Position-guided channel knowledge graph construction method

Abstract

The invention, which belongs to the technical field of wireless communication, provides a position-guided channel knowledge graph construction method comprising the following steps: constructing an unmanned aerial vehicle auxiliary communication scene comprising a ground base station, and constructing a channel knowledge graph model; designing a back propagation neural network, training the back propagation neural network by using enough training samples corresponding to a ground base station j so as to learn a recessive nonlinear relationship between the position of the unmanned aerial vehicle and a channel gain, and representing a channel knowledge graph model as a dominant derivable function of the position of the unmanned aerial vehicle and the channel gain; and introducing a transfer learning algorithm, and performing fine tuning on the trained back propagation neural network of the ground base station j by using a limited sample corresponding to the target ground base station j 'to obtain a back propagation neural network of the target ground base station j', thereby obtaining a channel knowledge graph of the target ground base station j '. The method is a classical optimization method suitable for unmanned aerial vehicle trajectory optimization and deployment problems.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communication technology, and specifically relates to a method for constructing a position-derivative channel knowledge graph. Background Art

[0002] In order to expand the coverage of wireless networks and enhance the flexibility of deployment, integrating drones for various purposes into cellular networks has become an important technical approach to realize the 6G Internet of Things. In the existing technology, the air-to-ground channel is only described in an average sense. When the altitude of the UAV is not high enough, the existing channel model cannot cover the entire radio propagation environment, for example, due to blockage and shadows caused by buildings or other shelters. Therefore, in practical applications, the existing channel assumptions will reduce the communication performance of drones.

[0003] The emergence of Channel Knowledge Map (CKM) can overcome the above defects. By storing and utilizing information about the physical environment, CKM provides a location-corresponding database containing rich information about the channel (e.g., channel gain, interference, and angle information) for realizing wireless communications based on environment perception. The channel information in CKM is obtained through offline simulation or online / offline prior measurements. Constructing information at any location by partially simulating or measuring channel information can reduce the cost of CKM construction.

[0004] Existing work on CKM construction can be roughly divided into two categories, interpolation-based methods and machine learning-based methods. Interpolation-based methods directly use the channel information of adjacent locations to calculate the channel information of the target location, including inverse distance weighted method, K nearest neighbor method and Kriging. However, this method requires a large number of samples to accurately construct a map, which usually requires huge storage capacity. In addition, this purely data-driven method ignores the influence of environmental geometric prior information. Machine learning-based methods aim to capture the direct mapping relationship from any location to its output channel knowledge, including expectation maximization, maximum likelihood estimation, deep Gaussian process and deep learning. However, deep Gaussian process and deep learning methods treat the training process as a black box, making it difficult to obtain a clear functional relationship between input and output. Based on the channel statistical model, expectation maximization and maximum likelihood estimation methods usually divide the available measurement data into different groups, each of which shares the same channel model parameters. Since different channel models usually have different parameters, the constructed CKM is a piecewise function of the location and is not differentiable with respect to the location. In problems such as drone trajectory and deployment design, these CKMs cannot be applied to classical solution methods based on optimization theory. Therefore, it is necessary to develop a position-differentiable channel knowledge graph construction method. Summary of the invention

[0005] The purpose of the present invention is to propose a method for constructing a position-differentiable channel knowledge graph, which is position-differentiable and suitable for classical optimization methods in UAV trajectory optimization and deployment problems.

[0006] The present invention is achieved through the following technical solutions:

[0007] A method for constructing a position-differentiable channel knowledge graph includes the following steps:

[0008] Step S1: construct a UAV-assisted communication scenario including a ground base station, and construct a channel knowledge graph model, which provides a mapping from the UAV position to the corresponding channel gain;

[0009] Step S2, designing a back propagation neural network, using sufficient training samples corresponding to the ground base station j, training the back propagation neural network to learn the implicit nonlinear relationship between the UAV position and the channel gain, and representing the channel knowledge graph model as an explicit differentiable function of the UAV position and the channel gain;

[0010] Step S3, introduce the transfer learning algorithm, use the finite samples corresponding to the target ground base station j′, fine-tune the trained back propagation neural network of the ground base station j′, obtain the back propagation neural network of the target ground base station j′, and then obtain the channel knowledge graph of the target ground base station j′.

