A deep learning-based qpsk receiver and an auxiliary model training method thereof

Abstract

The application relates to a QPSK receiver auxiliary model training method based on deep learning, which comprises the following steps: building a QPSK system model, using the QPSK system model to obtain a training data set, building a deep learning neural network model, generating a prediction value, training the deep learning neural network model using the training data set, optimizing a loss function of the deep learning neural network model, optimizing the loss function using a Nadam algorithm, updating parameters of the deep learning neural network model until the error rate of the comparison between the prediction value and an actual value is minimum, and taking the updated deep learning neural network model as a QPSK receiver auxiliary model. The model trained by the method adopts a deep learning neural network based on a reflection structure, considers the hidden correlation between values, can be used for demodulation in a demodulation link of a communication receiver, and the demodulation of the model can solve the problem that the recognition accuracy is not high due to noise interference under the condition of a low signal-to-noise ratio in a traditional hard decision method.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communications, and in particular relates to a QPSK receiver based on deep learning and an auxiliary model training method thereof. Background Art

[0002] In recent years, deep learning has been widely used in fields such as computer vision, autonomous driving, image recognition and classification, and object identification. Deep learning is favored because it is an end-to-end approach. Compared to other machine learning methods, it boasts a wide range of applications, excellent adaptability, portability, and a powerful learning capability, capable of learning deeper features from large amounts of data. In the field of wireless communications, integrating deep learning with wireless receivers is currently a hot topic in academic research, and it is highly likely that this approach will address some of the shortcomings of traditional wireless receivers.

[0003] Quadrature Phase Shift Keying (QPSK), a digital modulation method with strong interference resistance, high spectrum efficiency, and relatively simple circuit implementation, has been widely used in wireless communication systems. In actual communication, in addition to channel noise interference, Doppler shift occurs when two independent local oscillators are used in the transmitter and receiver, and when there is relative motion between the transmitter and receiver, resulting in a certain frequency deviation between the received and transmitted signals. Furthermore, due to the non-ideal nature of radio frequency equipment, the received IQ signal may exhibit IQ imbalance, i.e., an imbalance in the amplitude and / or phase between the I and Q channels. IQ imbalance can be described by a set of parameters (α, β), where α represents the amplitude imbalance and β represents the phase imbalance. All of these interferences can affect the receiver's reception accuracy. Under high signal-to-noise ratio conditions, traditional QPSK receivers often use hard decision methods to demodulate the received interference-distorted signal, followed by decoding and other processes to recover the information. This results in high bit error rates and low recognition accuracy.

[0004] Therefore, a solution is urgently needed to solve the problem that when a communication receiver receives a signal under high signal-to-noise ratio conditions, the recognition accuracy is low due to interference such as channel noise. Summary of the Invention

[0005] Based on the above-mentioned shortcomings and deficiencies in the prior art, one of the objects of the present invention is to at least solve one or more of the above-mentioned problems in the prior art. In other words, one of the objects of the present invention is to provide a deep learning-based QPSK receiver and its auxiliary model training method that meet one or more of the aforementioned needs.

[0006] In order to achieve the above-mentioned object of the invention, the present invention adopts the following technical solutions:

[0007] A QPSK receiver-assisted model training method based on deep learning specifically includes the following steps:

[0008] S1. Build a QPSK system model and use it to obtain a training data set.

[0009] S2. Build a deep learning neural network model, which is used to demodulate the coded and noise-affected signal and generate a prediction value;

[0010] S3. Use the training data set to train the deep learning neural network model and optimize the loss function of the deep learning neural network model;

[0011] S4. Use the Nadam algorithm to optimize the loss function and update the parameters of the deep learning neural network model until the error rate between the predicted value and the true value is minimized. The updated deep learning neural network model is used as the auxiliary model of the QPSK receiver.

