Signal-to-noise ratio adaptive signal modulation identification method

Abstract

The invention relates to a signal-to-noise ratio adaptive signal modulation identification method. The method comprises the following steps: inputting IQ information and a signal-to-noise ratio (SNR) estimated value of a signal into a signal modulation identification model, and identifying to obtain a modulation mode of the signal; the signal modulation identification model is an MCLDNN of which the feature fusion module is provided with an SNR attention module; the SNR attention module can receive the SNR estimated value and obtain a signal-to-noise ratio attention weight according to the SNR estimated value; receiving a signal multi-channel fusion feature map output by the multi-channel fusion convolutional layer of the MCLDNN; weighting the signal multi-channel fusion feature map by using a signal-to-noise ratio attention weight, connecting the weighted signal multi-channel fusion feature map with a signal multi-channel fusion feature map residual error, and outputting a multi-channel fusion feature combined with SNR attention; and the multi-channel fusion features combined with the SNR attention are input to a time sequence module of the MCLDNN.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of signal processing, and particularly relates to a signal modulation recognition method with adaptive signal-to-noise ratio. BACKGROUND

[0002] Modulation recognition technology in wireless communication is a key link in the field of signal processing, and its core goal is to automatically identify the modulation mode from the received signal. The recognition method can be divided into likelihood-based method and feature-based method.

[0003] The likelihood-based method constructs a probability model of the signal, calculates the likelihood function of the received signal belonging to each modulation type, and makes a classification decision by comparing the likelihood ratio. Typical methods include average likelihood ratio test (ALRT) and generalized likelihood ratio test (GLRT), but they have high requirements for prior knowledge, high computational complexity and limited generalization ability.

[0004] The feature-based method extracts features from the signal and realizes classification through a classifier (such as a machine learning model or a neural network).

[0005] Among them, the manual feature method has low computational complexity, but the feature design depends on domain knowledge, is difficult to adapt to new modulation or complex channel, and has limited generalization ability. Typical methods include the modulation recognition method based on SVM proposed by X. Zhang et al. of Harbin Engineering University, the modulation recognition method based on random forest proposed by Z. Zhang et al. of Harbin Engineering University, etc.

[0006] Deep learning feature-based methods can automatically extract multi-level features from raw signals, have strong high-dimensional feature expression ability, and are suitable for complex signals and channels, but have large computational complexity and poor interpretability. Typical methods include the modulation recognition method of the convolutional neural network (CNN) proposed by T.J. O'Shea et al. of the University of Maryland, Park Branch, the modulation recognition method of the two-layer gated recurrent unit (GRU) proposed by D. Hong et al. of the University of Science and Technology of China, the CLDNN (Convolutional Long Short-Term Memory Neural Network) modulation recognition method proposed by N.E. West et al. of the Oklahoma State University, the CGDNet network modulation recognition method proposed by Njoku et al. of the Korea National Institute of Science and Technology, the modulation recognition method based on the LSTM denoising autoencoder proposed by Ziqi Ke et al. of the University of Texas at Austin, the modulation recognition method based on the multi-column convolutional long short-term memory deep neural network (MCLDNN, Multi-Column Long Short-Term Memory Deep Neural Network) proposed by Jialang Xu et al. of the University of Electronic Science and Technology, and the modulation recognition method of the multi-channel modulation recognition model (RLITNN) based on multiple features proposed by Luo Zhongqiang et al. of Sichuan Light Industry University.

[0007] The MCLDNN method shown in formula (1) can extract time and space features of single and combined IQ symbols through one-dimensional convolution, two-dimensional convolution and LSTM, and has good convergence speed and average recognition accuracy. Figure 1

[0008] However, the existing MCLDNN method has the following problems:

[0009] (1) The model input is only IQ information, which leads to low recognition accuracy in a low SNR scenario;

[0010] (2) Directly splicing two features may dilute key features with redundant information.

[0011] The convolution kernel only covers the local spatial features of I / Q two channels, but does not consider more extensive cross-channel interaction; the kernel height of the subsequent convolution is 1, which cannot utilize the spatial information after multi-branch splicing, and the feature extraction capability is limited;

[0012] (3) Regularization is missing, which is prone to overfitting. The single-stage training strategy has a simple structure and is not conducive to extracting optimal parameters.

[0013] ​In the prior art, a life detection method and system based on radar signal processing are disclosed in Chinese Patent Publication No. CN 119148096 B, which relates to the technical field of radar signal processing and includes collecting radar echo signals of a target area, controlling a phased array radar transmitting antenna array to emit an orthogonal frequency division multiplexing double-frequency continuous wave radar signal, inputting the processed radar echo signals into an orthogonal demodulation module for synchronous coherent demodulation to obtain a baseband signal, inputting a feature vector into a multi-task deep learning network integrated with an attention mechanism for feature fusion and target identification to obtain a second processing signal, obtaining a target feature spectrum by calculating the time-frequency distribution of the second processing signal, determining whether a life exists in the target area according to the time-varying characteristics and amplitude fluctuation rules of the respiratory frequency component and the heartbeat frequency component, and if it is determined that a life exists in the target area, calculating the three-dimensional spatial coordinates of the life through a multi-base station cooperative positioning algorithm and outputting a detection result containing the three-dimensional spatial coordinates and life characteristic information. However, the above-mentioned comparative document does not improve the feature fusion and convolution processing method, and does not adjust the feature weight through SNR explicit modeling. SUMMARY

[0014] To solve the technical problems existing in the prior art, the purpose of the present application is to provide a signal modulation recognition method with adaptive signal-to-noise ratio, which can adjust the feature weight and improve the recognition accuracy of low SNR modulation mode.

