A Method, Device and System for Automatic Modulation Recognition in Wireless Communication

Abstract

The present invention provides a neural network system for automatic modulation recognition of wireless communication, including an input module, a residual unit, and an output module. Among them, the input module is used to receive the wireless communication signal to be recognized, obtain the features of the signal, and transmit the features to the residual unit. The residual unit includes a ConvBlackA unit and multiple ConvBlackB units. ConvBlackA contains three Conv2D layers, two BN layers, two Gaussian Dropout layers, and two PReLU layers. ConvBlackB adds one BN layer, one Gaussian Dropout layer, and one PReLU layer on the basis of ConvBlackA. The output module is used to receive the output of the residual unit and generate the modulation recognition result of the signal. Compared with the existing deep learning models under high signal-to-noise ratio, the proposed GuResNet has better modulation recognition accuracy than other DL models. Under low signal-to-noise ratio, the proposed signal-to-noise ratio perception mechanism can significantly improve the recognition performance.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technologies, and particularly to a method and system for high-precision intelligent modulation recognition in a low signal-to-noise ratio environment. Background Art

[0002] Automatic modulation recognition (AMR) technology has been widely used in both military and civilian fields, such as electronic countermeasure, signal detection, and spectrum detection. Existing automatic modulation recognition methods are generally divided into likelihood-based (LB) methods and feature-based (FB) methods. Specifically, likelihood-based methods construct a likelihood function under the assumption of known channel information (e.g., channel fading model). Then, the likelihood function is maximized to obtain the possible modulation categories of the received signal. However, it is difficult to obtain accurate channel information in the actual environment, which leads to the difficulty of widespread application of LB methods. Feature-based methods can significantly solve the above problems. Traditional feature-based methods include two steps: feature extraction and pattern recognition. For feature extraction, signal features are manually extracted, such as higher-order cumulants, cyclic spectrum features, and constellation diagram features. For pattern recognition, a suitable classifier (e.g., decision tree, support vector machine) is constructed based on the extracted features to recognize the features, thereby distinguishing different modulation methods of the signal. However, traditional FB methods rely heavily on manually extracted signal features, and these features may be of low quality, resulting in low modulation recognition accuracy.

[0003] In recent years, deep learning (DL) has been successfully applied to various fields, such as computer vision and speech analysis. Deep learning relies on multi-layer neural networks to automatically abstract and extract high-quality features, demonstrating strong classification and prediction capabilities. Therefore, feature-based methods based on deep learning have begun to be applied to AMR tasks to solve the problems existing in traditional FB methods. Existing research generally uses mature DL architectures for image classification tasks (e.g., convolutional neural networks), which perform well in additive white gaussian noise (AWGN) and multipath channels. However, in actual communication channels, there are usually some other natural and artificial impairments, such as carrier frequency offset, clock drift, and phase offset. And these impairments cause unknown scaling, translation, and flipping of the signal, thus increasing the difficulty of AMR. Existing research, due to directly adopting DL models without combining the characteristics of the signal itself (e.g., time-domain signal envelope features) for targeted design, is difficult to ensure high recognition accuracy in a channel environment containing impairments. Therefore, it is necessary to explore more effective DL architectures and combine signal characteristics to improve AMR performance.

[0004] The deep residual network (ResNet) mainly includes an input module, residual units, and an output module, and its performance in image classification tasks is better than that of some other DL models (e.g., GoogLeNet model, VGG model). Some studies have begun to apply residual networks for AMR in more practical communication scenarios, but they cannot provide good performance even at low signal-to-noise ratios (signal-to-noise ratio below 0 dB). The existing problems are as follows: For the input module, a large-sized 7×7 convolutional kernel is first used, making it difficult to learn the details of signal features. Immediately after the convolution operation is a pooling layer, which contains some downsampling operations and loses some signal features. For the residual units, residual networks usually simply stack multiple residual units (e.g., 16 residual units) to abstract high-level feature maps (e.g., signal envelope features) based on low-level features (e.g., edge, gradient features), resulting in a more complex network that is prone to overfitting. For the output module, a fully connected layer is usually used to convert the feature map into categories, greatly increasing the network training parameters and making the network have poor generalization ability.

