A decentralized graph federated learning method integrating cross-regional device information

Abstract

The present application relates to a decentralized graph federated learning method that integrates cross-regional device information. A graph structure is constructed based on the communication relationship between devices in factories in the same or different regions. The neighbor devices in the same region of each device and the first-order neighbor devices and second-order neighbor devices across regions are found based on the graph structure. For each device, the feature information of its first-order neighbor devices and second-order neighbor devices are aggregated, and the embedding information of the corresponding device in the first layer of the graph convolutional neural network is obtained based on the aggregated information received by each of the devices. The embedding information of the device in the first layer of the graph convolutional neural network is updated. The embedding information of the device in the second layer of the graph convolutional neural network is updated based on the embedding information of each device in the second layer of the graph convolutional neural network. The loss function on each device and the gradient corresponding to its loss function are calculated based on the embedding information of each device in the second layer of the graph convolutional neural network. The gradient is updated based on the gradient of the neighbor devices in the same region of the device. The device updates the local model parameters based on the corresponding updated gradient.

Description

Technical Field

[0001] The present application relates to the technical field of decentralized graph federated learning, and in particular to a decentralized graph federated learning method that integrates cross-regional device information. Background Art

[0002] Currently, in the field of industrial intelligent robots, information exchange between robot devices in factories is carried out through graph structure relationships, which is crucial for model training; however, factories in different regions often face challenges in data privacy, laws and regulations, and communication costs when exchanging information, and once these information exchanges are missing, the graph structure will be incomplete, resulting in a significant decline in the performance of the model trained by robot devices in the factory. In addition, model training in factories is usually handled by central servers. As the number of robots increases, this centralized topology puts tremendous pressure on network bandwidth and communication load, and lacks robustness. Once the central node is attacked or an accident occurs, the entire federated learning process will be seriously affected.

[0003] Therefore, it is necessary to design a federated learning method that can effectively solve the problems of information interaction privacy security and communication cost in the field of industrial intelligent robots, while also effectively solving the defects of centralized topological structures. Summary of the invention

[0004] Based on this, it is necessary to provide a decentralized graph federated learning method that integrates cross-regional device information. The method includes:

[0005] S1: Based on the communication relationship between devices in factories in the same area or different areas, a graph structure is constructed; and based on the graph structure, neighbor devices in the same area of ​​each device, as well as first-order neighbor devices and second-order neighbor devices across areas are found. The second-order neighbor devices are neighbor devices in the same area of ​​the first-order neighbor devices.

[0006] S2: for each of the devices, aggregate feature information of its first-order neighbor devices and second-order neighbor devices, and obtain embedding information of the corresponding device in the first layer of the graph convolutional neural network based on the aggregated information received by each of the devices, and update the embedding information of the device in the second layer of the graph convolutional neural network based on the embedding information of the device in the first layer of the graph convolutional neural network;

[0007] S3: According to the embedded information of each device in the second layer of the graph convolutional neural network, the loss function on each device and the gradient corresponding to the loss function are calculated, and the gradient is updated according to the gradient of the neighboring devices in the same area of ​​the device; the device updates the local model parameters based on the corresponding updated gradient.

[0008] Beneficial effects: This method solves the problems of information interaction privacy security and communication cost in the field of industrial intelligent robots, while improving the model performance and ensuring the stability of the system; this method not only optimizes the utilization of data, but also enhances the system's anti-attack capability and robustness, providing strong support for the efficient and safe operation of industrial intelligent robots. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0010] Figure 1 This is a flowchart of a decentralized graph federated learning method for integrating cross-regional device information in an embodiment of the present application.

[0011] Figure 2 Schematic diagram of the communication relationship between robot equipment in multiple factories in an embodiment of the present application. DETAILED DESCRIPTION

[0012] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present application, so the present application is not limited by the specific embodiments disclosed below.

[0013] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0014] like Figure 1 As shown, this embodiment provides a decentralized graph federated learning method for integrating cross-regional device information, the method comprising:

[0015] S1: Based on the communication relationship between devices in factories in the same area or different areas, a graph structure is constructed; and based on the graph structure, neighbor devices in the same area of ​​each device, as well as first-order neighbor devices and second-order neighbor devices across regions are found. The second-order neighbor devices are neighbor devices in the same area of ​​the first-order neighbor devices.

[0016] In this embodiment, the device is a robotic device.