[0015] To achieve the above-mentioned purpose of the application, the present application provides a signal modulation recognition method with adaptive signal-to-noise ratio, the steps of which are as follows:

[0016] Inputting the IQ information of the signal and the signal-to-noise ratio (SNR) estimation value into a signal modulation recognition model to recognize the modulation mode of the signal;

[0017] The signal modulation recognition model is an MCLDNN provided with an SNR attention module in a feature fusion module;

[0018] The SNR attention module can receive the SNR estimation value and obtain a signal-to-noise ratio attention weight according to the SNR estimation value, receive a signal multi-channel fusion feature map output by a multi-channel fusion convolution layer of the MCLDNN, weight the signal multi-channel fusion feature map using the signal-to-noise ratio attention weight and connect it with the signal multi-channel fusion feature map in residual, output a multi-channel fusion feature combined with SNR attention, and input the multi-channel fusion feature combined with SNR attention into a time sequence module of the MCLDNN.

[0019] According to one aspect of the present application, the SNR attention module includes an SNR encoding module and an attention feature fusion module;

[0020] The SNR encoding module comprises a first fully connected layer, a first Swish activation layer, a second fully connected layer, a second Swish activation layer and a Sigmoid gate layer connected in sequence;

[0021] The SNR encoding module is capable of receiving the SNR estimation value and outputting the signal-to-noise ratio attention weight;

[0022] The attention feature fusion module comprises a channel-by-channel multiplication module and a residual connection module;

[0023] The channel-by-channel multiplication module is capable of receiving the signal-to-noise ratio attention weight; and multiplying the signal multi-channel fusion and the signal-to-noise ratio attention weight channel by channel, so as to weight the signal multi-channel fusion feature map by the signal-to-noise ratio attention weight, and obtain a weighted signal multi-channel fusion feature map;

[0024] The residual connection module is capable of receiving the weighted signal multi-channel fusion feature map and the signal multi-channel fusion feature map, and adding them together.

[0025] According to one aspect of the present application, the SNR attention module further comprises a fusion feature extraction module;

[0026] The fusion feature extraction module comprises a first Conv2D, a third Swish activation layer and a second Conv2D connected in sequence; the convolution kernel of the first Conv2D and the second Conv2D is 3x3;

[0027] The fusion feature extraction module is capable of receiving the signal multi-channel fusion feature map and performing convolution to obtain a convolved signal multi-channel fusion feature map.

[0028] According to one aspect of the present application, the feature extraction module of the signal modulation recognition model comprises a first feature extraction branch and a second feature extraction branch;

[0029] The first feature extraction branch comprises a third Conv2D and a fourth Swish activation layer connected in sequence; the third Conv2D uses a 2x8 convolution kernel and has a channel number of 64;

[0030] The first feature extraction branch is capable of performing feature extraction on the I / Q multi-channel components in the IQ information to obtain a first branch feature;

[0031] The second feature extraction branch comprises a reshaping layer, a first splicing layer, a fourth Conv2D and a fifth Swish activation layer connected in sequence; the fourth Conv2D uses a 3x8 convolution kernel and has a channel number of 128;

[0032] The second feature extraction branch can firstly reshape the I channel component and the Q channel component in the IQ information respectively, then splice to obtain a composite signal, and then perform feature extraction on the load signal to obtain a second branch feature.

[0033] According to one aspect of the application, the feature fusion module of the signal modulation identification model further comprises a second splicing layer and a fifth Conv2D; the fifth Conv2D uses a 3*3 convolution kernel, and the number of channels is 256;

[0034] The second splicing layer can splice the first branch feature and the second branch feature on the channel latitude axis=-1 to obtain a preliminary fusion feature.

[0035] The fifth Conv2D is a multi-channel fusion convolution layer of the MCLDNN, which can perform feature extraction on the preliminary fusion feature to obtain a signal multi-channel fusion feature map.

[0036] According to one aspect of the application, the first Conv2D to the fifth Conv2D all apply L2 weight decay.

[0037] According to one aspect of the application, a first dropout layer is connected between the first Swish activation layer and the second fully connected layer; the dropout rate of the first dropout layer is 0.2.

[0038] A second dropout layer is connected between the fourth Swish activation layer and the second splicing layer; the dropout rate of the second dropout layer is 0.3.