[0005] In addition, in a low signal-to-noise ratio scenario, the signal is easily submerged in noise, making it difficult for deep learning models to effectively abstract signal features. Moreover, existing model training mechanisms usually do not distinguish signal-to-noise ratios when training deep learning models, increasing the complexity of model recognition. Summary of the Invention

[0006] In view of the above problems, according to the first aspect of the present invention, a neural network system for automatic modulation recognition in wireless communication is proposed, including an input module, residual units, and an output module. Among them, the input module is used to receive the wireless communication signal to be recognized, acquire the features of the signal, and transmit the features to the residual units. The residual units include one ConvBlackA unit and multiple ConvBlackB units. ConvBlackA contains three Conv2D layers, two BN layers, two Gaussian Dropout layers, and two PReLU layers. ConvBlackB adds one BN layer, one Gaussian Dropout layer, and one PReLU layer on the basis of ConvBlackA. The output module is used to receive the output of the residual units and generate the modulation recognition result of the signal.

[0007] In an embodiment of the present invention, the residual units include 1 ConvBlackA unit and 5 ConvBlackB units.

[0008] In an embodiment of the present invention, the input module includes: a 5×5 convolutional layer, a BN layer, and a PReLU layer.

[0009] In one embodiment of the present invention, the output module includes: a PReLU layer, a Gaussian Dropout layer, a GAP layer, and a Softmax layer.

[0010] According to a second aspect of the present invention, there is provided a method for training an automatic modulation recognition model of a wireless communication system, including

[0011] Step 200: Using signals with a signal-to-noise ratio less than or equal to a low signal-to-noise ratio threshold, extracting five-dimensional features of the signals using second-order Legendre polynomials, and forming a matrix with five rows using the five-dimensional features of multiple consecutive sampling points as a feature matrix, and using this feature matrix as a training sample to obtain multiple training samples corresponding to different low signal-to-noise ratio signals;

[0012] Step 300: Using the multiple training samples corresponding to different low signal-to-noise ratio signals to train the neural network system for automatic modulation recognition of the present invention to obtain multiple neural network models corresponding to different low signal-to-noise ratios, where each low signal-to-noise ratio neural network model corresponds to a low signal-to-noise ratio range.

[0013] In one embodiment of the present invention, before step 200, it further includes training the neural network system for automatic modulation recognition of the present invention using signal samples at all signal-to-noise ratios to obtain a general network model and hyperparameters, including: forming a matrix with two rows using the in-phase component and quadrature component features of multiple consecutive sampling points of the signal as a feature matrix, and using this feature matrix as a training sample and training using multiple training samples.

[0014] According to a third aspect of the present invention, there is provided an automatic modulation recognition method for a wireless communication system,

[0015] including:

[0016] Step 1000: For the received baseband complex signal, using a signal-to-noise ratio estimation algorithm and utilizing the time-domain and frequency-domain characteristics of the signal to estimate the signal-to-noise ratio of the signal;

[0017] Step 2000: Extracting signal features, when the signal-to-noise ratio is less than or equal to a predetermined low signal-to-noise ratio threshold, extracting five-dimensional features of the signal using second-order Legendre polynomials, forming a matrix with five rows using the five-dimensional features of multiple consecutive sampling points, using it as a feature matrix, and inputting the feature matrix into a low signal-to-noise ratio neural network model corresponding to the low signal-to-noise ratio range applicable to the signal trained by the training method of the automatic modulation recognition model of the wireless communication system of the present invention to identify the modulation method of the signal.

[0018] In one embodiment of the present invention, step 2000 further includes:

[0019] Step 2100: If the signal-to-noise ratio of the signal is greater than the low signal-to-noise ratio threshold, use the in-phase component and quadrature component features of multiple consecutive sampling points to form a matrix with 2 rows, and use it as a feature matrix. Input the feature matrix into the general model trained by the training method of the automatic modulation recognition model of the wireless communication system of the present invention to identify the modulation method of the signal.