[0012] As a preferred solution, the QPSK system model is used to obtain a training data set. Specifically, the QPSK system model is used to perform QPSK modulation on the Hamming-encoded data, and then oversampled using root raised cosine, and then Gaussian noise is added as interference. After matched filtering and undersampling, 56-bit features are obtained.

[0013] As a preferred solution, the deep learning neural network model demodulates the coded and noise-affected signal by specifically including:

[0014] Input the first fully connected layer, which uses the nonlinear activation function Relu;

[0015] Input LSTM layer, the LSTM layer has 128 neurons;

[0016] Input the second fully connected layer for output. The second fully connected layer has 56 neurons and uses the activation function sigmoid.

[0017] As a preferred solution, the loss function is binary cross entropy.

[0018] In a second aspect, the present invention provides a QPSK demodulation method based on a deep learning-assisted model, using a QPSK receiver-assisted model trained by any of the above-mentioned deep learning-based QPSK receiver-assisted model training methods, characterized in that it specifically includes:

[0019] Use the QPSK receiver-assisted model to demodulate the coded and noise-affected signal and generate predictions.

[0020] Perform Hamming decoding on the predicted value to restore the original code.

[0021] As a preferred solution, the Hamming decoding is (7,4) decoding, which outputs a 32-bit bit stream.

[0022] In a third aspect, the present invention provides a QPSK intelligent receiver based on a deep learning assisted model, which uses the above-mentioned QPSK demodulation method based on a deep learning assisted model.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] The model trained by the method of the present invention adopts a deep learning neural network, and a neural network model is designed by adopting a multi-label classification method to replace the demodulation link of a communication receiver.

[0025] The neural network model primarily utilizes an LSTM recurrent neural network, leveraging the memory structure of recurrent neural networks to improve predictive performance. The LSTM network is preceded and followed by fully connected layers, further enhancing the model's nonlinear expressiveness and learning capabilities. This allows for better demodulation and recovery of the original information, resolving the issue of low recognition accuracy in traditional hard-decision methods under high signal-to-noise ratio conditions due to interference from channel noise and other factors. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a flow chart of the deep learning-based QPSK receiver-assisted model training method of the present invention;

[0027] Figure 2 The figure is a comparison diagram of the bit error rates when the QPSK demodulation method of the present invention is used for demodulation and the traditional hard decision method is used for demodulation when the signal-to-noise ratio is 0-7dB.

[0028] Figure 3 The figure is a comparison of the bit error rates of the QPSK demodulation method of the present invention and the traditional hard decision method when the signal-to-noise ratio is 0-7dB and the normalized carrier frequency offset δf is set to 0.001, 0.002, and 0.004 respectively.

[0029] Figure 4 This is a comparison chart of the bit error rates of the QPSK demodulation method of the present invention and the traditional hard decision method when the signal-to-noise ratio is 0-7dB and the three IQ imbalance configurations are (5, -6), (-3, 10) and (-3, -2);

[0030] Figure 5 It is a structural diagram of the deep learning-based QPSK communication intelligent receiver of the present invention. DETAILED DESCRIPTION

[0031] To more clearly illustrate the embodiments of the present invention, specific embodiments of the present invention will be described below with reference to the accompanying drawings. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings and other embodiments can be obtained based on these drawings without inventive efforts.

[0032] Example: This application provides a QPSK receiver assisted model training method based on deep learning, the flow chart of which is as follows Figure 1 As shown, in one embodiment of the present application, the method specifically includes the following steps:

[0033] S1. Build a QPSK system model and use it to obtain a training data set.

[0034] S2. Build a deep learning neural network model, which is used to demodulate the coded and noise-affected signal and generate a prediction value;

[0035] S3. Use the training data set to train the deep learning neural network model and optimize the loss function of the deep learning neural network model;

[0036] S4. Use the Nadam algorithm to optimize the loss function and update the parameters of the deep learning neural network model until the error rate between the predicted value and the true value is minimized. The updated deep learning neural network model is used as the auxiliary model of the QPSK receiver.

