OFDM Channel Estimation and Signal Detection Method Based on Self-Normalizing Networks

Abstract

The present invention relates to the field of high-speed wireless communication, and in particular to an OFDM channel estimation and signal detection method based on a self-normalizing network. The present invention adopts an innovative logical design. First, a data set required for training the SNN network is generated through an OFDM system, and the data set is divided into a training set and a test set and data preprocessing is performed. Secondly, an SNN network is built. Subsequently, the data is input into the SNN network, the model is trained and learned, and the Adam optimizer is used to optimize the network parameters. The present invention can be applied to an OFDM system with a large number of subcarriers and a small number of pilots, effectively solving the problems of low channel estimation and signal detection accuracy and serious inter-carrier interference in existing and traditional algorithms for such systems. At the same time, a high channel estimation and signal detection accuracy and a low bit error rate can be achieved with a small number of pilot frequencies and cyclic prefixes, greatly saving spectrum resources.

Description

Technical Field

[0001] The present invention relates to the field of intelligent high-speed wireless communication, and particularly to an OFDM channel estimation and signal detection method based on a self-normalizing network. Background Art

[0002] OFDM has an important position in wireless communication systems due to its high spectral efficiency and strong ability to resist multipath channel fading. The wireless channel is affected by frequency-selective fading, shadow fading, etc., making the signal transmission path diverse and complex, resulting in a high bit error rate at the receiving end and interference to the received signal. To meet the needs of society and make efficient use of limited channel resources, channel estimation and signal detection are essential steps.

[0003] The process of estimating the model parameters of a certain channel model from the received signal is channel estimation. Usually, it can be divided into blind channel estimation, semi-blind channel estimation, and non-blind channel estimation according to whether pilots are used. In practical applications, the least squares (LS) estimation included in non-blind channel estimation is widely used due to its low complexity. However, it ignores the influence of inter-carrier interference and noise, so this has also become a reason hindering the further development of the LS algorithm.

[0004] In recent years, deep learning has set off a wave and become a popular tool used in research, achieving unprecedented results in fields such as medical health, language processing, target detection, and image recognition. Due to its powerful learning ability, the communication field has gradually combined with deep learning and is currently applied to technologies such as CSI feedback, channel estimation, and signal detection.

[0005] Among the technologies combining communication and deep learning, the most classic one is the algorithm combining DNN and OFDM systems. By treating the communication system as a whole in a data-driven manner, it is trained using 8 DNNs with the same structure and a large amount of labeled data. Experimental results show that this algorithm is superior to the LS algorithm. However, DNN generally performs poorly when there are many sub-carriers. It is necessary to increase the number of pilots to reduce the bit error rate (BER) and improve the accuracy. At the same time, its mean square error (MSE) is high and the convergence speed is slow.

[0006] Patent CN 112600772A proposes an OFDM channel estimation and signal detection method based on a data-driven neural network. Its essence is to add 8 layers of convolutional neural network (CNN) and 8 layers of BN layer in front of the DNN network for noise reduction processing, which can achieve a BER lower than that of the LS algorithm. However, its network has many layers and a large amount of calculation, which is prone to waste of time and resources.

[0007] The invention patent CN 111404849A proposes a method for OFDM channel estimation and signal detection based on deep learning, which uses CNN and Dense to form a channel estimation sub-network and a signal detection sub-network to achieve channel estimation and signal detection respectively. However, when the number of subcarriers is large, the network training time of this invention is long, and the problem of low accuracy is likely to occur, so it is not applicable to OFDM systems with a large number of subcarriers.

[0008] The invention patent CN 111510402B proposes a method for OFDM channel estimation based on deep learning, which improves the channel estimation quality through three layers of Dense. However, when the number of subcarriers is increased, the problems of gradient disappearance and overfitting are likely to occur. Summary of the Invention

[0009] In view of the above problems, the present invention proposes a method for OFDM channel estimation and signal detection based on the self-normalizing network SNN. The SNN adopts a fully connected layer, a Selu enhanced processing unit, and a Dropout Selu network layer to reduce the MSE, accelerate the network convergence, and improve the system bit error rate performance at the same time.

