A method for automatic modulation recognition of radio signals under restricted conditions

Abstract

The application discloses a method for automatically identifying radio signal modulation under restricted conditions, comprising the following steps: S1, collecting radio I / Q signals of different modulation types as original feature data sets, and converting each I / Q sample signal into an A / P signal through feature transformation, all A / P signals forming a transformed feature data set; wherein, the I / Q and A / P features of the same sample signal have the same label. The method focuses on two practical application scenarios of radio signal sample restriction and computing resource restriction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic modulation recognition of radio signals, and particularly relates to a method for automatic modulation recognition of radio signals under restricted conditions. BACKGROUND

[0002] Automatic modulation recognition is an important technology in the field of wireless communication, which can help the receiving end to quickly and accurately identify the modulation mode of the received signal. Different modulation modes correspond to different transmission performance and signal characteristics, so correctly identifying the modulation mode of the signal can provide important information such as data rate, bit error rate, power and other parameters, thereby helping to optimize the transmission and reception quality of the signal. Modulation recognition technology has a wide range of applications in military and civilian wireless communication systems, such as radar, mobile communication, satellite communication, radio monitoring and other fields.

[0003] At present, modulation recognition technology mainly includes traditional maximum likelihood classification method and feature-based recognition method. Among them, the maximum likelihood classification method requires a large number of assumptions about the received signal and channel conditions to obtain the probability density function, which gradually cannot adapt to the increasingly complex communication environment. And the recognition method based on traditional statistical features has high requirements for the selected features, and when the type of candidate modulation set changes, it is usually required to select new features, and different application scenarios also need to be adjusted and improved. Recently, some scholars have proposed using machine learning methods for modulation recognition. Because most machine learning methods, especially deep learning methods, have the ability to adaptively extract signal features, the recognition and classification performance has been significantly improved. At present, the automatic modulation recognition method based on deep learning has surpassed the traditional likelihood function method and the signal feature-based method, and has gradually developed into the mainstream direction in the field of automatic modulation. For the research of modulation recognition method, the existing methods almost prefer to be under the condition of supervision, and these algorithms have poor adaptability to unknown signals, high algorithm complexity, dependence on a large number of labeled samples and high algorithm support. Once the above conditions are not met, the performance of the modulation recognition algorithm will decrease sharply. SUMMARY

[0004] The technical problem to be solved by the present application is to overcome the defects of the prior art and provide a method for automatic modulation recognition of radio signals under restricted conditions.

[0005] In order to solve the above technical problems, the present application provides the following technical scheme:

[0006] The application discloses a radio signal automatic modulation recognition method under restricted conditions, mainly comprising the following steps: S1: collecting radio I / Q signals of different modulation types as original feature data sets, and converting each I / Q sample signal into an A / P signal through feature space transformation, so that all A / P signals form a transformed feature data set; wherein the I / Q and A / P features of the same sample signal have the same label.

[0007] S2: inputting the I / Q signal and the A / P signal in S1 as features, using an adaptive noise reduction module to reduce noise of the two sample signals, filtering noise and interference and other undesired signal components in the modulation signal, and finally obtaining radio modulation signals containing less noise.

[0008] S3: inputting the I / Q and A / P two-way noise reduction output signals in S2 into a pre-trained multi-information fusion network model for recognition, and outputting a predicted probability of a label; the multi-information fusion network model is obtained by training a labeled radio modulation signal data set, and the label vector represents the modulation type contained in the collected radio signal; the multi-information fusion network model comprises five parts, namely a standardization layer, a convolution feature extraction layer, a feature splicing layer, a feature fusion layer and an output layer.

[0009] In order to improve the influence of noise in the radio modulation signal on the recognition accuracy and the model complexity, an adaptive noise reduction module is introduced, so that the model can extract more feature information about the modulation mode, thereby improving the recognition accuracy and reducing the overall complexity of the model. The adaptive noise reduction module mainly comprises a convolution layer, an SE-Block and a soft threshold function; the convolution layer is used to obtain feature information of the radio modulation signal, mainly comprising a first convolution layer and a second convolution layer; in the absence of prior information, the SE-Block learns the noise threshold correlation coefficient in the modulation signal through parameter estimation, and the threshold is processed by the SE-Block in the noise reduction process, specifically including four steps of squeezing, excitation, scaling and soft threshold noise reduction, and the processing process is as follows

[0010]

[0011] The input X is represented as U through global average pooling, wherein U is a feature compression in the spatial dimension according to the input X, the two-dimensional space is compressed into a real number, and the weighting of the channel representation is obtained through a fully connected layer and a sigmoid activation:

[0012] S=F ex (U)=σ(ω(U))

[0013] Wherein, ω represents a fully connected layer, and sigma represents a sigmoid activation function. Finally, the noise reduction threshold is obtained by multiplying the channel representation U and the channel weighting S to obtain the noise reduction threshold τ of the adaptive noise reduction module.