[0020] According to the fourth aspect of the present invention, there is provided a computer-readable storage medium storing one or more computer programs, which are used to implement the automatic modulation recognition method of the wireless communication system of the present invention or the training method of the automatic modulation recognition model of the wireless communication system when executed.

[0021] According to the fifth aspect of the present invention, there is provided a computing system, including:

[0022] a storage device, and one or more processors;

[0023] wherein, the storage device is used to store one or more computer programs, and the computer programs are used to implement the automatic modulation recognition method of the wireless communication system of the present invention or the training method of the automatic modulation recognition model of the wireless communication system when executed by the processor.

[0024] In the present invention, at high signal-to-noise ratios, compared with existing deep learning models, the proposed GuResNet has better modulation recognition accuracy than other DL models; at low signal-to-noise ratios, the proposed signal-to-noise ratio perception mechanism can significantly improve the recognition performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The accompanying drawings here are incorporated into the specification and form a part of the specification, showing embodiments consistent with the present invention, and are used together with the specification to explain the principles of the present invention. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings based on these drawings without creative efforts. In the drawings:

[0026] Figure 1 shows the automatic modulation recognition flowchart of the embodiment of the present invention;

[0027] Figure 2 shows the general wireless communication system model of the present invention;

[0028] Figure 3 shows the deep residual neural network structure diagram of the embodiment of the present invention;

[0029] FIG. 4(a) shows the structure diagram of the residual block ConvBlackA of the deep residual neural network of the present invention;

[0030] Figure 4(b) shows the structural diagram of the residual block ConvBlackB of the deep residual neural network of the present invention;

[0031] Figure 5 Shows the relationship between the classification accuracy of the proposed method SG-NET and the baseline method and the SNR;

[0032] Figure 6 Shows the relationship between the classification accuracy of the proposed method SG-NET and the baseline method and the SNR estimation error. Detailed implementation manner

[0033] In view of the problems raised in the background art, the inventors conducted research and proposed a wireless communication automatic modulation recognition method, device and system with the goal of improving the modulation recognition accuracy. Figure 1 Shows the automatic modulation recognition flow chart of the present invention. The transmitting end of the wireless communication system modulates the signal and sends it out. The receiving end receives a large amount of modulated baseband complex signal data. The signal preprocessing extracts signal features from the complex signal, and then the trained deep residual network identifies the modulation signal type according to the signal features. Due to the signal being interfered by various noises such as AWGN, multipath effect, carrier frequency offset and clock drift, the signal appears unknown scaling, translation and flipping. Therefore, it is necessary to preprocess the signal, extract signal features, and reduce the influence of noise. When the SNR exceeds 0 dB, considering that the signal is less interfered by noise, no processing is performed on the signal; when the SNR is lower than 0 dB, considering that the signal is more interfered by noise, the second-order Legendre polynomial method is used to extract the signal interaction features. The present invention also designs a novel robust deep residual network model to improve the modulation signal type recognition efficiency.

[0034] The following is a detailed introduction to the present invention.

[0035] 1. Receive baseband complex signal

[0036] Figure 2 Shows the general wireless communication system model in the present invention. Signal b n ∈ {0, 1} is mapped to a new binary sequence through the source coding (including channel coding) of the transmitter, and this sequence is mapped to symbol s n . Through the digital-to-analog converter, s n is mapped to an analog continuous baseband signal where g T (t) is the signal pulse. Through up-conversion, the band-pass signal s(t) can be obtained

[0037]

[0038] where f c is the carrier frequency generated by the transmitter, and T b is the symbol period.

[0039] Considering the multipath fading channel h(t,τ), which includes some actual channel impairments such as carrier frequency offset Δ Lo (t) and sampling rate offset Δ clk (t), the general received signal r(t) is expressed as

[0040]

[0041] where n add (t) represents Gaussian white noise with a mean of 0 and a variance of , τ is the multipath delay of time t, and τ0 is the maximum delay spread.