[0017] Specifically, the graph structure is constructed by taking each of the devices as a node, the communication relationship between devices in the same area as a first edge, and the communication relationship between devices in different areas as a second edge;

[0018] Finding other devices with which each of the devices has a communication relationship in the same area according to the first edge as neighboring devices in the same area of ​​the device;

[0019] Other devices with which the devices have a communication relationship in different areas are found according to the second edge, as the first-order neighbor devices of the devices across areas.

[0020] Figure 2 The figure shows the communication relationship between the robot devices in factories A, B, and N, where the two robot devices connected by a solid line are neighbor devices in the same region, and the two robot devices connected by a dotted line are first-order neighbor devices across regions. Taking robot device B2 as an example, its neighbor device in the same region is B1, its first-order neighbor devices across regions are A2 and N1, and its second-order neighbor devices across regions are A1, An, and N2; the same-region neighbor devices, first-order neighbor devices across regions, and second-order neighbor devices of other robot devices in the figure are similar.

[0021] S2: For each of the devices, aggregate the feature information of its first-order neighbor devices and second-order neighbor devices, and obtain the embedding information of the corresponding device in the first layer of the graph convolutional neural network based on the aggregated information received by each of the devices, and update the embedding information of the device in the second layer of the graph convolutional neural network based on the embedding information of the device in the first layer of the graph convolutional neural network.

[0022] Furthermore, the aggregation process of feature information includes:

[0023] For any of the devices, the feature information of its first-order neighbor devices is preliminarily aggregated with the feature information of all second-order neighbor devices corresponding to its first-order neighbor devices to obtain preliminary aggregated information; the preliminary aggregation formula is:

[0024] ;

[0025] in, Indicates the device-based j Preliminary aggregation information of first-order neighbor devices; represents the activation function; Indicates the device’s j The set of second-order neighbor devices corresponding to the first-order neighbor devices; represents the gating function; Indicates jThe first-order neighbor device and its corresponding m The adjacency matrix of the second-order neighbor devices; Indicates j The first-order neighbor device corresponds to the m Characteristic information of second-order neighbor devices; Indicates j The characteristic information includes the relationship information between the node corresponding to the device and the first edge and / or the second edge, and the temperature, pressure, flow rate, and energy consumption of the device when working in the current communication round; the relationship information includes the edge weight of the first edge and / or the second edge, the connection relationship between the first edge and / or the second edge and the node, and the number of first edges and / or second edges connected to the node; the historical information includes the temperature, pressure, flow rate, and energy consumption of the device when working before the current communication round.

[0026] Edge weights are used to calculate the shortest path from the source point to all other points in the graph structure. The distance array is updated iteratively, the unvisited vertex with the smallest distance is selected, and then the distances of its adjacent vertices are updated. The formula is expressed as:

[0027] ;

[0028] in, Represents a vertex v The shortest path value of Represents a vertex u The shortest path value of u With Vertex v The weight of the edge (u,v) between them.

[0029] In this embodiment, the gating function The sigmoid function is used to control the output of information. The sigmoid function processes the input information and outputs a value between 0 and 1. It is used to control the contribution of the feature information of the first-order neighbor device in the aggregation process, which will help the model pay more attention to the feature information of the current task. And the activation function A softmax function is employed so that the output can be interpreted as the probability of a node being in each class.

[0030] All preliminary aggregated information of any device is uploaded to the repeater for secondary aggregation to obtain the aggregated information, and the aggregated information is sent to the corresponding device. The secondary aggregation formula is:

[0031] ;

[0032] in, Indicates that it is sent to i Aggregate information of individual devices; represents the activation function; Indicates the device’s j The set of second-order neighbor devices corresponding to the first-order neighbor devices; represents the gating function; Indicates j The first-order neighbor device and its corresponding m The adjacency matrix of the second-order neighbor devices; Indicates the device-based j Preliminary aggregation information of first-order neighbor devices; Indicates i Historical information of each device; Indicates j The first-order neighbor device corresponds to the m The characteristic information of the second-order neighbor devices.

[0033] Through two aggregations, the amount of data information is increased, which greatly guarantees the privacy and security of data information. In addition, since the characteristic information of all devices is only aggregated and sent once on the repeater, this avoids the large amount of communication costs of multiple devices uploading in multiple sending modes during conventional interactions, greatly ensuring the privacy and security of data.

