Signal modulation identification method and device based on deep learning

Abstract

The invention discloses a signal modulation identification method and device based on deep learning, and relates to the technical field of signal modulation identification, and the method comprises the steps: obtaining a to-be-identified modulation signal; removing noise from the modulation signal to be identified by using a Gaussian filter, and carrying out short-time Fourier transform on the modulation signal to be identified after noise removal to obtain a gray scale time-frequency graph; performing normalization processing on the gray scale time-frequency graph, and converting the normalized gray scale time-frequency graph into a color time-frequency graph by adopting a color mapping function; processing the color time-frequency graph by adopting a trained deep learning network to obtain a classification result of the modulation signal to be identified; wherein the trained deep learning network is obtained by training the initial deep learning network by taking data of a preset category as a training data set. According to the invention, the identification accuracy of the modulation mode in a low signal-to-noise ratio environment can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal modulation recognition, and particularly relates to a signal modulation recognition method and device based on deep learning. Background Art

[0002] With the rapid development of wireless communication technologies, such as the wide application of 5G, Internet of Things (IoT), and satellite navigation technologies, human society has entered a digital era of all things interconnected, and wireless communication technologies are reshaping modern society at an unprecedented speed. From remote medical treatment supported by 5G networks to smart homes driven by the IoT, the access of a large number of devices has made the spectrum resources increasingly tense and the electromagnetic environment increasingly complex. At the same time, in special scenarios such as deep space exploration, signals often fall into a low signal-to-noise ratio (SNR) dilemma due to long-distance transmission or strong interference, making it difficult for the receiving end to accurately identify the modulation type.

[0003] Under this background, the rapid development of modulation recognition technology has been promoted. Modulation recognition technology is mainly applied to non-cooperative communication scenarios. The core goal of the receiving end in non-cooperative communication scenarios is to automatically identify the modulation mode of the signal without prior information, and then demodulate and restore the original signal. Modulation recognition technology has become a key support in fields such as cognitive radio, spectrum management, and electronic reconnaissance.

[0004] Traditional modulation recognition methods mainly rely on expert experience, such as feature engineering based on spectral line features, high-order cumulants, or information entropy. Domain experts are required to design complex feature parameters and complete classification through classifiers such as decision trees to achieve communication signal modulation recognition. However, with the gradual deterioration of the electromagnetic environment, the characteristics of communication signals are easily masked by noise, the threshold setting is relatively rigid, and it is difficult to cover emerging modulation types, such as high-order QAM. Therefore, such methods are vulnerable in actual scenarios with severe noise interference and changing channel conditions. Especially when the SNR is lower than 0 dB, the signal and noise are highly coupled, and the recognition accuracy of traditional algorithms drops sharply, severely restricting the application in non-cooperative communication, emergency response, and other scenarios.

[0005] In recent years, deep learning has injected new impetus into modulation recognition. Through models such as Convolutional Neural Networks (CNNs), the system can directly extract deep time-frequency domain features from the original signal, avoiding the limitations of manual design. Existing research shows that the accuracy of deep learning-based solutions exceeds 90% at medium and high signal-to-noise ratios (SNR≥10 dB), but the performance drops sharply in low SNR environments, even below 50%. This shortcoming exposes the defects of current models: over-reliance on idealized data training, insufficient robustness to noise, and high model complexity, making it difficult to adapt to edge computing devices. Summary of the Invention

[0006] To solve the above problems existing in the prior art, the present invention provides a signal modulation recognition method and device based on deep learning. The technical problems to be solved by the present invention are realized through the following technical solutions: In the first aspect, the present invention provides a signal modulation recognition method based on deep learning, including: Obtain the modulation signal to be recognized; Use a Gaussian filter to remove noise from the modulation signal to be recognized, perform short-time Fourier transform on the noise-removed modulation signal to be recognized to obtain a grayscale time-frequency diagram; perform normalization processing on the grayscale time-frequency diagram, and use a color mapping function to convert the normalized grayscale time-frequency diagram into a color time-frequency diagram; Use the trained deep learning network to process the color time-frequency diagram to obtain the classification result of the modulation signal to be recognized; Among them, the trained deep learning network is obtained by training the initial deep learning network with data of a preset category as the training data set.