[0011] Furthermore, in step S1, the UAV-assisted communication scenario has an urban area divided into X×Y grids with equal spacing, and the ground base stations in the urban area are set as Indicates that the ground base station In the three-dimensional coordinate system, it is located at a fixed position (x j ,y j ,z j ),in, represents the horizontal position of the ground base station j, z j represents the height of the ground base station j. The UAV flies at a fixed height, and the horizontal position of the UAV is defined as q.

[0012] Furthermore, in step S1, in the UAV-assisted communication scenario, the ground base station j corresponds to the channel knowledge graph The channel knowledge graph is a function related only to the horizontal position of the UAV, and provides the channel gain corresponding to any given horizontal position q of the UAV at a constant height. 's mapping.

[0013] Furthermore, in step S2, the back propagation neural network includes an input layer, a nonlinear hidden layer, a linear hidden layer and an output layer, and the Tanh activation function is selected as the activation function of the nonlinear hidden layer. The back propagation neural network is trained given sufficient samples to learn the nonlinear hidden relationship between the drone position and the channel gain, and obtain the weight parameters and bias parameters of the back propagation neural network.

[0014] Furthermore, in step S2, the explicit differentiable function is expressed as in, is the output of the nonlinear hidden layer, and are the weight parameters and bias parameters of the nonlinear hidden layer, and are the weight parameters and bias parameters of the linear hidden layer, and are the weights and biases of the linear hidden layer neurons to the output layer, respectively.

[0015] Furthermore, in step S3, a transfer learning algorithm is used to use a source data set domain D S , a source learning task T S and a target dataset domain D T The knowledge in a target learning task T T The channel knowledge graph of each ground base station corresponds to the same environment, and the source dataset domain D S is the training dataset and test dataset of ground base station j, source learning task T S is a trained back propagation neural network, the target dataset domain D T is the training data set and test data set of the target ground base station j′, and the target learning task T T is the back propagation neural network of the ground base station j′, D S ≠D T , T S ≠T T , the training data set of the target ground base station j′ is a finite sample.

[0016] Furthermore, in step S3, the trained back propagation neural network is first tested using the test data set of the ground base station j, and then the tested back propagation neural network is refined using the training data set of the ground base station j′, the weight parameters and bias parameters of the nonlinear hidden layer are frozen, and then the back propagation neural network after retraining is tested using the test data set of the ground base station j′ to obtain an improved back propagation neural network.

[0017] The present invention has the following beneficial effects:

[0018] The present invention first constructs a UAV-assisted communication scenario and a channel knowledge graph model, then designs a back-propagation neural network, obtains a predicted channel knowledge graph after learning, and characterizes the predicted channel knowledge graph as an explicit differentiable function of the UAV position and the channel gain, and finally introduces a transfer learning algorithm to construct an improved channel knowledge graph under finite samples. The channel knowledge graph obtained in this way has high accuracy and is position-differentiable. When solving problems such as UAV trajectory optimization and deployment in real scenarios, it is suitable for classical optimization methods such as convex optimization theory; transfer learning is introduced in the back-propagation neural network, and the back-propagation neural networks at different ground base station positions share network parameters for common feature extraction, and train unknown network parameters based on finite samples, thereby avoiding measuring the channel information of each ground base station and a large number of sample positions, reducing the channel measurement cost and the network training cost, and at the same time enabling the network to have good generalization for the channel knowledge graphs corresponding to different ground base stations, and also avoiding the performance degradation caused by under-fitting of the back-propagation neural network under finite sample conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The present invention is further described in detail below with reference to the accompanying drawings.

[0020] Figure 1 It is a flow chart of the present invention.

[0021] Figure 2 This is a diagram of the drone-assisted communication scenario of the present invention.

[0022] Figure 3 A flowchart is constructed for channel knowledge graph based on back-propagation neural network under sufficient sample conditions.

