Adaptive learning method and system for dynamic environment of Internet of Things
By building a deep learning model in a graph neural network and using iterative training and mixed memory sets, the performance loss problem in dynamic channel environment is solved, and the communication capabilities of IoT devices and the adaptability of the model is achieved.
Patent Information
- Application Number
- CN202411256284.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-09
- Publication Date
- 2025-01-03
AI Technical Summary
The existing graph neural network-based solutions are difficult to adapt to changes in dynamic channel environments, resulting in performance losses of deep learning models and cannot meet the real-time device communication needs in the Internet of Things environment.
The graph neural network is used to build a deep learning model, and through iterative training and mixed memory sets, the minimum mean square error function is used for supervised learning, and the beam design method in a dynamic channel environment is constructed, the graph data set training mapping pair is optimized, and the predicted power allocation vector is generated. , and update the memory set through the performance function to calculate the negative ratio of rate to power to achieve the adaptability of the model.
It effectively overcomes the phenomenon of catastrophic forgetting, improves the adaptability and communication performance of the model in a dynamic environment, realizes continuous learning and optimization, and improves the communication capabilities of IoT devices.
Smart Images
Figure CN119250111A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of machine learning and mobile communication, and in particular relates to an adaptive learning method and system for a dynamic environment of the Internet of Things. Background Art
[0002] With the continuous development of deep learning, its ability to solve device communication problems in the IoT environment has attracted more and more attention from relevant researchers. Traditional iterative optimization algorithms often bring high computational complexity and are difficult to meet the real-time requirements of device interaction in the IoT environment. At the same time, deep learning solutions based on deep neural networks perform poorly when a large number of devices are connected and lack generalization capabilities. As an emerging deep learning framework, graph neural networks have better learning capabilities and generalization performance than deep neural networks. At the same time, they effectively utilize the underlying topology of massive connected devices in the IoT environment to provide more useful information for the model. However, existing solutions based on graph neural networks follow the assumption that test samples and training samples follow the same distribution. When the model adapts to the new distribution, it will cause catastrophic forgetting of the learned distribution and suffer great performance loss.
[0003] Therefore, the existing solutions cannot adapt to the dynamically changing channel environment. Therefore, it has become a major challenge to study the beam design method in a dynamic channel environment to ensure that the deep learning model can continuously learn and optimize wireless resources to achieve the performance optimization goal. Summary of the invention
[0004] In view of the problems existing in the prior art, the present invention provides an adaptive learning method and system for a dynamic environment of the Internet of Things, which can adapt to a dynamically changing channel environment and improve the communication capability of devices in the Internet of Things environment.
[0005] To solve the above technical problems, the present invention provides the following technical solutions: an adaptive learning method for a dynamic environment of the Internet of Things, comprising the following steps:
[0006] S1. Obtain the channel data of time slot t, use the iterative algorithm to generate the power vector label corresponding to the channel data, and construct the training mapping pair D of time slot t. t ;
[0007] S2, according to the training mapping of time slot t to D t , and the power vector label value corresponding to the channel data to construct the graph data set training mapping pair G of the current time slot t t ;
[0008] S3, based on the known memory graph dataset M t , the graph dataset training mapping pair G t and the memory graph dataset M tAfter stacking, randomly shuffle them to form a graph dataset training mapping pair set M t ∪G t , use the graph dataset to train the mapping pair set M t ∪G t The channel data of each sample in is taken as input, and the power vector label value corresponding to each channel data in the training graph dataset mapping pair set is taken as output to construct and iteratively train the graph neural network; at the same time, the minimum mean square error function is used as the loss function for supervised learning during the iterative training process until all the graph dataset training mapping pair sets M are traversed t ∪G t , obtain a graph neural network model and the optimal weight θ of the graph neural network model;
[0009] S4. Train the graph dataset to map the set M t ∪G t The channel information of each sample in is input into the graph neural network model obtained in step S3 to obtain a predicted power allocation vector;
[0010] S5, based on the predicted power allocation vector obtained in step S4, and the graph data set training mapping pair set M t ∪G t , using the performance function to calculate the predicted power allocation vector and rate and the map dataset training mapping pair set M t ∪G t The optimal and negative ratio of the rate, according to which M t , G t Select the mapping pairs that meet the preset conditions, stack and shuffle the mapping pairs to obtain a new memory graph dataset M t+1 , and the new memory graph dataset M t+1 Updated to the memory graph dataset M t , then return to step S1.