[0007] In the second aspect, the present invention further provides a signal modulation recognition device based on deep learning, including: A data acquisition module for obtaining the modulation signal to be recognized; A data processing module for using a Gaussian filter to remove noise from the modulation signal to be recognized, performing short-time Fourier transform on the noise-removed modulation signal to be recognized to obtain a grayscale time-frequency diagram; performing normalization processing on the grayscale time-frequency diagram, and using a color mapping function to convert the normalized grayscale time-frequency diagram into a color time-frequency diagram; A data recognition module for using the trained deep learning network to process the color time-frequency diagram to obtain the classification result of the modulation signal to be recognized; Among them, the trained deep learning network is obtained by training the initial deep learning network with data of a preset category as the training data set.

[0008] Advantages of the present invention: The signal modulation recognition method and device based on deep learning provided by the present invention improve the focusing ability of the deep learning network model on signal features and improve the recognition accuracy of multiple modulation methods in a low signal-to-noise ratio environment.

[0009] The following will further describe the present invention in detail with reference to the drawings and embodiments. Description of the Drawings

[0010] Figure 1 is a flowchart of a signal modulation recognition method based on deep learning provided by an embodiment of the present invention; Figure 2It is a schematic diagram of the time domain diagram of the modulation signal to be recognized provided by the embodiment of the present invention before and after denoising by a Gaussian filter; Figure 3 It is a schematic diagram of the QPSK signal before and after denoising by a Gaussian filter at 6 dB provided by the embodiment of the present invention; Figure 4 It is a schematic diagram of the CPFSK signal before and after denoising by a Gaussian filter at 6 dB provided by the embodiment of the present invention; Figure 5 It is a schematic diagram of the QPSK signal before and after denoising by a Gaussian filter at -6 dB provided by the embodiment of the present invention; Figure 6 It is a schematic diagram of the CPFSK signal before and after denoising by a Gaussian filter at -6 dB provided by the embodiment of the present invention; Figure 7 It is a schematic diagram of the deep learning network provided by the embodiment of the present invention. Detailed implementation manners

[0011] The following further describes the present invention in detail with reference to specific embodiments, but the implementation manners of the present invention are not limited thereto.

[0012] Please refer to Figure 1 , Figure 1 It is a flowchart of a signal modulation recognition method based on deep learning provided by the embodiment of the present invention. A signal modulation recognition method based on deep learning provided by the present invention includes: S101. Obtain the modulation signal to be recognized.

[0013] Specifically, in this embodiment, the modulation mode of the modulation signal to be recognized can be any one of 3 analog modulations (AM-DSB, AM-SSB, WBFM) or 8 digital modulations (8PSK, QPSK, BPSK, CPFSK, GFSK, PAM4, 16QAM, 64QAM).

[0014] S102. Use a Gaussian filter to remove noise from the modulation signal to be recognized, perform short-time Fourier transform on the denoised modulation signal to be recognized to obtain a gray time-frequency diagram; perform normalization processing on the gray time-frequency diagram, and use a color mapping function to convert the normalized gray time-frequency diagram into a color time-frequency diagram.

[0015] Specifically, in this embodiment, first, use a Gaussian filter to denoise the modulation signal to be recognized. The Gaussian filter is defined as: ; where represents the standard deviation of the Gaussian filter, Denote the radius of the Gaussian kernel. Optionally, the standard deviation of the Gaussian filter is set to 1, and the radius of the Gaussian kernel is set to 7.

[0016] The modulation signal to be recognized after denoising by the Gaussian filter It is expressed as: ; Wherein, Denote the modulation signal to be recognized.

