Imitation radio frequency division multiplexing channel estimation and symbol detection method and system based on deep learning

Abstract

The invention discloses an imitated radio frequency division multiplexing channel estimation and symbol detection method and system based on deep learning, and the method comprises the steps: S1, modulating a transmission symbol in a communication process through employing an imitated radio frequency division multiplexing technology, transmitting the modulated transmission symbol through a channel, and obtaining an imitated radio frequency division multiplexing receiving signal at a receiving end; s2, constructing a deep neural network model, taking an I path and a Q path in the simulated radio frequency division multiplexing receiving signal as the input of a neural network, taking a transmission symbol as the expected output of the neural network, and training the neural network, so that the neural network has the prediction capability on the transmission data symbol; and S3, after training of the deep neural network model is completed, inputting the simulated radio frequency division multiplexing signals received in real time in the communication process into the trained neural network model to obtain a predicted value of a transmission data symbol.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technologies, and in particular to a method and system for RF division multiplexing channel estimation and symbol detection based on deep learning. Background Art

[0002] With the development of science and technology, people's demand for wireless communication under high-speed mobility is gradually increasing, such as high-speed rail, vehicle networking, etc. Therefore, the next generation of wireless communication has put forward higher requirements for wireless communication under high-speed mobility. In 4G and 5G wireless communication technologies, orthogonal frequency division multiplexing has been widely used due to its good anti-multipath fading ability, high spectrum utilization and excellent performance under linear time-invariant channels. However, in linear time-varying channels, the orthogonality of orthogonal frequency division multiplexing will drop sharply, resulting in interference between subcarriers, which greatly reduces the performance of orthogonal frequency division multiplexing. In order to obtain good performance in linear time-varying channels and meet the needs of wireless communication under high-speed mobility, many new signal modulation technologies have been proposed, and orthogonal frequency division multiplexing was born in this context.

[0003] The RFDM technology is based on the discrete affine Fourier transform, which is a generalization of the discrete Fourier transform. It can achieve full diversity because its linear frequency modulation pulse parameters can adapt to the channel characteristics, thereby realizing the complete delay Doppler representation of the channel in the discrete affine Fourier transform domain. Experiments show that RFDM has excellent performance under linear time-varying channels and can meet the needs of wireless communications under high-speed mobility.

[0004] With the rapid development of computer computing power in recent years, deep learning technology has also developed rapidly. Deep learning technology is widely used in many fields, such as computer vision, natural language processing, etc. At the same time, deep learning technology is also widely used in the field of wireless communications, and its potential in the field of wireless communications is constantly being explored. However, there is still a lack of research on the application of deep learning in RFDM channel estimation and symbol detection. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a method for channel estimation and symbol detection of radio frequency division multiplexing based on deep learning, the method comprising:

[0006] Step S1, using the radio frequency division multiplexing technique to modulate the transmission symbols in the communication process, and after the channel transmission, the receiving end obtains the radio frequency division multiplexing receiving signal;

[0007] Step S2, constructing a deep neural network model, taking the I path and the Q path in the simulated RF division multiplexing received signal as the input of the deep neural network, and the transmission symbol as the expected output of the deep neural network, training the deep neural network to obtain a trained deep neural network model;

[0008] Step S3: inputting the RFDM signal received in real time during the communication process into the trained neural network model to obtain the predicted value of the transmission data symbol.

[0009] Optionally, in step S1, the transmission symbol consists of three parts: a pilot, a guard interval and data, wherein the pilot is known to both the sender and receiver and is used to estimate the channel when receiving the signal; the data carries information to be sent; and the guard interval is between the pilot and the data and is used to prevent interference between the pilot and the data.