[0034] Furthermore, the obtaining the embedding information of the corresponding device in the first layer of the graph convolutional neural network based on the aggregated information received by each of the devices includes:

[0035] Add the weight matrix of the first layer of the graph convolutional neural network in the area where the corresponding device is located to the calculation formula of the aggregation information, and calculate the embedding information of the first layer of the graph convolutional neural network of the corresponding device.

[0036] ;

[0037] in, Indicates i The embedding information of each device in the first layer of the graph convolutional neural network; represents the activation function; Indicates the device’s j The set of second-order neighbor devices corresponding to the first-order neighbor devices; represents the gating function; Indicates j The first-order neighbor device and its corresponding m The adjacency matrix of the second-order neighbor devices; Indicates the device-based j Preliminary aggregation information of first-order neighbor devices; Indicates i Historical information of each device; Indicates jThe first-order neighbor device corresponds to the m Characteristic information of second-order neighbor devices; Indicates i Device location The weight matrix of the first layer of the convolutional neural network.

[0038] Since traditional information embedding methods often perform poorly when processing sparse graph data, a graph convolutional neural network (GCN) is selected for information embedding in this embodiment. GCN can compensate for the sparsity of data by aggregating neighborhood information and improve the accuracy of embedding. In addition, the GCN structure is indirect and easy to expand and optimize, so that it can adapt to the embedding task of large-scale knowledge graphs, which is particularly important when processing large amounts of graph data. The local connection characteristics of GCN enable it to be processed in layers, and each layer can capture and contain different information. As the number of layers increases, the amount of information involved in the operation becomes richer. Therefore, after the first layer of GCN completes the information embedding, the subsequent layers can further utilize this information. This hierarchical information embedding method not only makes the entire network structure more modular, but also ensures that the work between each layer is independent of each other and does not interfere with each other, thereby improving the reliability of the system. Through this hierarchical strategy, GCN can effectively process complex graph structure data and provide an efficient and reliable solution for the embedding of large-scale graph data.

[0039] In this embodiment, updating the embedding information of the device in the second layer of the graph convolutional neural network based on the embedding information of the device in the first layer of the graph convolutional neural network includes:

[0040] The embedded information of the device in the first layer of the graph convolutional neural network is used as the initial state, and a long short-term memory network is used to update the embedded information of the device in the second layer of the graph convolutional neural network based on the feature information of the device.

[0041] Given that GCN performs well in processing static graph structures, but has certain limitations in processing dynamic graphs, this is mainly because the structure of dynamic graphs changes with time. Therefore, in this embodiment, a long short-term memory network (LSTM) is used to update the information embedding of the second layer of GCN. LSTM is an ideal choice for processing dynamic graphs because it can process sequence data and capture temporal features. At the same time, GCN is good at capturing the local structural properties of nodes and their relationships. Therefore, combining GCN with LSTM can more comprehensively understand and predict changes in the network, while effectively processing structural information that changes over time in dynamic graphs; in addition, the model combining GCN and LSTM is an end-to-end model, which means that it can simultaneously learn the temporal features and local structural properties of dynamic networks. This combination not only enhances the theoretical basis of the model, but also shows higher practicality and effectiveness in practical applications. Through this innovative combination, this embodiment enables GCN to better adapt to the complexity of dynamic networks and provides a powerful tool for network analysis and prediction. At the same time, due to the hierarchical characteristics of GCN, if the subsequent amount of information is too large, the hierarchical information embedding of GCN can continue, and LSTM can continue to be used to update the embedded information, which is crucial for the subsequent expansion and improvement of the system.

[0042] Furthermore, the process of updating using the long short-term memory network includes:

[0043] The embedded information of the device in the first layer of the graph convolutional neural network and the feature information of the device are input into the input gate of the long short-term memory network to obtain the first output and candidate state. The calculation formula is:

[0044] ;

[0045] ;

[0046] in, Indicates that the sigmoid layer in the input gate is t The first output at time +1; Represents the sigmoid activation function; Represents the weight matrix of the sigmoid layer in the input gate; Indicates i The embedding information of each device in the first layer of the graph convolutional neural network; Indicates t +1 moment i Characteristic information of each device; Represents the bias vector of the sigmoid layer in the input gate; Represents the tanh layer in the input gate t Candidate state of the cell at time +1; represents the hyperbolic tangent activation function; Represents the weight matrix of the tanh layer in the input gate; Represents the bias vector of the tanh layer in the input gate;