[0014] τ = F scale (U,S) = U x S

[0015] Soft threshold is a denoising method that can reconstruct the original expected signal from noisy data, and the denoising process can be represented as:

[0016]

[0017] Where x is the input feature, y is the output feature, and τ is the positive threshold. The output after soft threshold denoising is the feature obtained in step S2.

[0018] For the I / Q and A / P two-way denoising outputs in S2, a standardization layer is used to first normalize the modulation signal, then a convolution layer is used to obtain the spatio-temporal information in the modulation signal, then a concatenation layer is used to combine the multi-dimensional information of the I / Q and A / P sample signals, and finally a feature fusion layer and an output layer are used to complete the multi-dimensional information aggregation and modulation type prediction result output; different training sample quantities are set and the recognition accuracy and various indicators of the model are calculated, and the model parameters are adjusted until the network performance is optimal.

[0019] In an embodiment of the present application, the multi-information fusion network model comprises a convolution layer, a feature concatenation layer, a feature fusion layer and an output layer: wherein the I / Q route convolution feature extraction unit comprises a batch normalization layer, a first convolution layer, a ReLU activation layer, a maximum pooling layer, a second convolution layer and a ReLU activation layer connected in sequence, and the A / P route convolution feature extraction unit comprises a batch normalization layer, a first convolution layer, a ReLU activation layer, a maximum pooling layer, a second convolution layer and a ReLU activation layer connected in sequence; the feature concatenation layer is concatenated according to the sampling point dimension; the feature fusion layer comprises a first LSTM layer, a tanh activation layer, a second LSTM layer and a tanh activation layer connected in sequence; and the output layer comprises a full connection layer and a Softmax activation layer.

[0020] The kernel size of the first convolution layer in the multi-information fusion network model is 5, the filter size is 128, and the step is 1; the kernel size of the maximum pooling layer is 2, and the step is 2; the kernel size of the second convolution layer is 3, the filter size is 128, and the step is 1; the output dimension of the first LSTM layer is 256, the output dimension of the second LSTM layer is 256; and the output dimension of the full connection layer is 11.

[0021] The wireless signal automatic modulation recognition method under the restricted condition, characterized in that: the training of the neural network model comprises:

[0022] Collect radio modulation signal data set and label the modulation type of the radio signal;

[0023] The labeled data set is sequentially divided into a training set and a test set, and the training set is sequentially input into the adaptive noise reduction module for noise filtering, and a weak noise modulated signal is output;

[0024] The output signal of the noise reduction module is sequentially input into the standardization layer, the convolution feature extraction layer, the feature splicing layer, the feature fusion layer and the output layer in the multi-information fusion network, and the prediction probability of the label vector is output;

[0025] The weights of the adaptive noise reduction module and the multi-information fusion network are adjusted using the test set until the deviation of the recognition accuracy of the training and the recognition accuracy of the test is within a certain range, and the adaptive noise reduction multi-information fusion network model parameters are saved after training.

[0026] After the adaptive noise reduction multi-information fusion network model is trained, high-precision recognition can be realized under small sample conditions, and the model is parameter-compressed using the progressive building technique, and the weight parameters in the convolution layer, the fully connected layer and the LSTM layer in the sparse network that have less influence on the result are gradually compressed until the set sparse target is reached.

[0027] During model pruning, the sparse degree calculation formula is as follows

[0028] s t =s f +(s i -s f )(1-(t-t0) / nΔt) 3

[0029] Where s t is the initial sparse value, s f represents the target sparse value, t0 refers to the step of starting training at the building frequency t, the fine-tuning process is divided into n steps to gradually increase the sparsity, t∈{t0,t0+Δt,...,t0+nΔt}; During pruning, binary mask variables with the same tensor size and shape as in the original model are added to the fully connected layer, the convolution layer and the long short-term memory network layer in the adaptive noise reduction model. In the fine-tuning process, the weight mask of each layer is sorted, and the corresponding weights in the smallest part are set to 0 until the sparse target is reached.

[0030] After the model is progressively pruned, the parameters in the network are sparsified, the integer algorithm inference is realized by quantizing the network parameter type, and the problem of balancing model accuracy and device delay is solved.