[0023] Figure 4 Construct a flow chart for channel knowledge graph based on transfer learning under limited sample conditions.

[0024] Figure 5 It is the real channel knowledge graph of the ground base station 1 in the present invention.

[0025] Figure 6 It is the channel knowledge graph of ground base station 1 constructed based on the back propagation neural network method.

[0026] Figure 7 It is the channel knowledge graph of ground base station 1 constructed based on the K nearest neighbor interpolation method.

[0027] Figure 8 It is the channel knowledge graph of ground base station 1 constructed based on the Kriging method.

[0028] Fig. 9The error results between the channel knowledge graph constructed by the back propagation neural network method (BPNN), K nearest neighbor interpolation method (KNN), and Kriging method and the real channel knowledge graph under different samples.

[0029] Fig.10 It is the real channel knowledge graph of the ground base station 2 in the present invention.

[0030] Fig.11 It is the channel knowledge graph of the target ground base station 2 constructed based on the transfer learning method.

[0031] Fig.12 Channel knowledge graph of target ground base station 2 constructed for retraining back-propagation neural network. DETAILED DESCRIPTION

[0032] like Figure 1 As shown, the method for constructing a position-differentiable channel knowledge graph includes the following steps:

[0033] Step S1: construct a UAV-assisted communication scenario including a ground base station, and construct a channel knowledge graph model, which provides a mapping from the UAV position to the corresponding channel gain;

[0034] Specifically, Figure 2 As shown in the figure, in the UAV-assisted communication scenario, there is an urban area of ​​282m×282m, which is divided into X×Y=94×94 grids at 3m intervals. Consider a communication network of multiple groups of UAVs, where K UAV users send separate messages to their associated ground base stations (GBSs) on the same frequency band. Represents the set of ground base stations, ground base stations In the three-dimensional coordinate system, it is located at a fixed position (x j ,y j ,z j )=(93m,249m,3.24m), where: represents the horizontal position of the ground base station j, z j represents the height of the ground base station j, which is greater than or equal to 0. The UAV flies at a fixed height H = 50m. The horizontal position of the UAV is defined as

[0035] The location-specific channel knowledge graph (CKM) is used to store the location-specific wireless channel knowledge associated with the ground base station. In the UAV-assisted communication scenario, the ground base station j has a channel knowledge graph corresponding to the ground base station. The channel knowledge graph It is a function related only to the position of the UAV, providing the horizontal position of any given UAV at a constant height. The corresponding channel gain The mapping of If the channel information between any UAV and the ground base station in each grid is sampled, the cost is too high. CKM is actually constructed by interpolation technology based on measurement data. Therefore, the data stored in the constructed The channel gain in may be different from the true value Slightly different. For any GBSj, the constructed can be regarded as functions related only to the position q of the UAV k. Represents a real matrix of size a-by-b.

[0036] Step S2, designing a back propagation neural network, using sufficient training samples corresponding to the ground base station j, training the back propagation neural network to learn the implicit nonlinear relationship between the UAV position and the channel gain, and representing the channel knowledge graph model as an explicit differentiable function of the UAV position and the channel gain;

[0037] Specifically, the back-propagation neural network is Figure 2 As shown, it includes an input layer, a nonlinear hidden layer, a linear hidden layer and an output layer. The Tanh activation function is selected as the activation function of the nonlinear hidden layer. The number of neurons in the nonlinear hidden layer is M=40. Given sufficient samples, the back propagation neural network is trained to learn the nonlinear hidden relationship between the drone position and the channel gain, and the weight parameters and bias parameters of the back propagation neural network are obtained.

[0038] The output of the back propagation neural network is CKM, which is essentially the channel gain predicted by the network at the unmeasured position. For ground base station j, given enough sampling positions and corresponding measured channel gains, the back propagation neural network corresponding to the ground base station is trained to obtain the weights and bias parameters of the back propagation neural network.