[0047] The embedded information of the device in the first layer of the graph convolutional neural network and the feature information of the device are input into the forget gate of the long short-term memory network to obtain the second output, which is calculated as follows:

[0048] ;

[0049] in, Indicates that the forget gate is t The second output at time +1; Represents the weight matrix of the forget gate; Represents the bias vector of the forget gate;

[0050] The embedded information of the device in the first layer of the graph convolutional neural network, the characteristic information of the device, the first output, the second output and the candidate state are input to the output gate, the cell state is updated and the third output is obtained. The calculation formula is:

[0051] ;

[0052] ;

[0053] ;

[0054] in, Represents the long short-term memory network in t The third output at time +1; Indicates that the output gate is t Output at the moment; Indicates that the cells t +1 The updated status at the moment; Indicates that the cells t The updated status at all times; Indicates that the output gate is t Output at +1 moment; Represents the weight matrix of the output gate; represents the bias vector of the output gate;

[0055] The long short-term memory network is t The third output at time +1 is the i The embedded information of a device in the second layer of the graph convolutional neural network is recorded as .

[0056] S3: According to the embedded information of each device in the second layer of the graph convolutional neural network, the loss function on each device and the gradient corresponding to the loss function are calculated, and the gradient is updated according to the gradient of the neighboring devices in the same area of ​​the device; the device updates the local model parameters based on the corresponding updated gradient.

[0057] Specifically, the calculation process of updating the gradient includes:

[0058] Step 1: Each of the devices performs model training based on the embedded information of the second layer of the graph convolutional neural network to obtain global model parameters; based on the global model parameters and the data samples on each device, the loss function on each device is calculated respectively, and the calculation formula is:

[0059] ;

[0060] in, Indicates i The loss function on each device; represents the global model parameters; Indicates i The number of data samples per device; Indicates i On the device v Loss functions for data samples, including cross entropy loss; Indicates i On the device v The predicted value of a data sample; Indicates i On the device v data samples; Indicates i On the device v The real label of a data sample; the data sample includes the temperature, pressure, flow rate, and energy consumption of the device when it is working in the current communication round;

[0061] Step 2: Based on k The loss function calculation on each device in the round of communication k The gradient corresponding to the loss function on each device in round communication is calculated as:

[0062] ;

[0063] in, Indicates k Round 1 i The gradient of the loss function on each device; represents partial derivative; Indicates k Round 1 i The loss function on each device; Indicates i Local model parameters for each device;

[0064] Step 3: According to k The gradient calculations corresponding to the first k The confidence of the device in the round of communication to its neighboring devices in the same area is calculated as:

[0065] ;

[0066] in, Indicates k Round 1 i Each device r The confidence level of neighboring devices in the same area; represents the standardized DTW function; Indicates k Round 1 i Device r The gradient corresponding to the loss function on neighboring devices in the same area;

[0067] Step 4: Based on k In the round of communication, the device calculates the confidence of all its neighboring devices in the same area. k The average confidence of a device in round-by-round communication with its neighboring devices in the same area is calculated as:

[0068] ;

[0069] in, Indicates k The average confidence level of a device in its neighboring devices in the same area during round-robin communication; Indicates k Round 1 i The number of neighboring devices in the same area as the device;

[0070] Step 5: In k In the round of communication, i The confidence level of each neighboring device in the same area of ​​the device is i The average confidence comparison corresponding to the devices is:

[0071] When confidence Greater than or equal to the average confidence When i Device r Neighboring devices in the same area join the i A sample list of devices;

[0072] Otherwise, the i Device r Neighboring devices in the same area give up;

[0073] Through this screening process, neighbor devices in the same area with gradient similarity (low confidence) can be proposed, which ensures the aggregation of neighbor devices in the same area with high data similarity and reduces the waste of communication resources.

[0074] Step 6: k In the round of communication, based on i The gradient corresponding to the loss function of each neighboring device in the same area in the sample list of the device i The adaptive gradient weight allocation is performed on each device, and the calculation formula is:

[0075] ;

[0076] in, Indicates k Round 1 i Gradient weights for each device; Indicates the module; n Indicates i The number of neighboring devices in the same area in the sample list of each device; Indicates k Round 1 i The sample list of devices j The gradient corresponding to the loss function on neighboring devices in the same area;

[0077] The gradient size is used to adaptively adjust the gradient weight so that smaller gradients can obtain larger gradient weights to avoid being overwhelmed by the dominant gradient.