[0042] For the received signal r(t), it is sampled at a frequency to generate a discrete sequence r n = r(nT s ), where n ∈ {1, 2, …, M} is the sampling ordinal number of the sampling point, M is the number of samplings, and T s is the sampling period. Among them, the received signal r(t) includes two parts, the in-phase component r I and the quadrature component r Q , and the discrete sequence r n also includes the in-phase component r I (n) and the quadrature component r Q (n)

[0043] r n = r I (n) + jr Q (n), (3)

[0044] where r I (n) and r Q (n) form an IQ sequence. The data of M sampling points form r

[0045] r = {r1, r2, ..., r M} (4)

[0046] 2. Signal preprocessing

[0047] The present invention uses a low signal-to-noise ratio threshold to distinguish high signal-to-noise ratio and low signal-to-noise ratio. The low signal-to-noise ratio threshold can be 0 dB or 5 dB, or can be set to other values according to needs. Signals with a signal-to-noise ratio lower than the low signal-to-noise ratio threshold are of low signal-to-noise ratio, and signals higher than the low signal-to-noise ratio threshold are of high signal-to-noise ratio. In high signal-to-noise ratio scenarios, the signal is less affected by noise. To avoid increasing the complexity of the entire system, the present invention directly uses the in-phase component and quadrature component of the signal as features. The in-phase component and quadrature component of the nth sampling point are respectively used as the two elements of the nth column of the matrix. The in-phase components and quadrature components of M sampling points form a 2×M matrix, which serves as a training sample for the deep residual neural network. Preferably, M is 128.

[0048] In low signal-to-noise ratio scenarios, the signal is greatly affected by noise, resulting in the signal features being easily submerged in the noise and increasing the difficulty of modulation recognition. Therefore, the present invention innovatively uses Legendre polynomials to preprocess the signal, extracts the interaction features of the in-phase and quadrature components of the signal, and enhances the AMR performance under low signal-to-noise ratio. Among them, the Legendre polynomial can be expressed as

[0049]

[0050] where m is the order of the Legendre polynomial, and x = [x I , x Q is the received IQ signal data. If the Legendre polynomial has an even order, q ∈ [0, m / 2]; otherwise, q ∈ [0, (m - 1) / 2]. To extract the interaction features without making the entire method more complex, the present invention uses the second-order Legendre polynomial

[0051]

[0052] Therefore, the extracted second-order Legendre polynomial features are x I , x Q , a total of five dimensions. Among them, is the interaction feature of the extracted IQ signal. At the nth sampling point, in formula (3), r I (n) is x I , and r Q (n) is x Q . The above five features of the nth sampling point are used as the nth column of the matrix to form a 5×M matrix, which serves as a training sample for the deep residual neural network. Preferably, M is 128.

[0053] Compared with the original IQ data, using the second-order Legendre polynomial for feature augmentation increases the diversity of the data and further improves the robustness of model training, thereby increasing the recognition rate under low signal-to-noise ratio.

[0054] The present invention is a model trained separately for low signal-to-noise ratio (SNR) and high SNR. For the case of low SNR, in order to make the trained model more accurate, different models are trained for different SNR ranges. For example, one model is trained for an SNR range of 0 dB to -10 dB, one model is trained for -10 dB to -20 dB, and one model is trained for an SNR below -20 dB.

[0055] 3. Train CNN deep residual network

[0056] For the modulation recognition task, the recognition rate of directly using the traditional CNN deep residual network is relatively low. The present invention proposes a novel deep residual network model that significantly improves the recognition rate. First, the traditional CNN deep residual network is introduced below.

[0057] 3.1 Traditional CNN deep residual network

[0058] The basic structure of the traditional CNN consists of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The input layer is used for data preprocessing, such as normalization, data augmentation (for example, when the training data is limited, some new data can be generated from the existing training dataset through some transformations), random cropping, etc. CNN generally uses multiple convolutional layers and pooling layers, and the convolutional layer and the pooling layer are alternately arranged, that is, one convolutional layer is connected to one pooling layer, and after the pooling layer, another convolutional layer is connected, and so on. The convolutional layer is used to extract the features of the image, and each convolutional layer extracts different features, while the pooling layer is used to sample the features, such as mean pooling and max pooling. Usually, the neurons in the convolutional layer are only connected to some neuron nodes in the previous layer, and the fully connected layer also uses convolutional operations, but it is connected to some neuron nodes in the previous layer and is used to connect the extracted feature maps. The output layer is used to output the classification result.