[0037] As a preferred solution, in a further embodiment of step S1, the training data set is obtained using the QPSK system model. Specifically, the QPSK system model is used to perform QPSK modulation on the Hamming-encoded data, and then oversampled using root raised cosine, and then Gaussian noise is added as interference, and 56-bit features are obtained through matched filtering and undersampling.

[0038] The QPSK system model transmits a 32-bit stream of 01 bits, randomly generated by MATLAB. In this model, the transmitted data undergoes (7,4) Hamming encoding at the transmitter to generate 56 bits. This data is then QPSK modulated and oversampled using root-raised cosine, with 8 samples per symbol. Gaussian noise is then added for interference. Matched filtering and undersampling are performed, with one symbol sampled every 8 samples. This yields 56-bit features, and the 56-bit Hamming codes obtained by Hamming encoding serve as values. All features generated by this QPSK system model and their corresponding values ​​constitute the training dataset.

[0039] In a further embodiment of step S2, the demodulation of the coded and noise-affected signal by the deep learning neural network model specifically includes:

[0040] Input the first fully connected layer, which uses the nonlinear activation function Relu;

[0041] Input LSTM layer, the LSTM layer has 128 neurons;

[0042] Input the second fully connected layer for output. The second fully connected layer has 56 neurons and uses the activation function sigmoid.

[0043] As a further preferred solution, the predicted value is 56 bits.

[0044] Specifically, the first layer of the deep learning neural network is a fully connected layer with a total of 128 neurons. The activation function used is the nonlinear activation function Relu. The second layer is the main part of the network, using LSTM, with a total of 128 neurons, and the activation function used is Relu. With the help of the memory structure in the recursive neural network, the prediction performance is improved. The last layer is a fully connected layer, which serves as the output layer. It has a total of 56 neurons and the activation function used is sigmoid. The output result of the LSTM network is input into the sigmoid layer for calculation, so that each real number in the input vector is mapped to a real number between 0 and 1, and all values ​​in the output vector are in the interval [0,1], realizing the multi-label classification method.

[0045] In the above steps S3 and S4, the deep learning neural network is used to output the results of five iterations, and the result of the fifth iteration is taken as the output result, which is a 56-bit floating-point number. After a 01 judgment, a 56-bit predicted value (pre_label) is obtained, which is compared with the true value (true_label) to calculate the error rate. Based on the output results of the neural network, the loss function of the deep learning neural network model is calculated.

[0046] Then the Nadam algorithm is used to optimize the loss function, and the fully connected neural network model is trained through back propagation to update and optimize the parameter values ​​and weights of neurons in each layer.

[0047] During the update process, the model parameters are adjusted so that the predicted value (pre_label) obtained by 01 judgment of the output information is compared with the true value (true_label), and the calculated error rate is minimized, which means that the training of the deep learning neural network model is completed.

[0048] Specifically, the loss function is set to binary cross entropy. It is defined as:

[0049] Here, the batch size is N, y is a binary label 0 or 1, and p(y) is the probability that the output belongs to the y label.

[0050] The following provides a specific implementation process in an embodiment of the present application:

[0051] In this embodiment, step S1 specifies that for each training data set, the signal-to-noise ratio ranges from 0 to 7 dB, with intervals of 1 dB. The number of data samples for each signal-to-noise ratio is 200,000, resulting in a total of 1.6 million training samples (200,000 * 8). For each test data set, the signal-to-noise ratio ranges from 0 dB to 7 dB, with intervals of 0.5 dB. The number of samples for each signal-to-noise ratio is 100,000, resulting in a total of 1.5 million test samples (100,000 * 15). The data to be transmitted in step S1 is a 01 bit stream, with 4 information bits, divided into 8 groups, for a total of 32 bits. At the transmitter, the data to be transmitted is (7,4) Hamming encoded to obtain 56 bits. This is then modulated using QPSK and oversampled using root-raised cosine, with 8 samples per symbol. Gaussian noise is then added as interference. Matched filtering and undersampling are performed: one symbol is sampled for every 8 samples, resulting in 56-bit features. The 56-bit Hamming code obtained by Hamming encoding is used as the value. All features and their corresponding values ​​together form a dataset.