[0010] The present invention provides the following technical solutions: A method for OFDM channel estimation and signal detection based on a self-normalizing network, including the following steps:

[0011] Step 1: Use the OFDM system framework to generate a data set required for training the self-normalizing network;

[0012] Step 2: Divide the data set and perform data preprocessing, where the training set contains 80% of the data and the test set contains 20% of the data;

[0013] Step 3: Build a self-normalizing network for OFDM system channel estimation and signal detection;

[0014] Step 4: Set the hyperparameters required for training the self-normalizing network;

[0015] Step 5: Input the training set into the self-normalizing network for offline training;

[0016] Step 6: Output the accuracy curve and the loss curve, check whether the accuracy curve is fitted and whether the loss curve converges. If either curve does not meet the requirements, repeat step 4. If both curves meet the requirements, proceed to step 7;

[0017] Step 7: Input the test set into the OFDM channel estimation and signal detection model based on the self-normalizing network for online deployment.

[0018] The self-normalizing network for OFDM channel estimation and signal detection is built using TensorFlow 1.15.0 and PyCharm 2020.1 tools.

[0019] In step 1, the dataset required for training the self-normalizing network is generated by the OFDM system under a Rayleigh fading channel. The signal transmitted by the transmitter is used as a label, and the signal received by the receiver is used as the input of the self-normalizing network.

[0020] In step 2, the data preprocessing method is , where is the signal received by the receiver, is the signal after preprocessing.

[0021] In step 3, the self-normalizing network includes 3 BN layers, 4 fully connected layers (Dense), 3 Selu units, and 3 Dropout Selu layers. Among them, the first 3 of the 4 fully connected layers are connected to the BN layer in front. To improve the model performance, enhance the generalization ability and learning ability of the model, all 4 fully connected layers are adjacent to Selu units for enhancement processing, and the data after enhancement processing is input into the Dropout Selu layer.

[0022] In step 3, the number of neurons in the 3 hidden layers (fully connected layers) of the self-normalizing network is 2048, 2048, and 128, and the number of neurons in the output layer is 16. The Selu function is used as the activation function for the hidden layer. The mathematical expression of the Selu function is , where is approximately 1.051, is approximately 1.673. When > 0, the change in the input will be amplified by times; when ≤ 0, the negative half-axis information of the change in the input will be retained. The Sigmoid function is used as the activation function for the output layer. The mathematical expression of the Sigmoid function is , is 2.71828.

[0023] In step 3, in the self-normalizing network, the AlphaDropout is introduced as the random inactivation function in the Dropout Selu layer.

[0024] In step 4, the specific hyperparameters required for training the self-normalizing network are: the number of iterations (Epochs) is 100, the internal parameter update rate (Rate) of the Dropout Selu layer is 0.05, the learning rate (Lr) is 0.001, and the optimizer is Adam.

[0025] The loss function of the OFDM channel estimation and signal detection model based on the self-normalizing network is defined as:

[0026] , where is the output value of the model, is the label of the model, is the number of subcarriers in the OFDM system.

[0027] In step 7, the modulation method adopted by the OFDM channel estimation and signal detection model based on the self-normalizing network is QPSK, and this model is equivalently migrated for learning to achieve OFDM channel estimation and signal detection under different subcarriers, different pilots, and different signal-to-noise ratios.

[0028] Compared with the prior art, the present invention has the following features: 1. By adopting the self-normalizing network and combining it with the BN layer, the network convergence speed is accelerated, and at the same time, the recognition ability of the neural network is improved. 2. A Selu unit is added inside the self-normalizing network to strengthen the signal, which can increase the generalization ability and learning ability of the model and improve the network accuracy. 3. The Dropout Selu layer uses AlphaDropout as the random inactivation function and sets the parameter update speed to 0.05, effectively preventing the overfitting phenomenon caused by a large number of neurons. 4. The fully connected layer (Dense) uses Selu as the activation function, which can automatically normalize the data distribution to a mean of 0 and a variance of 1. 5. The self-normalizing network is embedded with Dense, avoiding the problems of gradient disappearance and gradient explosion. 6. The channel estimation and signal detection model based on the self-normalizing network can be applied to the OFDM system with a large number of subcarriers and a small number of pilots, greatly saving resources, and at the same time, it can ensure high channel estimation and signal detection accuracy, effectively solving the problem of inter-carrier interference. Description of the Drawings

[0029] Figure 1 is the structure diagram of the OFDM system based on the self-normalizing network.

