Automatic modulation identification method based on GF-IRANet

Abstract

The invention belongs to the field of communication confrontation reconnaissance, and discloses a GF-IRANet-based automatic modulation identification method, which comprises the following steps of: firstly, converting an RML 2016a modulation data set into a corresponding CW time-frequency image; preprocessing the CW time-frequency image through GF filtering to obtain a CW data set; and then feature extraction is carried out on the CW data set after GF filtering through IRANet, and modulation classification is realized. The invention provides a network model for applying an image classification technology to modulation identification. According to the method, the network structure is simplified, and the number of parameters is reduced on the premise of ensuring the recognition rate, so that the algorithm operation speed is improved, and the network classification speed is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of communication countermeasure reconnaissance and relates to an automatic modulation recognition method based on a GF-IRANet. BACKGROUND

[0002] Traditional electronic warfare adopts a combination of machines and manual work to determine the signal modulation type, and there is no doubt that the accuracy and efficiency of this method are not high. With the continuous development of the field of artificial intelligence, neural networks have gradually penetrated into various industries, and a large number of scholars have studied automatic modulation recognition technology based on deep learning technology. Everyone is looking for a method that is fast in classification speed and high in accuracy.

[0003] There are many methods for modulation recognition through deep learning at present. The papers “A hybrid model for automatic modulation classification based on residual neural networks and long short term memory” written by M. M. Elsagheer and S. M. Ramzy and “An autoencoder-based i / q channel interaction enhancement method for automatic modulation recognition” written by F. Zhang, C. Luo, J. Xu, and Y. Luo achieve good results in the field of modulation recognition by using IQ data as input and combining different structured network models to classify modulation data sets. However, IQ data can only reflect the joint characteristics of time and energy of signals, and lacks characteristics in the frequency domain. The more comprehensive the input features of neural networks are, the more comprehensive modulation information the network can capture during the training process, and thus accurate classification results can be obtained.

[0004] The paper “Deep architectures for modulation recognition” written by T. J. O’Shea uses an attention mechanism addition in modulation recognition to further enhance the feature values in the neural network. At the same time, the paper “Automatic modulation classification using resnext-gru with deep feature fusion” written by L. Li also proposes to use a denoising autoencoder to reduce the noise of the short-time Fourier transform time-frequency image, so as to improve the classification accuracy, which shows that suppressing noise can improve the modulation recognition effect to a certain extent.

[0005] Most modulation recognition techniques have poor performance, mainly due to incomplete data features and the influence of data set noise, which cannot meet the accuracy requirements of battlefield electromagnetic signal discrimination. Therefore, the present application proceeds from three aspects. First, preprocess the data set, integrate the original IQ data, complete the time-frequency conversion and noise reduction processing; second, provide more abundant input information for the network, input the IQ data and time-frequency image data into the network at the same time; third, improve the network structure, modify the residual attention network (Residual Attention Network, RANet) network structure to double input, and convert the modulation recognition problem to an image recognition problem. SUMMARY

[0006] The present application can improve the performance of automatic modulation recognition in electronic warfare reconnaissance based on effective signal features. The time-frequency image of a communication signal can reflect its internal modulation information characteristics and can be applied to identify signals of different modulation types. However, the time-frequency image of a low signal-to-noise ratio communication signal will be affected by noise, resulting in a low modulation recognition rate. Therefore, the present application proposes a method of guided filtering and noise reduction for time-frequency images, and inputs the filtered time-frequency image data into an improved residual attention network (Residual Attention Network, RANet) to realize modulation signal classification. In order to verify the effectiveness of the proposed method, the present application uses the RadioML2016.10a data set for testing. The simulation results show that the proposed method can improve the modulation recognition rate in a low signal-to-noise ratio scenario while maintaining high signal-to-noise ratio recognition effect.