[0039] For ground base station j, the explicit differentiable function corresponding to its CKM is expressed as in, is the output of the nonlinear hidden layer, and are the weight parameters and bias parameters of the nonlinear hidden layer, and are the weight parameters and bias parameters of the linear hidden layer, and are the weights and biases from the linear hidden layer neurons to the output layer. The channel knowledge graph model is represented as an explicit differentiable function of the drone position and channel gain, which is used to lay the foundation for research on resource optimization and terminal deployment in specific scenarios.

[0040] Select ground base station j=1 and show the construction result of its channel knowledge graph. The real channel knowledge graph of ground base station 1 (GBS1) is as follows: Figure 5 As shown in Figure 1, 2209 positions are randomly selected as training samples, and the channel knowledge graph is constructed using three different schemes: (1) the method based on back propagation neural network (BPNN) of the present invention; (2) K nearest neighbor interpolation method; (3) Kriging interpolation method. The construction results of the three methods are shown in Figure 1. Figure 6 , Figure 7 and Figure 8 In addition, we give the error results of the three methods reconstructing the channel knowledge graph and the real channel knowledge graph under different sample numbers, as shown in Fig. 9 It can be observed that the channel knowledge graph constructed by the present invention is closest to the real channel knowledge graph.

[0041] Step S3, introducing a transfer learning algorithm, using finite samples corresponding to the target ground base station j′ (j′=2), fine-tuning the trained back propagation neural network of the ground base station j (j=1), obtaining the back propagation neural network of the target ground base station j′, and then obtaining the channel knowledge graph of the target ground base station j′;

[0042] In theory, to generate a channel knowledge graph for multiple ground base stations, a large number of channel gain samples need to be obtained based on the location of each ground base station to train the back propagation neural network corresponding to each ground base station. However, although the different locations of base stations will lead to different channel gain distributions, the channel knowledge graphs of multiple ground base stations correspond to the same scattering environment, and the training networks of multiple ground base stations are similar.

[0043] For transfer learning algorithms, such as Figure 3 As shown, using a source dataset domain D S , a source learning task T S and a target dataset domain D T Improve a target learning task T T The channel knowledge graph of each ground base station corresponds to the same environment, where the source dataset domain D S is the training data set and test data set of ground base station 1, source learning task T S is the trained back propagation neural network of ground base station 1, and the target dataset domain D T is the training data set and test data set of the target ground base station 2, and the target learning task T T is the back propagation neural network of ground base station 2 (GBS2), D S ≠D T , T S ≠T T, the training data set of the target ground base station 2 is a finite sample. The training data set and the test data set of the corresponding ground base station are obtained by ray tracing technology or actual measurement. Since the features extracted by the back propagation neural network of the ground base station 1 can be mapped to the back propagation neural network of the ground base station 2, thereby providing prior information of the environment, the present invention chooses to freeze the first part of the back propagation neural network of the ground base station 1 to share some underlying features with the back propagation neural network of the ground base station 2, rather than retraining the entire back propagation neural network of the ground base station 2, which specifically includes the following steps:

[0044] Step S31, using the test data set of the ground base station 1 to test the trained back propagation neural network;

[0045] Step S32: Use the training data set of ground base station 2 to refine the back propagation neural network of the tested ground base station 1, and freeze the weight parameters of the nonlinear hidden layer. and bias parameters To mine shared environment knowledge and network parameters among different back-propagation neural networks;

[0046] Step S33: Use the test data set of ground base station 2 to test the retrained back propagation neural network to obtain the trained back propagation neural network of ground base station 2, and construct a mapping function of the channel knowledge graph of ground base station 2 based on the back propagation neural network.

[0047] Finally, the construction results of the channel knowledge graph under finite sample conditions are shown: the real channel knowledge graph of ground base station 2 is as follows: Fig.10 As shown in Figure 2, 800 positions are randomly selected as training samples, and the channel knowledge graph is constructed through two different schemes: (1) Fig.11 The construction method based on transfer learning proposed by the present invention is shown as follows; (2) Fig.12 The construction method of the retrained back propagation neural network is shown. It can be seen that the error between the channel knowledge graph constructed by introducing transfer learning and the real channel knowledge graph is 7.92dB, and the error between the channel knowledge graph constructed by retraining the back propagation neural network and the real channel knowledge graph is 9.84dB. It can be observed that under the condition of limited samples, the channel knowledge graph constructed based on the transfer learning method is closer to the real channel knowledge graph of ground base station 2.

