An artificial intelligence-based QPSK receiver and an auxiliary model training method thereof

Abstract

The application relates to an artificial intelligence-based QPSK receiver auxiliary model training method, which comprises the following steps: a QPSK system model is built, a training data set is obtained by using the QPSK system model, a rethink neural network model is built, a prediction label is generated, the rethink neural network model is trained by using the training data set, a loss function of the rethink neural network model is optimized, the loss function is optimized by using a Nadam algorithm, parameters of the rethink neural network model are updated until the error rate of comparison between the prediction label and an actual label is minimum, and the updated rethink neural network model is used as a QPSK receiver auxiliary model. The model trained by the method adopts a rethink neural network based on a rethink structure, the hidden correlation between labels is considered, demodulation can be carried out 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 traditional hard decision method in a low signal-to-noise ratio.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of wireless communication, and particularly relates to a QPSK receiver based on artificial intelligence and an auxiliary model training method thereof. BACKGROUND

[0002] In recent years, artificial intelligence has been widely applied in many fields such as computer vision, automatic control technology and natural language processing, and has made new breakthroughs: face recognition, driverless cars, smart speakers, machine translation and the like. Artificial intelligence has gradually penetrated into all aspects of people's life, which has attracted widespread attention from the academic and industrial circles. In the field of wireless communication, combining artificial intelligence with wireless communication receivers is one of the current research hotspots, and the defects of traditional wireless communication receivers can be repaired to some extent.

[0003] As a digital modulation mode with high spectrum utilization and strong anti-interference, the quadrature phase shift keying (QPSK) technology is widely used in wireless communication systems. The QPSK receiver assisted by artificial intelligence is replacing and improving the traditional QPSK receiver. The traditional QPSK receiver adopts a hard decision method when demodulating the coded and noisy signal under low signal-to-noise ratio conditions, and the final bit error rate is high, and the recognition accuracy is not high.

[0004] Therefore, there is an urgent need for a solution to solve the problem of low recognition accuracy caused by noise interference when the receiver receives signals under low signal-to-noise ratio conditions. SUMMARY

[0005] Based on the above-mentioned defects and shortcomings in the prior art, one of the purposes of the present application is to at least solve one or more of the above-mentioned problems in the prior art, or in other words, one of the purposes of the present application is to provide a QPSK receiver based on artificial intelligence and an auxiliary model training method thereof which meet one or more of the above-mentioned needs.

[0006] In order to achieve the above-mentioned purposes of the application, the following technical solutions are adopted:

[0007] A QPSK receiver based on artificial intelligence and an auxiliary model training method thereof, specifically comprising the following steps:

[0008] S1, a QPSK system model is built, and a training data set is obtained using the QPSK system model;

[0009] S2, a rethink neural network model is built, and the rethink neural network model is used to demodulate the coded and noisy signal to generate a predicted label;

[0010] S3, training the rethink neural network model using the training data set, optimizing the loss function of the rethink neural network model;

[0011] S4, using the Nadam algorithm to optimize the loss function, updating the parameters of the rethink neural network model until the error rate of the predicted label compared with the real label is minimum, and taking the updated rethink neural network model as the QPSK receiver auxiliary model.

[0012] As a preferred scheme, the training data set is obtained using the QPSK system model, specifically, the data after Hamming encoding is modulated by QPSK using the QPSK system model, then oversampled by root raised cosine, and then added with Gaussian noise as interference, and 56-bit features are obtained after matched filtering and undersampling.

[0013] As a preferred scheme, the demodulation of the rethink neural network model on the signal affected by encoding and noise specifically includes:

[0014] The input signal is iterated five times repeatedly;

[0015] The first layer network is input, and the first layer network uses a nonlinear activation function Relu;

[0016] The second layer network is input, and the second layer network is a long short-term memory neural network with 128 neurons and uses an activation function sigmoid;

[0017] The full connection layer is input for output, and the full connection layer has 56 neurons.

[0018] As a further preferred scheme, the predicted label is 56 bits.

[0019] In a second aspect, the present application provides a QPSK demodulation method based on an artificial intelligence auxiliary model, using the QPSK receiver auxiliary model trained by any one of the above artificial intelligence-based QPSK receiver auxiliary model training methods, characterized in that it specifically includes:

[0020] The QPSK receiver auxiliary model is used to demodulate the signal affected by encoding and noise to generate a predicted label;

[0021] The predicted label is Hamming decoded to restore the original code.