[0039] A third dropout layer is connected between the fifth Swish activation layer and the second splicing layer; the dropout rate of the third dropout layer is 0.3.

[0040] In the classification module of the signal modulation identification model, a fourth dropout layer is connected after the hidden layer, and the dropout rate of the fourth dropout layer is 0.5.

[0041] According to one aspect of the application, a first batch normalization layer is connected between the first Conv2D and the third Swish activation layer.

[0042] A second batch normalization layer is connected after the second Conv2D.

[0043] A third batch normalization layer is connected between the third Conv2D and the fourth Swish activation layer.

[0044] A fourth batch normalization layer is connected between the fourth Conv2D and the fifth Swish activation layer.

[0045] According to one aspect of the present application, the process of training the signal modulation recognition model comprises:

[0046] The first stage: training the SNR attention module, freezing other layers; until the training termination condition of the SNR attention module is met; in the process of training the SNR attention module, the learning rate is 1x10 -3 ;

[0047] The second stage: training the whole model of the signal modulation recognition model by using the cosine annealing learning rate strategy; until the training termination condition of the signal modulation recognition model is met; the initial learning rate of the cosine annealing learning rate strategy is 1x10 -4 .

[0048] The present application also provides an electronic device comprising one or more processors, one or more memories, and one or more computer programs; wherein the processor is connected with the memory, and the one or more computer programs are stored in the memory; when the electronic device is running, the processor executes the one or more computer programs stored in the memory, so that the electronic device executes the signal modulation recognition method with adaptive SNR.

[0049] The present application provides a signal modulation recognition method with adaptive SNR, which has the following beneficial effects compared with the prior art:

[0050] (1) The feature weight is adjusted by SNR explicit modeling, and the modulation mode recognition accuracy in low SNR is improved;

[0051] (2) The feature fusion and convolution processing method is improved, and the average modulation recognition rate is improved;

[0052] (3) The overfitting rate is reduced by designing a hybrid regularization architecture, and the recognition accuracy is improved by using a two-stage progressive training method, pre-training by SNR, and fine-tuning of the whole network. BRIEF DESCRIPTION OF DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0054] Figure 1 The schematic diagram of the principle of the existing MCLDNN method used in the prior art signal modulation recognition method according to the present application is shown;

[0055] Figure 2A structural diagram of a signal modulation identification model in a signal modulation identification method with SNR adaptivity according to an embodiment of the present application is schematically shown;

[0056] Figure 3 A structural diagram of an SNR attention module in a signal modulation identification method with SNR adaptivity according to an embodiment of the present application is schematically shown. DETAILED DESCRIPTION

[0057] The description of the embodiments of the present application should be combined with the corresponding drawings, which should be regarded as a part of the complete description. In the drawings, the shape or thickness of the embodiments can be exaggerated and simplified or facilitated for illustration. Moreover, parts of the structures in the drawings will be described separately, and it should be noted that the elements not shown or not described by words in the drawings are in the form known by those skilled in the art.

[0058] The description of the embodiments herein, any reference to direction and orientation, is only for the convenience of description and cannot be understood as any limitation on the scope of protection of the present application. The following description of the preferred embodiments will involve combinations of features, which can exist independently or in combination, and the present application is not particularly limited to the preferred embodiments. The scope of the present application is defined by the claims.

[0059] As shown in Figures 1 to 3 A signal modulation identification method with SNR adaptivity of the present application is as follows:

[0060] The IQ information of the signal and the SNR estimation value are input into the signal modulation identification model to identify the modulation mode of the signal;

[0061] The signal modulation identification model is an MCLDNN provided with an SNR attention module in the feature fusion module;

[0062] The SNR attention module can receive the SNR estimation value and obtain the SNR attention weight according to the SNR estimation value; and receive the signal multi-channel fusion feature map output by the multi-channel fusion convolution layer of the MCLDNN; then weight the signal multi-channel fusion feature map using the SNR attention weight and connect the signal multi-channel fusion feature map with the residual error to output the multi-channel fusion feature combined with the SNR attention; and input the multi-channel fusion feature combined with the SNR attention into the timing module of the MCLDNN.

[0063] In the present embodiment, the present embodiment can run on a general-purpose heterogeneous computing platform, as shown in Figure 2 The model includes the following core modules:

[0064] 1. An input module for acquiring multiple features;

[0065] 2. a feature extraction module for feature extraction with double-branch parallelism;

[0066] 3. a feature fusion module for feature fusion and SNR adaptive enhancement;

[0067] 4. a temporal module for spatiotemporal feature modeling and classification;

[0068] 5. a classification module for classification to identify the modulation mode of the signal.

[0069] The SNR attention module in the feature fusion module (EnhancedSNRBlock) enables the MCLDNN network to dynamically adjust feature weights in combination with SNR.

