A ground-to-air communication channel tracking method based on deep learning

Abstract

The present invention provides a ground-to-air communication channel tracking method based on deep learning, comprising the following steps: obtaining simulated channel information between each base station and each reference point through simulation; obtaining a set of simulated angular channel fingerprints between the kth base station and each reference point, representing the set, and constructing a hybrid expert neural network, wherein the expert layer is composed of K parallel sub-expert layers; training the hybrid expert neural network using the sets of simulated angular channel fingerprints obtained in step S2; and, during actual communication, estimating the channel at the current position of each drone using pilot signals transmitted by drones received by each base station, thereby obtaining the actual angular channel fingerprints between each base station and its observable drones. Based on the trained hybrid expert neural network, an estimate of the current position of the drone is obtained, and the corresponding channel is tracked to reconstruct the ground-to-air communication channel. The present invention can improve channel tracking accuracy and communication quality.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communications and relates to a ground-to-air communication channel tracking method based on deep learning. Background Art

[0002] Due to their high mobility and ease of deployment, drones have become a crucial component of future mobile communication networks. Drones can serve as new user devices in mobile networks, communicating directly with ground base stations, or as part of the network, acting as aerial relays or even base stations to provide services to users on the ground. Compared to traditional terrestrial communications, drones, thanks to their flexibility, can quickly adjust their position to improve channel quality when links are blocked by obstacles. This means that drone communication links can quickly respond and restore communication links in the event of a link disconnection. Therefore, drones, as aerial devices, can effectively expand wireless network coverage and improve service quality. Furthermore, leveraging their high mobility, drones can dynamically adjust their position and coverage area to adapt to various practical scenarios, holding broad application prospects in temporary communication scenarios such as emergency communications, hotspot coverage, and real-time monitoring. However, the high mobility of drones also poses challenges to communication systems: for drones moving at high speeds in real-world scenarios, channel aging caused by channel estimation errors and transmission and processing delays can have a significant impact on the achievable rate of the communication system.

[0003] UAV communications rely heavily on accurate channel state information (CSI). Due to the high mobility of UAVs, UAV channels have stronger time-varying characteristics compared to traditional communications, which requires system parameter adaptation to ensure optimal transmission performance. Therefore, obtaining and utilizing accurate CSI is crucial to ensuring the Quality of Service (QoS) of ground-to-air communication systems. An effective way to solve this problem is to use past CSI to predict future CSI. Existing channel prediction schemes are either based on geometric models, ignoring the known channel information at the base station, or obtain channel estimates based on pilot information and then predict the channel, ignoring the spatial location information of the UAV. Therefore, these schemes suffer from the defects of channel aging and low channel tracking accuracy. Summary of the Invention

[0004] The purpose of this invention is to propose a ground-to-air communication channel tracking method based on deep learning, to improve the channel tracking accuracy, effectively solve the channel aging problem, and improve the communication quality.

[0005] The present invention is achieved through the following technical solutions:

[0006] A ground-to-air communication channel tracking method based on deep learning comprises the following steps:

[0007] Step S1: Set simulation information corresponding to the actual scenario in the simulation software, where the simulation information includes K base stations and N reference points, and obtain simulated channel information between each base station and each reference point through simulation;

[0008] Step S2: By converting each simulated channel information into the angle domain, the simulated angle domain channel fingerprints between each base station and all reference points are obtained. The set of simulated angle domain channel fingerprints between the kth base station and each reference point is expressed as 1≤k≤K;

[0009] Step S3: constructing a hybrid expert neural network, which includes an input layer, an expert layer, a gating layer, and an output layer connected in sequence. The expert layer is composed of K parallel sub-expert layers, and the loss function is designed according to the mean square error function;

[0010] Step S4: Use the simulated angular channel fingerprint sets obtained in step S2 to train the hybrid expert neural network: the simulated angular channel fingerprint sets corresponding to all base stations are used as the data of the input layer, and the simulated angular channel fingerprint sets are used as the input layer. As the input data of the kth sub-expert layer, the gating layer splices the output results of the K sub-expert layers, and the output layer outputs the mapping relationship between the angular channel fingerprint and the spatial position;