[0031] The formula for converting float32 floating point to int8 integer during quantization is as follows

[0032]

[0033] The formula for fixed-point conversion to floating-point is as follows

[0034] R = (Q - Z) * S

[0035] wherein R represents input floating-point data, Q represents fixed-point data after quantization, Z represents the numerical value of the starting point, and S represents the scaling factor, the size of which is

[0036]

[0037] The adaptive noise reduction multi-information fusion network uses an Adam optimizer and a cross-entropy loss function for gradient updating in the training process.

[0038] Compared with the prior art, the present application has the following beneficial effects:

[0039] 1: The method focuses on two practical application scenarios of limited radio signal sample conditions and limited computing resource conditions. The method includes an adaptive noise reduction multi-information fusion small sample method (AMIF) and a lightweight method for model pruning and quantization. First, the I / Q and A / P components of the radio received signal are denoised to filter noise interference in the received signal and retain more useful feature information in the signal. The feature information is normalized through a BN layer, and then deep feature information in the signal is extracted through a one-dimensional convolution layer and a maximum pooling layer.

[0040] 2: The present application uses a splicing method to combine the I / Q and A / P feature information extracted by convolution, and uses two LSTM layers to fuse different dimensional spatiotemporal feature information in the radio signal, thereby enhancing the network's learning ability to mine signal modulation mode features under small sample conditions.

[0041] 3: The present application realizes network parameter sparseness and model lightweighting while maintaining the original network accuracy by pruning model parameters, and accelerates the model inference process by converting float parameter types in the original network to int types through quantization. The present application achieves high recognition accuracy, less inference time, and reduces the excessive dependence on computing resources under small sample conditions and limited computing resource conditions. BRIEF DESCRIPTION OF DRAWINGS

[0042] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, illustrate the present application together with the embodiments thereof, and explain the present application, but do not limit the present application. In the drawings:

[0043] Figure 1 is a schematic diagram of the present application;

[0044] Figure 2 is a schematic diagram of the adaptive noise reduction module;

[0045] Figure 3 This is a structural diagram of a multi-information fusion network model.

[0046] Figure 4 This is a comparison chart of the overall recognition accuracy of the present invention under different training data;

[0047] Figure 5 A comparison chart of recognition accuracy using different comparison methods.

[0048] Figure 6 Parameter comparison during the model pruning quantization process. Detailed Implementation

[0049] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0050] Example 1

[0051] like Figures 1-6 As shown, the present invention provides an automatic modulation identification method for radio signals under constrained conditions. Step 1: Collect the in-phase orthogonal I / Q components of the radio modulation signal, and use feature space transformation to obtain the amplitude and phase (A / P) components of the modulation signal to construct an experimental dataset. The constructed dataset includes a training set and a test set, wherein the dataset contains the I / Q components and A / P components of the same radio signal.

[0052] Step 2: Construct an adaptive noise reduction and multi-information fusion small sample modulation recognition framework based on deep neural networks;

[0053] Step 3: In order to reduce the impact of noise in radio modulated signals on the recognition results and network model complexity, a method is designed to embed the soft threshold noise reduction method into the channel attention mechanism to achieve adaptive noise reduction of the modulated signal.

[0054] Step 4: In order to extract the spatiotemporal feature information of radio modulation signals more efficiently, a feature extraction network model based on one-dimensional convolution is designed.

[0055] Step 5: To improve the feature learning and feature representation capabilities of radio modulated signals, a multi-information fusion network model based on the splicing method is designed.

[0056] Step 6: To optimize the proposed adaptive noise reduction and multi-information fusion few-shot network model, an optimization algorithm with dynamic learning rate adjustment and Early Stopping is used based on the Adam optimizer. During training, the constructed radio modulation signal training set and test set are input into the designed adaptive noise reduction and multi-information fusion few-shot network model to train and test the network. Finally, the optimal model parameters after training are saved.

[0057] Step 7, for the adaptive noise reduction multi-information fusion network model after training, in order to compress the number of parameters in the network model and realize network lightweight, the progressive pruning algorithm is used to sparsely the network model with less influence on the original recognition accuracy of the model;

[0058] Step 8, after the progressive pruning of the model, the parameters in the network are sparsified, in order to reduce the model storage size and inference time, the integer algorithm inference is realized by quantizing the network parameter type;

[0059] Step 1, constructing a wireless radio modulation signal data set under restricted conditions includes the following processing:

[0060] (1) The collected radio modulation signals mainly include analog modulation types: WBFM, AM-SSB, AM-DSB, digital modulation types: BPSK, QPSK, 8PSK, 16QAM, 64QAM, BFSK, CPFSK, PAM4, a total of 11 modulation types. The in-phase and quadrature component I / Q form of each modulation type radio signal is represented as

[0061]

[0062] Where i k represents the kth sample point of the in-phase component I, q k represents the kth sample point of the quadrature component Q, and the maximum sample point number N in the present application is 128.