[0078] Step 7: Based on k Round 1 i The gradient weight of the device, the gradient and gradient weight corresponding to each neighboring device in the same area in the sample list, and update the i The gradient corresponding to the device is obtained k+ In round 1 of communication i The update gradient corresponding to each device is calculated as:

[0079] ;

[0080] in, Indicates k+ In round 1 of communication i The update gradient corresponding to each device; Indicates k Round 1 i The sample list of devices j The gradient weights of neighboring devices in the same area.

[0081] Different from other classic federated learning algorithms, after each round of communication, the devices aggregate gradients instead of model parameters, which can reduce communication costs while still ensuring the progress of the training process.

[0082] Furthermore, the local model parameter update formula of the device is:

[0083] ;

[0084] in, Indicates k +1 round of communication i Local model parameters for each device; Indicates k Round 1 i Local model parameters for each device; represents the learning rate; Indicates k+ In round 1 of communication i The update gradient corresponding to each device.

[0085] By selecting neighbor devices with higher confidence from neighbor devices in the same area of ​​the device for aggregation, local aggregation of devices in the area is achieved, effectively solving the problems caused by the centralized topology structure.

[0086] The decentralized graph federated learning method provided in this example, which integrates cross-regional device information, aims to improve system stability, solve information island problems caused by laws and regulations or data privacy, improve model performance, and reduce communication overhead. This method provides a new solution for protecting data privacy and improving efficiency by optimizing distributed data processing and model training.

[0087] Specifically, this method enables devices with cross-regional communication relationships to obtain characteristic information of devices in neighboring areas through the interaction of cross-regional device information. In actual factories, this characteristic information is generally the temperature, pressure, flow, energy consumption and other data of robot equipment. When each device obtains these data of neighboring devices, it can adjust the parameters in its own training process according to these data to optimize energy consumption, efficiency and so on. In this process, the privacy and security of data are effectively guaranteed by the mode of 2 aggregations and 1 distribution, and the communication cost is effectively reduced compared with the traditional interaction mode. When the device obtains the characteristic information of these neighboring devices and uses these data information for training in the local area, the amount of training sample data can be greatly increased, thereby improving the accuracy of the final training of the model. In the local training stage, a decentralized approach is first adopted to ensure the stability of the system, effectively solving the single point failure of the system, and according to the choice of confidence, devices with low gradient similarity (low confidence) are effectively eliminated, so as to ensure that the aggregated neighboring devices in the same area are devices with high data similarity, while reducing the waste of communication resources. Finally, the gradient aggregation method is adopted to update the model parameters, and the gradient aggregation method reduces the communication cost to a certain extent. Therefore, this method not only enables the device to obtain more cross-regional data information while ensuring the stability of the system, but also improves data similarity through confidence selection and gradient aggregation, and reduces communication costs.

[0088] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0089] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be construed as limiting the scope of the patent application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent application shall be subject to the attached claims.

Claims

1. A decentralized graph federated learning method that integrates cross-regional device information, characterized in that: include: S1: Build a graph structure based on the communication relationship between devices in factories in the same or different areas; And based on the graph structure, find out the neighbor devices in the same area of ​​each device, as well as the first-order neighbor devices and second-order neighbor devices across regions. The second-order neighbor devices are the neighbor devices in the same area of ​​the first-order neighbor devices; S2: for each of the devices, aggregate feature information of its first-order neighbor devices and second-order neighbor devices, and obtain embedding information of the corresponding device in the first layer of the graph convolutional neural network based on the aggregated information received by each of the devices, and update the embedding information of the device in the second layer of the graph convolutional neural network based on the embedding information of the device in the first layer of the graph convolutional neural network; The aggregation process of feature information includes: For any of the devices, preliminarily aggregate the characteristic information of its first-order neighbor devices and the characteristic information of all second-order neighbor devices corresponding to its first-order neighbor devices to obtain preliminary aggregated information; the characteristic information includes the relationship information between the node corresponding to the device and the first edge and / or the second edge, and the temperature, pressure, flow, and energy consumption of the device when working in the current communication round; the relationship information includes the edge weight of the first edge and / or the second edge, the connection relationship between the first edge and / or the second edge and the node, and the number of the first edge and / or the second edge connected to the node; Upload all preliminary aggregated information of any device to the repeater for secondary aggregation to obtain the aggregated information, and send the aggregated information to the corresponding device; S3: According to the embedded information of each device in the second layer of the graph convolutional neural network, the loss function on each device and the gradient corresponding to the loss function are calculated, and the gradient is updated according to the gradient of the neighboring devices in the same area of ​​the device; the device updates the local model parameters based on the corresponding updated gradient.