[0070] The SNR attention module generates SNR attention weights matching the number of channels to weight the multi-channel fusion feature map output by the multi-channel fusion convolutional layer (fifth Conv2D). It can dynamically adjust the importance of features according to the signal-to-noise ratio (enhance useful features and suppress noise in low signal-to-noise ratio environments) to improve the robustness of the model in different signal-to-noise ratio environments. It realizes feature enhancement for environmental perception and improves accuracy in low signal-to-noise ratio scenarios.

[0071] In the SNR adaptive signal modulation recognition method, the SNR attention module includes an SNR encoding module and an attention feature fusion module.

[0072] The SNR encoding module includes a first fully connected layer, a first Swish activation layer, a second fully connected layer, a second Swish activation layer, and a Sigmoid gating layer connected in sequence.

[0073] The SNR encoding module can receive the SNR estimate and output the signal-to-noise ratio attention weight.

[0074] The attention feature fusion module includes a channel-wise multiplication module and a residual connection module.

[0075] The channel-wise multiplication module can receive the signal-to-noise ratio attention weight; and multiply the signal multi-channel fusion and the signal-to-noise ratio attention weight channel by channel, so as to weight the signal multi-channel fusion feature map using the signal-to-noise ratio attention weight, and obtain a weighted signal multi-channel fusion feature map.

[0076] The residual connection module can receive the weighted signal multi-channel fusion feature map and the signal multi-channel fusion feature map, and add them together.

[0077] In the SNR adaptive signal modulation recognition method, the SNR attention module also includes a fusion feature extraction module.

[0078] The fusion feature extraction module comprises a first Conv2D, a third Swish activation layer and a second Conv2D connected in sequence; the convolution kernel of the first Conv2D and the second Conv2D is 3*3;

[0079] The fusion feature extraction module can receive the signal multi-channel fusion feature map and perform convolution to obtain a convolutional signal multi-channel fusion feature map.

[0080] In the embodiment, the specific process is divided into four stages:

[0081] Stage 1: SNR feature encoding (environmental perception modeling)

[0082] 1. Input preprocessing:

[0083] Receive a scalar SNR estimate input (shape: [batch_size, 1]) representing the current signal quality.

[0084] 2. Nonlinear expansion:

[0085] Expand the dimension to `filters×expansion×2` (preferably expand to 256×0.5×2=256 dimensions) through the first fully connected layer, and use the Swish activation function (first Swish activation layer):

[0086] h1=Swish(W1·SNR+b1)

[0087] Establish a high-dimensional nonlinear mapping to capture the complex relationship between SNR and feature adjustment.

[0088] 3. Random regularization:

[0089] Apply 20% Dropout through the first Dropout layer to force the network to learn robust feature correlations:

[0090] h1'=Dropout(h1,rate=0.2)

[0091] 4. Feature compression:

[0092] Compress to `filters×expansion` dimensions (128 dimensions) through the second fully connected layer, and Swish activate again (second Swish activation layer):

[0093] h2=Swish(W2·h1'+b2)

[0094] 5. Attention generation:

[0095] The final third fully connected layer outputs a signal-to-noise ratio attention weight vector α with the same number of channels as the convolutional channels, which is constrained to (0, 1) by a Sigmoid gating layer using the Sigmoid function:

[0096] α = σ (W3·h2 + b3)

[0097] α = σ (W3·Swish (W2·Dropout (Swish (W1·SNR))))

[0098] where σ is the Sigmoid function that generates channel-level attention coefficients.

[0099] Stage 2: Dual-path convolutional feature extraction

[0100] 1. Main convolutional path:

[0101] The input signal multi-channel fusion feature map x ∈ R H×W×C After two 3 × 3 convolutional units (first Conv2D and second Conv2D):

[0102] y = BN (Conv2D (BN (Conv2D (x)))

[0103] Each convolutional layer is followed by batch normalization (first batch normalization layer and second batch normalization layer) and Swish activation (third Swish activation layer):

[0104] Swish (x) = x·sigmoid (βx)

[0105] where β is a learnable parameter that enhances the non-linear representation capability.

[0106] 2. Regularization mechanism:

[0107] Apply L2 weight decay (coefficient 1 × 10 -4 ) to each 3 × 3 convolutional kernel:

[0108] L reg = λ∑w i 2

[0109] Prevent overfitting through dual regularization constraints.

[0110] Stage 3: Environment adaptive feature modulation

[0111] 1. Attention reshaping:

[0112] Reshape the weight vector α output by the SNR encoder into a 4D tensor:

[0113] α' = reshape (α, [-1, 1, 1, C])

[0114] Realize the dimension alignment with the convolution feature map y (broadcast mechanism)

[0115] 2. Gating enhancement:

[0116] Perform channel-level feature scaling:

[0117]

[0118] wherein represents a channel-by-channel multiplication, strengthens important features, and suppresses noise-sensitive channels.

[0119] 3. Residual connection:

[0120] Preserve the original feature information and form an identity mapping branch:

[0121] y_final = y_att + y

[0122] This structure ensures effective gradient propagation and alleviates network degradation problems.