[0059] With the enhancement of computing power, the number of layers of neural networks is increasing. However, when the number of layers is too large, phenomena such as overfitting will occur. It is not that the more layers, the better the training effect. To solve this problem, the CNN deep residual neural network is proposed. For a given layer of the network (the nth layer), its input is no longer only the output of the previous layer (the n - 1th layer), but the sum of the output of the previous layer (the n - 1th layer) and the output of the layer before the previous layer (the n - 2th layer), that is, the output of the n - 2th layer bypasses the n - 1th layer and reaches the nth layer. Further, the bypass can be not limited to one layer and can also bypass multiple layers.

[0060] 3.2 CNN deep residual neural network GuResNet structure of the present invention

[0061] For the recognition of signal modulation modes, the present invention proposes a novel deep residual network model, named GuResNet. Similar to the traditional CNN deep residual neural network, GuResNet includes an input module, a residual unit, and an output module.Figure 3 The structure of GuResNet is shown. Compared with the traditional CNN deep residual neural network, the input module of the present invention particularly uses a convolutional kernel of size 5×5 and PReLU to extract low-level features (edge and gradient information) of the signal. In the residual unit of the present invention, a Gaussian Dropout layer is innovatively introduced to prevent the model from overfitting. For the output module, the present invention uses GAP to replace the fully connected layer in the traditional residual network, improving the generalization ability of the model.

[0062] ⑴ Input module: This module mainly includes three layers, namely the two-dimensional convolutional Conv2D layer, the BN layer, and the activation function layer. Specifically, for the Conv2D layer, the present invention uses a convolutional kernel of 5×5 to replace the convolutional kernel of 7×7 in the traditional residual network to learn the local correlation of samples and preliminarily abstract low-level features from the input signal, such as edge and gradient information, etc. Using a convolutional kernel of 5×5 can reduce the number of network parameters and enhance the non-linear expression ability of the network, thereby abstracting more useful low-level features of the signal. Preferably, the number of its kernels is 32. The BN layer can improve the generalization ability of the model to a certain extent. In addition, the present invention uses PReLU to replace the activation function ReLU in the traditional residual network to prevent the model from losing some signal features. This layer can be represented as

[0063]

[0064] where c i is the input feature of the activation function, and p i is the training parameter.

[0065] ⑵ Residual unit: This module includes six residual units, one ConvBlackA unit and five ConvBlackB units. As shown in Figure 4(a), ConvBlackA includes three Conv2D layers, two BN layers, two Gaussian Dropout layers and two PReLU layers. Preferably, the convolutional kernels of the three Conv2D layers are 1x1, 3x3, 1x1 from top to bottom, and the number of convolutional kernels is 32, 64, 32 from top to bottom. As shown in Figure 4(b), ConvBlackB adds a BN layer, a Gaussian Dropout layer and a PReLU layer compared with ConvBlackA to further improve the learning ability of the model. The number of ConvBlackB units included in the present invention is 5. The deeper network is not necessarily better with more layers. The selected one ConvBlackA unit and five ConvBlackB units of residual units in the present invention include multiple layers of networks, which is the optimal embodiment in terms of the number of layers.

[0066] The BN layer and the PReLU layer have the same functions as the corresponding layers in the input module. The Conv2D layer further learns the high-level features of the signal from the low-level features initially abstracted by the input layer, such as the signal envelope features.