[0030] Figure 2 is the self-normalizing network structure built.

[0031] Figure 3 is the model BER performance comparison diagram under the conditions that the number of subcarriers (K) is 64 and 256, the number of cyclic prefixes (C) is 32, and the number of pilots (P) is 8.

[0032] Figure 4 is the model BER performance comparison diagram under the conditions that the number of subcarriers (K) is 256, the number of cyclic prefixes (C) is 32, and the number of pilots (P) is 8, 64, and 256.

[0033] Figure 5 Comparison chart of the accuracy of SNN and DNN when the number of subcarriers (K) is 256, the number of cyclic prefixes (C) is 32, and the number of pilots (P) is 64.

[0034] Figure 6 Comparison chart of the mean square error (MSE) between SNN and DNN of the present invention. Detailed implementation manners

[0035] Next, the technical solutions in the detailed implementation manners of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the detailed implementation manners of the present invention. Obviously, the described detailed implementation manners are only a specific implementation manner of the present invention, rather than all the detailed implementation manners. Based on the detailed implementation manners in the present invention, all other specific implementation manners obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

[0036] It should be noted that the terms used herein are only for describing the specific implementation manners and are not intended to limit the exemplary implementation manners according to the present application. As used herein, unless otherwise clearly specified in the context, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "include" and / or "comprise" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof;

[0037] The OFDM channel estimation and signal detection method based on the self-normalizing network proposed by the present invention includes the following steps:

[0038] Step 1: Generate a data set required for training the self-normalizing network by using the OFDM system framework. The signal generation process includes the following steps:

[0039] Step 1.1: Transmit a binary bit stream;

[0040] Step 1.2: Symbol mapping. After the signal is mapped, it is:

[0041] (1);

[0042] Step 1.3: Inverse fast Fourier transform (IFFT). After the signal passes through the IFFT, it is:

[0043] (2)

[0044] Step 1.4: Add a cyclic prefix (CP). After the signal adds the CP, it is:

[0045] (3)

[0046] In the formula, L is the length of the CP;

[0047] Step 1.5: Serial-to-parallel conversion;

[0048] Step 1.6: Through the wireless channel, and the signal at this time is:

[0049] (4)

[0050] where is the channel impulse response, is the Gaussian white noise;

[0051] Step 1.7: Parallel-to-serial conversion;

[0052] Step 1.8: Remove the cyclic prefix (CP), and the signal after removing the CP is:

[0053] (5)

[0054] where is the cyclic convolution;

[0055] Step 1.9: Fast Fourier Transform (FFT), and the signal after passing through the FFT is:

[0056] (6).

[0057] Step 2: Divide the dataset and perform data preprocessing, where the training set contains 80% of the data and the test set contains 20% of the data; the way of data preprocessing is , where in the formula is the signal received at the receiving end, is the signal after preprocessing.

[0058] Step 3: Build a self-normalizing network for channel estimation and signal detection in the OFDM system. The self-normalizing network includes 3 BN layers, 4 fully connected layers (Dense), 3 Selu units and 3 Dropout Selu layers. Among them, the latter 3 fully connected layers are connected with BN layers in front. To improve the model performance, enhance the model generalization ability and learning ability, 4 fully connected layers are all adjacent to Selu units for strengthening processing, and the data after strengthening processing is input into the Dropout Selu layer.

[0059] The number of neurons in the 3 hidden layers (BN layers) in the self-normalizing network is 2048, 2048, 128, and the number of neurons in the output layer is 16. The hidden layer uses the Selu function as the activation function, and its mathematical expression is , where is 1.051, is 1.673. When > 0, a small change in the input will be amplified to times; when ≤ 0, the information on the negative half-axis of the input change will be retained. The output layer uses the Sigmoid function as the activation function, and its mathematical expression is . At the same time, the AlphaDropout is introduced as the random inactivation function in the Dropout Selu layer.