[0007] Technical scheme of the present application:

[0008] An automatic modulation recognition method based on GF-IRANet, the steps are as follows:

[0009] First, convert the RML 2016a modulation data set into corresponding CW time-frequency images; then preprocess the CW time-frequency images through GF filtering to obtain the CW data set; then extract features from the GF filtered CW data set through IRANet to realize modulation classification;

[0010] (a) Choi-Williams time-frequency representation

[0011] To enable efficient modulation feature extraction using a subsequent neural network, the higher the resolution of the modulated signal's time-frequency image, the better. Considering the challenges faced by common time-frequency image acquisition techniques such as the short-time Fourier transform (STFT), including the conflict between time and frequency resolution, the difficulty in selecting window function scales, and the challenge of removing cross-terms from the Wigner-Ville distribution (WVD), this method uses the Choi-Williams distribution (CWD), which boasts high time-frequency resolution and the ability to suppress cross-terms, to perform time-frequency transformation on the IQ sequence x of the RML 2016a modulation dataset, obtaining a CW time-frequency image. The expression for the CWD is:

[0012]

[0013] Among them, t is the time variable; is the angular frequency variable; is the integration variable; It is a time delay; is the attenuation coefficient, which determines the width of the kernel function and is set to 1; the superscript * means complex conjugate;

[0014] (b) Guided filtering (GF)

[0015] The CW time-frequency image is used as the input image P of the filter GF, is the output image of the filter GF, is the guiding image, then:

[0016]

[0017] in, is the ordinal number of the pixel; It is A local window; and The output images are and guided images Middle The value at each pixel; and is the coefficient calculated in the kth window, which is calculated by minimizing the output image and the input image The squared difference between them is estimated as follows:

[0018]

[0019] in, is the preset regularization parameter; the coefficient and The solution is as follows:

[0020]

[0021]

[0022] wherein, and are respectively the mean and variance of the guided image in ; is the number of pixels in units; is the mean of the input image in units; when is different, the value of in equation (2) will change; in order to solve this problem, the mean of and must be calculated first, so the output formula of the filter GF is changed to:

[0023]

[0024] wherein, , ;

[0025] (c) Improved RANet model

[0026] The residual attention network (RANet) is optimized and designed, and a modulation recognition model CW-GF-IRANet based on single-input CW time-frequency image and a dual-input (CW+IQ)-GF-IRANet modulation recognition model are proposed.

[0027] 1) The specific implementation of the modulation recognition model CW-GF-IRANet based on single-input CW time-frequency image is as follows:

[0028] In CW-GF-IRANet, the input is a CW time-frequency image, which is then sequentially subjected to GF filtering processing, attention block feature extraction, and classification recognition three processes.

[0029] The attention block includes a main branch and a mask branch; wherein the main branch is for feature extraction and processing of the CW time-frequency image; the main branch includes a convolution layer, a normalization layer, an activation layer, and a pooling layer, and the feature pattern is obtained after the main branch processing ​The feature pattern extracted by the backbone branch is a reference pattern, and needs to be enhanced by a subsequent mask branch;

[0030] The mask branch is mainly composed of a convolution layer, a pooling layer, an activation layer and a residual block, and uses a bottom-up-top-down structure to learn a mask with the same size as the output feature pattern of the backbone branch; in the back propagation process, the mask branch also acts as a gradient update filter to prevent the incorrect gradient from updating the parameters of the backbone branch; the residual block contains multiple convolution, normalization, activation and Dropout operations, and the feature pattern is obtained after the mask branch processing ;

[0031] Directly stacking attention modules will cause a significant performance decline, to solve this problem, attention residual learning is proposed; by modifying the output of the attention module as:

[0032]

[0033] The attention block finally utilizes the mask branch and the backbone branch to realize key feature enhancement, and then adds the backbone branch to obtain At this time, the key features of different modulation methods are further strengthened through attention block residual learning, and finally, full connection layer and SoftMax activation function are used to realize classification recognition; wherein the dropout rate of each dropout layer is 0.2.