[0063] (2) The radio modulation signal I / Q component is converted into an amplitude-phase A / P component form, and the A / P form is represented as

[0064]

[0065] Where a k represents the kth sample point of the amplitude component A, represents the kth sample point of the phase component P. The feature transformation between the A / P component form and the I / Q component form of the radio modulation signal can be represented as

[0066]

[0067] (3) Under the condition of small sample, the sample number of radio modulation signal is extremely limited. In order to obtain more feature information under the condition of limited sample number, the present application uses the I / Q component and A / P component of the same modulation type radio signal to construct the sample data set used in the present application Where the data set D includes N samples of 11 modulation types, and the i-th sample signal in the data set is represented as represents the one-hot encoding vector of the modulation type label of the i-th sample signal.

[0068] The adaptive noise reduction and multi-information fusion small sample modulation recognition framework based on the deep neural network constructed in step 2 comprises an adaptive noise reduction module and a multi-information fusion network model. The adaptive noise reduction module mainly comprises three modules: an adaptive noise reduction module in which a soft threshold function is embedded into a channel attention mechanism SE-Block, a convolutional feature extraction module comprising a one-dimensional convolution and a maximum pooling, and a multi-information fusion module using two layers of LSTM units for feature fusion.

[0069] The adaptive noise reduction and multi-information fusion small sample modulation recognition based on the deep neural network constructed comprises:

[0070] Radio modulation signal I / Q samples are obtained, and A / P components of each radio modulation signal are obtained through feature transformation to constitute a data set required for network model training and testing, and then the training set and the test set are divided;

[0071] The training set is input to the adaptive noise reduction module for noise reduction to obtain modulation signal data containing less noise;

[0072] The modulation signal I / Q and A / P components containing less noise are input to the multi-information fusion network for feature extraction, and the prediction probability of each label is output through full connection and Softmax activation function, and the final classification result is output after threshold decision;

[0073] The test set is input to the adaptive noise reduction and multi-information fusion network model, the trained model is verified, and the network weight is adjusted until the deviation of the classification result on the test set and the classification result on the training set is within the allowable range, and the model training is completed;

[0074] The adaptive noise reduction module applicable to radio modulation signals constructed in step 3 comprises the following processing: convolutional feature mapping, threshold calculation and soft threshold noise reduction, and is composed of three parts, wherein

[0075] (1) The convolutional feature mapping layer in the adaptive noise reduction module is composed of a first normalization layer, a first convolutional layer with a kernel size of 3x3 and a step size of 1x1, a second normalization layer, and a second convolutional layer with a kernel size of 3x3 and a step size of 1x1. After the above convolutional feature mapping calculation, ReLU function activation is used.

[0076] (2) The threshold calculation layer in the adaptive noise reduction module comprises a global pooling layer, a full connection layer with an output dimension of 1 using ReLU function activation, a full connection layer with an output dimension of 1 using Sigmoid function activation, and a self-defined product calculation layer.

[0077] (3) The soft threshold denoising layer in the adaptive denoising module mainly includes a tensor subtraction calculation subtract layer, a tensor maximum value calculation maximum layer, and a tensor multiplication calculation multiply layer, which are used to realize the soft threshold denoising calculation process.

[0078] The multi-information fusion network model of the adaptive radio modulation signal constructed in step 4 includes the following processes: a data standardization layer, a convolution feature extraction layer, a feature splicing layer, a feature fusion layer, and an output layer, which are composed of five parts.

[0079] (1) The data standardization layer of the multi-information fusion network model is realized by using a batch normalization BN layer. The input data is first normalized and then standardized to obtain the output features, wherein the normalization process is as follows

[0080]

[0081] x i represents the input data, represents the data after normalization processing, is the mean and variance of the input data x i The standardization process is as follows

[0082]

[0083] y i is the output of the BN layer, and γ and β represent the expansion parameters and translation parameters preset by the neural network.