[0011] Furthermore, in the aforementioned step S2, the graph dataset training mapping pair G t Follow the steps below to obtain:
[0012] S2.1. Map the training of time slot t to D t Take the channel matrix H n , H n As the point features and edge features corresponding to the nth graph dataset; where n is the sample index;
[0013] S2.2. Take the corresponding power vector label value p n As the hidden data corresponding to the nth graph dataset, after encapsulation of the DGL graph class, the graph dataset training mapping pair G is obtained. t ; Graph dataset training mapping pair G tEach device in the IoT environment is regarded as a node, and the connection relationship between devices is regarded as an edge. Each node and each edge contains feature information.
[0014] Furthermore, in the aforementioned step S3, the graph neural network includes an input layer and two graph neural network layers sequentially connected thereto, a flattening layer, a fully connected layer, an activation function layer, and a loss function layer.
[0015] Furthermore, in the aforementioned step S3, the graph neural network is used as the underlying learning framework during iterative training of the graph neural network, and the graph dataset training mapping pair G t As the input of the graph neural network, the model first extracts the point features of each node as the hidden state Initialization, where v represents a node; at the same time, the edge features of the corresponding edges are extracted, and messages are formed and aggregated according to the following formula:
[0016]
[0017] Where v1,v2∈V, V is a set of nodes, N(v1) is the set of neighbors of node v1, is the hidden state of node v1, is the hidden state of node v2, l-1 represents the number of layers, is the edge feature of the edge formed by node v1 and node v2, MLP is the fully connected layer, and θ is the set of trainable parameters.
[0018] After each node aggregates its own information with the information of its neighboring nodes, the hidden state of each node is updated according to the following formula:
[0019]
[0020] Where mean represents the mean function.
[0021] Furthermore, in the aforementioned step S3, when iteratively training the graph neural network model, the maximum number of iterations is set to M, the number of graph neural network layers is L, the optimizer is Adam, the learning rate is lr, and the graph data set training mapping pair set M is t ∪G t Divide into several batches, iterate each batch, use the gradient descent method to update the model weights in the iterative process, optimize the global loss function, and train the mapping pair set M for the graph dataset t ∪G t After the traversal is completed, the training process of the graph neural network model ends, the model converges, the global loss function reaches the minimum value, the model weight Θ reaches the optimal value, and the final hidden state matrix X(L) is obtained after the flattening layer and the fully connected layer. As the prediction output of the graph neural network model.
[0022] Furthermore, in the aforementioned step S5, the predicted power allocation vector and rate are calculated using the performance function and the training mapping pair set M of the graph data set. t ∪G t The optimal negative ratio of the sum rate is as follows:
[0023]
[0024] Where Θ is the system weight, H (n) is the channel sample data, p (n) is the power allocation vector that needs to be learned by the graph neural network model, is the power allocation vector predicted by the graph neural network model, is the sum rate of the predicted output, is the set of graph dataset training mapping pairs M obtained by weighted minimum mean square error algorithm t ∪G t Optimal sum rate, where n is the sample index.
[0025] Furthermore, in the aforementioned step S5, according to the negative ratio, t , G t Selecting a mapping pair that meets the preset conditions specifically includes the following sub-steps:
[0026] S5.1. For the sample data M calculated by the performance function t and G t The graph dataset mappings in G are sorted in descending order, starting from t Select the first Nf(t) graph dataset sample mapping pairs from M t Select the first f(t) graph dataset sample mapping pairs, where f(t) is a function of time slot t;
[0027] S5.2, G t Select the first Nf(t) graph dataset sample mapping pairs and select from M t Select the first f(t) sample mapping pairs of the graph dataset and shuffle them.