[0010] Optionally, in step S1, the transmission symbols in the communication process are modulated using the RFDM technique, and the receiving end obtains the RFDM received signal through channel transmission, and the process specifically includes:

[0011] The pseudo-RFDM received signal under integer Doppler frequency shift is expressed as:

[0012] 0≤p≤N-1,q=(p+loc i ) N·

[0013] Wherein, N is the number of linear frequency modulation subcarriers; x[q] is the transmission symbol modulated by modulation methods such as phase shift keying (PSK) and quadrature amplitude modulation (QAM); is the additive white Gaussian noise generated during signal transmission; P is the number of transmission channels; h i is the complex gain of the ith channel; l i is the normalized delay of the ith channel; loc i =(-α i +2Nc1l i ) N ; i is the normalized integer Doppler shift of the ith channel; to achieve full diversity of RFDM, in the values ​​of c1 and c2, c1=(2ɑ max +1) / 2N, c2 is any irrational number or a rational number much smaller than 1 / 2N,

[0014] Optionally, in step S2, the constructed deep neural network model includes three layers, an input layer, a hidden layer and an output layer, and the simulation parameters are set as follows: the neural network has 8 layers, and the number of neuron nodes in each layer is 64, 128, 64, 32, 16, 8, 4, and 1 respectively; the activation function of the hidden layer of the deep neural network is ReLu, the activation function of the output layer is Sigmoid, and Adam is used as the neural network optimizer.

[0015] The present invention also discloses a deep learning-based RFDM channel estimation and symbol detection system, the system comprising: a signal modulation module, a model training module and a symbol detection module;

[0016] The signal modulation module is used to modulate the transmission symbols in the communication process using the radio frequency division multiplexing technology, and the receiving end obtains the radio frequency division multiplexing receiving signal after the channel transmission;

[0017] The model training module is used to construct a deep neural network model, take the I path and Q path in the simulated radio frequency division multiplexing received signal as the input of the deep neural network, and take the transmission symbol as the expected output of the deep neural network, and train the deep neural network to obtain a trained deep neural network model;

[0018] The symbol detection module is used to input the RFDM signal received in real time during the communication process into the trained neural network model to obtain the predicted value of the transmission data symbol.

[0019] Optionally, the transmission symbol in the signal modulation module consists of three parts: a pilot, a guard interval and data, wherein the pilot is data known to both the sender and receiver and is used to estimate the channel when receiving the signal; the data carries information to be sent; and the guard interval is between the pilot and the data and is used to prevent interference between the pilot and the data.

[0020] Optionally, the workflow of the signal modulation module specifically includes:

[0021] The pseudo-RFDM received signal under integer Doppler frequency shift is expressed as:

[0022]

[0023] Wherein, N is the number of linear frequency modulation subcarriers; x[q] is the transmission symbol modulated by modulation methods such as phase shift keying (PSK) and quadrature amplitude modulation (QAM); is the additive white Gaussian noise generated during signal transmission; P is the number of transmission channels; h i is the complex gain of the ith channel; l i is the normalized delay of the ith channel; loc i=(-ɑ i +2Nc1l i ) N ; i is the normalized integer Doppler shift of the ith channel; to achieve full diversity of RFDM, in the values ​​of c1 and c2, c1=(2α max +1) / 2N, c2 is any irrational number or a rational number much smaller than 1 / 2N,

[0024] Optionally, in the model training module, the constructed deep neural network model includes three layers, an input layer, a hidden layer and an output layer, and the simulation parameters are set as follows: the neural network has 8 layers, and the number of neuron nodes in each layer is 64, 128, 64, 32, 16, 8, 4, and 1 respectively; the activation function of the hidden layer of the deep neural network is ReLu, the activation function of the output layer is Sigmoid, and Adam is used as the neural network optimizer.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] (1) The present invention is designed based on the application of deep learning in RFDM technology, filling the gap in deep learning in RFDM channel estimation and symbol detection.

[0027] (2) The present invention combines the two steps of channel estimation and symbol detection into one, and only one step is required to obtain the predicted value of the input data, thus simplifying the system process.