2. The decentralized graph federated learning method for integrating cross-regional device information according to claim 1 is characterized in that: S1 includes: The graph structure is constructed by taking each of the devices as a node, taking the communication relationship between devices in the same area as a first edge, and taking the communication relationship between devices in different areas as a second edge; Finding other devices with which each of the devices has a communication relationship in the same area according to the first edge as neighboring devices in the same area of ​​the device; Other devices with which the devices have a communication relationship in different areas are found according to the second edge, as the first-order neighbor devices of the devices across areas.

3. The decentralized graph federated learning method for integrating cross-regional device information according to claim 1 is characterized in that: The preliminary aggregation formula is: ； in, Indicates the device-based j Preliminary aggregation information of first-order neighbor devices; represents the activation function; Indicates the device’s j The set of second-order neighbor devices corresponding to the first-order neighbor devices; represents the gating function; represents the j The first-order neighbor device and its corresponding m The adjacency matrix of the second-order neighbor devices; Indicates j The first-order neighbor device corresponds to the m Characteristic information of second-order neighbor devices; Indicates j The historical information of the first-order neighbor devices includes the temperature, pressure, flow rate, and energy consumption of the devices when they were working before the current communication round.

4. The decentralized graph federated learning method for integrating cross-regional device information according to claim 3 is characterized in that: The secondary aggregation formula is: ； in, Indicates that it is sent to i Aggregate information of individual devices; represents the activation function; Indicates the device’s j The set of second-order neighbor devices corresponding to the first-order neighbor devices; represents the gating function; Indicates j The first-order neighbor device and its corresponding m The adjacency matrix of the second-order neighbor devices; Indicates the device-based j Preliminary aggregation information of first-order neighbor devices; Indicates i Historical information of each device; Indicates j The first-order neighbor device corresponds to the m The characteristic information of the second-order neighbor devices.

5. The decentralized graph federated learning method for integrating cross-regional device information according to claim 4 is characterized in that: The step of obtaining the embedding information of the corresponding device in the first layer of the graph convolutional neural network based on the aggregated information received by each of the devices includes: Add the weight matrix of the first layer of the graph convolutional neural network in the area where the corresponding device is located to the calculation formula of the aggregation information, and calculate the embedding information of the first layer of the graph convolutional neural network of the corresponding device. ； in, Indicates i The embedding information of each device in the first layer of the graph convolutional neural network; represents the activation function; Indicates the device’s j The set of second-order neighbor devices corresponding to the first-order neighbor devices; represents the gating function; Indicates j The first-order neighbor device and its corresponding m The adjacency matrix of the second-order neighbor devices; Indicates the device-based j Preliminary aggregation information of first-order neighbor devices; Indicates i Historical information of each device; Indicates j The first-order neighbor device corresponds to the m Characteristic information of second-order neighbor devices; Indicates i Device location The weight matrix of the first layer of the convolutional neural network.

6. The decentralized graph federated learning method for integrating cross-regional device information according to claim 1 is characterized in that: The updating of the embedding information of the device at the second layer of the graph convolutional neural network based on the embedding information of the device at the first layer of the graph convolutional neural network includes: The embedded information of the device in the first layer of the graph convolutional neural network is used as the initial state, and a long short-term memory network is used to update the embedded information of the device in the second layer of the graph convolutional neural network based on the feature information of the device.