[0017] In this embodiment, after being processed by the Gaussian filter, the out-of-band noise can be effectively suppressed, such as Figure 2 shown Figure 2 is a schematic diagram of the time domain diagram of the modulation signal to be recognized provided by the embodiment of the present invention before and after denoising by the Gaussian filter.

[0018] In order to enable the model to obtain the time-frequency information of the modulation signal to be recognized, the denoised modulation signal to be recognized is subjected to short-time Fourier transform, and the short-time Fourier transform is defined as: ; Wherein, Denote the index of the time frame, Denote each frame of signal, Denote the window function, with the center located at the time frame ; Denote the index of the frequency; optionally, the window function selects the Hamming window, the window length is 40, the window shift is 2, the frequency resolution is 200 kHz, and a time-frequency diagram of 100×100 pixels is generated.

[0019] After performing short-time Fourier transform on the denoised modulation signal to be recognized, a grayscale time domain diagram is obtained, then the grayscale time domain diagram is normalized, and then it is converted into a color diagram through the Parula color mapping function, which will expand the signal feature dimension and is beneficial to improving the recognition accuracy of the model, such as Figures 3 - 6 shown Figure 3 is a schematic diagram of the QPSK signal before and after denoising by the Gaussian filter at 6 dB provided by the embodiment of the present invention, Figure 4 is a schematic diagram of the CPFSK signal before and after denoising by the Gaussian filter at 6 dB provided by the embodiment of the present invention, Figure 5 is a schematic diagram of the QPSK signal before and after denoising by the Gaussian filter at -6 dB provided by the embodiment of the present invention, Figure 6 is a schematic diagram of the CPFSK signal before and after denoising by the Gaussian filter at -6 dB provided by the embodiment of the present invention.

[0020] S103. Process the color time-frequency diagram by using the trained deep learning network to obtain the classification result of the modulation signal to be recognized; Among them, the trained deep learning network is obtained by training the initial deep learning network with the data of preset categories as the training data set.

[0021] Specifically, as Figure 7 shown, Figure 7 is a schematic diagram of the deep learning network provided by an embodiment of the present invention. In this embodiment, the overall structure of the deep learning network includes a feature extraction backbone, a feature pyramid multi-scale fusion module, an attention enhancement module, and a classification output module; among them, the feature extraction backbone includes a first convolution module, a first residual block, and a second residual block. After shallow feature extraction of the color time-frequency map through the first convolution module, deep feature extraction is performed through the first residual block and the second residual block; the feature pyramid multi-scale fusion module includes a first channel adjustment module, a second channel adjustment module, and an upsampling module; after the attention enhancement module fuses multi-level features, a SimAM attention module is embedded at the output end. Through an adaptive channel weighting mechanism, it focuses on the key feature regions of the time-frequency map, suppresses noise interference, and further improves the feature extraction ability of the model; the classification output module compresses the spatial dimension of the high-dimensional features after attention enhancement through global average pooling (Global Average Pooling, GAP), and then maps them to an eleven-dimensional classification space through a fully connected layer, and finally realizes the classification and recognition of eleven modulation signals such as QPSK, BPSK, CPFSK... AM-DSB.

[0022] In this embodiment, the trained deep learning network includes a trained first convolution module, a trained first residual block, a trained second residual block, a trained first channel adjustment module, a trained second channel adjustment module, a trained upsampling module, a trained second convolution module, a trained attention mechanism module, a trained global average pooling layer, and a trained fully connected layer; using the trained deep learning network to process the color time-frequency map, the classification result of the modulation signal to be recognized is obtained, including: Using the trained first convolution module to perform feature extraction on the color time-frequency map to obtain a first feature; using the trained first channel adjustment module to perform channel adjustment on the first feature to obtain a processed first feature; optionally, the first convolution module includes a convolutional layer, a batch normalization layer, and an activation layer. The size of the color time-frequency map input to the first convolution module is 100×100×3. The convolutional layer of the first convolution module uses 32 convolutional kernels, the size of the convolutional kernels is 3×3, and the padding mode is "same" to ensure that the output size remains unchanged or is close to the original size; Using the trained first residual block to perform feature extraction on the first feature to obtain a second feature; using the trained second channel adjustment module to perform channel adjustment on the second feature to obtain a processed second feature; Feature extraction is performed on the second feature using the trained second residual block to obtain a third feature; an upsampling operation is performed on the third feature using the trained upsampling module to obtain the processed third feature; The processed first feature, the processed second feature, and the processed third feature are fused multiple times to obtain a fused feature; Feature extraction is performed on the fused feature using the trained second convolutional module to obtain a fourth feature; The trained attention mechanism module is used to process the fourth feature, fusing spatial information and channel information to obtain a fifth feature; The trained global average pooling layer is used to process the fifth feature to obtain a sixth feature; The trained fully connected layer is used to process the sixth feature to obtain the classification result of the modulation signal to be recognized.