[0034] 2) The specific implementation of the dual-input (CW+IQ)-GF-IRANet modulation recognition model is as follows:

[0035] The dual-input (CW+IQ)-GF-IRANet modulation recognition model increases one IQ input on the basis of the CW-GF-IRANet, and adopts the same network structure; the IQ image generated by the RML 2016a modulation dataset and the CW time-frequency image are respectively processed through the GF filter, the feature extraction is realized through the attention block, the key features are captured and enhanced through the backbone branch and the mask branch, and the internal structure of the attention block is not changed in the whole process, and the connection layer, the convolution layer and the full connection layer are added in the classification recognition process; the connection layer one-dimensionally flattens the features output by the attention block, then the one-dimensional features generated by the two attention blocks of the dual-input (CW+IQ)-GF-IRANet modulation recognition model are superimposed, and then the convolution layer, the full connection layer and the SoftMax are used for feature integration and classification; wherein the parameters of each layer are consistent with the CW-GF-IRANet.

[0036] The beneficial effects of the present application are:

[0037] (1) The residual attention network (RANet) for image classification is used to realize modulation recognition, and the application proposes a network model for applying image classification technology to modulation recognition. The application simplifies the network structure, reduces the number of parameters under the premise of ensuring the recognition rate, thereby improving the algorithm running speed and the network classification speed.

[0038] (2) In order to ensure the robustness of the algorithm under low signal-to-noise ratio conditions, the application uses guided filtering (Guided Filter, GF) technology to perform noise reduction processing on the Choi-Williams time-frequency graph. This operation can highlight the time-frequency features of low signal-to-noise ratio signals, so that the improved RANet (IRANet) can more comprehensively capture the features of different modulation signals.

[0039] (3) Since there is similarity between the time-frequency features of 16QAM and 64QAM, AM-DSB and WBFM, the neural network will face challenges when determining which modulation mode a certain time-frequency feature belongs to. Therefore, the application further proposes a dual-stream network (Dual-Stream Network), which inputs the time-frequency image and another signal transformation (IQ data) into the improved RANet to improve the overall recognition rate. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 are BPSK time-frequency graphs before and after guided filtering; wherein (a) is not denoised at a signal-to-noise ratio of -4dB, (b) is GF denoised at a signal-to-noise ratio of -4dB, (c) is not denoised at a signal-to-noise ratio of -2dB, (d) is GF denoised at a signal-to-noise ratio of -2dB, (e) is not denoised at a signal-to-noise ratio of 0dB, and (f) is GF denoised at a signal-to-noise ratio of 0dB;

[0041] Figure 2 is a single-input CW-GF-IRANet network model;

[0042] Figure 3 is a dual-input IQ-CW-GF-IRANet network model;

[0043] Figure 4 is the recognition rate before and after guided filtering denoising;

[0044] Figure 5 is a comparison of single-input and dual-input recognition rates;

[0045] Figure 6 is a comparison with mainstream modulation recognition methods. DETAILED DESCRIPTION

[0046] The specific embodiments of the application are further described below in combination with the drawings and technical solutions.

[0047] 1. Experimental equipment and environmental configuration

[0048] Hardware configuration: graphics card GTX 3060 12G; memory 32G; CPU i9

[0049] Software system: Windows 11

[0050] Programming language: Python3.10

[0051] Deep learning framework: Tensorflow-gpu 2.8.0

[0052] 2. Experimental method, an automatic modulation recognition method based on GF-IRANet, the steps are as follows:

[0053] First, the RML 2016a modulation dataset is converted into the corresponding CW time-frequency image; then the CW time-frequency image is preprocessed by GF filtering to obtain the CW dataset; then the GF filtered CW dataset is extracted by IRANet to realize modulation classification;

[0054] (a) Choi-Williams time-frequency representation

[0055] In order to efficiently extract modulation features by the neural network later, the higher the resolution of the modulation signal time-frequency image is, the better. Considering the challenges faced by common time-frequency image acquisition techniques such as short-time Fourier transform (STFT), such as the conflict between time resolution and frequency resolution, the difficulty of window function scale selection, and the challenge of removing cross terms from the Wigner-Ville distribution (WVD), this method processes the IQ sequence x of the RML 2016a modulation dataset by Choi-Williams distribution (CWD) with high time-frequency resolution and cross-term suppression capability to obtain the CW time-frequency image. The expression of CWD is:

[0056]

[0057] Where t is the time variable; is the angular frequency variable; is the integral variable; is the time delay; is the attenuation coefficient, which determines the width of the kernel function, and is 1; the meaning of the upper index with * is complex conjugate;

[0058] (b) Guided filtering method GF (Guided filtering, GF)

[0059] The CW time-frequency image is taken as the input image P of the filter GF, is the output image of the filter GF, is the guiding image, then:

[0060]

[0061] in, is the ordinal number of the pixel; It is A local window; and The output images are and guided images Middle The value at each pixel; and is the coefficient calculated in the kth window, which is calculated by minimizing the output image and the input image The squared difference between them is estimated as follows:

[0062]

[0063] in, is the preset regularization parameter; the coefficient and The solution is as follows:

[0064]

[0065]

[0066] in, and They are Middle guide image The mean and variance of Therefore The number of pixels in units; Therefore The input image is a unit The mean of When different, the The value of will change; in order to solve this problem, we need to first find and The average value of the filter GF is Transformed into:

[0067]

[0068] in, , ;

[0069] (c) Improved RANet model

[0070] The residual attention network (RANet) is optimized and designed, and a modulation recognition model CW-GF-IRANet based on single-input CW time-frequency images and a dual-input (CW+IQ)-GF-IRANet modulation recognition model are proposed.

[0071] 1) Single-input CW-GF-IRANet (see Table 1)

[0072] In the single-input modulation model, the input data set is a CW time-frequency image with a dimension of (32, 32, 3), which then undergoes GF filtering processing, attention block feature extraction, and classification recognition in turn.

[0073] The attention block is the core part of the CW-GF-IRANet and mainly consists of a main branch and a mask branch. The main branch is used to extract and process the features of the input data and provide basic features for the subsequent attention mechanism. The main branch includes a convolution layer, a normalization layer, an activation layer, and a pooling layer. The convolution layer uses 32 convolution kernels with a size of (2, 2) and same padding. The activation layer applies a ReLU activation function. The pooling layer has a pooling kernel size of (2, 2) and same padding for downsampling. After processing by the main branch, a feature pattern with a dimension of (16, 16, 32) is obtained. The feature pattern extracted by the main branch is the reference pattern, which needs to be enhanced by the mask branch.

[0074] The mask branch mainly consists of a convolution layer, a pooling layer, an activation layer, and a residual block. It uses a bottom-up top-down structure to learn a mask with the same size as the output feature of the main branch. This mask is used to weight the output feature of the main branch, similar to the control gate in the Highway Network. During backpropagation, the mask branch also acts as a gradient update filter to prevent incorrect gradients (resulting from noisy labels) from updating the main branch parameters, making the network robust to noisy labels. The residual block contains multiple convolution, normalization, activation, and Dropout operations to adjust the channel number and feature map. .

[0075] Directly stacking attention modules can cause a significant performance drop. To solve this problem, the paper proposes attention residual learning. The output of the attention module is modified as follows:

[0076]

[0077] This method can retain the good properties of the original features, while giving the features the ability to bypass the mask branch and propagate forward directly to the top layer, weakening the feature selection ability of the mask branch, so that the stacked attention module can gradually refine the feature map and improve network performance.

[0078] The attention block extracts the features of different modulation signals. The key features at this point are further enhanced through residual attention learning. Finally, modulation classification is achieved using a fully connected layer (128 neurons, SeLU activation function) and a SoftMax layer (11 neurons, SoftMax activation function, L2 regularization). The dropout rate of each dropout layer is 0.2.

