A deep learning-assisted OFDM channel estimation and signal detection method

Abstract

The present invention relates to the field of high-speed wireless communication, and in particular to a deep learning-assisted OFDM channel estimation and signal detection method. The present invention first generates a data set through an OFDM system, then uses a self-normalizing network (SNN) to build a channel estimation model, and uses a convolutional neural network (CNN) and a bidirectional gated recurrent unit (BiGRU) to build a signal detection model; secondly, the data set is input into the neural network, the model is trained, and the network is optimized using an Adam optimizer and a mean square error loss function, and finally compared with LS, BiLSTM, GRU and DNN+BiLSTM models. The present invention effectively solves the problems of low channel estimation and signal detection accuracy, severe network loss and poor bit error rate performance in OFDM systems with a small number of cyclic prefixes and pilots. The simulation results show that in the case of fewer cyclic prefixes and pilots, the performance of this method is better than that of LS, BiLSTM, GRU and DNN+BiLSTM models.

Description

Technical Field

[0001] The present invention relates to the field of intelligent communications, and in particular to a deep learning-assisted OFDM channel estimation and signal detection method for high-speed wireless communications. Background Art

[0002] OFDM plays an important role in wireless communication systems due to its high spectrum efficiency and strong resistance to multipath channel fading. Wireless channels are affected by frequency selective fading, shadow fading, etc., which makes the signal transmission path diverse and complex, resulting in a high bit error rate at the receiving end and interference with the received signal. In order to meet social needs and efficiently utilize limited channel resources, channel estimation and signal detection are essential steps.

[0003] OFDM is the core technology of 4G communication and also the key technology of 5G communication. Its strong ability to resist multipath channel fading makes OFDM an indispensable part in the field of wireless communication. In terms of channel estimation, the least squares (LS) estimation is a popular algorithm used today. Its low complexity is a significant feature of LS, but it also ignores the interference between subcarriers, which makes LS show poor bit error rate performance when the number of pilots and cyclic prefixes is small.

[0004] At present, the combination of communication technology and deep learning has attracted the attention of many scholars, which has led to the rapid development of intelligent communication technology. The neural networks commonly used in deep learning include deep neural networks (DNN), convolutional neural networks (CNN), recurrent neural networks (RNN), etc. Among them, the long short-term memory network (LSTM) in RNN is deeply loved by communication scholars, but the internal structure of the LSTM network is complex, which limits its development.

[0005] Invention patent CN 112600772A combines CNN and DNN, which alleviates the impact of parameters to a certain extent. However, CNN and DNN cannot retain the above information in forward propagation, and the channel estimation and signal detection accuracy in OFDM systems with a small number of pilots or cyclic prefixes are poor.

[0006] Invention patent CN 111865863 A proposes an OFDM signal detection method based on RNN, which uses the LSTM model to learn time series. However, the LSTM structure is complex, takes a long time to train, and lacks the ability to learn the information of the previous moment. When applied to OFDM systems with a high number of subcarriers, it exhibits poor bit error rate performance.

[0007] Invention patent CN 111510402B uses a fully connected neural network to replace the channel estimation part in the OFDM system. However, the network parameters designed by this invention are too many, the calculation complexity is large, and the DNN has a poor ability to remember information. There is also the risk of gradient vanishing and gradient exploding. Summary of the invention

[0008] In view of the above problems, the present invention proposes a deep learning-assisted OFDM channel estimation and signal detection method, which adopts a self-normalizing network (SNN) to construct a channel estimation model, and adopts a convolutional neural network (CNN) and a bidirectional gated recurrent unit (BiLSTM) to construct a signal detection model. SNN effectively solves the problems of gradient explosion and gradient disappearance; CNN denoises the data while reducing the computational complexity; BiGRU establishes the connection between the output at the current moment and the states at the previous and next moments.

