in the industrial Internet of Things due to data labeling difficulties and uneven resource utilization, the present invention provides an industrial Internet of Things equipment detection method and system based on a federated graph neural network. Contrastive learning is introduced into the federated graph neural network and applied to industrial Internet of Things equipment fault detection. It can be trained without relying on labeled data, adapting to the difficulty of equipment data labeling in the industrial Internet of Things, and enhancing the accuracy and reliability of equipment fault detection.

[0005] In the first aspect, the present invention provides an industrial Internet of Things equipment detection method based on a federated graph neural network, which adopts the following technical solution: A method for detecting industrial Internet of Things devices based on a federated graph neural network, comprising: Obtain the real-time operating status data and sensor measurement values ​​of each device as input data; Use the input data as the device node features to build a topology graph, introduce the knowledge graph based on the topology graph, and build multi-dimensional edge relationships in the topology graph; The client device status is integrated with the knowledge graph information to obtain the integrated node features, and the attention mechanism is introduced to adaptively adjust the weight of the client local model; Feature enhancement is performed based on the fused node features, including generating an enhanced view for each node, then introducing a projection head to map the node representation to a new space, and finally calculating the contrast loss between the enhanced views; Establish a dynamic scheduling strategy for federated graph neural network clients based on reinforcement learning, including a dynamic client selection mechanism based on the ε-greedy strategy and define the performance score of the client in the tth round; The client local models after the specified federated training rounds are aggregated into a global model, and the global model is used to perform node classification tasks for graph structured data to complete industrial Internet of Things equipment fault detection.

[0006] Furthermore, the construction of multi-dimensional edge relationships in the topological graph includes obtaining input data, regarding each device as a node of the topological graph, with the node feature being the corresponding input data vector, using an adjacency matrix to represent the connection relationship, and updating the adjacency matrix as a weighted adjacency matrix using cosine similarity as the weight of the edge, introducing a knowledge graph containing equipment failure modes and maintenance history information, obtaining a knowledge graph node feature matrix based on knowledge nodes and knowledge node feature vectors, connecting device nodes with related knowledge nodes in the knowledge graph, and obtaining multi-dimensional edge relationships in the topological graph.

[0007] Furthermore, the fusion of the client device status and the knowledge graph information includes using a weighted adjacency matrix to represent the edge relationship of the physical and operational associations between devices, determining the weight of the edge between the device node and the knowledge node according to the relevance of the device and the knowledge, and aggregating the features of the neighbor nodes by weighted averaging. The aggregation formula is expressed as: , in, It is l +1 layer node feature representation, Is a node In the l The feature vector of the layer, is the adjacency matrix, representing the nodes and The weight of the edge between Is a node The degree of the node The number of adjacent nodes), and It is lThe learning weight matrices of the layers correspond to device information and knowledge graph information respectively.

[0008] Furthermore, the attention mechanism is introduced to adaptively adjust the weight of the client local model, including calculating the attention coefficient according to the node features. After the attention coefficient is introduced, the neighbor node features are re-aggregated to obtain an updated node feature representation. The attention coefficient calculation formula is: , in, Is a node With Node The attention weights between is the learning vector used to calculate the attention weights, and are the device features and knowledge features of the lth layer, represents the vector concatenation operation, It is represented as the activation function.

[0009] Furthermore, the feature enhancement based on the fused node features includes generating two enhanced views for each node, then introducing a projection head to map the final node representation output by the encoder to a new space, calculating the contrast loss according to the enhanced view, and minimizing the contrast loss. The contrast loss calculation formula is: , in, and are the representations of the two enhanced views of node v after the projection head, is the cosine similarity function, is a parameter that controls the smoothness of the loss, is the set of all nodes in the mini-batch, is the indicator function, when 1 if the value is 0, otherwise it is 0.

[0010] Furthermore, the establishment of a dynamic scheduling strategy for the federated graph neural network client based on reinforcement learning includes calculating the value of probability ε according to the current training round, and randomly selecting a client local model for training with probability ε, selecting the client local model with the best historical performance with probability 1-ε, and calculating the performance score of the tth round according to the average loss of the current round. Finally, a gradient-based dynamic adjustment mechanism is introduced in combination with the performance score. The formula of the dynamic adjustment mechanism is: , in, Is the client local model In the t The gradient norm of the wheel, is the weight coefficient of the gradient information.

