A vehicle-to-ground wireless channel prediction method based on position information coding and periodic nonlinear neural network

Abstract

The present application belongs to the field of communication technology and neural network, and provides a train-ground wireless channel prediction method based on position information coding and periodic nonlinear neural network, which is suitable for the optimization of train-ground wireless communication system. The specific steps are as follows: S1: collecting the position information of the train and the wireless channel instance, position coding the collected channel instance, and mapping the channel instance to a high-dimensional Fourier feature space; S2: integrating the periodic nonlinear neural network, constructing the generative deep neural network, and learning and modeling the channel instance based on the position information coding and the periodic nonlinear neural network; S3: applying the trained model to real-time channel prediction to support the dynamic optimization of the wireless communication system. The present application improves the accuracy and efficiency of train-ground wireless channel prediction, and has significant technical advantages.

Description

Technical Field

[0001] The present invention belongs to the fields of communication technology and neural networks, and in particular relates to a vehicle-ground wireless channel prediction method based on position information coding and periodic nonlinear neural network. Background Art

[0002] Intelligent channel prediction plays a crucial role in artificial intelligence (AI) optimization or AI-native communication networks in intelligent high-speed railways. In particular, in vehicle-to-ground (T2G) wireless communication scenarios, the rapidly time-varying channel state information (CSI) obtained by traditional channel estimation schemes often becomes outdated. Therefore, efficient and accurate channel prediction is essential without consuming additional radio resources. Given that the generation and variation of channels are caused by specific electromagnetic wave propagation and have clear physical meaning, they are well-suited for characterization and prediction using neural networks. In real-world communication scenarios, the area served by a base station and the number and location of scatterers are relatively fixed. Therefore, the channel at any fixed location within the base station's service area can be considered quasi-static. Because all wireless communication processes occurring within a specific area are subject to the same scattering environment, the static channel characteristics at different locations within the area are highly correlated and therefore predictable. Furthermore, in T2G wireless communication scenarios, the T2 wireless channel can be calculated by compensating for the Doppler shift of the static channel. Therefore, it is possible to use randomly sampled channel data within the wireless environment to learn an implicit representation of the scattering environment, efficiently and accurately predicting the channel impulse response at any location within the environment.

[0003] There are two major challenges in this work: first, the channel response is highly non-smooth in space, and conventional learning structures cannot accurately characterize the changes in the channel in space with limited samples as training data; second, this problem requires reconstructing high-dimensional information (such as channel impulse response) from extremely low-dimensional data (such as position coordinates), which is essentially a complex generation problem that requires the network to have the ability to learn channel structure.

[0004] Machine learning is one effective approach to addressing these issues. Numerous studies have demonstrated the ability of neural networks to directly exploit correlations between data and learn information about the scattering environment. However, traditional deep neural networks (DNNs) often require large amounts of training data and numerous hidden parameters to accomplish such prediction tasks. Furthermore, traditional DNNs are prone to overfitting or large prediction errors. Furthermore, the phase component of electromagnetic wave propagation is highly sensitive to position changes. Even small positional variations between adjacent sampling points can result in significant phase shifts, significantly impacting the DNN's final fitting accuracy. Therefore, it is necessary to design a specialized learning architecture that fully exploits the physical properties of electromagnetic wave propagation. Summary of the Invention

[0005] The purpose of the present invention is to overcome the above existing challenges and provide a vehicle-ground wireless channel prediction method based on position information encoding and periodic nonlinear neural network. Channel information is obtained based on the neural network to improve the efficiency and accuracy of channel prediction.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] A vehicle-ground wireless channel prediction method based on position information coding and periodic nonlinear neural network is characterized by comprising the following steps:

[0008] S1: collecting the train's location information and wireless channel instances, and performing position coding on the collected channel instances;

[0009] Collect train location information and wireless channel instances;

[0010] Perform positional encoding (PE) on the collected channel instances and map them into a high-dimensional Fourier feature space.

[0011] S2: Based on position information encoding and periodic nonlinear neural network, a generative deep neural network is constructed to learn and model channel instances;

[0012] S3: Apply the generative deep neural network model to perform real-time channel prediction.