[0009] The present invention provides the following technical solution: a deep learning-assisted OFDM channel estimation and signal detection method, comprising the following steps:

[0010] Step 1: Build an OFDM system and generate a data set;

[0011] Step 2: Construct a channel estimation model based on a self-normalizing network (SNN);

[0012] Step 3: Build a signal detection model based on convolutional neural network (CNN) and bidirectional gated recurrent unit (BiGRU);

[0013] Step 4: Set the hyperparameters required for the three neural networks: SNN, CNN, and BiGRU;

[0014] Step 5: Use the data set to train three neural networks: SNN, CNN, and BiGRU;

[0015] Step 6: Calculate the accuracy curve, network loss curve and training time. If the accuracy curve does not fit or the loss curve does not converge, reset the hyperparameters and proceed to step 4. If the accuracy curve fits and the loss curve converges, proceed to step 7.

[0016] Step 7: Input the test set into the deep learning-assisted OFDM channel estimation and signal detection model for online deployment.

[0017] In step 1, the initial signal transmitted by the transmitter is modulated, pilot inserted, serial-to-parallel converted, inverse fast Fourier transformed, cyclic prefix added, and parallel-to-serial converted to generate the required data set.

[0018] In step 2, in the SNN-based channel estimation model, the input data is first estimated by the least squares method (LS), and its expression is:

[0019]

[0020] in is the channel impulse response coefficient corresponding to the nth subcarrier, x p (n) is the pilot signal corresponding to the nth subcarrier transmitted, y p (n) is the pilot signal corresponding to the received n-th subcarrier.

[0021] Split The real and imaginary parts of are used as the input of SNN. The SNN channel estimation model is constructed using three scaled exponential linear units (SELU), deep neural network (DNN) and three BN layers, and the Selu function is selected as the hidden layer activation function, and its expression is:

[0022]

[0023] Among them, λ is 1.051, α is 1.673, and on the positive x-axis, the input data will be expanded to λ times; on the negative x-axis, some information will be retained.

[0024] In step 3, in the signal detection model based on CNN+BiGRU, the input data is firstly subjected to zero forcing (ZF) detection, and its expression is:

[0025]

[0026] in is the signal corresponding to the predicted nth subcarrier, y D (n) is the signal corresponding to the received n-th subcarrier, is the channel impulse response corresponding to the nth subcarrier output by the SNN-based channel estimation model.

[0027] Secondly, a signal detection model is constructed using 1 CNN layer, 1 BN layer and 5 BiGRU layers. CNN denoises the data through local connections and weight sharing, while reducing the computational complexity. The BN layer standardizes the data. Relu is selected as the activation function of CNN, and its expression is:

[0028] Relu(x)=max(0,x)

[0029] CNN is connected to a 5-layer BiGRU and uses Tanh as the activation function, which is expressed as:

[0030]

[0031] Finally, the Adam optimizer is used to optimize the neural network, and the mean square error function is selected as the loss function, which is expressed as:

[0032]

[0033] in, is the output value of the model, i.e. the transmitted signal, f(x) is the label of the model, i.e. the received signal, and N is the number of subcarriers.

[0034] In step 4, the specific hyperparameters are as follows: in SNN, the number of neurons in the hidden layer is 300 and 100 respectively, and the number of neurons in the output layer is 128; in CNN, 512 convolution kernels are set, and the size of each convolution kernel is 64; in BiGRU, the number of neurons is 64, 50, 32, 24, and 16 respectively; the number of epochs is 100, and steps_per_epoch is 50. An epoch refers to inputting all the data into the network to complete a forward calculation and back propagation. Since the cycle for completing an epoch training is long (the amount of data is large), the computer cannot handle inputting all the data at once, so it is divided into multiple batches. So why do we need to train multiple epochs? We know that when training a network, it is not enough to iteratively train all the data once, and repeated training is required for the network to converge.

[0035] In step 5, the data set is divided into a training set and a test set in a ratio of 8:2, and the neural network is trained under conditions of different numbers of pilot frequencies and different numbers of cyclic prefixes.

[0036] In step 7, the modulation method adopted by the deep learning-assisted OFDM channel estimation and signal detection model is 64QAM. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a flow chart of a specific implementation mode of the present invention.

[0038] Figure 2 This is the SNN+CNN+BiGRU network structure diagram of the present invention.

[0039] Figure 3 This is a bit error rate comparison chart when the number of subcarriers is 64, the number of cyclic prefixes is 4, and the number of pilots is 4.