[0067] Next, the Gaussian Dropout layer is shown in detail. The present invention considers that the channel is time-varying and there is interference from some multiplicative noise. Therefore, the present invention innovatively introduces a Gaussian Dropout layer in the traditional residual unit to reduce the interference of noise. By multiplying a random variable with a Gaussian distribution, the neurons in the residual unit are more insensitive to noise, thereby improving the AMR performance of the model. Among them, after adding Gaussian Dropout, the forward propagation operation can be expressed as

[0068]

[0069]

[0070]

[0071] where * represents the multiplication of the corresponding elements of the vectors s (v) and y (v) , v ∈ {1,..., V} is the hidden layer index, y (v) is the output vector of the v-th layer, s (v) is a vector, where the j-th element is a Gaussian random variable with a mean of 1 and a variance of , where e is a training parameter, z (v+1) is the output vector of the (v + 1)-th layer, is the i-th element of z (v+1) , is the input of the (v + 1)-th layer, is the i-th weight vector of the weight matrix w (v+1) of the (v + 1)-th layer, b i (v+1) is the i-th element of the bias vector b (v+1) of the (v + 1)-th layer.

[0072] (3) Output module: As Figure 3 shown, this module includes a PReLU layer, a Gaussian Dropout layer, a GAP layer, and a Softmax layer. Among them, the present invention uses a GAP layer to replace the fully connected layer in the traditional residual network, reducing the training parameters of the model, making the GuResNet architecture more robust, and thus having a more excellent prediction accuracy, while Softmax is the activation function layer of the output layer Dense layer.

[0073] 4. Automatic modulation recognition model training method

[0074] For the training of the deep residual network model, first, a general GuResNet network model and hyperparameters are trained using the original signal samples at all signal-to-noise ratios. That is, the in-phase components and quadrature components of multiple consecutive sampling points are used to form a matrix with 2 rows as a feature matrix, and this feature matrix is used as a training sample. Multiple training samples are used for training. The hyperparameters of the trained network model include batch size, optimization function, initial learning rate, and maximum number of training times, as shown in Table 1. These hyperparameters are used to guide the network learning and adjust the network parameters to obtain a good prediction accuracy.

[0075] Table 1 Hyperparameters of GuResNet

[0076] Hyperparameter Value Batch size 64 Optimization function Adam Initial learning rate <![CDATA[2×10 -4 > Maximum number of training times 500

[0077] When the signal-to-noise ratio is greater than the low signal-to-noise ratio threshold, since the signal is less affected by noise interference, the present invention directly performs AMR on the original signal samples using the general GuResNet architecture, hyperparameters, and parameters, and satisfactory recognition performance can be obtained. In order not to increase the complexity of the entire method, the model is no longer trained.

[0078] When the signal-to-noise ratio is less than or equal to the low signal-to-noise ratio threshold, the signal is easily submerged in noise, resulting in the DL model being unable to effectively abstract the signal features and recognize the modulation type. Therefore, in order to further improve the modulation recognition performance in the low signal-to-noise ratio scenario, different network parameters of GuResNet are retrained based on the perceived different signal-to-noise ratios. Before training, it is necessary to determine the number of models to be trained and the signal-to-noise ratio range applicable to each model, and then obtain a certain number of baseband complex signals, and then train according to the following steps:

[0079] Step A1: For the received baseband complex signal, use a signal-to-noise ratio estimation algorithm to estimate the signal-to-noise ratio of the signal by utilizing the time-domain and frequency-domain characteristics of the signal;

[0080] Step A2: Use the second-order Legendre polynomial to extract the five-dimensional features of the signal. The five-dimensional features of multiple consecutive sampling points are used to form a matrix with 5 rows as a feature matrix, and this feature matrix is used as a training sample; obtain multiple training samples;

[0081] Step A3: According to the signal-to-noise ratio of the signal estimated in Step A1, use the training samples obtained in Step A2 for this signal to train the model applicable to the signal-to-noise ratio of this signal.

[0082] 5. Automatic modulation recognition method

[0083] The above-trained model can be used for automatic modulation recognition, and the recognition method is as follows:

[0084] Step B1: Same as Step A1;

[0085] Step B2: Extract signal features. If the signal-to-noise ratio (SNR) of the signal is greater than the low SNR threshold, use the in-phase and quadrature components features of multiple consecutive sampling points to form a matrix with 2 rows, and use it as a feature matrix. If the SNR is less than or equal to the low SNR threshold, use the second-order Legendre polynomial to extract the five-dimensional features of the signal, and use the five-dimensional features of multiple consecutive sampling points to form a matrix with 5 rows, and use it as a feature matrix;

[0086] Step B3: Input the feature matrix into the model applicable to the SNR of the signal for modulation mode recognition. Among them, when the SNR is greater than the low SNR threshold, use the general model for modulation mode recognition.