[0084] (2) The convolution feature extraction layer in the multi-information fusion network model is composed of two parts, wherein the I / Q path convolution feature extraction unit is composed of a first convolution layer with a kernel size of 5 and a step size of 1, a ReLU activation layer, a maximum pooling layer, a second convolution layer with a kernel size of 3 and a step size of 1, and a ReLU activation layer. The A / P path convolution feature extraction unit is composed of a first convolution layer with a kernel size of 5 and a step size of 1, a ReLU activation layer, a maximum pooling layer, a second convolution layer with a kernel size of 5 and a step size of 1, and a ReLU activation layer.

[0085] (3) The feature splicing layer in the multi-information fusion network model splices the I / Q and A / P components of the radio modulation signal after feature extraction according to the sampling point dimension direction, and outputs a matrix containing multi-dimensional feature information.

[0086] (4) The feature fusion layer in the multi-information fusion network model is composed of a first LSTM layer with an output dimension of 256, which is set to output all sequences and uses a tanh function for activation. The second LSTM layer has an output dimension of 256, uses a tanh function for activation, and is composed of a Dropout layer.​

[0087] (5) The output layer in the multi-information fusion network model consists of Dense and Softmax activation layers with an output dimension of 11.

[0088] Step 6: To optimize the proposed adaptive noise reduction multi-information fusion few-shot network model, a callback function is added to optimize the learning rate during training, based on the Adam optimizer. In the experiment, the training set loss is set to not decrease within 6 epochs, i.e., the original learning rate is reduced by a factor of 10. At the same time, EarlyStopping in the callback function is used to supervise the training process. The training process is stopped when the network model loss does not decrease within 11 epochs, and the current optimal model is saved.

[0089] Step 7: To reduce the excessive reliance on computational resources by the adaptive denoising few-shot method, while simultaneously reducing the number of parameters in the adaptive denoising model and accelerating inference, this invention performs progressive parameter pruning on the trained model parameters, thereby sparsening the parameters of convolutional layers, LSTM layers, and fully connected layers in the model. The progressive pruning process can be represented as follows:

[0090] s t =s f +(s i -s f (1-(t-t0) / nΔt) 3

[0091] Among them, s t It is the initial sparse value, s f Let t0 represent the target sparsity value, where t0 refers to the training step starting at pruning frequency t. The fine-tuning process consists of n steps, gradually increasing the sparsity, where t∈{t0,t0+Δt,...,t0+nΔt}. During pruning, binary mask variables with the same tensor size and shape as those in the original model are added to the adaptive denoising model. During fine-tuning, the weight masks of each layer are sorted, and the corresponding weights in the minimum parts are set to 0 until the sparsity target is achieved.

[0092] Step 8: To further optimize the parameter types of the adaptive noise reduction model, integer quantization is performed on the parameters in the progressively pruned network. This quantization enables integer algorithm inference, thus balancing network accuracy and device latency. The formula for converting float32 floating-point numbers to int8 integers during quantization is shown below.

[0093]

[0094] Where R represents the input floating-point data, Q represents the quantized fixed-point data, Z represents the starting point value, and S represents the scaling factor, the size of which is...

[0095]

[0096] The application is further illustrated with reference to the accompanying drawings.

[0097] The application comprises the following steps:

[0098] Step 1, collecting in-phase quadrature I / Q components of radio modulation signals, and obtaining A / P components of the modulation signals through feature transformation to construct an experimental data set;

[0099] Step 2, constructing an adaptive noise reduction multi-information fusion small sample modulation recognition framework based on a deep neural network, to realize adaptive noise reduction and efficient feature extraction of radio modulation signals.

[0100] Step 3, inputting the training set and test set of the experimental data set into the constructed small sample network model to train and test the model, and saving the optimal model and parameters after training is completed.

[0101] Step 4, for the small sample network model after training is completed, in order to compress the number of parameters in the network model, realize network lightweight, and reduce the model storage size and inference time, through a progressive pruning algorithm, the weight parameters with less influence in the sparse network model are pruned while the original recognition accuracy of the model is ensured, and integer algorithm inference is realized through quantization of network parameter types.

[0102] Step 1, the collected I / Q components and A / P components of 11 types of wireless single signals are used as data set features, wherein the data set includes a training set and a test set, and the training set and the test set are divided according to 7:3.

[0103] Step 2 includes the following processing:

[0104] (1) In order to reduce the influence of noise in radio modulation signals on the recognition result and the complexity of the network model, an adaptive noise reduction method embedding a soft threshold noise reduction method into a channel attention mechanism is designed;

[0105] (2) In order to more efficiently extract the space-time feature information of radio modulation signals, a feature extraction network model based on one-dimensional convolution is designed;

[0106] (3) In order to improve the feature learning and feature expression ability of radio modulation signals, an information fusion network model based on a splicing method is designed;

[0107] Step 3, inputting the training set and test set into the adaptive noise reduction multi-information fusion small sample network model to train and test the network, and finally saving the model and parameters after training is completed.