[0011] Step S5: During the actual communication process, the pilot signal sent by the UAV received by each base station is used to estimate the channel of each UAV's current position, and the actual channel information between each base station and the UAV it can observe is obtained. The actual channel information is converted into the angle domain to obtain the actual angular domain channel fingerprint between each base station and the UAV it can observe. The actual angular domain channel fingerprint set between the kth base station and the UAV it can observe is expressed as

[0012] Step S6: Use the actual angular channel fingerprint between the current UAV and each base station that can observe the UAV as the input of the hybrid expert neural network trained in step S4 to obtain estimated information about the current UAV position. Based on the estimated information, the trajectory of the UAV is predicted based on the extended Kalman filter algorithm, and the channel corresponding to the UAV is tracked according to the predicted trajectory to reconstruct the ground-to-air communication channel.

[0013] Furthermore, in step S2, the base station uses a uniform planar array, and the simulated angular channel fingerprint G between the kth base station and the i-th reference point is ki for The simulated angular channel fingerprint set between the kth base station and all reference points is: in, Indicates M x dimensional discrete Fourier transform matrix, Indicates M y dimensional discrete Fourier transform matrix, M x 、M y Represents the number of antennas in the x and y directions, H ki represents the simulated channel information between the kth base station and the i-th reference point.

[0014] Furthermore, in step S3, the sub-expert layer is a deep convolutional neural network, including a convolutional layer, a pooling layer and a fully connected layer, and the sub-expert layer outputs the positioning results of each reference point.

[0015] Furthermore, in step S3, the loss function is designed as Among them, K is the number of sub-expert layers, represents the positioning result of the kth sub-expert layer, represents the final positioning result output by the hybrid expert neural network, and x represents the actual position of the reference point.

[0016] Furthermore, in step S5, during the actual communication process, the UAV first sends a pilot signal s for channel estimation. If the kth base station receives the signal, the signal received by the kth base station is Use the LS channel estimation algorithm to estimate the actual channel information between the kth base station and the drone The actual channel information is converted into the angle domain to obtain the corresponding actual angle domain channel fingerprint. Among them, P b Indicates the transmit power, N k represents noise, and τ represents the pilot signal length.

[0017] Furthermore, in step S6, predicting the trajectory of the UAV based on the estimated information and reconstructing the ground-to-air communication channel specifically includes the following steps:

[0018] Step S61: Obtain the state vector c at time n based on the current UAV position estimation information. n =[x n ,v n ] T =[x xn ,x yn ,x zn ,v xn ,v yn ,v zn ] T , where x n Represents the position estimation information, v n Indicates speed;

[0019] Step S62: Using a near constant velocity model, obtain the state evolution model c n+1 =Wc n +n c , where W represents the state transfer matrix, n c is Gaussian noise;

[0020] Step S63: Speed ​​v n As prior information, according to the formula The measurement model r(c n ),in, Indicates delay, represents the Doppler shift, represents the azimuth, represents the elevation angle, c represents the speed of light, f c represents the carrier frequency, z m represents the Gaussian noise vector;

[0021] Step S64: Predict the trajectory of the drone using the EKF algorithm:

[0022] Predicting drone status

[0023] Measurement model linearization: r represents the measurement model;

[0024] Prediction MSE matrix M n :M n∣n-1 =WM n-1 W H +Q c ;Q c Indicates n c The covariance matrix of

[0025] Calculate the Kalman gain matrix:

[0026] Status correction:

[0027] Update the MSE matrix: M n =(IK n R n )M n∣n-1 ;

[0028] Step S65: Reconstruct the ground-to-air communication channel based on the UAV trajectory obtained in step S64: Among them, β represents the large-scale fading coefficient, α represents the channel complex gain, and a x (·), a y (·) represents the antenna array response vector in the x and y directions, respectively. represent elevation and azimuth angles respectively.

[0029] Furthermore, the simulation software in step S1 is Wireless Insite simulation software.