[0087] The following are the tests conducted by the present invention and the comparisons with other methods.

[0088] The present invention uses the RML2016.10b_dict dataset to obtain time-domain complex baseband signals, which include ten modulation modes, namely:

[0089] · Digital modulation: BPSK, QPSK, 8PSK, 16QAM, 64QAM, CPFSK, GFSK, PAM4

[0090] · Analog modulation: AM-DSB, WBFM

[0091] This dataset contains 1,200,000 samples, the SNR ranges from -20 dB to +18 dB, the SNR interval is 2 dB, and each sample contains 128 sample points. In addition, the signal data in this dataset contains some actual signal impairments, such as AWGN, multipath fading, carrier frequency offset, and phase offset.

[0092] The present invention compares the proposed method (SG-NET) with some other baseline algorithms (MaxConvNet and DrCNN). Among them, MaxConvNet contains eight layers, two convolutional layers, two max-pooling layers, and four fully connected layers, and DrCNN contains six layers, two convolutional layers, and four fully connected layers.

[0093] Figure 5It shows the relationship between the classification accuracy of the method SG-NET proposed by the present invention (SG-NET consists of the deep residual network model GuResNet and its signal-to-noise ratio perception mechanism) and the baseline methods (MaxConvNet and DrCNN) varying with SNR. It can be seen from the figure that when the SNR exceeds 0 dB, the recognition accuracy of GuResNet can reach more than 90%, which is nearly 6% and 10% higher than those of the DrCNN and MaxConvNet methods respectively. Among them, since the analog signal only has one carrier tone during the silent period, confusion will also occur between WBFM and AM-DSB at high signal-to-noise ratios, and the recognition rate no longer increases significantly. In addition, in the low signal-to-noise ratio scenario, the proposed signal-to-noise ratio perception mechanism can obtain an average accuracy gain of 30%. In particular, when the SNR is between -14 dB and -8 dB, the method proposed by the present invention can obtain a performance gain of 37% compared with other methods. This means that the SG-NET proposed by the present invention has significant advantages at low signal-to-noise ratios.

[0094] To evaluate the robustness of the SG-NET proposed by the present invention, the SNR estimation error is defined as the difference between the true SNR and the estimated SNR, and the classification accuracy is a function of the SNR estimation error, as Figure 6 shown. Among them, data at high and low signal-to-noise ratios are selected to test these methods. It can be seen from the figure that compared with MaxConvNet and DrCNN, the SG-NET proposed by the present invention always has a higher noise tolerance at high signal-to-noise ratios. In addition, when the SNR estimation error is within 2 dB and -2 dB, SG-NET still has better performance than the baseline methods at low signal-to-noise ratios. However, when the absolute value of the SNR estimation error exceeds 2 dB, compared with the baseline methods, it is more difficult for the SG-NET proposed by the present invention to accurately identify the signal modulation mode. The reason is that there are some differences in signal characteristics at different SNRs, and SG-NET needs to be retrained at different SNRs to obtain excellent performance. Currently, the absolute value of the SNR estimation error can be reduced to less than 1 dB. Therefore, the SG-NET proposed by the present invention can be applied to practical scenarios with high robustness.

[0095] It should be noted that although the above steps are described in a specific order, it does not mean that the steps must be executed in the above specific order. In fact, some of these steps can be executed concurrently or even in a different order, as long as the required functions can be achieved.

[0096] The present invention can be a system, a method, and / or a computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions thereon for causing a processor to implement various aspects of the present invention.

[0097] A computer-readable storage medium can be a tangible device that retains and stores instructions for use by an instruction execution device. A computer-readable storage medium may include, for example, but is not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer-readable storage medium include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disc (DVD), a memory stick, a floppy disk, a mechanically encoded device, such as a punched card or raised structures in grooves having instructions stored thereon, and any suitable combination of the foregoing.