[0023] In this embodiment, the trained first residual block includes a first convolutional layer, a first batch normalization layer, a first activation layer, a second convolutional layer, a second batch normalization layer, and a second activation layer; performing feature extraction on the first feature using the trained first residual block to obtain a second feature includes: Feature extraction is performed on the first feature using the first convolutional layer to obtain a first output, the first output is processed using the first batch normalization layer to obtain a second output, the second output is processed using the first activation layer to obtain a third output, feature extraction is performed on the third output using the second convolutional layer to obtain a fourth output, and the fourth output is processed using the second batch normalization layer to obtain a fifth output; optionally, the first convolutional layer uses 32 convolutional kernels, the size of the convolutional layer is 3×3, the stride is 1, and the padding mode is "same", and the second convolutional layer uses 32 convolutional kernels, the size of the convolutional layer is 3×3, the stride is 1, and the padding mode is "same"; The first feature and the fifth output are concatenated to obtain a sixth output; The second activation layer is used to process the sixth output to obtain a second feature; the skip connection can fuse low-level detailed features and high-level abstract features across layers, effectively alleviate the gradient disappearance problem, and optimize the feature distribution through the end activation function to achieve efficient information transmission.

[0024] In this embodiment, the trained second residual block includes a first branch, a second branch, and a third activation layer; the first branch includes a third convolutional layer, a third batch normalization layer, a fourth activation layer, a fourth convolutional layer, and a fourth batch normalization layer, and the second branch includes a fifth convolutional layer and a fifth batch normalization layer; performing feature extraction on the second feature using the trained second residual block to obtain a third feature includes: The third convolutional layer is used to extract features from the second feature to obtain the seventh output. The third batch normalization layer is used to process the seventh output to obtain the eighth output. The fourth activation layer is used to process the eighth output to obtain the ninth output. The fourth convolutional layer is used to extract features from the ninth output to obtain the tenth output. The fourth batch normalization layer is used to process the tenth output to obtain the eleventh output; Optionally, the third convolutional layer uses 64 convolutional kernels, the size of the convolutional layer is 3×3, the stride is 2, and the padding mode is "same". The fourth convolutional layer uses 64 convolutional kernels, the size of the convolutional layer is 3×3, the stride is 2, and the padding mode is "same". The fifth convolutional layer is used to extract features from the second feature to obtain the twelfth output. The fifth batch normalization layer is used to process the twelfth output to obtain the thirteenth output; Optionally, the fifth convolutional layer uses 64 convolutional kernels, the size of the convolutional layer is 1×1, and the padding mode is "same". The eleventh output and the thirteenth output are concatenated to obtain the fourteenth output; The third activation layer is used to process the fourteenth output to obtain the third feature; Skip connections can fuse low-level detailed features and high-level abstract features across layers, effectively alleviating the vanishing gradient problem, and optimizing the feature distribution through the end activation function to achieve efficient information transmission.