[0022] As a preferred scheme, the Hamming decoding is (7, 4) decoding, and a 32-bit bit stream is output.

[0023] In a third aspect, the present application provides a QPSK intelligent receiver based on an artificial intelligence auxiliary model, using the above QPSK demodulation method based on an artificial intelligence auxiliary model.

[0024] Compared with the prior art, the present application has the beneficial effects that:

[0025] The model trained by the method of the present application adopts a rethink neural network based on a reflection structure, considers the hidden correlation between labels, simultaneously uses a multi-label classification method to improve the prediction performance, inputs information after Hamming coding, QPSK modulation, root raised cosine oversampling and Gaussian noise processing, and then matching filtering and undersampling, into the rethink neural network, so as to be able to demodulate in the demodulation link of the communication receiver, and the demodulation of the model can solve the problem of low recognition accuracy caused by noise interference under the condition of low signal-to-noise ratio in the traditional hard decision method. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 is a flowchart of the QPSK receiver auxiliary model training method based on artificial intelligence of the present application;

[0027] Figure 2 is a bit error rate comparison diagram of the demodulation method of the present application and the traditional hard decision method;

[0028] Figure 3 is a structural schematic diagram of the rethink neural network of the present application. DETAILED DESCRIPTION

[0029] In order to more clearly illustrate the embodiments of the present application, the specific embodiments of the present application will be described below with reference to the accompanying drawings. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained from these drawings without creative labor, and other embodiments can also be obtained.

[0030] Embodiment: The present application provides a QPSK receiver auxiliary model training method based on artificial intelligence, and a flowchart is as shown in Figure 1 In an embodiment of the present application, the method specifically comprises the following steps:

[0031] S1, a QPSK system model is built, and a training data set is obtained using the QPSK system model;

[0032] S2, a rethink neural network model is built, which is used for demodulating signals affected by coding and noise to generate predicted labels;

[0033] S3, the rethink neural network model is trained using the training data set, and the loss function of the rethink neural network model is optimized;

[0034] S4, using Nadam algorithm to optimize the loss function, updating the parameters of the rethink neural network model until the error rate of the predicted label compared with the real label is minimum, and taking the updated rethink neural network model as the QPSK receiver auxiliary model.

[0035] As a preferred scheme, in a further embodiment of step S1, the training data set is obtained using a QPSK system model, specifically, the data after Hamming encoding is modulated by QPSK using the QPSK system model, then oversampled by root raised cosine, and then added with Gaussian noise as interference, and then matched filtering and undersampling are performed to obtain 56-bit features.

[0036] In the QPSK system model, the data to be transmitted is (7, 4) Hamming encoded to obtain 56 bits at the sending end, then QPSK modulated, oversampled by root raised cosine, with 8 samples per symbol, then added with Gaussian noise as interference, and then matched filtering and undersampling are performed: one symbol is taken every 8 samples, then 56-bit features are obtained, and the 56-bit Hamming code obtained by Hamming encoding is taken as the label. All the features generated by the above QPSK system model and the corresponding labels together constitute the training data set.

[0037] In a further embodiment of step S2, the demodulation of the signal affected by encoding and noise by the rethink neural network model specifically includes:

[0038] Performing five repeated vector iterations on the input signal;

[0039] Inputting the first layer network, and the first layer network using a nonlinear activation function Relu;

[0040] Inputting the second layer network, and the second layer network being a long short-term memory neural network with 128 neurons and using an activation function sigmoid;

[0041] Inputting the full connection layer for output, and the full connection layer having 56 neurons.

[0042] As a further preferred scheme, the predicted label is 56 bits.

[0043] Specifically, the structure diagram of the rethink neural network is as follows: Figure 3As shown, the rethink network first iterates 5 times using a repeat vector (RepeatVector) according to the input shape, and then inputs to the first layer network, which has a total of 128 neurons, and uses a nonlinear activation function Relu. Then it is input to the second layer network, which uses a long short-term memory neural network (LSTM) and has a total of 128 neurons, and uses a sigmoid activation function. The last layer is a fully connected layer, which is the output layer and has a total of 56 neurons. The above network structure inputs the output result of the LSTM network to 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].