[0108] Step 4 is for the small sample modulation identification network model after training, because the network model is more complex, the requirement for computing resources is higher, and there is a certain redundancy in the model parameters, therefore, the application compresses the parameter quantity and the size of the model by using the progressive pruning method and the parameter quantization method, thereby speeding up the model inference and reducing the dependence on computing resources.

[0109] As shown in Figure 1 The whole structure block diagram of the automatic radio signal modulation identification under the limited condition designed by the application is shown. The proposed model first collects radio I / Q signals of different modulation types as original features, and converts each I / Q sample signal into an A / P signal through feature transformation, and synthesizes modulation identification data sets from all I / Q and A / P signal samples. Then, the I / Q signal and the A / P signal are input as features, and an adaptive noise reduction module is used to reduce the noise of the two sample signals, the noise reduction module filters the noise and interference and other unexpected signal components in the modulation signal, and finally obtains a radio modulation signal containing less noise. For the I / Q and A / P two-way noise reduction output signals, input into the multi-information fusion network model for identification, and the model finally outputs the prediction probability of the label; after the adaptive noise reduction multi-information fusion network model is trained, it can realize high-precision identification under the condition of small samples, and by using the progressive construction technology to compress the parameters of the model, and by quantizing the network parameter type to realize integer algorithm reasoning, the parameter quantity and the storage size of the model are compressed.

[0110] As shown in Figure 2 The structure diagram of the adaptive noise reduction module designed by the application is shown. As can be seen from Figure 2 The adaptive noise reduction module is composed of a feature mapping layer, a threshold calculation layer and a soft threshold noise reduction layer. Among them, the convolution feature mapping layer is composed of a first normalization layer, a first convolution layer with a kernel size of 3x3 and a step size of 1x1, a first ReLU activation layer; a second normalization layer, a second convolution layer with a kernel size of 3x3 and a step size of 1x1, a second ReLU activation layer. The threshold calculation layer is composed of a global pooling layer; a fully connected layer with an output dimension of 1, using ReLU function activation; a fully connected layer with an output dimension of 1, using Sigmoid function activation; in addition, it also includes a self-defined product calculation layer. The soft threshold noise reduction layer is mainly composed of a tensor subtraction calculation layer, a tensor maximum value calculation layer, and a tensor multiplication calculation layer, which are used to realize the soft threshold noise reduction calculation process.

[0111] As shown in Figure 3 The structure diagram of the multi-information fusion network model of the application is shown. As can be seen from Figure 3It can be seen that the small sample multi-information fusion network model is composed of five parts of data standardization layer, convolution feature extraction layer, feature splicing layer, feature fusion layer and output layer. Among them, the data standardization layer uses the batch standardization BN layer to realize, that is, to normalize and standardize the input data. The convolution feature extraction layer is composed of I / Q first convolution layer with kernel size of 5 and step size of 1, ReLU activation layer, maximum pooling layer, second convolution layer with kernel size of 3 and step size of 1, and ReLU activation layer. The A / P convolution feature extraction unit is composed of first convolution layer with kernel size of 5 and step size of 1, ReLU activation layer, maximum pooling layer, second convolution layer with kernel size of 5 and step size of 1, and ReLU activation layer. The feature splicing layer in the model splices the I / Q and A / P two-way outputs according to the sampling point dimension direction, and outputs a matrix containing multi-dimensional feature information. The feature fusion layer is composed of a first LSTM layer with an output dimension of 256, which sets all sequences to be output and uses a tanh function to activate; a second LSTM layer with an output dimension of 256, which uses a tanh function to activate and a Dropout layer. The output layer of the model is composed of a and a Dense layer with an output dimension of 11 and a Softmax activation layer.

[0112] As shown in Figure 4 The comparative experiment results of the adaptive noise reduction multi-information fusion small sample network model constructed by the application are shown. In order to evaluate the classification accuracy of the model under different training data, Figure 4 The classification accuracy of the high-precision small sample method SCLS based on the twin network under different training data is compared. In the experiment, 1%, 2%, 3%, 5%, 10%, 20%, 30%, 40%, 50%, 60% and 70% of the constructed data set are used as the training set, and the remaining sample signals are used as the test set. As shown in the figure, under the 1%~10%dataset data setting, the overall recognition accuracy of the SCLS model increases from 31.9% to 49.1%, while the recognition accuracy of the AMIF model of the application increases from 45.89% to 54.61%, which is higher than that of the SCLS model under the 5 training sample quantity settings, and the AMIF model is nearly 14% higher than the SCLS under the 1%dataset, and the overall recognition accuracy is improved by 4% under the 5%dataset. At the same time, with the continuous increase of the sample quantity, the overall recognition accuracy of the AMIF model is improved by 5.5%~9.3% compared with the SCLS method.