[0028] (3) The present invention fully utilizes the advantages of deep neural networks and demonstrates excellent performance in RFDM symbol detection; at the same time, the system has high robustness and can maintain good performance even in scenarios with large interference. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0030] Figure 1 It is a flow chart of a method for channel estimation and symbol detection of radio frequency division multiplexing based on deep learning of the present invention;

[0031] Figure 2 It is a schematic diagram of input symbols of the deep learning-based RFDM channel estimation and symbol detection method of the present invention;

[0032] Figure 3It is a system block diagram of the deep learning-based RFDM channel estimation and symbol detection method of the present invention;

[0033] Figure 4 It is a schematic diagram of the architecture of the deep learning-based RFDM channel estimation and symbol detection method of the present invention;

[0034] Figure 5 Schematic diagram of bit error rate performance when ζ=0 in an embodiment of the present invention;

[0035] Figure 6 Schematic diagram of bit error rate performance when ζ=2 in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0037] Embodiment 1

[0038] A deep learning-based RFDM channel estimation and symbol detection method, such as Figure 1 As shown, the method includes:

[0039] Step S1: modulate the transmission symbols in the communication process using the RFDM technique, transmit the modulated symbols through a channel, and obtain a RFDM received signal at the receiving end.

[0040] A large number of transmission symbols containing pilot, data and protection interval are modulated by RFDM and transmitted through a specific channel to obtain RFDM output signals. A large number of input and output signals constitute a data set for deep neural network training.

[0041] In this embodiment, the RFDM received signal under integer Doppler frequency shift is expressed as:

[0042]

[0043] Wherein, N is the number of linear frequency modulation subcarriers; x[q] is the input signal modulated by modulation methods such as phase shift keying (PSK) and quadrature amplitude modulation (QAM); is the additive white Gaussian noise generated during signal transmission; P is the number of transmission channels; h i is the complex gain of the ith channel; l i is the normalized delay of the ith channel; loc i =(-α i +2Nc1l i ) N ; α iis the normalized integer Doppler shift of the ith channel; to achieve full diversity of RFDM, in the values ​​of c1 and c2,

[0044] c1=(2α max +1) / 2N, c2 is any irrational number or a rational number much smaller than 1 / 2N,

[0045] Under fractional Doppler shift, the received signal will be expressed as:

[0046]

[0047] Among them, a i is the fractional part of the Doppler shift. For the case of fractional Doppler shift, c1=(2(α max +ζ)+1) / 2N, ζ is a positive integer, which is used to reduce the interference between different channels in the case of fractional Doppler frequency shift, but it will also increase the size of the guard interval and reduce the spectrum utilization.

[0048] like Figure 2 As shown, this embodiment provides a diagram of a symbol structure of radio frequency division multiplexing transmission to illustrate the symbol distribution of the signal transmitted by the system. It is composed of three symbols: pilot, data, and guard interval. Among them, the pilot is known to both the sender and the receiver, and is used to estimate the channel when receiving the signal; the data carries the information to be sent; the guard interval is between the pilot and the data, and is used to prevent interference between the pilot and the data.

[0049] Step S2, construct a deep neural network model, take the I path and Q path in the simulated RF division multiplexing received signal as the input of the neural network, and the transmission symbol as the expected output of the neural network, and train the neural network so that it has the ability to predict the transmission data symbol.

[0050] In the symbol detection at the receiving end: the deep neural network trained in the offline training phase is deployed to the receiving end, and the I and Q paths of the received RFDM signal are input into the trained neural network model. The predicted value of the transmission symbol can be obtained without estimating the channel in advance.

[0051] Construct a training data set for the deep neural network. The data set consists of the RFDM received signal and the transmission symbol. The input of the deep neural network is the I and Q paths of the RFDM received signal, and the transmission symbol is the expected output of the deep neural network. In order to improve the accuracy of symbol detection, multiple deep neural networks are used, and each deep neural network is responsible for detecting a data symbol. After determining the structure of the deep neural network, you can use the data set for training. Figure 3As shown, the input of this embodiment is the real part and the imaginary part of the received symbol of the simulated radio frequency division multiplexing. By deploying multiple deep neural networks, each network is responsible for outputting a data symbol, thereby increasing the prediction accuracy of the system.

[0052] Step S3: After the deep neural network model is trained, the simulated RF division multiplexing signal received in real time during the communication process is input into the trained neural network model to obtain the predicted value of the transmission symbol.