[0052] In step S2, the first layer of the deep learning network is a fully connected layer with a total of 128 neurons. The activation function used is the nonlinear activation function Relu. The second layer is the main part of the network, using LSTM, with a total of 128 neurons, and the activation function used is Relu. With the help of the memory structure in the recursive neural network, the prediction performance is improved. The last layer is a fully connected layer, which serves as the output layer. It has a total of 56 neurons and the activation function used is sigmoid. The output result of the LSTM network is input into the sigmoid layer for calculation, so that each real number in the input vector is mapped to a real number between 0 and 1, and all values ​​in the output vector are in the interval [0,1].

[0053] In steps S3 and S4, the 56-bit floating-point number output by the deep learning neural network is used to perform a 01 judgment to obtain a 56-bit predicted value (pre_label), which is compared with the true value (true_label) to calculate the error rate. Based on the output results of the neural network, the loss function of the deep learning neural network model is calculated.

[0054] Then the Nadam algorithm is used to optimize the loss function, and the fully connected neural network model is trained through back propagation to update and optimize the parameter values ​​and weights of neurons in each layer.

[0055] The loss function is set to binary cross entropy. It is defined as:

[0056] Here, the batch size is N, y is a binary label 0 or 1, and p(y) is the probability that the output belongs to the y label.

[0057] During the update process, the model parameters are adjusted so that the predicted value (pre_label) obtained by 01 judgment of the output information is compared with the true value (true_label), and the calculated error rate is minimized, which means that the training of the deep learning neural network model is completed.

[0058] The model trained using this method is based on a deep learning-based intelligent QPSK communication receiver. By employing a multi-label classification approach, a neural network model was designed to replace the demodulation stage of the communication receiver. The neural network model primarily utilizes an LSTM (Long-Term Long-Term Memory) architecture, leveraging the memory structure of recurrent neural networks to improve prediction performance. The LSTM network is preceded and followed by fully connected layers to further enhance the model's nonlinear representation capabilities and learning capabilities. The interference-distorted signal transmitted through the channel is fed into the intelligent QPSK communication receiver for demodulation, followed by Hamming decoding to recover the original information. This overcomes the problem of low recognition accuracy in traditional hard-decision methods under high signal-to-noise ratio conditions due to interference such as channel noise.

[0059] In a second aspect, the present application provides a QPSK demodulation method based on a deep learning-assisted model, using a QPSK receiver-assisted model trained by any of the above-mentioned deep learning-based QPSK receiver-assisted model training methods, characterized in that it specifically includes:

[0060] Use the QPSK receiver-assisted model to demodulate the coded and noise-affected signal and generate predictions.

[0061] Perform Hamming decoding on the predicted value to restore the original code.

[0062] In a preferred embodiment, the Hamming decoding is (7,4) decoding, outputting a 32-bit bit stream. Specifically, the 56-bit predicted value (pre_label) is subjected to (7,4) Hamming decoding to obtain a 32-bit bit stream, which is the restored information bits. By comparing the restored information bits with the original information bits, the bit error rate can be calculated, which is used to determine the accuracy of the QPSK demodulation method of the present application.

[0063] The above implementation uses MATLAB R2022b and pycharm 2020, tensorflow2.4.0CPU to simulate the above steps for simulation experiments, and calculates and plots the bit error rate comparison chart calculated by demodulation using the QPSK demodulation method of this application and the traditional hard decision method when the signal-to-noise ratio is 0-7dB, as the signal-to-noise ratio increases, as shown in the figure. Figure 2As shown. The normalized carrier frequency offset δf (relative to the symbol rate) is set to 0.001, 0.002, and 0.004 respectively. The bit error rate comparison chart calculated by demodulating using the QPSK demodulation method of the present application and the traditional hard decision method is shown in the figure below. Figure 3 The comparison of the bit error rates calculated by using the QPSK demodulation method of the present application and the traditional hard decision method for demodulation under three IQ imbalance configurations: (5, -6), (-3, 10) and (-3, -2) is shown in the figure. Figure 4 shown.