[0011] Furthermore, the client local models after the specified federated training rounds are aggregated into a global model, including introducing an adaptive learning rate mechanism to update the learning rate of each client local model, defining a reset cycle, resetting the performance scores of all client local models to an initial state after training for the reset cycle, and aggregating the client local models after the specified federated training rounds into a global model. The adaptive learning rate formula is: , in, is the learning rate of the previous round, Represents the learning rate adjustment factor, which controls the magnitude of the adjustment. represents a symbolic function that determines whether to increase or decrease the learning rate based on the direction of performance change. and Represent the performance scores of the current round and the previous round, The function limits the learning rate to range, used to prevent the learning rate from becoming too large or too small.

[0012] In the second aspect, an industrial Internet of Things equipment detection system based on a federated graph neural network includes: Data acquisition module: obtains the real-time operating status data of each device and the measurement value of the sensor as input data; The topology graph module is configured to: construct a topology graph using input data as device node features, introduce a knowledge graph based on the topology graph, and construct a multi-dimensional edge relationship in the topology graph; The model module is configured to: fuse the client device status with the knowledge graph information to obtain the fused node features, and introduce the attention mechanism to adaptively adjust the weight of the client local model; The feature enhancement module is configured to: perform feature enhancement based on the fused node features, including generating an enhanced view for each node, then introducing a projection head to map the node representation to a new space, and finally calculating the contrast loss between the enhanced views; The regulation module is configured to: establish a dynamic scheduling strategy for federated graph neural network clients based on reinforcement learning, including adopting a dynamic client selection mechanism based on the ε-greedy strategy, and define the performance score of the client in the tth round; The transformation module is configured to aggregate the client local models after the specified federated training rounds into a global model, use the global model to perform node classification tasks of graph structure data, and complete industrial Internet of Things equipment fault detection.

[0013] In a third aspect, the present invention provides a computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, a method for detecting industrial Internet of Things devices based on a federated graph neural network.

[0014] In a fourth aspect, the present invention provides a terminal device comprising a processor and a computer-readable storage medium, the processor being used to implement various instructions; the computer-readable storage medium being used to store multiple instructions, the instructions being suitable for being loaded and executed by the processor, a method for detecting industrial Internet of Things devices based on a federated graph neural network.

[0015] In summary, the present invention has the following beneficial technical effects: 1. The present invention constructs a topological graph by acquiring the real-time operation status data of the equipment and the sensor measurement values, and introduces the knowledge graph to construct a multi-dimensional edge relationship, which comprehensively depicts the complex relationship between the equipment and provides rich information for fault detection. For example, by connecting the equipment with the knowledge nodes such as the fault mode and maintenance history in the knowledge graph, the possibility of equipment failure can be judged from multiple dimensions, and the detection accuracy is improved by 3%-5% compared with the traditional method.

[0016] 2. The present invention integrates the client device status with the knowledge graph information, and introduces the attention mechanism to adaptively adjust the weights, so that the model can focus on key information and more accurately capture the characteristics of equipment failures. By calculating the attention coefficient, reasonable weights are assigned to the characteristics of different neighbor nodes, thereby enhancing the model's ability to identify equipment failures.

[0017] 3. The present invention performs feature enhancement based on the fused node features, generates enhanced views, introduces projection heads and calculates contrast loss, thereby improving the performance of the model in the absence of labeled data. The calculation and minimization of contrast loss enables the model to learn more robust features and enhances the accuracy of judging equipment failures.

[0018] 4. The present invention establishes a client dynamic scheduling strategy based on reinforcement learning, adopts the ε-greedy strategy to select clients, balances the relationship between exploration and utilization, dynamically adjusts the selection probability according to the current training round, randomly explores new clients to obtain diverse information in the early stage, and selects more clients with good performance in the later stage to optimize the model, thereby improving the model training efficiency and generalization ability.

[0019] 5. The present invention uses a periodic reset mechanism to prevent the model from being overly dependent on a small number of clients, so that underestimated clients have the opportunity to re-participate in training and evaluation, adapt to the dynamic environment of the Industrial Internet of Things, and further optimize the federated learning process.