[0028] The present invention also provides an adaptive learning system for the dynamic environment of the Internet of Things, including: a data stream module, a memory set module, a graph data set construction module, a hybrid module, and a graph neural network module;
[0029] The data stream module is used to obtain the channel data of time slot t, use the iterative algorithm to generate the power vector label corresponding to the channel data, and construct the training mapping pair D of time slot t. t ;
[0030] Memory set module, used to store graph dataset sample data M t ;
[0031] The graph dataset construction module is used to map D according to the training map of time slot t. t , and the power vector label value corresponding to the channel data to construct the graph data set training mapping pair G of the current time slot t t ;
[0032] A hybrid module for leveraging known memory graph datasets M t , the graph dataset training mapping pair G t and the memory graph dataset M t After stacking, randomly shuffle them to form a graph dataset training mapping pair set M t ∪G t ;
[0033] Graph neural network module, used to train the mapping pair set M with a graph dataset t ∪G t The channel data in the graph are taken as input, and the power vector label value corresponding to each channel data in the set is mapped as output for model training to obtain the power vector prediction value corresponding to each channel data.
[0034] Compared with the prior art, the beneficial technical effects of the present invention using the above technical solution are as follows:
[0035] (1) The present invention uses graph neural networks to construct a deep learning model, effectively utilizing the permutation equivariance characteristics of graph neural networks, deeply mining the underlying topological structure between devices in the Internet of Things environment, and improving the actual communication performance of the model.
[0036] (2) The present invention constructs a memory set during the training process of the graph neural network model, and trains the model by mixing the memory set and the current sample set, so that the model overcomes the catastrophic forgetting phenomenon and shows good adaptability in the dynamically changing Internet of Things environment.
[0037] (3) The present invention defines the performance loss function by using the negative ratio of the sum rate, and uses the performance loss function as a criterion for selecting memory set samples, thereby effectively embedding the problem structure into the model training process. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 This is a model structure diagram based on graph neural network.
[0039] Figure 2 This is a flowchart of a model training method based on graph neural network in the Internet of Things environment. DETAILED DESCRIPTION
[0040] In order to better understand the technical content of the present invention, specific embodiments are given and described as follows in conjunction with the accompanying drawings.
[0041] Various aspects of the invention are described herein with reference to the accompanying drawings, in which many illustrative embodiments are shown. The embodiments of the invention are not limited to those described in the accompanying drawings. It should be understood that the invention is implemented by any of the various concepts and embodiments described above, as well as the concepts and embodiments described in detail below, because the concepts and embodiments disclosed in the invention are not limited to any implementation. In addition, some aspects disclosed in the invention may be used alone or in any appropriate combination with other aspects disclosed in the invention.
[0042] refer to Figure 2 The present invention provides an adaptive learning method for a dynamic environment of the Internet of Things, comprising the following steps:
[0043] S1. Obtain the channel data of time slot t, use the iterative algorithm to generate the power vector label corresponding to the channel data, and construct the training mapping pair D of time slot t. t ;
[0044] S2, according to the training mapping of time slot t to D t , and the power vector label value corresponding to the channel data to construct the graph data set training mapping pair G of the current time slot t t ; Graph dataset training mapping pair G t Follow the steps below to obtain:
[0045] S2.1. Map the training of time slot t to D t Take the channel matrix H n , H n As the point features and edge features corresponding to the nth graph dataset; where n is the sample index;
[0046] S2.2. Take the corresponding power vector label value p n As the hidden data corresponding to the nth graph dataset, after encapsulation of the DGL graph class, the graph dataset training mapping pair G is obtained. t ; Graph dataset training mapping pair G t Each device in the IoT environment is regarded as a node, and the connection relationship between devices is regarded as an edge. Each node and each edge contains feature information.