[0060] Step 4: Set the hyperparameters required for training the self-normalizing network, specifically: the number of iterations (Epochs) is 100, the internal parameter update rate (Rate) of the Dropout Selu layer is 0.05, the learning rate (Lr) is 0.001, and the optimizer is Adam.

[0061] Step 5: Input the training set into the self-normalizing network for offline training. The loss function adopted by the network is defined as: , where is the output value of the model, and is the label of the model.

[0062] Step 6: Output the accuracy curve and the loss curve, and check whether the accuracy curve is fitted and the loss curve converges. If any of the curves does not meet the requirements, repeat Step 4. If both curves meet the requirements, proceed to Step 7.

[0063] Step 7: Input the test set into the OFDM channel estimation and signal detection model based on the self-normalizing network for online deployment. The modulation method adopted by the model is QPSK, and this model is equivalently transferred for learning to achieve OFDM channel estimation and signal detection under different subcarriers, different pilots, and different signal-to-noise ratios.

[0064] The self-normalizing network is built using the Tensorflow 1.15.0 and Pycharm 2020.1 tools.

[0065] Simulation Experiment 1

[0066] The simulation parameters are as follows:

[0067]

[0068] Based on the above experimental conditions, compared with the LS algorithm and the algorithm corresponding to the DNN, the system BER comparison diagram is as shown in Figure 3 . When (K, C, P) is (64, 32, 8), the SNN is superior to the DNN and the LS. Compared with the DNN, the performance is improved by 4 dB. When (K, C, P) is (256, 32, 8), the SNN is superior to the DNN and the LS. Compared with the DNN, the performance is improved by 3 dB. And as the number of subcarriers decreases, the improvement effect is more obvious.

[0069] Simulation Experiment 2

[0070] The simulation parameters are as follows:

[0071]

[0072] Based on the above experimental conditions, compared with the algorithms corresponding to the LS algorithm and DNN, the system BER comparison diagram is as Figure 4 shown. When (K, C, P) is (256, 32, 8), the performance of the traditional algorithm LS is not ideal, and the SNN is better than the DNN, with a performance improvement of 3 dB. When (K, C, P) is (256, 32, 64), although the performance of the DNN is better than that of the LS, the effect is approximately the same as that of the LS at 25 dB. The SNN is far better than the existing technology, and compared with the DNN, it is improved by about 5 dB. When (K, C, P) is (256, 32, 256), the performance of the SNN is significantly better than the existing technology. Compared with the DNN, it is improved by about 7 dB, and the SNN with a pilot of 64 is better than the DNN with a pilot of 256 after 10 dB, which is more resource-saving.

[0073] Figure 5 It is the accuracy comparison diagram of the SNN and DNN when (K, C, P) is (256, 32, 64). At this time, the accuracy of the DNN is relatively low, and the number of pilots needs to be increased to handle the case of high subcarriers, while the SNN can achieve high accuracy without increasing the pilots.

[0074] Figure 6 The MSE comparison between the SNN and DNN in all experiments is given. Obviously, when (K, C, P) is (64, 32, 8), the DNN converges after about 40 times, and the SNN converges after about 20 times; when (K, C, P) is (256, 32, 8), the DNN converges after about 60 times, and the SNN converges after about 40 times; when (K, C, P) is (256, 32, 64), the DNN converges after about 70 times, and the SNN converges after about 40 times; when (K, C, P) is (256, 32, 256), the DNN converges after about 40 times, and the SNN converges after about 30 times; in any case, the MSE of the SNN is less than that of the DNN.

[0075] In summary, the OFDM channel estimation and signal detection method based on the self-normalizing network proposed by the present invention can be applied in the field of high-speed wireless communication. The present invention adopts an innovative logical design. First, a data set required for training the network is generated through an OFDM system, and the data set is divided into a training set and a test set and data preprocessing is performed. Subsequently, the data is input into the self-normalizing network, the model is trained and learned, and the Adam optimizer is used to optimize the network parameters. The present invention effectively solves the problems existing in the OFDM system with a large number of subcarriers and a small number of pilots, such as low channel estimation and signal detection accuracy and severe interference between subcarriers. Higher accuracy and lower BER can be achieved with fewer pilot numbers and cyclic prefix numbers, greatly saving limited resources. The simulation results show that this method has stronger capabilities compared with LS and DNN.