[0020] 6. The present invention uses random feature masking and edge dropout operations in feature enhancement to enable the model to learn more robust feature representations and enhance its adaptability to noise and data changes. When there is noise in the data or some features are missing, the model can still accurately determine equipment failures. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a schematic diagram of the overall process of an industrial Internet of Things equipment detection method based on a federated graph neural network according to Example 1 of the present invention. DETAILED DESCRIPTION

[0022] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0023] Example 1 Reference Figure 1 , a method for detecting industrial Internet of Things devices based on a federated graph neural network in this embodiment includes: Obtain the real-time operating status data and sensor measurement values ​​of each device as input data; Use the input data as the device node features to build a topology graph, introduce the knowledge graph based on the topology graph, and build multi-dimensional edge relationships in the topology graph; The client device status is integrated with the knowledge graph information to obtain the integrated node features, and the attention mechanism is introduced to adaptively adjust the weight of the client local model; Feature enhancement is performed based on the fused node features, including generating an enhanced view for each node, then introducing a projection head to map the node representation to a new space, and finally calculating the contrast loss between the enhanced views; Establish a dynamic scheduling strategy for federated graph neural network clients based on reinforcement learning, including a dynamic client selection mechanism based on the ε-greedy strategy and define the performance score of the client in the tth round; The client local models after the specified federated training rounds are aggregated into a global model, and the global model is used to perform node classification tasks for graph structured data to complete industrial Internet of Things equipment fault detection.

[0024] Specifically, a method for detecting industrial Internet of Things devices based on a federated graph neural network includes the following steps: like Figure 1 As shown, S1, obtain the real-time operating status data of each device and the measurement value of the sensor as input data; The input data mainly includes the real-time operating status data of each device and the measurement values ​​of the sensors. The real-time operating status data covers multiple aspects such as temperature, vibration, pressure, etc. These data can directly reflect the current working status of the equipment. The sensor measurement values ​​are accurate records of the physical quantities related to the equipment through various sensors, providing detailed information for the monitoring of the equipment status. These input data are the basis for building topological maps and introducing knowledge graphs. Their accuracy and completeness are crucial to the reliability of subsequent analysis. The input data includes the real-time operating status data of the equipment (such as temperature, vibration, pressure, etc.) and the sensor measurement values. Assume that there are n devices in total. For the i-th device, its input data can be expressed as a vector ,where ,d is the feature dimension, and different dimensions correspond to different operating states or ,measurement values.

[0025] S2. Use the input data as the device node features to build a topology graph, introduce the knowledge graph based on the topology graph, and build multi-dimensional edge relationships in the topology graph; Consider each device as a node in the topology graph, and the node feature is the input data vector , the node feature matrix of the entire topological graph can be expressed as , represented as a Matrix, where n is the number of devices and d is the dimension.

[0026] The connection of the edge is determined based on the physical relationship, operation dependency or synergy between devices. The connection relationship of the graph can be represented by an adjacency matrix A. If there is an edge between device i and device j, then ;otherwise . And usually (The node itself has no self-loops), A is a The symmetric matrix of . Then the edge weights are calculated. The edge weights reflect the similarity or interaction frequency between devices. Euclidean distance and cosine similarity are commonly used to calculate similarity. Taking cosine similarity as an example, the similarity between device i and device j is The calculation formula is: , in, represents the vector dot product, represents the magnitude of a vector, and are the node feature vectors of devices i and j respectively, then the edge weight It can be defined based on similarity, for example, simply , update the adjacency matrix A to a weighted adjacency matrix, then Represents the weight of the edge between devices i and j.

[0027] Finally, the knowledge graph is introduced, which contains information such as equipment failure mode and maintenance history. Assume that there are m knowledge nodes in the knowledge graph, and each knowledge node also has its feature vector ,in, is the characteristic dimension of the knowledge node, and the node characteristic matrix of the knowledge graph is , is a , where m is the number of knowledge nodes in the knowledge graph, They are k knowledge nodes in The value of a feature dimension.

[0028] Connecting the device nodes in the topology graph with the knowledge nodes in the knowledge graph can be represented by an association matrix R. If the device node i is connected to the knowledge node k, then ;otherwise , R is a The matrix of .