[0123] Map the SNR scalar to the channel attention vector through the fully connected network, and realize dynamic feature weighting;

[0124] Use Swish activation function and Dropout regularization to improve the robustness of the model to SNR changes;

[0125] The Sigmoid gating mechanism adjusts the intensity of the convolution feature response, and the feature enhancement effect in the low SNR scene is significant.

[0126] Multi-feature fusion and SNR modulation cooperate to improve the feature expression ability of complex signals.

[0127] In the signal modulation recognition method adaptive to signal-to-noise ratio, the feature extraction module of the signal modulation recognition model includes a first feature extraction branch and a second feature extraction branch.

[0128] The first feature extraction branch includes a third Conv2D and a fourth Swish activation layer connected in sequence; the third Conv2D uses a 2×8 convolution kernel and has 64 channels;

[0129] The first feature extraction branch can extract features from the I / Q multi-channel components in the signal IQ information to obtain first branch features.

[0130] The second feature extraction branch includes a reshaping layer, a first splicing layer, a fourth Conv2D and a fifth Swish activation layer connected in sequence; the fourth Conv2D uses a 3×8 convolution kernel and has 128 channels.

[0131] The second feature extraction branch can firstly reshape the I channel component and the Q channel component in the IQ information respectively, then splice to obtain a composite signal, and then perform feature extraction on the load signal to obtain a second branch feature.

[0132] In the embodiment, the time-frequency feature and the quadrature component double-branch are extracted in parallel, and 2*8 and 3*8 convolution kernels are combined to capture multi-scale information.

[0133] In the signal modulation recognition method adaptive to signal-to-noise ratio, the feature fusion module of the signal modulation recognition model further includes a second splicing layer and a fifth Conv2D; the fifth Conv2D uses a 3*3 convolution kernel, and the number of channels is 256;

[0134] The second splicing layer can splice the first branch feature and the second branch feature on the channel latitude axis=-1 to obtain a preliminary fusion feature.

[0135] The fifth Conv2D is a multi-channel fusion convolution layer of the MCLDNN, and can perform feature extraction on the preliminary fusion feature to obtain a signal multi-channel fusion feature map.

[0136] In the embodiment, the channel dimension splicing is used instead of spatial splicing in feature fusion, the first feature extraction branch (64 channels) and the second feature extraction branch (128 channels) are spliced on axis=-1 (channel dimension) to obtain a fusion feature of [None, 2, 128, 192], and the complementary information of the multi-branch is retained.

[0137] In the embodiment and the previous embodiment, the 2*8 and (1, 8) kernel are replaced by 2*8, 3*8 and 3*3 kernel in the selection of the convolution kernel. The 3*8 kernel expands the receptive field to 3 rows through the height direction padding (padding='same') to cover all the spatial information of the two input branches; the 3*3 kernel performs spatial refinement on the fusion feature, the small size characteristic of which significantly reduces the parameter amount and enhances the feature expression ability, and in combination with the SNR adaptive module, the recognition rate in the low signal-to-noise ratio scene is improved.

[0138] In the signal modulation recognition method adaptive to signal-to-noise ratio, L2 weight attenuation is applied to the first Conv2D to the fifth Conv2D.

[0139] In the embodiment, L2 regularization (regularization constraint parameter=L2 (1*10 -4 )) is added to all convolution layers to constrain the model complexity and inhibit overfitting.

[0140] In the signal modulation recognition method adaptive to signal-to-noise ratio, a first dropout layer is connected between the first Swish activation layer and the second fully connected layer; the dropout rate of the first dropout layer is 0.2.

[0141] The fourth Swish activation layer is connected with the second splicing layer through a second dropout layer, and a dropout rate of the second dropout layer is 0.3.

[0142] The fifth Swish activation layer is connected with the second splicing layer through a third dropout layer, and a dropout rate of the third dropout layer is 0.3.

[0143] In the classification module of the signal modulation recognition model, a fourth dropout layer is connected after a hidden layer, and a dropout rate of the fourth dropout layer is 0.5.

[0144] In the embodiment, Dropout (0.3 and 0.5) is used in the convolution layer and the full connection layer to reduce the risk of overfitting. The dropout rate of the channel and the classification head is differentiated to suppress overfitting.

[0145] For example, Dropout (0.3) is added after Conv2D+batch normalization+activation in the feature extraction module to suppress overfitting of the convolution layer (the first feature extraction branch and the second feature extraction branch).

[0146] Dropout (0.5) is used after the hidden layer (Dense (512)) of the classification head, because the full connection layer has more parameters and needs stronger regularization.

[0147] Dropout (0.2) is used in the SNR encoding module in the SNR attention module to prevent overfitting of the attention mechanism.

[0148] In the signal modulation recognition method with adaptive SNR, a first batch normalization layer is connected between a first Conv2D and a third Swish activation layer.

[0149] A second batch normalization layer is connected after a second Conv2D.

[0150] A third batch normalization layer is connected between a third Conv2D and a fourth Swish activation layer.

[0151] A fourth batch normalization layer is connected between a fourth Conv2D and a fifth Swish activation layer.