[0013] Furthermore, step S1 includes:

[0014] S11: Collect train location information

[0015] The train location information includes but is not limited to the acquisition of various train location information, including: train location information obtained by transponder, inertial navigation, GPS, Beidou, etc.;

[0016] S12: Collecting a wireless channel instance

[0017] The wireless channel instance includes one or more pieces of uplink / downlink historical communication data and historical channel information generated in the process of train and base station communication; the historical communication data includes channel state information (CSI); and the historical channel information includes historical channel state information and historical channel impulse response (CIR).

[0018] S13: Position encoding is performed on the collected channel instance, as follows:

[0019] G(d k )=[…,a k cos(2πσ j / m d k ),a k sin(2πσ j / m d k ),…] T for j=0,…,m-1.

[0020] wherein σ and j / m are scale factors of position encoding, and the frequency of each dimension is adjusted by the scale factors σ and j / m to generate a series of low-to-high frequencies. Here, σ is an adjustable scale factor, j represents the index of encoding, and m is the total number of frequencies.

[0021] Further, step S2 includes:

[0022] S21: Integrating a periodic nonlinear neural network and position encoding to construct a generative deep neural network;

[0023] The generative deep neural network includes a position information encoding layer and a periodic nonlinear neural network layer.

[0024] The position information encoding layer expands input position, time and frequency parameters into high-dimensional features through position encoding and provides the features to the periodic nonlinear neural network layer for real-time channel prediction.

[0025] The periodic nonlinear neural network refers to a neural network using a nonlinear activation function, and the present application uses a sinusoidal representation network (SIREN), as follows:

[0026]

[0027]

[0028] In SIREN, x represents the input of the neural network, Φ(x) represents the output of the neural network, Represents a specific connection between two layers, Represents the i-th layer of the network, which is composed of the weight matrix enter Bias And sine nonlinear activation function.

[0029] S22: Learn and model the channel instances.

[0030] The overall training process is as follows:

[0031] First, data preparation is performed by loading a real channel impulse response (CIR) dataset and decomposing the complex values ​​into real and imaginary parts. Input parameters such as position, time, and frequency are expanded into high-dimensional features through position encoding.

[0032] Initialize the model and initialize the weights and biases of SIREN. Use a special weight initialization method to ensure that the model can effectively capture the high-frequency characteristics of the channel. The weight initialization scheme is as follows:

[0033]

[0034] Among them, d in is the input dimension, and U is the uniform distribution function.

[0035] During the training process, a supervised learning method is used, using wireless channel instances (i.e., channel impulse response (CIR)) as labels, and the prediction accuracy and generalization ability of the model are improved by optimizing the loss function. The loss function is:

[0036]

[0037] Among them, q is the output dimension, y is the training label, is the predicted value. MSE represents the average distance between the target value and the predicted value across all output dimensions.

[0038] Back propagation calculates gradients and updates network parameters.

[0039] Furthermore, in step S3, the neural network model trained in step S2 is used to perform real-time channel prediction, inputting the train position, environment and channel parameters, and outputting the predicted real-time channel impulse response (CIR), including the real and imaginary parts of the CIR, to provide support for dynamic adjustment of the communication system.

[0040] Furthermore, the real-time channel prediction includes predicting the future wireless channel impulse response based on the train's real-time position information, running speed and other dynamic data, and dynamically adjusting the parameters of the communication system to optimize the signal quality.

[0041] By mathematically analyzing the physical process of the channel impulse response, the present invention found that for a certain electromagnetic wave propagation path in a specific area, its response amplitude is only related to the propagation path length and the number of reflections, while its phase changes periodically with the path length. Therefore, in the framework proposed by the present invention, the position information encoding provides a high-dimensional Fourier feature, which helps the model capture the high-frequency information of the input signal in the feature space and enhances the model's perception of the channel multipath characteristics. The periodic nonlinear neural network further processes the Fourier features of the position encoding. Due to the periodic characteristics of its sinusoidal activation function, the periodic nonlinear neural network can better fit the high dynamic phase characteristics of the channel instance mapped to the Fourier feature space, thereby accurately predicting the channel response in both the frequency domain and the time domain.