[0029] Based on the weighted adjacency matrix A calculated previously, this part of the edge relationship is represented to reflect the direct physical and operational associations between devices. For the edges between device nodes and knowledge nodes, their weights can also be defined. For example, the weight is determined based on the correlation between the device and the knowledge. Assume that the correlation between device i and knowledge node k is obtained by a certain similarity calculation method. , update the association matrix R to be a weighted association matrix, Represents the weight of the edge between device i and knowledge node k.

[0030] S3, fuse the client device status with the knowledge graph information to obtain the fused node features, and introduce the attention mechanism to adaptively adjust the weight of the client local model; Furthermore, in the client local model, the interactive information between device nodes and knowledge graph nodes is considered at the same time, and the features of neighbor nodes are aggregated by weighted average to achieve the fusion of device status information and knowledge graph information: , in, It is l +1 layer node feature representation, Is a node In the l The feature vector of the layer, is the adjacency matrix, representing the nodes and The weight of the edge between Is a node The degree of the node The number of adjacent nodes), and It isl The learning weight matrices of the layers correspond to device information and knowledge graph information respectively.

[0031] In order to enhance the expressive power of graph neural networks, this method uses the attention mechanism to adaptively adjust the weights according to the relationship between different nodes, taking into full account the importance of possible differences between device nodes and knowledge nodes in the graph: , in, Is a node With Node The attention weights between is the learning vector used to calculate the attention weights, and Yes, yes l The device characteristics and knowledge characteristics of the layer, represents the vector concatenation operation, It represents the learning weight matrix of the device information corresponding to the lth layer. and The nodes are and The device state feature vector at layer l represents, Represents the learned weight matrix corresponding to the knowledge graph information of the lth layer. In this way, the client local model can not only capture the dependencies between devices, but also use the information from the knowledge graph to enhance the understanding of device failure modes.

[0032] S4, feature enhancement based on the fused node features, including generating an enhanced view for each node, then introducing a projection head to map the node representation to a new space, and finally calculating the contrast loss between the enhanced views; In order to solve the problem of scarce client-side labeled data in the industrial Internet of Things, contrastive learning is introduced into the client local model. Contrastive learning can enhance the model's feature learning ability through self-supervision, thereby improving performance in the absence of labeled data. First, two enhanced views are generated for each node v and : , in, represents the data enhancement operation, is a hyperparameter of the enhancement strategy. Specific enhancement methods include random feature masking and edge dropout. This enhancement can be seen as a different "observation" of the same node, which helps the model learn more robust features. Next, in order to more effectively process these enhanced views, this method introduces a projection head , mapping the node representations to a new space where the contrastive loss is applied: , in, is the final node representation output by the encoder (L represents the last layer), and is the weight matrix of the projection head, and is the bias vector, is the activation function, and the projection head is a two-layer neural network, where and The first layer is formed. and This structure helps to separate the representation learning and contrastive learning tasks, so that the backbone network can retain more information. On this basis, this method further defines the contrast loss. For a pair of enhanced views , the contrast loss can be expressed as: , in, and are the representations of the two enhanced views of node v after the projection head, is the cosine similarity function, is a parameter that controls the smoothness of the loss, is the set of all nodes in the mini-batch, is the indicator function, when 1 if the value is 0, otherwise it is 0.

[0033] By designing such a loss function, this method encourages the representations of different views from the same node to be similar, while ensuring that the representations from different nodes are different. Ultimately, the overall contrastive learning goal of this model is to minimize the contrastive loss of all node pairs in the mini-batch, which is defined as follows: , in, is the size of the mini-batch. In this way, this method can take advantage of contrastive learning, improve the performance of the model in dealing with scenarios with a lack of labeled data, and further improve the model performance.