[0053] like Figure 4 As shown, this embodiment predicts the RFDM transmission symbols by deploying a pre-trained deep neural network at the signal receiving end. This embodiment includes two stages, namely an offline training stage and an online deployment stage. In the offline training stage, a deep neural network is constructed, and the neural network is trained using a training data set consisting of transmitted and received RFDM signals, so that the output of the neural network is closest to the actual transmitted data symbols. In the online deployment stage, the trained deep neural network model is deployed at the receiving end, and the user inputs the RFDM signal received in real time during the communication process into the pre-trained neural network, and the predicted value of the transmission symbol can be output.

[0054] Embodiment 2

[0055] The present invention discloses a deep learning-based radio frequency division multiplexing channel estimation and symbol detection system, the system comprising: a signal modulation module, a model training module and a symbol detection module;

[0056] The signal modulation module is used to modulate the transmission symbols in the communication process using the RFDM-like technology, and after the channel transmission, the receiving end obtains the RFDM-like receiving signal.

[0057] A large number of transmission symbols containing pilot, data and protection interval are modulated by RFDM and transmitted through a specific channel to obtain RFDM output signals. A large number of input and output signals constitute a data set for deep neural network training.

[0058] In this embodiment, the RFDM received signal under integer Doppler frequency shift is expressed as:

[0059]

[0060] Wherein, N is the number of linear frequency modulation subcarriers; x[q] is the transmission symbol modulated by modulation methods such as phase shift keying (PSK) and quadrature amplitude modulation (QAM); is the additive white Gaussian noise generated during signal transmission; P is the number of transmission channels; h i is the complex gain of the ith channel; l iis the normalized delay of the ith channel; loc i =(-α i +2Nc1l i ) N ; α i is the normalized integer Doppler shift of the ith channel; to achieve full diversity of RFDM, in the values ​​of c1 and c2,

[0061] c1=(2α max +1) / 2N, c2 is any irrational number or a rational number much smaller than 1 / 2N,

[0062] Under fractional Doppler shift, the received signal will be expressed as:

[0063]

[0064] Among them, a i is the fractional part of the Doppler shift. For the case of fractional Doppler shift, c1=(2(α max +ζ)+1) / 2N, ζ is a positive integer used to reduce the interference between different channels in the case of fractional Doppler frequency shift, but it will also increase the size of the guard interval and reduce the spectrum utilization.

[0065] The model training module is used to construct a deep neural network model, take the I path and Q path in the simulated RF division multiplexing received signal as the input of the deep neural network, and use the transmission symbol as the expected output of the deep neural network to train the deep neural network to obtain a trained deep neural network model.

[0066] In order to improve the accuracy of symbol detection, we used multiple deep neural networks with the same structure. The constructed deep neural network model consists of three layers: input layer, hidden layer and output layer. The simulation parameters are set as follows: the neural network has 8 layers, and the number of neuron nodes in each layer is 64, 128, 64, 32, 16, 8, 4, and 1 respectively; the activation function of the hidden layer of the deep neural network is ReLu, the activation function of the output layer is Sigmoid, and Adam is used as the neural network optimizer. Each neural network is responsible for detecting a transmission symbol. After determining the structure of the deep neural network, you can use the data set for training.

[0067] The symbol detection module is used to input the RFDM signal received in real time during the communication process into the trained neural network model to obtain the predicted value of the transmission symbol.

[0068] After the neural network training is completed, the trained neural network can be deployed to the receiving end. When the user receives the RFDM signal, it is input into the neural network to directly predict the transmission symbol without estimating the channel in advance.

[0069] Embodiment 3

[0070] In order to illustrate the technical advancement of the method of this embodiment, the bit error rate performance of the deep learning-based RFDM channel estimation and symbol detection method proposed in this embodiment is simulated on a Python platform.