[0047] S3, based on the known memory graph dataset M t , the graph dataset training mapping pair G t and the memory graph dataset M t After stacking, randomly shuffle them to form a graph dataset training mapping pair set M t ∪G t , use the graph dataset to train the mapping pair set M t ∪G tThe channel data of each sample in is taken as input, and the power vector label value corresponding to each channel data in the training graph dataset mapping pair set is taken as output to construct and iteratively train the graph neural network; at the same time, the minimum mean square error function is used as the loss function for supervised learning during the iterative training process until all the graph dataset training mapping pair sets M are traversed t ∪G t , obtain the graph neural network model and the optimal weight Θ of the graph neural network model.
[0048] Among them, the graph neural network includes an input layer and two graph neural network layers sequentially connected to it, a flattening layer, a fully connected layer, an activation function layer, and a loss function layer.
[0049] In the first step, set the maximum number of iterations to M, the number of graph neural network layers to L, the optimizer to Adam, the learning rate to lr, and map the graph dataset training to the set M t ∪G t Divide into several batches, iterate each batch, and use the gradient descent method to update the model weights and optimize the global loss function during the iteration process. When training the graph neural network model, the graph neural network is used as the underlying learning framework. The graph dataset training mapping is used for G t As the input of the graph neural network, the model first extracts the point features of each node as the hidden state Initialization, where v represents a node. At the same time, the edge features of the corresponding edges are extracted, and messages are formed and aggregated according to the following formula:
[0050]
[0051] Where v1,v2∈V, V is a set of nodes, N(v1) is the set of neighbors of node v1, is the hidden state of node v1, l-1 represents the number of layers, is the edge feature of the edge formed by node v1 and node v2, MLP is the fully connected layer, and θ is the set of trainable parameters.
[0052] In the second step, after each node aggregates its own information with the information of neighboring nodes, the hidden state of each node is updated according to the following formula:
[0053]
[0054] Where mean represents the mean function.
[0055] The third step is to train the mapping pair set M on the graph dataset. t ∪G t After the traversal is completed, the training process of the graph neural network model ends, the model converges, the global loss function reaches the minimum value, the model weight Θ reaches the optimal value, and the final hidden state matrix X(L) After being processed by the flattening layer and the fully connected layer, we get As the prediction output of the graph neural network model.
[0056] S4. Train the graph dataset to map the set M t ∪G t The channel information of each sample in is input into the graph neural network model obtained in step S3 to obtain a predicted power allocation vector;
[0057] S5, based on the predicted power allocation vector obtained in step S4, and the graph data set training mapping pair set M t ∪G t , using the performance function to calculate the predicted power allocation vector and rate and the map dataset training mapping pair set M t ∪G t The optimal and negative ratio of the rate, according to which M t , G t Select the mapping pairs that meet the preset conditions, stack and shuffle the mapping pairs to obtain a new memory graph dataset M t+1 , and the new memory graph dataset M t+1 Updated to the memory graph dataset M t , then return to step S1.
[0058] Step S5 specifically includes: firstly, using the performance function to calculate the predicted power allocation vector and rate and the training mapping pair set M of the graph data set. t ∪G t The optimal negative ratio of the sum rate is as follows:
[0059]
[0060] Where Θ is the system weight, H (n) is the channel sample data, p (n) is the power allocation vector that needs to be learned by the graph neural network model, is the power allocation vector predicted by the graph neural network model, is the sum rate of the predicted output, is the set of graph dataset training mapping pairs M obtained by weighted minimum mean square error algorithm t ∪G t Optimal sum rate, where i is the sample index.
[0061] The second step is to calculate the M in the sample data after the performance function is calculated. t and G t The graph dataset mappings in G are sorted in descending order, starting from t Select the first Nf(t) graph dataset sample mapping pairs from M tSelect the first f(t) graph dataset sample mapping pairs, where f(t) is a function of time slot t;
[0062] Step 3: G t Select the first Nf(t) graph dataset sample mapping pairs and select from M t Select the first f(t) graph dataset sample mapping pairs from the stack and shuffle them to obtain a new memory graph dataset M t+1 , and the new memory graph dataset M t+1 Updated to the memory graph dataset M t , then return to step S1.