[0034] S5. Establish a dynamic scheduling strategy for federated graph neural network clients based on reinforcement learning, including adopting a dynamic client selection mechanism based on the ε-greedy strategy and defining the performance score of the client in the tth round; In addition to introducing contrastive learning on the client side to enhance feature learning capabilities, this method also designs a dynamic scheduling strategy for federated graph neural network clients based on reinforcement learning during the server aggregation process. The client selection and training process is optimized through reinforcement learning, thereby achieving more efficient resource allocation and model updates, and improving the adaptability and learning effect of the overall system in heterogeneous environments. This strategy aims to balance the trade-off between exploration and utilization to adapt to the complexity of graph structured data in the industrial Internet of Things and the dynamic changes in client performance. In the first round of federated learning, the system randomly selects clients to participate in training with a probability of ε, and selects the client with the best historical performance with a probability of 1-ε. Specifically, for a total of M clients, in the tth round of training, this method selects m clients (m < M) to participate, and the selection process is as follows: , in, Indicates the client to be selected. represents random selection, Indicates that the client with the best performance is selected. The value is adjusted dynamically with the training round: , in, represents the initial exploration rate, represents the minimum exploration rate, represents the maximum number of training rounds, Indicates the current training round. In the process of dynamic client selection, the performance evaluation of the client plays a vital role, so this method defines the client Performance score at round t , to quantify its performance and contribution in each round of training: , in, Is the client In the The average loss after rounds of local training, Is the client In the The accuracy after a round of local training, Is the client The amount of local data, It is the largest amount of data among all clients. and is the weight coefficient, which is used to balance the importance of different factors and satisfy To quantify the client performance, this method further defines the loss and accuracy Calculation method of loss The calculation of is intended to reflect the performance of the model in local client training, which can be specifically expressed as: , in, Is the client The batch size is is the loss function, is the current model, and are input features and labels respectively. Accuracy The calculation of is as follows: , in Is the client The size of the dataset, represents the indicator function. After each round of training, the performance score of each client is updated. When the best performing client needs to be selected, the model will select the client with the highest performance score. This performance evaluation method comprehensively considers the model's performance on the client and the client's data contribution, thus providing a more comprehensive client quality measure in a federated environment.

[0035] When selecting clients, this method not only considers the performance score of each client , and also introduces a dynamic adjustment mechanism based on the gradient calculated during the training process of each client. Specifically, the size of the client's gradient can be used to measure its contribution to the global model. If a client's gradient update is small (which means that the client's contribution to the global model is low), its participation frequency can be reduced. To this end, this method introduces gradient information when selecting the client: in, Is the client In the t The gradient norm of the wheel, is the weight coefficient of the gradient information. Finally, the comprehensive performance score of the client can be expressed as: In this way, this method can comprehensively consider the client's historical performance, computational contribution, and gradient changes when dynamically selecting clients to optimize the model training process.

[0036] S6. Aggregate the client local models after the specified federated training rounds into a global model, use the global model to perform node classification tasks for graph structure data, and complete industrial Internet of Things equipment fault detection.

[0037] During the training of the federated graph neural network, different clients have significantly different learning processes and convergence speeds due to differences in their local data distribution and computing power. Using a uniform fixed learning rate may cause some clients to train too fast or too slow, affecting the convergence and performance of the overall model. To solve this problem, this method designs an adaptive learning rate mechanism. The core idea of ​​the adaptive learning rate is to dynamically adjust the learning rate of each client based on its historical performance. Specifically, for the client , this method updates its learning rate in round t as follows: , in, is the learning rate of the previous round, Represents the learning rate adjustment factor, which controls the magnitude of the adjustment. represents a symbolic function that determines whether to increase or decrease the learning rate based on the direction of performance change. and Represent the performance scores of the current round and the previous round, The function limits the learning rate to The learning rate range is used to prevent the learning rate from becoming too large or too small. This adaptive learning rate mechanism can automatically adjust according to performance changes during training without human intervention. Clients with good performance can accelerate learning through a larger learning rate, while clients with poor performance can stabilize the learning process through a smaller learning rate. Through personalized learning rates, the differences in learning speeds between different clients can be balanced to a certain extent, reducing the negative impact of data heterogeneity on model training.

[0038] Dynamic client selection and adaptive learning rate mechanisms can significantly improve the performance of the model, but in the long-term training process, the model may over-rely on a few clients with excellent performance, causing other potentially valuable clients to be neglected for a long time. To avoid this problem, this method introduces a periodic reset mechanism. Specifically, this method defines a reset period T. After every T rounds of training, this method resets the performance scores of all clients to the initial state: , in, It is a predefined initial performance score, usually set to 0 or a neutral value. By introducing a periodic reset mechanism, underestimated clients get the opportunity to be reselected and evaluated, so that they can better adapt to the dynamic environment in the industrial Internet of Things. The reset period T is a hyperparameter that needs to be adjusted according to the specific application scenario. A shorter period is conducive to rapid adaptation to changes, while a longer period helps to make better use of high-quality clients. After completing the specified federated training rounds, the global federated graph neural network model training is completed. The client uses the trained global model to perform node classification tasks of graph structure data, thereby realizing the industrial Internet of Things equipment fault detection task.