[0040] Figure 4 This is a comparison chart of bit error rates when the number of subcarriers is 64, the number of cyclic prefixes is 4, and the number of pilots is 64.

[0041] Figure 5This is a comparison chart of bit error rates when the number of subcarriers is 64, the number of cyclic prefixes is 16, and the number of pilots is 4.

[0042] Figure 6 This is a comparison chart of the mean square error (MSE) between the method of the present invention and BiLSTM, GRU, and DNN+BiLSTM;

[0043] Figure 7 This is a comparison chart of the accuracy of the method of the present invention, BiLSTM, GRU, and DNN+BiLSTM. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the specific embodiments of the present invention to clearly and completely describe the technical solutions in the specific embodiments of the present invention. Obviously, the specific embodiment described is only one specific embodiment of the present invention, not all specific embodiments. Based on the specific embodiments of the present invention, all other specific embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "include" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or their combinations;

[0046] The deep learning-assisted OFDM channel estimation and signal detection method proposed in the present invention comprises the following steps:

[0047] Step 1: Build an OFDM system. The initial signal transmitted by the transmitter undergoes modulation, pilot insertion, serial-to-parallel conversion, inverse fast Fourier transform, cyclic prefix addition, and parallel-to-serial conversion to generate the required data set.

[0048] Step 2: Construct a channel estimation model based on a self-normalizing network (SNN). In the SNN-based channel estimation model, the input data is first estimated using the least squares (LS) method, and its expression is:

[0049]

[0050] in is the channel impulse response coefficient corresponding to the nth subcarrier, xp(n) is the pilot signal corresponding to the transmitted nth subcarrier, and yp(n) is the pilot signal corresponding to the received nth subcarrier.

[0051] Split The real and imaginary parts of are used as the input of SNN. The SNN channel estimation model is constructed using three scaled exponential linear units (SELU), deep neural network (DNN) and three BN layers, and the Selu function is selected as the hidden layer activation function, and its expression is:

[0052]

[0053] Among them, λ is 1.051, α is 1.673, and on the positive x-axis, the input data will be expanded to λ times; on the negative x-axis, some information will be retained, that is, the input data that satisfies the hidden layer activation function will be retained.

[0054] Step 3: Construct a signal detection model based on convolutional neural network (CNN) and bidirectional gated recurrent unit (BiGRU). In the signal detection model based on CNN+BiGRU, the input data is firstly subjected to zero forcing (ZF) detection, and its expression is:

[0055]

[0056] in is the signal corresponding to the predicted n-th subcarrier, yD(n) is the signal corresponding to the received n-th subcarrier, is the channel impulse response corresponding to the nth subcarrier output by the SNN-based channel estimation model.

[0057] Secondly, a signal detection model is constructed using 1 CNN layer, 1 BN layer and 5 BiGRU layers. CNN denoises the data through local connections and weight sharing, while reducing the computational complexity. The BN layer standardizes the data. Relu is selected as the activation function of CNN, and its expression is:

[0058] Relu(x)=max(0,x)

[0059] CNN is connected to 5 layers of BiGRU, and the calculation formula of the GRU network is:

[0060] z t =Sigmoid(w z [h t-1 ,x t ]+b z )

[0061] r t =Sigmoid(w r [h t-1 ,x t ]+b r )

[0062]

[0063]

[0064] Among them, z t is the update gate, r t To reset the gate, To update the information generated by the gate, h t is the output at time t, w is the weight, and b is the bias. The forward propagation and backward propagation GRU units construct BiGRU, and its calculation formula is:

[0065]

[0066]

[0067]

[0068] Among them, the state at time t is h t , the output weight of GRU forward propagation is w t1 , the output weight of GRU back propagation is w t2 .

[0069] Tanh is used as the activation function in BiGRU, and its expression is:

[0070]

[0071] Finally, the Adam optimizer is used to optimize the neural network, and the mean square error function is selected as the loss function, which is expressed as:

[0072]

[0073] in, is the output value of the model, i.e. the transmitted signal, f(x) is the label of the model, i.e. the received signal, and N is the number of subcarriers.