[0064] On the third aspect, the present application provides a QPSK intelligent receiver based on a deep learning assisted model, which uses the above-mentioned QPSK demodulation method based on a deep learning assisted model.

[0065] The receiver uses a QPSK communication intelligent receiver based on deep learning at the receiving end. A neural network model is designed by adopting the multi-label classification method to replace the demodulation link of the communication receiver. Its structural diagram is shown in the figure below. Figure 5 As shown, the neural network model primarily utilizes LSTM, leveraging the memory structure of recurrent neural networks to improve prediction performance. The LSTM network is preceded and followed by fully connected layers to further enhance the model's nonlinear representation and learning capabilities. The interference-distorted signal transmitted through the channel is sent to a QPSK communication intelligent receiver for demodulation, and then Hamming decoding is performed to recover the original information. This overcomes the problem of low recognition accuracy in traditional hard-decision methods under high signal-to-noise ratio conditions due to interference such as channel noise.

[0066] It should be noted that the above embodiments are only detailed descriptions of the preferred embodiments and principles of the present invention. For ordinary technicians in this field, there will be changes in the specific implementation methods based on the ideas provided by the present invention, and these changes should also be regarded as the scope of protection of the present invention.

Claims

1. A QPSK receiver-assisted model training method based on deep learning, characterized in that: The specific steps include: S1. Build a QPSK system model and use the QPSK system model to obtain a training data set; S2. Building a deep learning neural network model, wherein the deep learning neural network model is used to demodulate the coded and noise-affected signal and generate a prediction value; S3. Using the training data set to train the deep learning neural network model, and optimizing the loss function of the deep learning neural network model; S4. Optimize the loss function using the Nadam algorithm, update the parameters of the deep learning neural network model until the error rate between the predicted value and the true value is minimized, and use the updated deep learning neural network model as the QPSK receiver auxiliary model; The QPSK system model is used to obtain a training data set. Specifically, the QPSK system model is used to perform QPSK modulation on the Hamming-encoded data, and then oversampled using root raised cosine, and then Gaussian noise is added as channel interference. 56-bit features are obtained through matched filtering and undersampling. The 56-bit Hamming code obtained by Hamming encoding is used as the value. All features and corresponding values ​​generated by the above QPSK system model together constitute the training data set. The deep learning neural network model demodulates the coded and noise-affected signal in the following ways: Input the first fully connected layer, which uses a nonlinear activation function Relu; Input LSTM layer, the LSTM layer has 128 neurons; The second fully connected layer is input for output. The second fully connected layer has 56 neurons and uses the activation function sigmoid, so that each real number in the input vector is mapped to a real number between 0 and 1, and all values ​​in the output vector are in the interval [0, 1].

2. The QPSK receiver-assisted model training method based on deep learning according to claim 1, wherein: The predicted value is 56 bits.

3. The QPSK receiver-assisted model training method based on deep learning according to claim 1, wherein: The loss function is binary cross entropy.

4. A QPSK demodulation method based on a deep learning-assisted model, using a QPSK receiver-assisted model trained by the deep learning-based QPSK receiver-assisted model training method according to any one of claims 1 to 3, characterized in that: Specifically include: Demodulating the coded and noise-affected signal using the QPSK receiver-assisted model to generate a prediction value; Perform Hamming decoding on the predicted value to restore the original code.

5. The QPSK demodulation method based on a deep learning-assisted model according to claim 4, wherein: The Hamming decoding is (7,4) decoding, which outputs a 32-bit bit stream.

Images