[0039] Example 2 The difference between this embodiment and embodiment 1 is that this embodiment provides an industrial Internet of Things equipment detection system based on a federated graph neural network, including: Data acquisition module: obtains the real-time operating status data of each device and the measurement value of the sensor as input data; The topology graph module is configured to: construct a topology graph using input data as device node features, introduce a knowledge graph based on the topology graph, and construct a multi-dimensional edge relationship in the topology graph; The model module is configured to: fuse the client device status with the knowledge graph information to obtain the fused node features, and introduce the attention mechanism to adaptively adjust the weight of the client local model; The feature enhancement module is configured to: perform feature enhancement based on the fused node features, including generating an enhanced view for each node, then introducing a projection head to map the node representation to a new space, and finally calculating the contrast loss between the enhanced views; The regulation module is configured to: establish a dynamic scheduling strategy for federated graph neural network clients based on reinforcement learning, including adopting a dynamic client selection mechanism based on the ε-greedy strategy, and define the performance score of the client in the tth round; The transformation module is configured to aggregate the client local models after the specified federated training rounds into a global model, use the global model to perform node classification tasks of graph structure data, and complete industrial Internet of Things equipment fault detection.

[0040] A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, a method for detecting industrial Internet of Things devices based on a federated graph neural network.

[0041] A terminal device includes a processor and a computer-readable storage medium, the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded and executed by the processor. A method for detecting industrial Internet of Things devices based on a federated graph neural network.

[0042] The above are all preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for detecting industrial Internet of Things devices based on a federated graph neural network, characterized in that: include: Obtain the real-time operating status data and sensor measurement values ​​of each device as input data; Use the input data as the device node features to build a topology graph, introduce the knowledge graph based on the topology graph, and build multi-dimensional edge relationships in the topology graph; The client device status is integrated with the knowledge graph information to obtain the integrated node features, and the attention mechanism is introduced to adaptively adjust the weight of the client local model; Feature enhancement is performed based on the fused node features, including generating an enhanced view for each node, then introducing a projection head to map the node representation to a new space, and finally calculating the contrast loss between the enhanced views; Establish a dynamic scheduling strategy for federated graph neural network clients based on reinforcement learning, including a dynamic client selection mechanism based on the ε-greedy strategy and define the performance score of the client in the tth round; The client local models after the specified federated training rounds are aggregated into a global model, and the global model is used to perform node classification tasks for graph structured data to complete industrial Internet of Things equipment fault detection.

2. According to claim 1, a method for detecting industrial Internet of Things devices based on a federated graph neural network is characterized in that: The method of constructing a multi-dimensional edge relationship in a topological graph includes obtaining input data, regarding each device as a node of the topological graph, wherein the node feature is a corresponding input data vector, using an adjacency matrix to represent the connection relationship, and updating the adjacency matrix as a weighted adjacency matrix using cosine similarity as the weight of the edge, introducing a knowledge graph containing device failure modes and maintenance history information, obtaining a knowledge graph node feature matrix based on knowledge nodes and knowledge node feature vectors, connecting device nodes with relevant knowledge nodes in the knowledge graph, and obtaining a multi-dimensional edge relationship in the topological graph.

3. According to claim 2, a method for detecting industrial Internet of Things devices based on a federated graph neural network is characterized in that: The fusion of the client device status and the knowledge graph information includes using a weighted adjacency matrix to represent the edge relationship of the physical and operational associations between devices, determining the weight of the edge between the device node and the knowledge node according to the relevance of the device and the knowledge, and aggregating the features of the neighbor nodes by weighted averaging. The aggregation formula is expressed as: ， in, is the node feature representation of the l+1th layer, Is a node The feature vector at layer l, is the adjacency matrix, representing the nodes and The weight of the edge between Is a node The degree of node The number of adjacent nodes of and is the learning weight matrix of the lth layer, corresponding to device information and knowledge graph information respectively.