[0113] As shown in Figure 5The AMIF model is tested under a small sample data setting and compared with six methods, including a modulation recognition small sample SCLS method, a machine learning KNN method, and advanced model methods CNN2, LSTM, CLDNN, and PETCGDNN in modulation recognition. The recognition accuracy of the AMIF model is the highest under 20 signal-to-noise ratios (SNRs) at 1% dataset, which is about 60% higher than that of the CLDNN model and about 35% higher than that of the SCLS model at 6 dB. At 2% dataset, the recognition accuracy of the AMIF model is the highest among the six compared methods, which is about 58% higher than that of the KNN model at more than 0 dB. At 5% dataset, the recognition accuracy of the AMIF model is also higher than that of the other methods. At 10% dataset, the AMIF method still has the highest recognition accuracy among all the compared methods under different SNRs. In summary, the AMIF model realizes high recognition accuracy and high robustness because the different features of the signal samples are fused and the signal samples are denoised, thereby fusing more effective signal space feature information. The work of the application provides a possibility for the application of the modulation recognition method in a small sample actual scene.

[0114] As Figure 6 shown, the recognition accuracy of the AMIF model changes under different SNRs after progressive pruning. When the sparsity is 0 to 0.8, the recognition accuracy of the model does not change significantly between -20 dB and 18 dB. When the sparsity is 0.9, the recognition accuracy of the model decreases by about 2% between -2 dB and 18 dB. When the sparsity is 0.95, the non-zero parameters of the AMIF model are only 5% of the original model, but the recognition accuracy of the model decreases by about 13% at high SNR.

[0115] As shown in FIG. 1, the parameter indicators of the AMIF model change after pruning and quantization. When the sparsity is 0 to 0.8, the recognition accuracy of the model is maintained above 91%, and the recognition accuracy of the model is higher than that of the unpruned model when the sparsity is 0.5, 0.6, and 0.7. When the sparsity is 0.5, the maximum recognition accuracy and the average recognition accuracy of the model reach the highest values of 92.14% and 62.09%, respectively, and the size of the model is only 17.4% of the original model. When the sparsity is greater than or equal to 0.9, the recognition accuracy of the model decreases rapidly, and the size of the model decreases to 5% and 10% of the original model. This is because important weights in the model are pruned, resulting in a decrease in the recognition accuracy of the model. In summary, when the sparsity of the AMIF model is 0.8, the recognition accuracy, parameter quantity, and model size of the model can achieve a better balance. The work of the application also provides a possibility for the further deployment and application of the modulation recognition method in a resource-constrained edge computing scene.

[0116] Table 1 Comparison of parameters in AMF model pruning quantization process

[0117]

[0118] NNZ Parameters: Number of non-zero parameters

[0119] Finally, it should be noted that the above only for the preferred embodiments of the present application, and is not intended to limit the application, although in the foregoing detailed description of the application has been made, for those skilled in the art, it still can be modified, or part of the technical features of the equivalent replacement of the technical solutions described in the foregoing embodiments. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application, shall be included within the scope of the present application.