[0063] refer to Figure 1 ,The present invention also provides an adaptive learning system for the dynamic environment of the Internet of Things, including: a data stream module, a memory set module, a graph data set construction module, a hybrid module, and a graph neural network module;
[0064] The data stream module is used to obtain the channel data of time slot t, use the iterative algorithm to generate the power vector label corresponding to the channel data, and construct the training mapping pair D of time slot t. t ;
[0065] Memory set module, used to store graph dataset sample data M t ;
[0066] The graph dataset construction module is used to map D according to the training map of time slot t. t , and the power vector label value corresponding to the channel data to construct the graph data set training mapping pair G of the current time slot t t ;
[0067] A hybrid module for leveraging known memory graph datasets M t , the graph dataset training mapping pair G t and the memory graph dataset M t After stacking, randomly shuffle them to form a graph dataset training mapping pair set M t ∪G t ;
[0068] Graph neural network module, used to train the mapping pair set M with a graph dataset t ∪G t The channel data in the graph are taken as input, and the power vector label value corresponding to each channel data in the set is mapped as output for model training to obtain the power vector prediction value corresponding to each channel data.
[0069] Although the present invention has been described above with preferred embodiments, it is not intended to limit the present invention. A person skilled in the art of the present invention may make various modifications and improvements without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be determined by the definition of the claims.
Claims
1. An adaptive learning method for the dynamic environment of the Internet of Things, characterized in that: The following steps are involved: S1. Obtain the channel data of time slot t, use the iterative algorithm to generate the power vector label corresponding to the channel data, and construct the training mapping pair D of time slot t. t ; S2, according to the training mapping of time slot t to D t , and the power vector label value corresponding to the channel data to construct the graph data set training mapping pair G of the current time slot t t ; S3, based on the known memory graph dataset M t , the graph dataset training mapping pair G t and the memory graph dataset M t After stacking, randomly shuffle them to form a graph dataset training mapping pair set M t ∪G t , use the graph dataset to train the mapping pair set M t ∪G t The channel data of each sample in is taken as input, and the power vector label value corresponding to each channel data in the training graph dataset mapping pair set is taken as output to construct and iteratively train the graph neural network; at the same time, the minimum mean square error function is used as the loss function for supervised learning during the iterative training process until all the graph dataset training mapping pair sets M are traversed t ∪G t , obtain a graph neural network model and the optimal weight Θ of the graph neural network model; S4. Train the graph dataset to map the set M t ∪G t The channel information of each sample in is input into the graph neural network model obtained in step S3 to obtain a predicted power allocation vector; S5, based on the predicted power allocation vector obtained in step S4, and the graph data set training mapping pair set M t ∪G t , using the performance function to calculate the predicted power allocation vector and rate and the map dataset training mapping pair set M t ∪G t The optimal and negative ratio of the rate, according to which M t , G t Select the mapping pairs that meet the preset conditions, stack and shuffle the mapping pairs to obtain a new memory graph dataset M t+1 , and the new memory graph dataset M t+1 Updated to the memory graph dataset M t , then return to step S1.
2. The adaptive learning method for the dynamic environment of the Internet of Things according to claim 1 is characterized in that: In step S2, the graph dataset training mapping pair G t Follow the steps below to obtain: S2.
1. Map the training of time slot t to D t Take the channel matrix H n , H n As the point features and edge features corresponding to the nth graph dataset; where n is the sample index; S2.
2. Take the corresponding power vector label value p n As the hidden data corresponding to the nth graph dataset, after encapsulation of the DGL graph class, the graph dataset training mapping pair G is obtained. t ; Graph dataset training mapping pair G t Each device in the IoT environment is regarded as a node, and the connection relationship between devices is regarded as an edge. Each node and each edge contains feature information.
3. The adaptive learning method for the dynamic environment of the Internet of Things according to claim 1 is characterized in that: In step S3, the graph neural network includes an input layer and two graph neural network layers sequentially connected to it, a flattening layer, a fully connected layer, an activation function layer, and a loss function layer.