[0079] 2) Dual input (CW+IQ)-GF-IRANet (see Table 2)

[0080] To achieve optimal classification results, a dual-input IRANet, or (CW+IQ)-GF-IRANet, is proposed. This adds an IQ input to the CW-GF-IRANet, using the same network structure. The main idea remains the same as the single-input model: the IQ image is filtered using GF filtering, and features are extracted using the T-SM-B module. Key features are captured and enhanced using the backbone and mask branches. The internal structure of the attention block remains unchanged throughout this process. However, a connection layer, a convolutional layer, a fully connected layer, and a dropout layer are added to the modulation recognition and classification block. The connection layer flattens the features output by the attention block in one dimension, then superimposes the one-dimensional features generated by the two attention blocks. The features are then integrated and classified using convolutional, fully connected, and softmax layers. The parameters of each layer remain the same as those of the single-input CW-GF-IRANet.

[0081] Table 1 CW-GF-IRANet network model

[0082]

[0083] Table 2 (CW+IQ)-GF-IRANet network model

[0084] 3. Experimental Verification

[0085] (1) Experimental dataset

[0086] The following experiments used the open-source RadioML2016.10a dataset. The RadioML2016.10a dataset contains 220,000 modulated signals classified into 20 categories, with signal-to-noise ratios ranging from -20 dB to +18 dB. Eleven common modulation schemes include WBFM, AM-DSB, AM-SB, BPSK, CPFSK, GFSK, 4PAM, 16QAM, 64QAM, QPSK, and 8PSK. The 1,000 signals in each of these modulation schemes have different signal-to-noise ratios. Each signal in the dataset consists of 128 complex floating-point time I / Q samples generated in an environment with Gaussian noise and multipath fading. To verify the effectiveness of this strategy, the dataset was split into training and test sets with an 8:2 ratio. During the simulation, for each modulation signal at each signal-to-noise ratio, 800 signals were randomly selected for training and 200 for testing.

[0087] (2) Experimental results

[0088] a. Performance comparison of GF before and after noise reduction

[0089] Since various noise signals generated in actual scenarios will interfere with the modulation recognition effect, effective noise suppression technology is crucial to improving the modulation recognition accuracy. Figure 4 The modulation recognition rate curves before and after GF denoising show that the recognition rate is higher between -8dB and 0dB compared to before GF denoising. The GF denoising method proposed in this paper can improve the robustness of the automatic modulation recognition system under low signal-to-noise ratio conditions, making it more suitable for modulation recognition in real-world channel environments. In low signal-to-noise ratio environments, the GF denoising method effectively reduces image noise.

[0090] b. Performance comparison between single-input and dual-input networks

[0091] In order to solve the problem of low recognition rate of modulation with single feature, a dual input method is proposed. From the perspective of signal recognition, this method can extract signal features from several dimensions, which is beneficial to modulation recognition. Figure 5 It is easy to observe that the combination of IQ and CW time-frequency images produces a higher recognition gain than a single CW time-frequency image. By integrating IQ and CW features for feature re-extraction in the convolutional layer and merging and summing them at the output, the probability of modulation recognition by the network is greatly improved.

[0092] c. Performance comparison of different methods

[0093] This paper compares the performance of dual-input (CW+IQ)-GF-IRANet and single-input CW-GF-IRANet with some benchmark methods[8]

[17] . The simulation results are shown in Figure 6The graph shows that the performance of the method of the present application is about 10% higher than that of the reference algorithm in the low SNR region of -18dB-2dB, while the performance of the two proposed methods is slightly higher than that of the reference algorithm in the high SNR range of 0dB-18dB. The above performance can be attributed in part to the GF filtering technique, which suppresses irrelevant information in low signal-to-noise ratio images to minimize noise, and the remaining attention block of the network, which attracts more attention to the salient features of the image, creating opportunities for modulation recognition rate improvement.