[0071] The simulation parameters are set as follows: the carrier frequency and bandwidth are 4 GHz and 32 kHz respectively; the number of linear frequency modulation subcarriers N = 32; the number of guard intervals in the input symbol to prevent interference between the pilot and the data is 2Q = 2(2(α max +ζ)+1)(l max +1)-1), where there are Q guard intervals on the left and right of the pilot. For the case of integer Doppler shift, the value of ζ is always 0; for the case of fractional Doppler shift, the increase of ζ can reduce the interference between different channels, but it will also increase the number of guard intervals, thereby reducing the spectrum efficiency. The maximum delay of the channel is 31.25us, and the maximum delay after normalization is 1; the maximum Doppler shift of the channel is 1kHz, and the maximum Doppler shift after normalization is 1; the number of channels P = 2; the complex gain of the channel is generated by an independent complex random variable with zero mean and 1 / P variance; the normalized delay of the channel is 1 = [0, 1]; the normalized Doppler shift of the channel in the i-th channel is v i =v max cosθ i , where v max represents the maximum Doppler shift, θ i Uniformly distributed on [-π,π]; the signal-to-noise ratio (SNR) of the pilot p =30dB. In terms of the setting of neural network parameters, including input layer, hidden layer and output layer, the neural network has 8 layers. From the input layer to the output layer, the number of neuron nodes in each layer is 64, 128, 64, 32, 16, 8, 4, and 1 respectively; the activation function of the hidden layer of the neural network is ReLu, the activation function of the output layer is Sigmoid, and Adam is used as the optimizer of the neural network. Specifically, Figure 5 and Figure 6 The performance of the method is demonstrated in the case of fractional Doppler shift, where Figure 5 The bit error rate performance under ζ = 0 is shown. Figure 6The bit error rate performance under ζ = 2 is shown. The bit error rate performance of the deep learning-based RFDM channel estimation and symbol detection method is obtained from simulation results and compared with the existing embedded channel estimation method and the performance under ideal channel state information. Among them, both schemes use linear minimum mean square error (LMMSE) for symbol detection.

[0072] Through simulation, it can be seen that this embodiment has the following technical advancements:

[0073] Compared with the traditional embedded channel estimation and symbol detection technology, the present invention has higher robustness. In the case of less interference, such as ζ=2, the present invention has comparable performance with the scheme, but when the interference increases, such as ζ=0, the present invention can still maintain better performance, while the performance of the traditional method has declined significantly.

[0074] This embodiment discloses a method for channel estimation and symbol detection of simulated RF division multiplexing based on deep learning, which includes the following steps: Step S1, use simulated RF division multiplexing technology to modulate the transmission symbol in the communication process, and transmit it through the channel to obtain a simulated RF division multiplexing received signal at the receiving end; Step S2, build a deep neural network model, use the I and Q paths in the simulated RF division multiplexing received signal as the input of the neural network, and the transmission symbol as the expected output of the neural network, and train the neural network to enable it to have the ability to predict the transmission data symbol. Among them, in order to improve the accuracy of symbol detection, we use multiple deep neural networks, each of which is responsible for detecting a data symbol. The method of using multiple neural networks has greatly improved the system's ability to detect simulated RF division multiplexing symbols; Step S3, after the deep neural network model is trained, the received simulated RF division multiplexing signal is input into the trained neural network model, and the predicted value of the transmission symbol can be obtained without pre-estimating the channel.

[0075] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for channel estimation and symbol detection of radio frequency division multiplexing based on deep learning, characterized in that: The method comprises: Step S1, using the radio frequency division multiplexing technique to modulate the transmission symbols in the communication process, and after the channel transmission, the receiving end obtains the radio frequency division multiplexing receiving signal; Step S2, constructing a deep neural network model, taking the I path and the Q path in the simulated RF division multiplexing received signal as the input of the deep neural network, and the transmission symbol as the expected output of the deep neural network, training the deep neural network to obtain a trained deep neural network model; Step S3: inputting the RFDM signal received in real time during the communication process into the trained neural network model to obtain the predicted value of the transmission data symbol.

2. The method for channel estimation and symbol detection based on deep learning for radio frequency division multiplexing according to claim 1, characterized in that: In step S1, the transmission symbol consists of three parts: a pilot, a guard interval and data, wherein the pilot is known to both the sender and the receiver and is used to estimate the channel when receiving the signal; the data carries the information to be sent; and the guard interval is between the pilot and the data and is used to prevent interference between the pilot and the data.