[0042] Compared with the prior art, the present invention has the following advantages:

[0043] 1) The network proposed in this paper simulates the physical process of CIR generation, thus integrating prior information into the network structure, which greatly reduces the complexity of network learning and its dependence on data volume.

[0044] 2) The convergence speed of the network proposed in this invention is faster than that of various control networks. Thanks to the convergence characteristics of the sinusoidal representation network (SIREN), and the addition of position information encoding, the periodic nonlinear neural network avoids the problem of spectral deviation in traditional neural network fitting during the fitting process. The model's ability to perceive the multipath characteristics of electromagnetic wave propagation is effectively enhanced, and the network can quickly obtain high-frequency information in the input signal in the initial stage. Since the network training algorithm is backpropagation, shallow networks with fewer parameters have a faster convergence speed.

[0045] 3) The proposed network achieves smaller convergence losses than various control networks at varying sampling rates. The unique network structure, which mimics the CIR generation process, reduces the network's dependence on the training dataset sampling rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flow chart of the channel prediction method of the present invention;

[0047] Figure 2 Schematic diagram of a neural network of the present invention;

[0048] Figure 3 Schematic diagram of the convergence curve of the neural network of the present invention;

[0049] Figure 4 Schematic diagram showing the performance comparison between the neural network of the present invention and various control methods;

[0050] Figure 5 This is a schematic diagram comparing the neural network prediction results of the present invention with the original data;

[0051] Figure 6 This is a schematic diagram comparing the predicted results of the control method of the present invention with the original data;

[0052] Figure 7 This is a schematic diagram comparing the predicted results of the control method of the present invention with the original data. DETAILED DESCRIPTION

[0053] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] like Figure 1 The present invention relates to a vehicle-ground wireless channel prediction method based on position information coding and periodic nonlinear neural network, which is characterized by comprising the following steps:

[0055] A vehicle-ground wireless channel prediction method based on position information coding and periodic nonlinear neural network is characterized by comprising the following steps:

[0056] S1: collecting the train's location information and wireless channel instances, and performing position coding on the collected channel instances;

[0057] Collect train location information and wireless channel instances;

[0058] Perform positional encoding (PE) on the collected channel instances and map them into a high-dimensional Fourier feature space.

[0059] S2: Based on position information encoding and periodic nonlinear neural network, a generative deep neural network is constructed to learn and model channel instances;

[0060] S3: Apply neural network models for real-time channel prediction.

[0061] Furthermore, step S1 includes:

[0062] S11: Collect train location information

[0063] The train location information includes but is not limited to the acquisition of various train location information, including: train location information obtained by transponder, inertial navigation, GPS, Beidou, etc.;

[0064] S12: Collect wireless channel examples

[0065] The wireless channel instance includes one or more uplink / downlink historical communication data and historical channel information generated in the process of train and base station communication; the historical communication data includes channel state information (CSI); the historical channel information includes historical channel state information and historical channel impulse response (CIR). It should be noted that CIR is not only used for direct calculation of parameters related to CSI, but also provides an intuitive representation of the multipath characteristics of the wireless channel.

[0066] The quasi-static representation of the CIR can be expressed as:

[0067]

[0068] Where N represents the number of propagation paths, a k is the loss function of the kth path, f is the carrier frequency, v c is the speed of light, d k is the length of the propagation path, and τ is the propagation delay. The formula of CIR shows that the channel instance is formed by a series of path responses. Each path response can be divided into two parts: the attenuation factor a k = λ / 4πd k and the phase shift

[0069] Estimating the Doppler shift: In the train-ground wireless communication scenario, there is a Doppler effect between the base station and the train, which causes a frequency shift on each propagation path. For a given path, the Doppler shift is estimated as:

[0070]

[0071] Where v k is the relative speed of the kth path.

[0072] The presence of the Doppler shift will cause the received CIR to change dynamically over time, which can be expressed as:

[0073]

[0074] Further, in order to alleviate the impact of the Doppler shift and restore the quasi-static nature of the CIR, a Doppler compensation mechanism is applied at the receiver. The Doppler compensation mechanism includes but is not limited to Doppler compensation based on speed and position information.