[0030] The present invention has the following beneficial effects:

[0031] 1. The present invention adopts a hybrid expert neural network, the expert layer of which is composed of multiple parallel sub-expert layers, to achieve accurate positioning of the UAV in actual scenarios by using multi-station joint positioning; the hybrid expert neural network is trained using data simulated in advance on simulation software, and in actual scenarios, the UAV is positioned using a convergent hybrid expert neural network through uplink pilot signals, which can not only reduce the complexity of the positioning algorithm, but also enhance the stability of the ground-to-air communication channel; based on the positioning results, the UAV motion trajectory and channel are tracked in real time, which helps to reduce the impact of channel estimation errors and transmission and processing delays on system performance, thereby effectively solving the channel aging problem and improving communication quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The present invention will be further described in detail below with reference to the accompanying drawings.

[0033] Figure 1 Flowchart of the present invention.

[0034] Figure 2 This is a model diagram of the hybrid expert neural network of the present invention.

[0035] Figure 3 This is a diagram of the accuracy simulation results of the present invention.

[0036] Figure 4 Graph showing simulation results of the channel tracking performance of the present invention. DETAILED DESCRIPTION

[0037] like Figure 1 As shown, the ground-to-air communication channel tracking method based on deep learning includes the following steps:

[0038] Step S1: Set simulation information corresponding to the actual scenario in the simulation software, where the simulation information includes K base stations and N reference points, and obtain simulated channel information between each base station and each reference point through simulation;

[0039] Specifically, in this embodiment, Wireless INsite simulation software is used. The simulation information includes environmental information, signal bandwidth, signal frequency, number of antennas, and polarization mode that demonstrate the actual communication scenario. The environmental information includes buildings, obstacles, K base stations, and N reference points. The reference points correspond to drones in actual communication. More specifically, the signal bandwidth is set to 1 MHz, the signal frequency is set to 1 GHz, the base station uses a uniform planar array antenna, specifically a 16*16 UPA omnidirectional antenna, the reference point spacing is 1 m, the simulation area size is 200 m*200 m, and the drone uses a single omnidirectional antenna.

[0040] Step S2: By converting each simulated channel information into the angle domain, the simulated angle domain channel fingerprint between each base station and each reference point is obtained. The set of simulated angle domain channel fingerprints between the kth base station and all reference points is expressed as 1≤k≤K;

[0041] Specifically, the base station uses a uniform planar array, and the simulated angular channel fingerprint G between the kth base station and the i-th reference point is ki for The simulated angular channel fingerprint set between the kth base station and all reference points is: Among them, M x and M y Represents the number of antennas in the x-direction and y-direction respectively, Indicates M x dimensional discrete Fourier transform matrix, Indicates M y dimensional discrete Fourier transform matrix, The (p,q)th element of can be expressed as represents the Kronecker product, H ki represents the simulated channel information between the kth base station and the i-th reference point, 1≤i≤N;

[0042] Step S3, construct Figure 2 The hybrid expert neural network shown in the figure includes an input layer, an expert layer, a gating layer, and an output layer connected in sequence. The expert layer is composed of K parallel sub-expert layers and is used to extract the spatial features of the angular channel fingerprint and learn the mapping relationship between these features and the position. The loss function is designed based on the mean square error function.

[0043] Specifically, the data input to the input layer is the angular channel fingerprint of the reference points corresponding to all base stations, which is expressed as in,

[0044] The sub-expert layer is a deep convolutional neural network, including convolutional layer, pooling layer and fully connected layer. The sub-expert layer outputs the positioning results of each reference point;

[0045] The input data of the kth expert layer is the simulated angular channel fingerprint set between the kth base station and all reference points: Perform convolution and pooling operations in sequence;

[0046] Convolution: Use a convolution kernel to slide on the input data of the previous layer with a fixed step size. Each slide convolves the local input data with the convolution kernel to obtain the output value at the corresponding position. Through repeated sliding convolution operations, all the features of the input layer are obtained. Relevant parameters include convolution kernel size, number of convolution layers and sliding step size. Each expert layer consists of 3 convolution layers, 3 pooling layers and 4 fully connected layers. The number of convolution kernels is 8, 16 and 32 respectively, and the convolution kernel sizes are 8, 4 and 2 respectively.