4. The adaptive learning method for the dynamic environment of the Internet of Things according to claim 1 is characterized in that: In step S3, the graph neural network is used as the underlying learning framework during iterative training of the graph neural network, and the graph dataset training mapping pair G t As the input of the graph neural network, the model first extracts the point features of each node as the hidden state Initialization, where v represents a node; at the same time, the edge features of the corresponding edges are extracted, and messages are formed and aggregated according to the following formula: Where v1,v2∈V, V is a set of nodes, N(v1) is the set of neighbors of node v1, is the hidden state of node v1, is the hidden state of node v2, l-1 represents the number of layers, is the edge feature of the edge formed by node v1 and node v2, MLP is the fully connected layer, and θ is the set of trainable parameters. After each node aggregates its own information with the information of its neighboring nodes, the hidden state of each node is updated according to the following formula: Where mean represents the mean function.
5. The adaptive learning method for the dynamic environment of the Internet of Things according to claim 1, characterized in that: In step S3, when iteratively training the graph neural network model, the maximum number of iterations is set to M, the number of graph neural network layers is L, the optimizer is Adam, the learning rate is lr, and the graph data set training mapping pair set M is t ∪G t Divide into several batches, iterate each batch, use the gradient descent method to update the model weights in the iterative process, optimize the global loss function, and train the mapping pair set M for the graph dataset t ∪G t After the traversal is completed, the training process of the graph neural network model ends, the model converges, the global loss function reaches the minimum value, the model weight Θ reaches the optimal value, and the final hidden state matrix X (L) After being processed by the flattening layer and the fully connected layer, we get As the prediction output of the graph neural network model.
6. The adaptive learning method for the dynamic environment of the Internet of Things according to claim 1, characterized in that: In step S5, the predicted power allocation vector and rate are calculated using the performance function and the training mapping pair set M of the graph data set. t ∪G t The optimal negative ratio of the sum rate is as follows: Where Θ is the system weight, H (n) is the channel sample data, p (n) is the power allocation vector that needs to be learned by the graph neural network model, is the power allocation vector predicted by the graph neural network model, is the sum rate of the predicted output, is the set of graph dataset training mapping pairs M obtained by weighted minimum mean square error algorithm t ∪G t Optimal sum rate, where n is the sample index.
7. The adaptive learning method for the dynamic environment of the Internet of Things according to claim 1, characterized in that: In step S5, according to the negative ratio, t , G t Selecting a mapping pair that meets the preset conditions specifically includes the following sub-steps: S5.
1. For the sample data M calculated by the performance function t and G t The graph dataset mappings in G are sorted in descending order, starting from t Select the first Nf(t) graph dataset sample mapping pairs from M t Select the first f(t) graph dataset sample mapping pairs, where f(t) is a function of time slot t; S5.2, G t Select the first Nf(t) graph dataset sample mapping pairs and select from M t Select the first f(t) sample mapping pairs of the graph dataset and shuffle them.
8. An adaptive learning system for the dynamic environment of the Internet of Things, characterized in that: include: Data flow module, memory set module, graph dataset construction module, hybrid module, graph neural network module; The data stream module is used to obtain the channel data of time slot t, use the iterative algorithm to generate the power vector label corresponding to the channel data, and construct the training mapping pair D of time slot t. t ; Memory set module, used to store graph dataset sample data M t ; The graph dataset construction module is used to map D according to the training map of time slot t. t , and the power vector label value corresponding to the channel data to construct the graph data set training mapping pair G of the current time slot t t ; A hybrid module for leveraging known memory graph datasets M t , the graph dataset training mapping pair G t and the memory graph dataset M t After stacking, randomly shuffle them to form a graph dataset training mapping pair set M t ∪G t ; Graph neural network module, used to train the mapping pair set M with a graph dataset t ∪G t The channel data in the graph are taken as input, and the power vector label value corresponding to each channel data in the set is mapped as output for model training to obtain the power vector prediction value corresponding to each channel data.