[0098] The foregoing is described around the present disclosure to enable any ordinary person skilled in the art to implement or use the present disclosure. For those of ordinary skill in the art, various modifications to the present disclosure are obvious, and the general principles defined herein can also be applied to other variations without departing from the spirit or scope of the present disclosure. In addition, unless otherwise stated, all or part of any aspect and / or embodiment can be used with all or part of any other aspect and / or embodiment. Therefore, the present disclosure is not limited to the examples and designs described herein, but is consistent with the broadest scope of the principles and novel features disclosed herein.

Claims

1. A neural network system for automatic modulation recognition in wireless communication, comprising an input module, a residual unit, and an output module; Among them, The input module includes a 5×5 convolutional layer, a BN layer, and a PReLU layer, and is used to receive the wireless communication signal to be recognized, obtain the features of the signal, and transmit the features to the residual unit; The residual unit includes one ConvBlackA unit and multiple ConvBlackB units. The ConvBlackA unit contains three Conv2D layers, two BN layers, two Gaussian Dropout layers, and two PReLU layers. The ConvBlackB unit adds one BN layer, one Gaussian Dropout layer, and one PReLU layer on the basis of the ConvBlackA unit; The output module includes a PReLU layer, a Gaussian Dropout layer, a GAP layer, and a Softmax layer, and is used to receive the output of the residual unit and generate the modulation recognition result of the signal.

2. The neural network system according to claim 1, wherein the residual unit includes 5 ConvBlackB units.

3. A training method for an automatic modulation recognition model of a wireless communication system, comprising Step 200: Using signals with a signal-to-noise ratio less than or equal to the low signal-to-noise ratio threshold, extract the five-dimensional features of the signals using the second-order Legendre polynomial, and use the five-dimensional features of multiple consecutive sampling points to form a matrix with 5 rows as a feature matrix. Take this feature matrix as a training sample, and obtain multiple training samples corresponding to different low signal-to-noise ratio signals; Step 300: Use the multiple training samples corresponding to different low signal-to-noise ratio signals to train the system according to claim 1 or 2 to obtain multiple neural network models corresponding to different low signal-to-noise ratios, where each low signal-to-noise ratio neural network model corresponds to a low signal-to-noise ratio range.

4. According to the method described in claim 3, before step 200, it further includes training the system described in claim 1 or 2 using signal samples at all signal-to-noise ratios to obtain a general neural network model and hyperparameters, including: Use the in-phase component and quadrature component features of multiple consecutive sampling points of the signal to form a matrix with 2 rows as a feature matrix. Take this feature matrix as a training sample and use multiple training samples for training.

5. An automatic modulation recognition method for a wireless communication system, comprising: Step 1000: For the received baseband complex signal, use a signal-to-noise ratio estimation algorithm, and utilize the time-domain and frequency-domain characteristics of the signal to estimate the signal-to-noise ratio of the signal; Step 2000: Extract signal features. When the signal-to-noise ratio is less than or equal to the predetermined low signal-to-noise ratio threshold, use the second-order Legendre polynomial to extract the five-dimensional features of the signal, and use the five-dimensional features of multiple consecutive sampling points to form a matrix with 5 rows, and take it as a feature matrix. Input the feature matrix into the low signal-to-noise ratio neural network model corresponding to the low signal-to-noise ratio range applicable to the signal trained by the method according to claim 3 or 4 to identify the modulation method of the signal.

6. The method according to claim 5, wherein step 2000 further includes: Step 2100: If the signal-to-noise ratio of the signal is greater than the low signal-to-noise ratio threshold, use the in-phase component and quadrature component features of multiple consecutive sampling points to form a matrix with two rows, and use it as a feature matrix. Input the feature matrix into the general model trained by the method described in claim 4 to identify the modulation method of the signal.

7. A computer-readable storage medium storing one or more computer programs which, when executed, are configured to implement the method according to any one of claims 3-6.

8. A computing system, comprising: a storage device, and one or more processors; wherein the storage device is configured to store one or more computer programs which, when executed by the processor, are configured to implement the method according to any one of claims 3-6.

Images