[0047] The convolution operation is expressed as: Among them, X l-1 represents the input of the previous layer, f(·) represents the activation function, Represents the parameters in the convolution kernel in the network, * represents the convolution operation, Indicates the bias of the layer;

[0048] Pooling: The pooling layer further extracts features from the data generated by the convolutional layer to achieve dimensionality reduction and reduce network training parameters.

[0049] The pooling operation is expressed as: Y l =down(X l ), where down(·) represents the pooling operation, using maximum pooling;

[0050] Finally, the fully connected layer is used to further process the pooled data. The processing process is expressed as: in, and They represent the activation function, weight vector, and bias vector of the g-th layer network in the fully connected layer. Except for the last layer, which uses linear activation function, the activation functions of the remaining layers all use relu. The number of neurons in the fully connected layer is 256, 128, 32, and 3 respectively. The user position obtained by the k-th sub-expert layer after the above operation is

[0051] The gating layer first splices the output results of different sub-expert layers to obtain the output results of multiple sub-expert layers It then passes through multiple fully connected layers to learn the weights of different sub-expert layers, and finally outputs the results through the output layer. The calculation process of the vector X by the gating layer and the output layer can be expressed as: in, and The activation function, weight vector and bias vector of the t-th layer network in the fully connected layer respectively;

[0052] According to the mean square error (MSE) function, the loss function for the hybrid expert neural network is designed. for Among them, K is the number of sub-expert layers, represents the positioning result of the i-th sub-expert layer, represents the final positioning result output by the hybrid expert neural network, and x represents the actual position of the reference point.

[0053] Step S4: Use the simulated angular channel fingerprint sets obtained in step S2 to train the hybrid expert neural network: the simulated angular channel fingerprint sets corresponding to all base stations are used as the data of the input layer, and the simulated angular channel fingerprint sets are used as the input layer. As the input data of the kth sub-expert layer, the gating layer concatenates the output results of the K sub-expert layers and passes them through multiple fully connected layers. The output layer then outputs the mapping relationship between the angular channel fingerprint and the spatial position.

[0054] Step S5: During the actual communication process, the pilot signal sent by the UAV received by each base station is used to estimate the channel of each UAV's current position, and the actual channel information between each base station and the UAV it can observe is obtained. The actual channel information is converted into the angle domain to obtain the actual angular domain channel fingerprint between each base station and the UAV it can observe. The actual angular domain channel fingerprint set between the kth base station and the UAV it can observe is expressed as

[0055] Specifically, in the actual communication process, the UAV first sends a pilot signal s for channel estimation. If the kth base station receives the signal (that is, the kth base station can observe the UAV), then the signal received by the kth base station is Use the LS channel estimation algorithm to estimate the actual channel information between the kth base station and the drone The actual channel information is converted into the angle domain to obtain the corresponding actual angle domain channel fingerprint. The specific conversion process is the same as step S2, where P b Indicates the transmission power. In this embodiment, P is set b =1W,N k represents noise, τ represents the pilot signal length, and in this embodiment, τ=100.

[0056] Step S6: Using the actual angular channel fingerprint between the current UAV and each base station that can observe the UAV as input to the hybrid expert neural network trained in step S4, an estimate of the UAV's current position is obtained. Based on the estimate, the UAV's trajectory is predicted using the extended Kalman filter algorithm. The channel corresponding to the UAV is tracked based on the predicted trajectory to reconstruct the ground-to-air communication channel.

[0057] "Predicting the trajectory of the UAV based on the estimated information using an extended Kalman filter algorithm, tracking the channel corresponding to the UAV based on the predicted trajectory, and reconstructing the ground-to-air communication channel" specifically includes the following steps:

[0058] Step S61: Obtain the state vector c at time n based on the current UAV position estimation information. n =[x n ,v n ] T =[x xn ,x yn ,x zn ,v xn ,v yn ,v zn ] T , where x n Represents the position estimation information, x xn ,x yn ,x zn Represents x n The x, y, and z axis coordinates, v n Indicates speed, v xn ,v yn ,v zn Respectively represent v n The x, y, and z axis coordinates, x n With v n The relationship between them is expressed as: where n d 、n v represents Gaussian noise, ΔT represents the time variation;