3. The method for RFDM channel estimation and symbol detection based on deep learning according to claim 1, characterized in that: In step S1, the transmission symbols in the communication process are modulated using the RFDM technique, and the receiving end obtains the RFDM received signal through channel transmission, which specifically includes: The pseudo-RFDM received signal under integer Doppler frequency shift is expressed as: 0≤p≤N-1,q=(p+loc i ) N Wherein, N is the number of linear frequency modulation subcarriers; x[q] is the transmission symbol modulated by modulation methods such as phase shift keying (PSK) and quadrature amplitude modulation (QAM); is the additive white Gaussian noise generated during signal transmission; P is the number of transmission channels; h i is the complex gain of the ith channel; l i is the normalized delay of the ith channel; loc i =(-α i +2Nc1l i ) N ; α i is the normalized integer Doppler shift of the ith channel; to achieve full diversity of RFDM, in the values ​​of c1 and c2, c1=(2α max +1) / 2N, c2 is any irrational number or a rational number much smaller than 1 / 2N, 4. The method for RFDM channel estimation and symbol detection based on deep learning according to claim 1, characterized in that: In step S2, the constructed deep neural network model includes three layers, an input layer, a hidden layer and an output layer, and the simulation parameters are set as follows: the neural network has 8 layers, and the number of neuron nodes in each layer is 64, 128, 64, 32, 16, 8, 4, and 1 respectively; the activation function of the hidden layer of the deep neural network is ReLu, the activation function of the output layer is Sigmoid, and Adam is used as the neural network optimizer.

5. A deep learning-based RFDM channel estimation and symbol detection system, wherein the symbol detection system is used to implement the symbol detection method according to any one of claims 1 to 4, characterized in that: The system includes: a signal modulation module, a model training module and a symbol detection module; The signal modulation module is used to modulate the transmission symbols in the communication process using the radio frequency division multiplexing technology, and the receiving end obtains the radio frequency division multiplexing receiving signal after the channel transmission; The model training module is used to construct a deep neural network model, take the I path and Q path in the simulated radio frequency division multiplexing received signal as the input of the deep neural network, and take the transmission symbol as the expected output of the deep neural network, and train the deep neural network to obtain a trained deep neural network model; The symbol detection module is used to input the RFDM signal received in real time during the communication process into the trained neural network model to obtain the predicted value of the transmission data symbol.

6. The deep learning-based RFDM channel estimation and symbol detection system according to claim 5, characterized in that: The transmission symbol in the signal modulation module consists of three parts: a pilot, a guard interval and data, wherein the pilot is known to both the sender and the receiver and is used to estimate the channel when receiving the signal; the data carries the information to be sent; and the guard interval is between the pilot and the data and is used to prevent interference between the pilot and the data.

7. The deep learning-based RFDM channel estimation and symbol detection system according to claim 5, characterized in that: The workflow of the signal modulation module specifically includes: The pseudo-RFDM received signal under integer Doppler frequency shift is expressed as: 0≤p≤N-1,q=(p+loc i ) N . Wherein, N is the number of linear frequency modulation subcarriers; x[q] is the transmission symbol modulated by modulation methods such as phase shift keying (PSK) and quadrature amplitude modulation (QAM); is the additive white Gaussian noise generated during signal transmission; P is the number of transmission channels; h i is the complex gain of the ith channel; l i is the normalized delay of the ith channel; loc i =(-α i +2Nc1l i ) N ; α i is the normalized integer Doppler shift of the ith channel; to achieve full diversity of RFDM, in the values ​​of c1 and c2, c1=(2α max +1) / 2N, c2 is any irrational number or a rational number much smaller than 1 / 2N, 8. The deep learning-based RFDM channel estimation and symbol detection system according to claim 5, characterized in that: In the model training module, the constructed deep neural network model includes three layers, an input layer, a hidden layer and an output layer. The simulation parameters are set as follows: the neural network has 8 layers, and the number of neuron nodes in each layer is 64, 128, 64, 32, 16, 8, 4, and 1 respectively; the activation function of the hidden layer of the deep neural network is ReLu, the activation function of the output layer is Sigmoid, and Adam is used as the neural network optimizer.

Images