[0152] In the embodiment, batch normalization is added after convolution to enable stable training of the signal modulation recognition model and accelerate convergence.

[0153] In the signal modulation recognition method with adaptive SNR, the process of training the signal modulation recognition model includes:

[0154] The first stage: training the SNR attention module, freezing other layers; until the training termination condition of the SNR attention module is met; during the training of the SNR attention module, the learning rate is 1x10 -3 ;

[0155] The second stage: training the full model of the signal modulation recognition model by using a cosine annealing learning rate strategy; until the training termination condition of the signal modulation recognition model is met; the initial learning rate of the cosine annealing learning rate strategy is 1x10 -4 .

[0156] In the embodiment, the training strategy is to pre-train the SNR attention module and then fine-tune the full model, combined with the cosine annealing learning rate, to accelerate the convergence and improve the recognition rate.

[0157] Stage 1: freeze the non-SNR module and focus on training the SNR encoder;

[0158] Stage 2: unfreeze the full model and fine-tune by using the cosine annealing strategy.

[0159] The fitting rate can be reduced, and the recognition rate of two-stage training is better than that of single-stage training above-8dB.

[0160] The RadioML2016.10a(-20dB to +18dB) test data set is used in the above training process.

[0161] The complete process is as follows:

[0162] The method of the application can run on a general-purpose heterogeneous computing platform, such as Figure 2 As shown, the model includes the following core modules:

[0163] (1) Multi-feature acquisition:

[0164] The multi-feature input layer includes 4 inputs, wherein:

[0165] Input 1: I / Q multi-channel components after complex signal decomposition (2 channels x 128 sampling points x 1 channel);

[0166] Input 2 / 3: I channel component and Q channel component of the quadrature signal (128 sampling points x 1 channel);

[0167] SNR input: scalar signal-to-noise ratio value.

[0168] (2) Dual-branch parallel feature extraction:

[0169] The first feature extraction branch performs time-frequency feature processing:

[0170] The 64-dimensional spatial feature is extracted through a 2*8 convolution kernel, normalized by batch (BN) and activated by Swish, and 30% Dropout is applied.

[0171] The second feature extraction branch fuses the orthogonal signals:

[0172] The I / Q channel components are spliced along the channel dimension, and 128-dimensional features are extracted through a 3*8 convolution kernel, and the same structure is regularized.

[0173] (3) Feature fusion and SNR adaptive enhancement:

[0174] Multi-feature fusion: Splice the double-branch features along the channel dimension, and extract the features through a 3*3 convolution;

[0175] SNR modulation: Use the SNR attention module to combine the fused 256-channel features with the SNR to generate channel-level signal-to-noise ratio attention weights, and perform channel-by-channel weighting on the signal multi-channel fusion feature map.

[0176] (4) Spatio-temporal feature modeling and classification:

[0177] Time series conversion: The spatial features are reshaped into time series, and the time series dependence is captured through a double-layer LSTM network;

[0178] Classification decision: The full connection layer outputs the class probability distribution, and the Softmax activation is applied, and 50% Dropout is applied to improve the generalization ability.

[0179] The method of the application is trained and tested on the RadioML2016.10a dataset under the RTX 3060, Tensorflow-GPU 2.6.0 environment, and the MCLDNN model is reproduced and compared, which is a conventional method in the art to verify the theoretical performance of the algorithm.

[0180] The method is verified, and the signal recognition rate of -8dB is 53.41% when the SNR estimation value is an ideal value (the estimation error is 0), which is increased by 14.55% compared with the benchmark method, the signal recognition rate of -4dB is 78.45%, which is increased by 10.22% compared with the benchmark method, the signal recognition rate of 0dB is 90%, which is increased by 1.82% compared with the benchmark method, and the average value of the recognition rate of -20dB to 18dB is 66.05%, which is increased by 4% compared with the benchmark method, which proves that the recognition performance of the method has obvious advantages under the condition of low signal-to-noise ratio. When the SNR estimation value has an estimation error of ±1dB, the signal recognition rate under the condition of -8dB true value is 52.5%, the signal recognition rate under the condition of -4dB true value is 77.32%, the signal recognition rate under the condition of 0dB true value is 89.45%, and the average value of the recognition rate of -20dB to 18dB is 65.66%, which proves that the method has good effect under the condition that the SNR estimation has a certain error. The above data are performance reference values under the experimental conditions of the application, and in actual application, the parameters can be adapted and optimized in combination with factors such as hardware configuration. Based on the principle of the application, those skilled in the art can reasonably expect the performance under typical error conditions.

[0181] The method of the application is deployed in an electronic device comprising a hardware-accelerated computing unit that implements parallel computation acceleration of a neural network model.

[0182] The SNR estimation method of the application is not specifically limited, and any prior art that can provide a signal-to-noise ratio estimation value can be applied to the application, including but not limited to an estimation method based on cyclic stationary characteristics and an estimation method based on deep learning.