[0059] Step S62: Using a near constant velocity model, obtain the state evolution model c n+1 =Wc n +n c , where W represents the state transfer matrix, n c is Gaussian noise, and the covariance matrix of the Gaussian noise is

[0060] Step S63: Speed ​​v n As prior information, according to the formula The measurement model r(c n ),in, Indicates delay, represents the Doppler shift, represents the azimuth, represents the elevation angle, c represents the speed of light, f c represents the carrier frequency, z m represents the Gaussian noise vector;

[0061] Step S64: Predict the trajectory of the drone using the EKF algorithm:

[0062] Predicting drone status

[0063] Measurement model linearization: r represents the measurement model;

[0064] Prediction MSE matrix M n :M n∣n-1 =WM n-1 W H +Q c ;Q c Indicates n c The covariance matrix of

[0065] Calculate the Kalman gain matrix:

[0066] Status correction:

[0067] Update the MSE matrix: M n =(IK n R n )M n∣n-1 ;

[0068] Step S65: Reconstruct the ground-to-air communication channel based on the UAV trajectory obtained in step S64: Among them, β represents the large-scale fading coefficient, α represents the channel complex gain, and a x (·), a y (·) represents the antenna array response vector in the x and y directions, respectively. represent elevation and azimuth angles respectively.

[0069] The antenna array response vector in this embodiment is expressed as:

[0070] Among them, M x 、M y They represent the number of transmitting antennas in the x and y directions respectively, d represents the distance between two adjacent antennas on the antenna panel, and λ represents the wavelength of the electromagnetic wave.

[0071] Figure 3In the figure, the horizontal axis represents the positioning error, ranging from 0 to 20 meters, and the vertical axis represents the percentage of test data within the error range. The CDF curves (cumulative distribution function curves of positioning error) are plotted for positioning using two base stations (tx2 and tx3) and the combined positioning method (the present invention). It can be seen that the CDF curve corresponding to the present invention converges to 1 more quickly, achieving 90% accuracy within a 2.3-meter error range, a 20% improvement in positioning accuracy compared to positioning using a single base station.

[0072] Figure 4 In the figure, the horizontal axis is the signal-to-noise ratio (SNR), ranging from -15 to 30 dB, and the vertical axis is the normalized mean square error (NMSE). The simulation experiment compared and analyzed three cases: the LS channel estimation algorithm, the position estimation-based channel algorithm (outdate), and the present invention. It can be seen that the NMSE of the present invention is the lowest within the entire SNR range. This is because the LS algorithm and the position estimation-based algorithm cannot solve the channel aging problem. The present invention first tracks the drone's trajectory and then reconstructs the channel based on the drone's future position, which can improve the accuracy of ground-to-air communication channel tracking.

[0073] The above description is merely a preferred embodiment of the present invention and therefore cannot be used to limit the scope of the present invention. In other words, equivalent changes and modifications made according to the scope of the patent application and the contents of the specification should still fall within the scope of the patent of the present invention.

Claims

1. A ground-to-air communication channel tracking method based on deep learning, characterized by: The steps include: Step S1: Set simulation information corresponding to the actual scenario in the simulation software, where the simulation information includes K base stations and N reference points, and obtain simulated channel information between each base station and each reference point through simulation; Step S2: By converting each simulated channel information into the angle domain, the simulated angle domain channel fingerprints between each base station and all reference points are obtained. The set of simulated angle domain channel fingerprints between the kth base station and each reference point is expressed as 1≤k≤K; Step S3: constructing a hybrid expert neural network, which includes an input layer, an expert layer, a gating layer, and an output layer connected in sequence. The expert layer is composed of K parallel sub-expert layers, and the loss function is designed according to the mean square error function; Step S4: Use the simulated angular channel fingerprint sets obtained in step S2 to train the hybrid expert neural network: the simulated angular channel fingerprint sets corresponding to all base stations are used as the data of the input layer, and the simulated angular channel fingerprint sets are used as the input layer. As the input data of the kth sub-expert layer, the gating layer splices the output results of the K sub-expert layers, and the output layer outputs the mapping relationship between the angular channel fingerprint and the spatial position; Step S5: During the actual communication process, the pilot signal sent by the UAV received by each base station is used to estimate the channel of each UAV's current position, and the actual channel information between each base station and the UAV it can observe is obtained. The actual channel information is converted into the angle domain to obtain the actual angular domain channel fingerprint between each base station and the UAV it can observe. The actual angular domain channel fingerprint set between the kth base station and the UAV it can observe is expressed as Step S6: Use the actual angular channel fingerprint between the current UAV and each base station that can observe the UAV as the input of the hybrid expert neural network trained in step S4 to obtain estimated information about the current UAV position. Based on the estimated information, the trajectory of the UAV is predicted based on the extended Kalman filter algorithm, and the channel corresponding to the UAV is tracked according to the predicted trajectory to reconstruct the ground-to-air communication channel.