[0074] Step 4: Set the hyperparameters required for SNN, CNN, and BiGRU neural networks. The specific hyperparameters are as follows: In SNN, the number of hidden layer neurons is 300 and 100 respectively, and the number of output layer neurons is 128; in CNN, 512 convolution kernels are set, and the size of each convolution kernel is 64; in BiGRU, the number of neurons is 64, 50, 32, 24, and 16 respectively; Epochs is 100, and steps_per_epoch is 50;

[0075] Step 5: Use the data set to train three neural networks: SNN, CNN, and BiGRU. The data set is divided into a training set and a test set in a ratio of 8:2. The neural network is trained under different numbers of pilots and different numbers of cyclic prefixes;

[0076] Step 6: Calculate the accuracy curve, network loss curve and training time. If the accuracy curve does not fit or the loss curve does not converge, reset the hyperparameters and proceed to step 4. If the accuracy curve fits and the loss curve converges, proceed to step 7.

[0077] Step 7: Input the test set into the deep learning-assisted OFDM channel estimation and signal detection model for online deployment. The modulation method used by the deep learning-assisted OFDM channel estimation and signal detection model is 64QAM, and this model is compared with the traditional technology LS, the existing technology BiLSTM, GRU and DNN+BiLSTM.

[0078] Simulation experiment 1

[0079]

[0080] Based on the above experimental conditions, the system bit error rate comparison chart is shown below: Figure 3 As shown in the figure, due to the small number of pilots and cyclic prefixes, the performance of the traditional LS algorithm is poor, while the SNN+CNN+BiGRU model proposed in the present invention is better than the existing technology, with a performance improvement of 2.7dB.

[0081] Simulation experiment 2

[0082]

[0083] Based on the above experimental conditions, the system bit error rate comparison chart is shown below: Figure 4 As shown in the figure, by increasing the number of pilots, the bit error rate performance of various models is improved, but the SNN+CNN+BiGRU model is significantly better than the existing technology, with a performance improvement of 3.8dB.

[0084] Simulation experiment 3

[0085]

[0086] Based on the above experimental conditions, the system bit error rate comparison chart is shown below: Figure 5 As shown in the figure. At 25dB, the performance of the SNN+CNN+BiGRU model is close to that of GRU and DNN+BiLSTM, but still better than the existing technology; from 0 to 25dB, the SNN+CNN+BiGRU model is generally better than LS, BiLSTM, GRU and DNN+BiLSTM, with a performance improvement of 2.8dB.

[0087] Figure 6 The MSE of the SNN+CNN+BiGRU model proposed in this invention is compared with that of BiLSTM, GRU and DNN+BiLSTM. Figure 6 It can be seen that the MSE of SNN+CNN+BiGRU is lower than that of BiLSTM, GRU and DNN+BiLSTM, indicating that its network structure is better than the existing model. BiLSTM contains a large number of parameters, which increases the training time and leads to a higher MSE. BiGRU only contains update gates and reset gates, and adds CNN to denoise the data, which makes up for the shortcomings of BiLSTM and accelerates network convergence. Due to the long input sequence, GRU cannot remember data information well, and BiGRU increases the sensitivity to information. The DNN structure is prone to gradient explosion or gradient disappearance problems, and SNN effectively avoids such shortcomings and improves network accuracy.

[0088] Figure 7 The accuracy of the SNN+CNN+BiGRU model proposed in the present invention is compared with that of BiLSTM, GRU and DNN+BiLSTM. Compared with the prior art, the channel estimation and signal detection accuracy of SNN+CNN+BiGRU is higher. The reason is that BiGRU obtains the information of the previous moment through the forward GRU and the information of the next moment through the reverse GRU. The forward GRU and the reverse GRU are linearly superimposed, which improves the accuracy of channel estimation and signal detection.

[0089] In summary, the OFDM channel estimation and signal detection method assisted by deep learning proposed in the present invention can be applied in the field of high-speed wireless communication. The present invention adopts an innovative logic design. First, a data set is generated through an OFDM system, and then a channel estimation model is constructed using a self-normalizing network, and a signal detection model is constructed using a convolutional neural network and a bidirectional gated cyclic unit; secondly, the data set is input into the neural network, the model is trained, and the network is optimized using the Adam optimizer and the mean square error loss function, and finally compared with the LS, BiLSTM, GRU and DNN+BiLSTM models. The present invention effectively solves the problems of low channel estimation and signal detection accuracy, severe network loss and poor bit error rate performance in OFDM systems with a small number of cyclic prefixes and pilots. The simulation results show that in the case of fewer cyclic prefixes and pilots, the performance of this method is better than that of the LS, BiLSTM, GRU and DNN+BiLSTM models.