7. The decentralized graph federated learning method for integrating cross-regional device information according to claim 6 is characterized in that: The process of updating using the LSTM network includes: The embedded information of the device in the first layer of the graph convolutional neural network and the feature information of the device are input into the input gate of the long short-term memory network to obtain the first output and candidate state. The calculation formula is: ； ； in, Indicates that the sigmoid layer in the input gate is t The first output at time +1; Represents the sigmoid activation function; Represents the weight matrix of the sigmoid layer in the input gate; Indicates i The embedding information of each device in the first layer of the graph convolutional neural network; Indicates t +1 moment i Characteristic information of each device; Represents the bias vector of the sigmoid layer in the input gate; Indicates the tanh layer in the input gate t Candidate state of the cell at time +1; represents the hyperbolic tangent activation function; Represents the weight matrix of the tanh layer in the input gate; Represents the bias vector of the tanh layer in the input gate; The embedded information of the device in the first layer of the graph convolutional neural network and the feature information of the device are input into the forget gate of the long short-term memory network to obtain the second output, which is calculated as follows: ； in, Indicates that the forget gate is t The second output at time +1; Represents the weight matrix of the forget gate; Represents the bias vector of the forget gate; The embedded information of the device in the first layer of the graph convolutional neural network, the characteristic information of the device, the first output, the second output and the candidate state are input to the output gate, the cell state is updated and the third output is obtained. The calculation formula is: ； ； ； in, Represents the long short-term memory network in t The third output at time +1; Indicates that the output gate is t Output at the moment; Indicates that the cells t +1 The updated status at the moment; Indicates that the cells t The updated status at all times; Indicates that the output gate is t Output at +1 moment; Represents the weight matrix of the output gate; represents the bias vector of the output gate; The long short-term memory network is t The third output at time +1 is the i The embedded information of a device in the second layer of the graph convolutional neural network is recorded as .

8. The decentralized graph federated learning method for integrating cross-regional device information according to claim 1 is characterized in that: In S3, the calculation process of updating the gradient includes: Step 1: Each of the devices performs model training based on the embedded information of the second layer of the graph convolutional neural network to obtain global model parameters; based on the global model parameters and the data samples on each device, the loss function on each device is calculated respectively, and the calculation formula is: ； in, Indicates i The loss function on each device; represents the global model parameters; Indicates i The number of data samples per device; Indicates i On the device v Loss functions for data samples, including cross entropy loss; Indicates i On the device v The predicted value of a data sample; Indicates i On the device v data samples; Indicates i On the device v The real label of a data sample; the data sample includes the temperature, pressure, flow rate, and energy consumption of the device when it is working in the current communication round; Step 2: Based on k The loss function calculation on each device in the round of communication k The gradient corresponding to the loss function on each device in round communication is calculated as: ； in, Indicates k Round 1 i The gradient of the loss function on each device; represents partial derivative; Indicates k Round 1 i The loss function on each device; Indicates i Local model parameters for each device; Step 3: According to k The gradient calculations corresponding to the first k The confidence of the device in the round of communication to its neighboring devices in the same area is calculated as: ； in, Indicates k Round 1 i Each device r The confidence level of neighboring devices in the same area; represents the standardized DTW function; Indicates k Round 1 i Device r The gradient corresponding to the loss function on neighboring devices in the same area; Step 4: Based on k In the round of communication, the device calculates the confidence of all its neighboring devices in the same area. k The average confidence of a device in round-by-round communication with its neighboring devices in the same area is calculated as: ； in, Indicates k The average confidence level of a device in its neighboring devices in the same area during round-robin communication; Indicates k Round 1 i The number of neighboring devices in the same area as the device; Step 5: In k In the round of communication, i The confidence level of each neighboring device in the same area of ​​the device is i The average confidence comparison corresponding to the devices is: When confidence Greater than or equal to the average confidence When i Device r Neighboring devices in the same area join the i A sample list of devices; Otherwise, the i Device r Neighboring devices in the same area give up; Step 6: k In the round of communication, based on i The gradient of the loss function on each neighboring device in the same area in the sample list of the device is i The adaptive gradient weight allocation is performed on each device, and the calculation formula is: ； in, Indicates k Round 1 i Gradient weights for each device; Indicates the module; n Indicates i The number of neighboring devices in the same area in the sample list of each device; Indicates k Round 1 i The sample list of devices j The gradient corresponding to the loss function on neighboring devices in the same area; Step 7: Based on k Round 1 i The gradient weight of the device, the gradient and gradient weight corresponding to each neighboring device in the same area in the sample list, and update the i The gradient corresponding to the device is obtained k+ In round 1 of communication i The update gradient corresponding to each device is calculated as: ； in, Indicates k+ In round 1 of communication i The update gradient corresponding to each device; Indicates k Round 1 i The sample list of devices j The gradient weights of neighboring devices in the same area.

9. The decentralized graph federated learning method for integrating cross-regional device information according to claim 8, characterized in that: The local model parameter update formula of the device is: ； in, Indicates k +1 round of communication i Local model parameters for each device; Indicates k Round 1 i Local model parameters for each device; represents the learning rate; Indicates k+ In round 1 of communication i The update gradient corresponding to each device.

Images