2. The method for tracking ground-to-air communication channels based on deep learning according to claim 1, wherein: In step S2, the base station uses a uniform planar array, and the simulated angular channel fingerprint G between the kth base station and the i-th reference point is ki for The simulated angular channel fingerprint set between the kth base station and all reference points is: in, Indicates M x dimensional discrete Fourier transform matrix, Indicates M y dimensional discrete Fourier transform matrix, M x 、M y Represents the number of antennas in the x and y directions, H ki represents the simulated channel information between the kth base station and the i-th reference point.

3. The deep learning-based ground-to-air communication channel tracking method according to claim 1, characterized in that: In step S3, the sub-expert layer is a deep convolutional neural network, including a convolutional layer, a pooling layer and a fully connected layer, and the sub-expert layer outputs the positioning results of each reference point.

4. The method for tracking ground-to-air communication channels based on deep learning according to claim 3, wherein: In step S3, the loss function is designed as Among them, K is the number of sub-expert layers, represents the positioning result of the kth sub-expert layer, represents the final positioning result output by the hybrid expert neural network, and x represents the actual position of the reference point.

5. The method for tracking ground-to-air communication channels based on deep learning according to claim 2, wherein: In step S5, during the actual communication process, the UAV first sends a pilot signal s for channel estimation. If the kth base station receives the signal, the signal received by the kth base station is Use the LS channel estimation algorithm to estimate the actual channel information between the kth base station and the drone The actual channel information is converted into the angle domain to obtain the corresponding actual angle domain channel fingerprint. Among them, P b Indicates the transmit power, N k represents noise, and τ represents the pilot signal length.

6. A method for tracking ground-to-air communication channels based on deep learning according to claim 1, 2, 3 or 4, characterized in that: In step S6, predicting the trajectory of the UAV based on the estimated information and reconstructing the ground-to-air communication channel specifically includes the following steps: Step S61: Obtain the state vector c of the drone at time n based on the current drone position estimation information. n =[x n ,v n ] T =[x xn ,x yn ,x zn ,v xn ,v yn ,v zn ] T , where x n Represents the position estimation information, v n Indicates speed; Step S62: Using a near constant velocity model, obtain the state evolution model c n+1 =Wc n +n c , where W represents the state transfer matrix, n c is Gaussian noise; Step S63: Speed ​​v n As prior information, according to the formula The measurement model r(c n ),in, Indicates delay, represents the Doppler shift, represents the azimuth, represents the elevation angle, c represents the speed of light, f c represents the carrier frequency, z m represents the Gaussian noise vector; Step S64: Predict the trajectory of the UAV using the EKF algorithm: Predicting drone status Measurement model linearization r represents the measurement model; Prediction MSE matrix M n :M n∣n-1 =WM n-1 W H +Q c ;Q c Indicates n c The covariance matrix of Calculate the Kalman gain matrix: Status correction: Update the MSE matrix: M n =(IK n R n )M n∣n-1 ; Step S65: Reconstruct the ground-to-air communication channel based on the UAV trajectory obtained in step S64: Among them, β represents the large-scale fading coefficient, α represents the channel complex gain, and a x (·), a y (·) represents the antenna array response vector in the x and y directions, respectively. represent elevation and azimuth angles respectively.

7. A method for tracking ground-to-air communication channels based on deep learning according to claim 1, 2, 3 or 4, characterized in that: The simulation software in step S1 is WirelessInsite simulation software.

Images