Claims

1. A method for automatic modulation and identification of radio signals under constrained conditions, mainly comprising the following steps: S1: Collect radio I / Q signals of different modulation types as the original feature dataset, and convert each I / Q sample signal into an A / P signal through feature transformation. All A / P signals form a transformed feature dataset; wherein, the I / Q and A / P features of the same sample signal have the same label; S2: Using the I / Q signal and A / P signal in S1 as feature inputs, the adaptive noise reduction module is used to denoise the two sample signals. The noise reduction module filters out noise and unwanted signal components in the modulated signal, and finally obtains a radio modulated signal with less noise. S3: For the I / Q and A / P noise-reduced output signals in S2, the input is fed into a pre-trained multi-information fusion network model for identification, and the predicted probability of the output label vector is generated. The multi-information fusion network model is trained using a labeled radio modulation signal dataset, and the label vector represents the modulation type contained in the collected radio signal. The multi-information fusion network model consists of five parts: a normalization layer, a convolutional feature extraction layer, a feature concatenation layer, a feature fusion layer, and an output layer. For the two noise reduction outputs of I / Q and A / P in S2, the normalization layer is used to normalize the modulation signal first, then the convolutional feature extraction layer is used to obtain the spatiotemporal information in the modulation signal, then the feature splicing layer is used to combine the multidimensional information of the I / Q and A / P sample signals, and finally the feature fusion layer and output layer are used to complete the aggregation of multidimensional information and output the modulation type prediction result. The convolutional feature extraction layer includes I / Q-path convolutional feature extraction units and A / P-path convolutional feature extraction units. The I / Q-path convolutional feature extraction unit comprises a batch normalization layer, a first convolutional layer, a ReLU activation layer, a max pooling layer, a second convolutional layer, and a ReLU activation layer, all connected in sequence. The A / P-path convolutional feature extraction unit comprises a batch normalization layer, a first convolutional layer, a ReLU activation layer, a max pooling layer, a second convolutional layer, and a ReLU activation layer, all connected in sequence. The feature concatenation layer concatenates features according to the sampling point dimension. The feature fusion layer comprises a first LSTM layer, a tanh activation layer, a second LSTM layer, and a tanh activation layer, all connected in sequence. The output layer comprises a fully connected layer and a Softmax activation layer.

2. The method for automatic modulation and identification of radio signals under constrained conditions according to claim 1, characterized in that: In step S2, the adaptive noise reduction module consists of convolutional layers, SE-Block, and a soft thresholding function. The convolutional layers are used to acquire feature information of the radio modulation signal, including a first convolutional layer and a second convolutional layer. Without prior information, SE-Block learns the noise threshold correlation coefficient in the modulation signal through parameter estimation. During the noise reduction process, the threshold is processed by SE-Block, specifically including four steps: compression, excitation, scaling, and soft thresholding. The input X is then processed by global average pooling to generate a representation U for the feature channels, where U is a feature compression of the input X according to its spatial dimension, compressing the two-dimensional space into a real number. The weighting of the channel representations is obtained through a fully connected layer and sigmoid activation. in, Indicates a fully connected layer. The sigmoid activation function is represented; finally, the noise reduction threshold is obtained by multiplying the channel representation U and the channel weighting S to obtain the noise reduction threshold of the adaptive noise reduction module. : Soft thresholding is a denoising method that can reconstruct the original desired signal from noisy data. Its denoising process can be represented as follows: in, These are input features. It is the output feature. It is a positive threshold; the output after noise reduction by soft thresholding is used as the feature obtained in step S2.

3. The method for automatic modulation and identification of radio signals under constrained conditions according to claim 1, characterized in that: The first convolutional layer has a kernel size of 5, a filter size of 128, and a stride of 1; the max pooling layer has a kernel size of 2 and a stride of 2; the second convolutional layer has a kernel size of 3, a filter size of 128, and a stride of 1; the first LSTM layer has an output dimension of 256; the second LSTM layer has an output dimension of 256; and the fully connected layer has an output dimension of 11.

4. The method for automatic modulation and identification of radio signals under constrained conditions according to claim 1, characterized in that: The training of the multi-information fusion network model includes: Collect a dataset of radio modulation signals and label the modulation types of the radio signals; The labeled dataset is divided into training set and test set in sequence. The training set is then input into the adaptive noise reduction module for noise filtering, and the output is a signal with weak noise modulation. The output signal of the noise reduction module is sequentially input into the normalization layer, convolutional feature extraction layer, feature concatenation layer, feature fusion layer and output layer in the multi-information fusion network model, and the predicted probability of the label vector is output. The weights of the adaptive denoising module and the multi-information fusion network are adjusted using the test set until the deviation between the recognition accuracy of the training and the recognition accuracy of the test is within a predetermined range. After training, the parameters of the adaptive denoising multi-information fusion network model are saved.

5. The method for automatic modulation and identification of radio signals under constrained conditions according to claim 4, characterized in that: After the adaptive noise reduction multi-information fusion network model is trained, it can achieve high-precision recognition under small sample conditions. The model is then compressed by using progressive pruning technology, and the weight parameters in the convolutional layer, fully connected layer and LSTM layer of the sparse network that have a smaller impact on the result are reduced until the set sparse target is achieved.

6. The method for automatic modulation and identification of radio signals under constrained conditions according to claim 5, characterized in that: The adaptive noise reduction multi-information fusion network model uses the Adam optimizer and cross-entropy loss function for gradient updates during training.

Images