[0075] According to formulas (1), (2), and (3), the compensated CIR can be expressed as:

[0076] h comp (f,τ) = h(f-Δf,τ)·e j2πΔft(4)

[0077] Furthermore, for quasi-static CIR, according to the Friis equation, the signal energy received by a certain signal transmission path can be written as:

[0078]

[0079] Among them, d k represents the path length of the direct path, P TX Indicates the transmission energy of TX, G TX represents the transmitting antenna gain, G RX represents the receiving antenna gain, and λ = c / f is the wavelength of the electromagnetic wave. The electromagnetic field strength excited by a certain propagation path at the RX end can be written as:

[0080]

[0081] When all antenna gains are set to 1, the square root of the received signal energy to the transmitted energy can be written as:

[0082]

[0083] For phase shift The phase shift information can be decomposed into real and imaginary parts using sine and cosine components in polar coordinate form, that is:

[0084]

[0085] S13: Position encoding is performed on the collected channel instances as follows:

[0086]

[0087] Where σ and j / m are the scale factors of the positional encoding. The frequency of each dimension is adjusted by the scale factors σ and j / m to generate a series of frequencies from low to high. Here σ is an adjustable scale factor, j represents the index of the encoding, and m is the total number of frequencies.

[0088] From the above formulas (1) and (4), it can be concluded that the wireless channel instance (i.e., channel impulse response (CIR)) in vehicle-ground wireless communication can be decomposed into multiple electromagnetic wave propagation paths, where each path (LOS path and NLOS path) has a loss factor inversely proportional to the path length and a phase shift proportional to the path length. The channel impulse response function exhibits certain radial and periodic characteristics of amplitude and phase, respectively, where:

[0089] The radial characteristic, i.e. the amplitude, is inversely proportional to the propagation path length (based on the Friis equation).

[0090] Periodic characteristics: On the electromagnetic wave propagation path, the phase offset of the signal changes periodically with the path length and carrier frequency.

[0091] Therefore, the channel impulse response can be regarded as an implicit function of the train position information coordinates.

[0092] Furthermore, step S2 includes:

[0093] S21: Integrating periodic nonlinear neural networks and positional encoding to construct generative deep neural networks;

[0094] Position encoding, or position information encoding, provides high-dimensional Fourier features, helping the model capture high-frequency information of the input signal in the feature space and enhancing the model's ability to perceive channel multipath characteristics. A periodic nonlinear neural network further processes the Fourier features of the position encoding.

[0095] like Figure 2 , the generative deep neural network includes a position information encoding layer and a periodic nonlinear neural network layer;

[0096] The position information encoding layer expands the input position, time, and frequency parameters into high-dimensional features through position encoding and provides them to the periodic nonlinear neural network layer for real-time channel prediction;

[0097] According to formulas (1), (7), and (8), in the generation problem of channel impulse response, the channel impulse response function is considered to be two parts, which respectively show radial and periodic characteristics of amplitude and phase. This characteristic is called partial implicit periodicity. This type of function can be expressed as:

[0098]

[0099] d k =(P x , P y , P z ) (11)

[0100] The CIR is decomposed into the radial variation of the amplitude F(d k ) and periodic changes in phase d k The three-dimensional coordinates P of the receiver and transmitter x , P y , P z Decision, F(d k ) is about d k The inverse function of , describing the attenuation of the signal with distance, The position code obtained in step S1 is about A periodic function of . For a about d kThe implicit periodic function of Has partial implicit periodicity.

[0101] Specifically, the periodic nonlinear neural network refers to a neural network using a nonlinear activation function. The present invention uses a sinusoidal representation network (SIREN), as follows:

[0102]

[0103] In SIREN, x represents the input of the neural network, Φ(x) represents the output of the neural network, Represents a specific connection between two layers, Represents the i-th layer of the network, which is composed of the weight matrix enter Bias Compared to traditional ReLU and Tanh activation functions, SIREN uses the sine activation function to accurately model the implicit periodicity of channel phase changes due to differences in sampling distance and propagation path length, as well as the radial attenuation of signal amplitude that varies with path loss and propagation distance.