[0183] The hardware adaptability of the application is as follows:

[0184] 1) The hardware adaptability of the application includes but is not limited to:

[0185] Optional hardware carrier 1 (embedded AI platform)

[0186] After optimization (such as model compression, architecture adjustment, and hardware acceleration configuration), it can be deployed on an embedded device with a dedicated AI computing unit. The device needs to meet the following conditions:

[0187] Support neural network quantization inference (such as INT8 / FP16 precision);

[0188] The computing power is sufficient to run the model of the application (a typical value is greater than or equal to 0.5TOPS INT8).

[0189] 2) Optional hardware carrier 2 (general-purpose computing device)

[0190] It can be deployed on a general-purpose computing device comprising a graphics processing unit (GPU). The GPU needs to meet the following conditions:

[0191] Supporting CUDA or similar programming model (such as OpenCL);

[0192] Memory capacity sufficient to store model parameters and computation cache (typical value >= 6GB);

[0193] Equipped with matrix operation acceleration unit (such as NVIDIA Tensor Core or equivalent technology).

[0194] According to an aspect of the present application, an electronic device is provided, comprising: one or more processors, one or more memories, and one or more computer programs; wherein the processor is connected with the memory, and the one or more computer programs are stored in the memory, and when the electronic device is running, the processor executes the one or more computer programs stored in the memory, so that the electronic device executes the signal modulation recognition method with adaptive signal-to-noise ratio as in the above technical solution.

[0195] According to an aspect of the present application, a computer readable storage medium is provided for storing computer instructions, which are executed by a processor to implement the signal modulation recognition method with adaptive signal-to-noise ratio as in the above technical solution.

[0196] The computer readable storage medium can include any medium capable of storing or transmitting information. Examples of the computer readable storage medium include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, optical fiber media, radio frequency (RF) links, etc. Code segments can be downloaded via a computer network such as the Internet, an intranet, etc.

[0197] The signal modulation recognition method with adaptive signal-to-noise ratio of the present application has the following steps: inputting the IQ information of the signal and the signal-to-noise ratio (SNR) estimation value into the signal modulation recognition model to identify the modulation mode of the signal; the signal modulation recognition model is an MCLDNN provided with an SNR attention module in the feature fusion module; the SNR attention module can receive the SNR estimation value and obtain the signal-to-noise ratio attention weight according to the SNR estimation value; and receive the signal multi-channel fusion feature map output by the multi-channel fusion convolution layer of the MCLDNN; then use the signal-to-noise ratio attention weight to weight the signal multi-channel fusion feature map and connect it with the signal multi-channel fusion feature map residual error, output the multi-channel fusion feature combined with the SNR attention; and input the multi-channel fusion feature combined with the SNR attention into the timing module of the MCLDNN.

[0198] Moreover, it should be noted that the present application can be provided as a method, an apparatus, or a computer program product. Therefore, the present application embodiments can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application embodiments can take the form of a computer program product on one or more computer-usable storage media (including disks, diskettes, tapes, optical, silicon, solid substrate, etc.) embodying computer-readable instructions.

[0199] The present application embodiments are described with reference to the flowchart illustrations and / or block diagrams of the methods, terminal devices (systems) and computer program products according to the present application embodiments. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing terminal devices to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal devices, create means for implementing the functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams.

[0200] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing terminal devices to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions means which implement the function specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams.

[0201] It should also be noted that the terms "comprising", "including", or any other variant thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article or terminal device that comprises a list of elements does not include only those elements in the list, but can also include other elements not expressly listed or inherent to such process, method, article or terminal device. Without further limitation, an element defined by the statement "comprising a means" does not exclude the presence of additional identical elements in the process, method, article or terminal device including the stated element.

[0202] Finally, it should be noted that the above description is of preferred embodiments of the application, and that although preferred embodiments of the application have been described, modifications and adaptations taken within the spirit of the application as described are to be considered as within the scope of the application. Therefore, the following claims are intended to cover all such modifications and adaptations that come within the scope of the embodiments of the application.

Claims

1. A signal modulation recognition method with adaptive signal-to-noise ratio, characterized in that: Here are the steps: Inputting the IQ information and the signal-to-noise ratio (SNR) estimation value of the signal into the signal modulation recognition model to identify the modulation mode of the signal; The signal modulation recognition model is a MCLDNN with an SNR attention module in the feature fusion module; The SNR attention module can receive the SNR estimation value and obtain the signal-to-noise ratio attention weight according to the SNR estimation value; and receive the signal multi-channel fusion feature map output by the multi-channel fusion convolution layer of the MCLDNN; then use the signal-to-noise ratio attention weight to weight the signal multi-channel fusion feature map and residually connect it with the signal multi-channel fusion feature map to output the multi-channel fusion feature combined with SNR attention; and input the multi-channel fusion feature combined with SNR attention into the timing module of the MCLDNN.