4. According to claim 3, a method for detecting industrial Internet of Things devices based on a federated graph neural network is characterized in that: The attention mechanism is introduced to adaptively adjust the weight of the client local model, including calculating the attention coefficient according to the node features. After the attention coefficient is introduced, the neighbor node features are re-aggregated to obtain an updated node feature representation. The attention coefficient calculation formula is: ， in, Is a node With Node The attention weights between is the learning vector used to calculate the attention weights, and are the device features and knowledge features of the lth layer, represents the vector concatenation operation, It is represented as the activation function.

5. According to claim 4, a method for detecting industrial Internet of Things devices based on a federated graph neural network is characterized in that: The feature enhancement based on the fused node features includes generating two enhanced views for each node, then introducing a projection head to map the final node representation output by the encoder to a new space, calculating the contrast loss according to the enhanced view, and minimizing the contrast loss. The contrast loss calculation formula is: ， in, and are the representations of the two enhanced views of node v after the projection head, is the cosine similarity function, is a parameter that controls the smoothness of the loss, is the set of all nodes in the mini-batch, is the indicator function, when 1 if the value is 0, otherwise it is 0.

6. According to claim 5, a method for detecting industrial Internet of Things devices based on a federated graph neural network is characterized in that: The establishment of a dynamic scheduling strategy for the client of a federated graph neural network based on reinforcement learning includes calculating the value of probability ε according to the current training round, randomly selecting a client local model for training with probability ε, selecting the client local model with the best historical performance with probability 1-ε, calculating the performance score of the tth round according to the average loss of the current round, and finally introducing a gradient-based dynamic adjustment mechanism in combination with the performance score. The formula of the dynamic adjustment mechanism is: ， in, Is the client local model The gradient norm at the tth round, is the weight coefficient of the gradient information.

7. The method for detecting industrial Internet of Things devices based on a federated graph neural network according to claim 6 is characterized in that: The client local models after the specified federation training rounds are aggregated into a global model, including introducing an adaptive learning rate mechanism to update the learning rate of each client local model, defining a reset cycle, resetting the performance scores of all client local models to an initial state after training for the reset cycle, and aggregating the client local models after the specified federation training rounds into a global model. The adaptive learning rate formula is: ， in, is the learning rate of the previous round, Represents the learning rate adjustment factor, which controls the magnitude of the adjustment. represents a symbolic function that determines whether to increase or decrease the learning rate based on the direction of performance change. and Represent the performance scores of the current round and the previous round, The function limits the learning rate to within the range.

8. An industrial Internet of Things equipment detection system based on a federated graph neural network, characterized in that: include: Data acquisition module: obtains the real-time operating status data of each device and the measurement value of the sensor as input data; The topology graph module is configured to: construct a topology graph using input data as device node features, introduce a knowledge graph based on the topology graph, and construct a multi-dimensional edge relationship in the topology graph; The model module is configured to: fuse the client device status with the knowledge graph information to obtain the fused node features, and introduce the attention mechanism to adaptively adjust the weight of the client local model; The feature enhancement module is configured to: perform feature enhancement based on the fused node features, including generating an enhanced view for each node, then introducing a projection head to map the node representation to a new space, and finally calculating the contrast loss between the enhanced views; The regulation module is configured to: establish a dynamic scheduling strategy for federated graph neural network clients based on reinforcement learning, including adopting a dynamic client selection mechanism based on the ε-greedy strategy, and define the performance score of the client in the tth round; The transformation module is configured to aggregate the client local models after the specified federated training rounds into a global model, use the global model to perform node classification tasks of graph structure data, and complete industrial Internet of Things equipment fault detection.

9. A computer-readable storage medium storing a plurality of instructions, characterized in that: The instructions are suitable for being loaded by a processor of a terminal device and executing an industrial Internet of Things device detection method based on a federated graph neural network as described in claim 1.

10. A terminal device, comprising a processor and a computer-readable storage medium, wherein the processor is used to implement each instruction; and the computer-readable storage medium is used to store multiple instructions, characterized in that: The instructions are suitable for being loaded by a processor and executing an industrial Internet of Things device detection method based on a federated graph neural network as described in claim 1.

Images