[0104] S22: Learn and model the channel instances.

[0105] The overall training process is as follows:

[0106] First, data preparation is performed by loading a real channel impulse response (CIR) dataset and decomposing the complex values ​​into real and imaginary parts. Input parameters such as position, time, and frequency are expanded into high-dimensional features through position encoding.

[0107] Initialize the model and initialize the weights and biases of SIREN. Use a special weight initialization method to ensure that the model can effectively capture the high-frequency characteristics of the channel. The weight initialization scheme is as follows:

[0108]

[0109] Among them, d in is the input dimension, and U is the uniform distribution function.

[0110] During the training process, a supervised learning method is used, using wireless channel instances (i.e., channel impulse response (CIR)) as labels, and the prediction accuracy and generalization ability of the model are improved by optimizing the loss function. The loss function is:

[0111]

[0112] Among them, q is the output dimension, y is the training label, is the predicted value. MSE represents the average distance between the target value and the predicted value across all output dimensions.

[0113] Back propagation calculates gradients and updates network parameters.

[0114] Furthermore, in step S3, the neural network model trained in step S2 is used to perform real-time channel prediction, inputting the train position, environment and channel parameters, and outputting the predicted real-time channel impulse response (CIR), including the real and imaginary parts of the CIR, to provide support for dynamic adjustment of the communication system.

[0115] Furthermore, the real-time channel prediction includes predicting the future wireless channel impulse response based on the train's real-time position information, running speed and other dynamic data, and dynamically adjusting the parameters of the communication system to optimize the signal quality.

[0116] By mathematically analyzing the physical process of the channel impulse response, the present invention found that for a certain electromagnetic wave propagation path in a specific area, its response amplitude is only related to the propagation path length and the number of reflections, while its phase changes periodically with the path length. Therefore, in the framework proposed by the present invention, the position information encoding provides a high-dimensional Fourier feature, which helps the model capture the high-frequency information of the input signal in the feature space and enhances the model's perception of the channel multipath characteristics. The periodic nonlinear neural network further processes the Fourier features of the position encoding. Due to the periodic characteristics of its sinusoidal activation function, the periodic nonlinear neural network can better fit the high dynamic phase characteristics of the channel instance mapped to the Fourier feature space, thereby accurately predicting the channel response in both the frequency domain and the time domain.

[0117] Algorithm performance comparison: The prediction accuracy of the vehicle-ground wireless channel prediction method based on position information encoding and periodic nonlinear neural network (PE_SIREN) of the present invention is compared with that of the sine representation network (SIREN), feedforward neural network (FNN), LSTM network, and CGAN network as follows:

[0118] 1) If Figure 3 As shown, when the sampling rate is 80%, the convergence speed and prediction accuracy of the training of the method of the present invention (PE_SIREN), the sine representation network (SIREN) and the feedforward neural network (FNN) can be compared from the curve. Among them, the X-axis represents the training step size, ranging from 1 to 200,000. The y-axis represents the mean square error (MSE) on the training data set. Under the same sampling rate, the method of the present invention (PE_SIREN) can achieve higher prediction accuracy and maintain a faster convergence speed than the sine representation network and the feedforward neural network. At the same time, the addition of position information encoding has a significant improvement in the performance of deep neural networks.

[0119] 2) If Figure 4 As shown, the present invention compares the five methods mentioned at three sampling rates. All five methods effectively solve the CIR prediction problem to a certain extent and achieve better performance with more training data. Among them, without a periodic activation function, the feedforward neural network method performs worse than the other four methods at all sampling rates. The results of the LSTM and CGAN networks show that these two network structures are unable to learn scattering environments because the prediction results rely on the channel correlation between adjacent sampling points. At a sampling rate of 80%, the NMSEs of the feedforward neural network, the sinusoidal representation network, and the method of the present invention (PE_SIREN) are 0.0243754, 0.0103265, and 0.00959294, respectively. At a sampling rate of 40%, the NMSEs of the three solutions are 0.0448695, 0.0145865, and 0.0128514, respectively. At a sampling rate of 20%, the NMSEs are 0.0729376, 0.0225399, and 0.0179453, respectively. These results demonstrate that the PE_SIREN learning architecture delivers superior performance compared to traditional learning architectures and popular networks. Furthermore, even at lower sampling rates, the prediction accuracy of our proposed method (PE_SIREN) remains unchanged. This further validates the feasibility of our proposed approach.