2. The signal modulation recognition method with adaptive signal-to-noise ratio according to claim 1, characterized in that: The SNR attention module includes an SNR encoding module and an attention feature fusion module; The SNR encoding module includes a first fully connected layer, a first Swish activation layer, a second fully connected layer, a second Swish activation layer and a Sigmoid gating layer connected in sequence; The SNR encoding module is capable of receiving the SNR estimation value and outputting the signal-to-noise ratio attention weight; The attention feature fusion module includes a channel-by-channel multiplication module and a residual connection module; The channel-by-channel multiplication module is capable of receiving the signal-to-noise ratio attention weight; and multiplying the signal multi-channel fusion and the signal-to-noise ratio attention weight channel by channel, thereby weighting the signal multi-channel fusion feature map using the signal-to-noise ratio attention weight to obtain a weighted signal multi-channel fusion feature map; The residual connection module can receive the weighted signal multi-channel fusion feature map and the signal multi-channel fusion feature map, and add the two.

3. The signal modulation recognition method with adaptive signal-to-noise ratio according to claim 2, characterized in that: The SNR attention module also includes a fusion feature extraction module; The fusion feature extraction module includes a first Conv2D, a third Swish activation layer, and a second Conv2D connected in sequence; the convolution kernels of the first Conv2D and the second Conv2D are both 3×3; The fusion feature extraction module can receive the signal multi-channel fusion feature map and perform convolution to obtain the convolved signal multi-channel fusion feature map.

4. The signal modulation recognition method with adaptive signal-to-noise ratio according to claim 2 or 3, characterized in that: The feature extraction module of the signal modulation recognition model includes a first feature extraction branch and a second feature extraction branch; The first feature extraction branch includes a third Conv2D and a fourth Swish activation layer connected sequentially; the third Conv2D uses a 2×8 convolution kernel and has 64 channels; The first feature extraction branch can extract features of the I / Q multi-channel components in the IQ information to obtain first branch features; The second feature extraction branch includes a sequentially connected reshaping layer, a first concatenation layer, a fourth Conv2D layer, and a fifth Swish activation layer; the fourth Conv2D layer uses a 3×8 convolution kernel and has 128 channels; The second feature extraction branch can first reshape the I channel component and the Q channel component in the IQ information respectively and then splice them to obtain a composite signal, and then perform feature extraction on the load signal to obtain the second branch feature.

5. The signal modulation recognition method with adaptive signal-to-noise ratio according to claim 4, characterized in that: The feature fusion module of the signal modulation recognition model also includes a second splicing layer and a fifth Conv2D; the fifth Conv2D uses a 3×3 convolution kernel and has 256 channels; The second splicing layer can splice the first branch feature and the second branch feature on the channel latitude axis=-1 to obtain a preliminary fusion feature; The fifth Conv2D serves as a multi-channel fusion convolutional layer of MCLDNN, which can extract features from the preliminary fusion features to obtain a signal multi-channel fusion feature map.

6. The signal modulation recognition method with adaptive signal-to-noise ratio according to claim 5, characterized in that: The first Conv2D to the fifth Conv2D all apply L2 weight attenuation.

7. The signal modulation recognition method with adaptive signal-to-noise ratio according to claim 5 or 6, characterized in that: The first dropout layer is connected between the first Swish activation layer and the second fully connected layer; the dropout rate of the first dropout layer is 0.2; A second dropout layer is connected between the fourth Swish activation layer and the second splicing layer; the dropout rate of the second dropout layer is 0.3; A third dropout layer is connected between the fifth Swish activation layer and the second splicing layer; the dropout rate of the third dropout layer is 0.3; In the classification module of the signal modulation recognition model, a fourth dropout layer is connected after the hidden layer, and the dropout rate of the fourth dropout layer is 0.

5.

8. The signal modulation recognition method with adaptive signal-to-noise ratio according to claim 7, characterized in that: The first batch of normalization layers are connected between the first Conv2D and the third Swish activation layer; The second Conv2D is connected with the second batch of normalization layers; The third batch of normalization layers is connected between the third Conv2D and the fourth Swish activation layer; The fourth batch normalization layer is connected between the fourth Conv2D and the fifth Swish activation layer.

9. The signal modulation recognition method with adaptive signal-to-noise ratio according to claim 4, characterized in that: The process of training a signal modulation recognition model includes: Phase 1: Train the SNR attention module and freeze other layers until the training termination condition of the SNR attention module is met. During the training of the SNR attention module, the learning rate is 1×10 -3 ; The second stage: adopt the cosine annealing learning rate strategy to train the full model of the signal modulation recognition model until the training termination condition of the signal modulation recognition model is met; the initial learning rate of the cosine annealing learning rate strategy is 1×10 -4 .

10. An electronic device, characterized in that: include: One or more processors, one or more memories, and one or more computer programs; wherein the processor is connected to the memory, the one or more computer programs are stored in the memory, and when the electronic device is running, the processor executes the one or more computer programs stored in the memory to enable the electronic device to perform the signal-to-noise ratio adaptive signal modulation identification method according to any one of claims 1 to 9.

Images