[0076] Although the specific embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these specific embodiments without departing from the principle and spirit of the invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. OFDM channel estimation and signal detection method based on self-normalizing network, characterized in that It includes the following steps: Step 1: Generate the dataset required for training the self-normalizing network using the OFDM system framework; Step 2: Divide the dataset and perform data preprocessing, where the training set contains 80% of the data and the test set contains 20% of the data; Step 3: Build a self-normalizing network for OFDM system channel estimation and signal detection; Step 4: Set the hyperparameters required for training the self-normalizing network; Step 5: Input the training set into the self-normalizing network for offline training; Step 6: Output the accuracy curve and the loss curve, check whether the accuracy curve is fitted and the loss curve converges. If either curve does not meet the requirements, repeat Step 4. If both curves meet the requirements, proceed to Step 7; Step 7: Input the test set into the OFDM channel estimation and signal detection model based on the self-normalizing network for online deployment; In Step 3, the self-normalizing network includes 3 BN layers, 4 fully-connected Dense layers, 3 Selu units, and 3 Dropout Selu layers. Among the 4 fully-connected layers, the first 3 of the 4 fully-connected layers are connected to the BN layer in front, and each of the 4 fully-connected layers is adjacent to a Selu unit for enhancement processing. The data after enhancement processing is then input into the Dropout Selu layer.

2. The OFDM channel estimation and signal detection method based on a self-normalizing network according to claim 1, wherein , The self-normalizing network for OFDM channel estimation and signal detection in Step 3 is built using the Tensorflow 1.15.0 and Pycharm 2020.1 tools.

3. The OFDM channel estimation and signal detection method based on self-normalizing network according to claim 1, characterized in that , In Step 1, the dataset required for training the self-normalizing network is generated by the OFDM system in a Rayleigh fading channel. The signal transmitted by the transmitter is used as the label, and the signal received by the receiver is used as the input of the self-normalizing network.

4. The OFDM channel estimation and signal detection method based on the self-normalizing network according to claim 1, wherein , In step 2, the data preprocessing method is , where is the signal received by the receiving end, is the signal after preprocessing.

5. The OFDM channel estimation and signal detection method based on a self-normalizing network according to claim 1, characterized in that , In step 3, the number of neurons in the three hidden layers of the self-normalizing network is 2048, 2048, 128, 128. The hidden layers are fully connected layers. The number of neurons in the output layer is 16. The Selu function is used as the activation function for the hidden layers. The mathematical expression of the Selu function is , where is 1.051, is 1.

673. When > 0, the change in the input will be expanded to times; when ≤ 0, the negative half-axis information of the input change will be retained. The Sigmoid function is used as the activation function for the output layer. The mathematical expression of the Sigmoid function is , where is 2.71828.

6. The OFDM channel estimation and signal detection method based on a self-normalizing network according to claim 5, characterized in that , In Step 3, in the self-normalizing network, the Dropout Selu layer introduces AlphaDropout as the random inactivation function.

7. The OFDM channel estimation and signal detection method based on a self-normalizing network according to claim 6, wherein , In Step 4, the hyperparameters required for training the self-normalizing network include: the number of iterations Epochs is 100, the internal parameter update rate Rate of the Dropout Selu layer is 0.05, and the learning rate Lr is 0.

001.

8. The OFDM channel estimation and signal detection method based on the self-normalizing network according to claim 1, characterized in that , The loss function of the OFDM channel estimation and signal detection model based on the self-normalizing network is defined as: , where is the output value of the model, is the label of the model, and among them is the number of subcarriers in the OFDM system.

9. The OFDM channel estimation and signal detection method based on the self-normalizing network according to claim 1, wherein , In Step 7, the modulation method adopted by the OFDM channel estimation and signal detection model based on the self-normalizing network is QPSK, and this model is subjected to equivalent transfer learning to achieve OFDM channel estimation and signal detection under different subcarriers, different pilots, and different signal-to-noise ratios.

Images