[0120] 3) If Figure 5 、 6 Figures 7 and 8 show a comparison of the predicted results of the three networks with the original measured power delay profile (PDP) at a low sampling rate, 1 km from the base station. It can be calculated that at 4 km, the MSEs of the predicted results of the proposed method (PE_SIREN), the sinusoidal representation network, and the feedforward neural network compared to the original data (Original_PDP) are 11.227120 dB, 14.919109 dB, and 27.023975 dB, respectively.

[0121] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A vehicle-ground wireless channel prediction method based on position information coding and periodic nonlinear neural network, characterized in that: The following steps are involved: S1: Collect train location information and wireless channel instances, and perform position encoding on the collected channel instances; map the channel instances to a high-dimensional Fourier feature space; The wireless channel instance includes one or more uplink / downlink historical communication data and historical channel information generated during the communication between the train and the base station; the historical communication data includes channel state information CSI; the historical channel information includes historical channel state information and historical channel impulse response; The collected channel instances are position-coded as follows: in, and is the scale factor of the position encoding, through the scale factor and To adjust the frequency of each dimension, generating a series of frequencies from low to high; is an adjustable scale factor, Indicates the index of the encoding, is the total number of frequencies; Indicates the The loss function of each path; S2: Based on position information encoding and periodic nonlinear neural network, a generative deep neural network is constructed to learn and model channel instances; Step S2 includes: S21: Integrating periodic nonlinear neural networks and positional encoding to construct generative deep neural networks; The generative deep neural network includes a position information encoding layer and a periodic nonlinear neural network layer; The position information encoding layer expands the input position, time, and frequency parameters into high-dimensional features through position encoding and provides them to the periodic nonlinear neural network layer for real-time channel prediction; The periodic nonlinear neural network refers to a neural network using a nonlinear activation function; S22: learning and modeling channel instances; The overall training process is as follows: First, data preparation is performed by loading a real channel impulse response (CIR) dataset and decomposing the complex value into real and imaginary parts. Input parameters such as position, time, and frequency are expanded into high-dimensional features through position encoding. Model initialization, initializing the weights and biases of the periodic nonlinear neural network; the weight initialization scheme is as follows: in, is the input dimension, is the uniform distribution function; During the training process, a supervised learning method is used, using wireless channel instances, namely channel impulse responses (CIRs), as labels. The prediction accuracy and generalization ability of the model are improved by optimizing the loss function. The loss function is: in, is the output dimension, is the training label, is the predicted value; Represents the average distance between the target value and the predicted value in all output dimensions; Back propagation calculates gradients and updates network parameters; S3: Apply the generative deep neural network model to perform real-time channel prediction.

2. The method according to claim 1, characterized in that In step S1, the train location information includes the acquisition of multiple types of train location information, and the multiple types of train location information include the train location information acquired through transponders, inertial navigation, GPS, and Beidou.

3. The method according to claim 1, characterized in that The periodic nonlinear neural network is a sinusoidal representation network SIREN, as follows: Among them, in SIREN, represents the input of the neural network, represents the output of the neural network, Represents a specific connection between two layers, Represents the network layer, which consists of a weight matrix ,enter , bias as well as It consists of non-linear activation functions.

4. The method according to claim 1, characterized in that In step S3, the neural network model trained in step S2 is used to perform real-time channel prediction. The train position, environment, and channel parameters are input and the predicted real-time channel impulse response (CIR) is output, including the real and imaginary parts of the CIR, to provide support for dynamic adjustment of the communication system.

5. The method according to claim 1, characterized in that: In step S3, the real-time channel prediction includes predicting the future wireless channel impulse response based on the real-time position information and running speed dynamic data of the train, and dynamically adjusting the parameters of the communication system to optimize the signal quality.

Images