Claims

1. An automatic modulation recognition method based on GF-IRANet, characterized in that: Here are the steps: First, the RML 2016a modulation dataset is converted into its corresponding CW time-frequency image. Then, the CW time-frequency image is preprocessed using GF filtering to obtain a CW dataset. Subsequently, IRANet is used to extract features from the GF-filtered CW dataset to achieve modulation classification. (a) Choi-Williams time-frequency representation The IQ sequence x of the RML 2016a modulation dataset is transformed into a CW time-frequency image using the Choi-Williams distribution with high time-frequency resolution and cross-term suppression capability. The expression of CWD is: , where t is the time variable; is the angular frequency variable; is the integration variable; It is a time delay; is the attenuation coefficient, which determines the width of the kernel function and is set to 1; the superscript * means complex conjugate; (b) Guided Filtering Method GF The CW time-frequency image is used as the input image P of the filter GF, is the output image of the filter GF, is the guiding image, then: ,in, is the ordinal number of the pixel; It is A local window; and The output images are and guided images Middle The value at each pixel; and is the coefficient calculated in the kth window, which is calculated by minimizing the output image and the input image The squared difference between them is estimated as follows: ,in, is the preset regularization parameter; the coefficient and The solution is as follows: ,in, and They are Middle guide image The mean and variance of Therefore The number of pixels in units; Therefore The input image is a unit The mean of When different, the The value of will change; in order to solve this problem, we need to first find and The average value of the filter GF is Transformed into: ,in, , ; (c) Improved RANet model The residual attention network is optimized and designed, and a modulation recognition model CW-GF-IRANet based on a single-input CW time-frequency image and a dual-input (CW+IQ)-GF-IRANet modulation recognition model are proposed; 1) The modulation recognition model CW-GF-IRANet based on single-input CW time-frequency image is specifically implemented as follows: In CW-GF-IRANet, the input is a CW time-frequency image, which then goes through three processes: GF filtering, attention block feature extraction, and classification recognition; The attention block includes a trunk branch and a mask branch; the trunk branch extracts and processes features of the CW time-frequency image; the trunk branch includes a convolution layer, a normalization layer, an activation layer, and a pooling layer. After processing by the trunk branch, the feature pattern is obtained. The feature pattern extracted by the main branch is the reference pattern, which needs to be enhanced by the subsequent mask branch; The mask branch is mainly composed of convolutional layers, pooling layers, activation layers, and residual blocks. It uses a bottom-up-top-down structure to learn a mask with the same size as the main branch output feature pattern. During the back propagation process, the mask branch also acts as a gradient update filter to prevent incorrect gradients from updating the main branch parameters. The residual block contains multiple convolution, normalization, activation, and Dropout operations. After the mask branch is processed, the feature pattern is obtained. ; Directly stacking attention modules will lead to a significant performance degradation. To solve this problem, attention residual learning is proposed; by modifying the output of the attention module to: , the attention block finally utilizes the mask branch Branch from the trunk Feature fusion realizes key feature enhancement and then integrates with the main branch Add together to get At this time, the key features that can distinguish different modulation modes are further enhanced through residual learning of the attention block, and finally the classification and recognition are achieved using the fully connected layer and the SoftMax activation function; the dropout rate of each dropout layer is 0.2; 2) The specific implementation of the dual-input (CW+IQ)-GF-IRANet modulation recognition model is as follows: The dual-input (CW+IQ)-GF-IRANet modulation recognition model adds an IQ input to the CW-GF-IRANet model and adopts the same network structure. The IQ image and CW time-frequency image generated by the RML 2016a modulation dataset are processed by GF filtering respectively. Feature extraction is achieved through the attention block, and key features are captured and enhanced using the backbone branch and mask branch. The internal structure of the attention block remains unchanged throughout the process. During the classification and recognition process, connection layers, convolutional layers, and fully connected layers are added. The connection layers flatten the features output by the attention block in one dimension. The one-dimensional features generated by the two attention blocks of the dual-input (CW+IQ)-GF-IRANet modulation recognition model are then superimposed. The features are then integrated and classified through convolutional layers, fully connected layers, and softmax layers. The parameters of each layer remain consistent with those of CW-GF-IRANet.

Images