[0090] Although specific embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the specific embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A deep learning-assisted OFDM channel estimation and signal detection method, characterized in that The following steps are included: Step 1: Build an OFDM system and generate a data set; Step 2: Construct a channel estimation model based on the self-normalizing network SNN; Step 3: Construct a signal detection model based on convolutional neural network CNN and bidirectional gated recurrent unit BiGRU; Step 4: Set the hyperparameters required for the three neural networks: SNN, CNN, and BiGRU; Step 5: Use the data set to train three neural networks: SNN, CNN, and BiGRU; Step 6: Calculate the accuracy curve, network loss curve and training time. If the accuracy curve does not fit or the loss curve does not converge, reset the hyperparameters and proceed to step 4. If the accuracy curve fits and the loss curve converges, proceed to step 7. Step 7: Input the test set into the deep learning-assisted OFDM channel estimation and signal detection model for online deployment; In step 2, in the SNN-based channel estimation model, the input data is first estimated by the least squares method LS, which is expressed as: in is the channel impulse response coefficient corresponding to the nth subcarrier, is the pilot signal corresponding to the nth subcarrier transmitted, is the pilot signal corresponding to the received nth subcarrier; Split The real and imaginary parts of are used as the input of SNN; The SNN channel estimation model is constructed using three scaled exponential linear units (SELUs), a deep neural network (DNN) and three BN layers, and the Selu function is selected as the hidden layer activation function, and its expression is: in, is 1.051, is 1.

673. On the positive x-axis, the input data will be expanded to times; In step 3, in the signal detection model based on CNN+BiLSTM, the input data is subjected to zero-forcing ZF detection, and its expression is: in is the signal corresponding to the predicted nth subcarrier, is the signal corresponding to the received nth subcarrier, is the channel impulse response corresponding to the nth subcarrier output by the SNN-based channel estimation model; A signal detection model is constructed using 1 CNN layer, 1 BN layer and 5 BiGRU layers. CNN denoises the data through local connections and weight sharing, while reducing computational complexity. The BN layer standardizes the data. Relu is selected as the activation function of CNN, and its expression is: CNN is connected to a 5-layer BiGRU and uses Tanh as the activation function, which is expressed as Finally, the Adam optimizer is used to optimize the neural network, and the mean square error function is selected as the loss function, which is expressed as follows: in, is the output value of the model, that is, the sending signal, is the label of the model, i.e., the received signal, N is the number of subcarriers.

2. The deep learning-assisted OFDM channel estimation and signal detection method according to claim 1, characterized in that: In step 1, the initial signal transmitted by the transmitter is modulated, pilot inserted, serial-to-parallel converted, inverse fast Fourier transformed, cyclic prefix added, and parallel-to-serial converted to generate the required data set.

3. The deep learning-assisted OFDM channel estimation and signal detection method according to claim 1, characterized in that: In step 4, the hyperparameters are as follows: in SNN, the number of neurons in the hidden layer is 300 and 100 respectively, and the number of neurons in the output layer is 128; in CNN, 512 convolution kernels are set, and the size of each convolution kernel is 64; in BiGRU, the number of neurons is 64, 50, 32, 24, and 16 respectively; the number of epochs is 100, and steps_per_epoch is 50.

4. The deep learning-assisted OFDM channel estimation and signal detection method according to claim 1, characterized in that: In step 5, the data set is divided into a training set and a test set in a ratio of 8:2, and the neural network is trained under different numbers of pilot frequencies and different numbers of cyclic prefixes.

5. The deep learning-assisted OFDM channel estimation and signal detection method according to claim 1, characterized in that: In step 7, the modulation method used by the deep learning-assisted OFDM channel estimation and signal detection model is 64QAM.

Images