[0044] The above steps S3 and S4 use the rethink neural network to output the results of five iterations, and take the result of the fifth iteration as the output result, which is a 56-bit floating point number. After 01 decision, a 56-bit prediction label (pre_label) is obtained. By comparing the prediction label with the true label (true_label), the error rate is calculated, and the loss function of the rethink neural network model is calculated according to the output result of the neural network.

[0045] Then the Nadam algorithm is used to optimize the loss function, and the back propagation training is performed on the fully connected neural network model to update the parameter values and weights of each layer of neurons.

[0046] During the updating process, the model parameters are adjusted, the prediction label (pre_label) obtained by 01 decision of the output information is compared with the true label (true_label), and the error rate is calculated to be the smallest, that is, the training of the rethink neural network model is completed.

[0047] The following provides a specific implementation process in an embodiment of the present application:

[0048] In this embodiment, in step S1, we define the training set to be 1.8 million, of which 20 million data with signal-to-noise ratio -6~2dB (interval 1dB), and the test set is 850,000, of which 50,000 data with signal-to-noise ratio -6~2dB (interval 0.5dB). In step S1, a piece of data to be transmitted is a 01 bit stream, the information bit is 4 bits, a total of 8 groups, that is, a total of 32 bits. In the sending end, the data to be transmitted is (7, 4) Hamming coded to obtain 56 bits, and then QPSK modulation is performed, and root raised cosine is used for oversampling, and the number of samples per symbol is 8. Then Gaussian noise is added as interference, and after matched filtering and undersampling: one symbol is taken every 8 samples, then 56-bit features are obtained, and the 56-bit Hamming code obtained by Hamming coding is used as the label. All features and corresponding labels together form a data set.

[0049] In step S2, the rethink network first iterates 5 times using a repeat vector (RepeatVector) according to the input shape, and then inputs into the first layer network with 128 neurons, and the activation function used is the nonlinear activation function Relu. Then input into the second layer network, which adopts the long short-term memory neural network (LSTM) with 128 neurons, and the activation function used is sigmoid. The last layer is a fully connected layer as the output layer with 56 neurons. The above network structure inputs the output result of the LSTM network 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].

[0050] In steps S3 and S4, the rethink neural network outputs the results of five iterations, and takes the result of the fifth iteration as the output result, which is a 56-bit floating point number. After 01 decision, a 56-bit prediction label (pre_label) is obtained. By comparing the prediction label with the true label (true_label), the error rate is calculated, so as to calculate the loss function of the rethink neural network model according to the output result of the neural network.

[0051] Then the Nadam algorithm is used to optimize the loss function, and the back propagation training is performed on the fully connected neural network model to update the parameter values and weights of each layer of neurons.

[0052] During the updating process, the model parameters are adjusted, the prediction label (pre_label) obtained by 01 decision of the output information is compared with the true label (true_label), and the minimum error rate is calculated, that is, the training of the rethink neural network model is completed.

[0053] The model trained by the above method uses the rethink neural network based on the rethink structure, considers the hidden correlation between the labels, and uses the multi-label classification method to improve the prediction performance. The information after Hamming coding, QPSK modulation, root raised cosine oversampling, Gaussian noise processing, matching filtering and undersampling is input into the rethink neural network, so that the demodulation can be performed in the demodulation link of the communication receiver. The demodulation of the model can solve the problem that the traditional hard decision method has low recognition accuracy under low signal-to-noise ratio conditions due to noise interference.

[0054] In the second aspect, the application provides a QPSK demodulation method based on an artificial intelligence auxiliary model, which uses the QPSK receiver auxiliary model trained by any one of the QPSK receiver auxiliary model training methods based on artificial intelligence, and is characterized in that it specifically comprises:

[0055] The QPSK receiver auxiliary model is used to demodulate the encoded and noise-affected signal to generate a predicted label;

[0056] The predicted label is subjected to Hamming decoding to restore the original code.