[0048] The above description is only a preferred embodiment of the present invention, and therefore cannot be used to limit the scope of implementation of the present invention. That is, equivalent changes and modifications made according to the scope of the patent application of the present invention and the contents of the specification should still fall within the scope covered by the patent of the present invention.

Claims

1. A method for constructing a position-differentiable channel knowledge graph, characterized in that: The steps include: Step S1: construct a UAV-assisted communication scenario including a ground base station, and construct a channel knowledge graph model, which provides a mapping from the UAV position to the corresponding channel gain; Step S2, designing a back propagation neural network, using sufficient training samples corresponding to the ground base station j, training the back propagation neural network to learn the implicit nonlinear relationship between the UAV position and the channel gain, and representing the channel knowledge graph model as an explicit differentiable function of the UAV position and the channel gain; Step S3, introduce the transfer learning algorithm, use the finite samples corresponding to the target ground base station j′, fine-tune the trained back propagation neural network of the ground base station j′, obtain the back propagation neural network of the target ground base station j′, and then obtain the channel knowledge graph of the target ground base station j′.

2. According to claim 1, a method for constructing a position-differentiable channel knowledge graph is characterized in that: In step S1, the UAV-assisted communication scenario has an urban area divided into X×Y grids with equal spacing, and the ground base stations in the urban area are set as Indicates that the ground base station In the three-dimensional coordinate system, it is located at a fixed position (x j ,y j ,z j ),in, represents the horizontal position of the ground base station j, z j represents the height of the ground base station j. The UAV flies at a fixed height and its horizontal position is defined as q.

3. The method for constructing a position-differentiable channel knowledge graph according to claim 2, characterized in that: In step S1, in the UAV-assisted communication scenario, the ground base station j corresponds to the channel knowledge graph The channel knowledge graph is a function related only to the horizontal position of the UAV, and provides the channel gain corresponding to any given horizontal position q of the UAV at a constant height. 's mapping.

4. The method for constructing a position-differentiable channel knowledge graph according to claim 3, characterized in that: In step S2, the back propagation neural network includes an input layer, a nonlinear hidden layer, a linear hidden layer and an output layer. The Tanh activation function is selected as the activation function of the nonlinear hidden layer. The back propagation neural network is trained given sufficient samples to learn the nonlinear hidden relationship between the drone position and the channel gain, and obtain the weight parameters and bias parameters of the back propagation neural network.

5. The method for constructing a position-differentiable channel knowledge graph according to claim 4, characterized in that: In step S2, the explicit differentiable function is expressed as in, is the output of the nonlinear hidden layer, and are the weight parameters and bias parameters of the nonlinear hidden layer, and are the weight parameters and bias parameters of the linear hidden layer, and are the weights and biases of the linear hidden layer neurons to the output layer, respectively.

6. The method for constructing a position-differentiable channel knowledge graph according to claim 5, characterized in that: In step S3, a transfer learning algorithm is used to use a source data set domain D S , a source learning task T S and a target dataset domain D T Improve a target learning task T T The channel knowledge graph of each ground base station corresponds to the same environment, and the source dataset domain D S is the training dataset and test dataset of ground base station j, source learning task T S is a trained back propagation neural network, the target dataset domain D T is the training data set and test data set of the target ground base station j′, and the target learning task T T is the back propagation neural network of the ground base station j′, D S ≠D T , T S ≠T T , the training data set of the target ground base station j′ is a finite sample.

7. The method for constructing a position-differentiable channel knowledge graph according to claim 6, characterized in that: In step S3, the trained back propagation neural network is first tested using the test data set of the ground base station j, and then the tested back propagation neural network is refined using the training data set of the ground base station j′, the weight parameters and bias parameters of the nonlinear hidden layer are frozen, and then the back propagation neural network after retraining is tested using the test data set of the ground base station j′ to obtain an improved back propagation neural network.

Images