[0057] In a preferred embodiment, the Hamming decoding is (7, 4) decoding, and a 32-bit bit stream is output. Specifically, the 56-bit predicted label (pre_label) is subjected to (7, 4) Hamming decoding to obtain a 32-bit bit stream, which is the restored information bit. The restored information bit is compared with the original information bit to calculate the bit error rate, which is used to determine the accuracy of the QPSK demodulation method of the application.

[0058] The above implementation uses MATLAB R2021b and pycharm 2020, tensorflow1.15.0 (python3.6) to simulate the above steps to perform simulation experiments, and calculate and draw a comparison chart of the bit error rates calculated by using the rethink neural network for demodulation and the traditional hard decision method for demodulation as the signal-to-noise ratio increases, as shown in Figure 2 .

[0059] In a third aspect, the application provides a QPSK intelligent receiver based on an artificial intelligence auxiliary model, which uses the above-mentioned QPSK demodulation method based on an artificial intelligence auxiliary model.

[0060] The receiver uses a rethink neural network based on a rethink structure, considers the hidden correlation between labels, and uses a multi-label classification method to improve the prediction performance. The information after Hamming encoding, QPSK modulation, root raised cosine oversampling, and Gaussian noise processing, and then matching filtering and undersampling is input into the rethink neural network, so that the receiver can demodulate in the demodulation link of the communication receiver. The demodulation of the receiver can solve the problem of low recognition accuracy caused by noise interference under the condition of low signal-to-noise ratio in the traditional hard decision method.

[0061] It should be noted that the above embodiments are only detailed descriptions of the preferred embodiments and principles of the application. For ordinary skilled persons in the art, the specific implementation methods will be changed according to the idea provided by the application, and these changes should be regarded as the protection scope of the application.

Claims

1. An artificial intelligence-based QPSK receiver-assisted model training method, characterized in that, Specifically comprising the following steps: S1, building a QPSK system model, using the QPSK system model to obtain a training data set; S2, building a rethink neural network model, the rethink neural network model is used for demodulating a signal affected by encoding and noise to generate a predicted label; S3, training the rethink neural network model using the training data set, and optimizing the loss function of the rethink neural network model; S4, using the Nadam algorithm to optimize the loss function, updating the parameters of the rethink neural network model until the error rate of the predicted label compared with the true label is minimum, and the updated rethink neural network model is used as a QPSK receiver auxiliary model; Using the QPSK system model to obtain a training data set is specifically using the QPSK system model to modulate the data after Hamming encoding, then using root raised cosine for oversampling, then adding Gaussian noise as interference, and obtaining 56-bit features after matched filtering and undersampling; Wherein, a data to be transmitted by the QPSK system model is a 01 bit stream, the information bits are 4 bits, a total of 8 groups, i.e. a total of 32 bits; In the QPSK system model, the data to be transmitted is (7, 4) Hamming encoded at the sending end to obtain 56 bits, then QPSK modulated, and root raised cosine is used for oversampling, the number of samples per symbol is 8, then Gaussian noise is added as interference, and after matched filtering and undersampling: one symbol is taken every 8 samples, then 56-bit features are obtained, and the 56-bit Hamming code obtained by Hamming encoding is used as a label; All features generated by the above QPSK system model and the corresponding labels together constitute a training data set.

2. The AI-based QPSK receiver assistance model training method of claim 1, wherein, The demodulation of the rethink neural network model to the signal affected by encoding and noise specifically comprises: Five repeated vector iterations are performed on the input signal; Input the first layer network, the first layer network uses a nonlinear activation function Relu; Input the second layer network, the second layer network is a long short-term memory neural network with 128 neurons, using the activation function sigmoid; Input the full connection layer for output, the full connection layer has 56 neurons.

3. The AI-based QPSK receiver assistance model training method of claim 2, wherein, The predicted label is 56 bits. 4.A QPSK demodulation method based on an artificial intelligence aided model, using a QPSK receiver aided model trained by the artificial intelligence based QPSK receiver aided model training method of any one of claims 1-3, characterized in that, Specifically comprising: Using the QPSK receiver auxiliary model to demodulate the signal affected by encoding and noise to generate a predicted label; Hamming decoding is performed on the predicted label to restore the original code.

5. The QPSK demodulation method based on an artificial intelligence auxiliary model according to claim 4, characterized in that, The Hamming decoding is (7, 4) decoding, and a 32-bit bit stream is output.

Images