[0025] In this embodiment, the SimAM attention module is a three-dimensional attention mechanism that integrates spatial information and channel information, which helps to make better choices in the visual processing process. The SimAM attention mechanism can directly calculate the minimum energy function to quantify the importance of each neuron, where the energy function is defined as: ; where, represents and 's binary label, represents the true value of the target neuron, represents the true values of other neurons, represents the estimated value of the target neuron, represents the neurons in the same channel as the target neuron, represents the number of energy functions, represents the regularization parameter, represents the target neuron of a single channel in the input feature, represents the serial number of the target neuron, and represent the weight and offset respectively. Finding the linear separability of from other neurons is equivalent to minimization. For Introducing the regularization term and simplifying gives: ; Solving the above equation gives: ; ; where, , ; The minimum energy function of the trained attention mechanism module has the following expression: ; As can be seen from the above equation, when is lower, it indicates that the correlation between the target neuron and adjacent neurons is lower, and the importance is greater. Therefore, during the use of the SimAM attention mechanism, only the mean and variance need to be calculated, and no new parameters need to be introduced to capture the important information in the feature map.

[0026] In this embodiment, the trained first channel adjustment module, the trained second channel adjustment module, and the trained upsampling module constitute a feature pyramid network; in order to capture the modulation features at different scales of the time-frequency map, a feature pyramid network is introduced into the model. The multi-level features output by the feature extraction backbone are aligned in spatial resolution through upsampling operations, and then the channel numbers are aligned through 1×1 convolution for channel adjustment. After that, the multi-level features are fused to form pyramid features containing rich context.

[0027] In this embodiment, the trained global average pooling layer compresses the spatial dimension of each channel into a scalar to prepare features for the final fully connected layer.

[0028] In this embodiment, the output of the trained fully connected layer has 11 categories, and the activation function is softmax.

[0029] In this embodiment, the training process of the trained deep learning network includes: Obtaining data of multiple preset categories; the data of the preset categories are modulation signals, and the modulation methods of the modulation signals include analog modulation and digital modulation.

[0030] Optionally, the publicly available RadioML2016.10a dataset is adopted. This dataset includes 3 kinds of analog modulation (AM-DSB, AM-SSB, WBFM) and 8 kinds of digital modulation (8PSK, QPSK, BPSK, CPFSK, GFSK, PAM4, 16QAM, 64QAM), totaling 11 kinds of modulation signals. Each kind of signal is evenly distributed at 2dB intervals in the signal-to-noise ratio range from -20dB to 18dB (a total of 20 signal-to-noise ratio levels). Each signal-to-noise ratio contains 1000 samples, and the total number of samples is 220000. It is divided into a training set (154000 samples), a validation set (22000 samples), and a test set (44000 samples) according to the ratio of 7:1:2. The original IQ data dimension is 2×128 (real and imaginary components), which is converted into a complex tensor form and subjected to zero-mean normalization to eliminate the acquisition bias of the device. The specific parameter information of the dataset is shown in Table 1.

[0031] Table 1 Main parameters of RadioML2016.10a

[0032] Denoise the data of multiple preset categories, perform short-time Fourier transform on the denoised data of multiple preset categories to obtain multiple grayscale time-frequency map data samples; perform normalization processing on multiple grayscale time-frequency map data samples, and use a color mapping function to convert the normalized multiple grayscale time-frequency map data samples into multiple color time-frequency map data samples.

[0033] Divide multiple color time-frequency map data samples into a training set, a validation set, and a test set according to a preset ratio, and label the samples in the training set, validation set, and test set to obtain the true labels of the samples in the training set, validation set, and test set.

[0034] Input some samples in the training set into the th deep learning network to be trained for training, and obtain the prediction results output during the th training process.

[0035] According to the prediction results output during the th training process and the true labels of the samples of the th deep learning network to be trained, calculate the loss, and use it as the loss of the th training process.

[0036] According to the loss of the th training process, perform backpropagation to update the network parameters of the th deep learning network to be trained, and obtain the The deep learning network to be trained next; at the same time, every time the preset verification interval value is reached, the samples in the verification set are input into the deep learning network to be trained, and the predicted results are obtained. Combining the true labels of the samples in the verification set, model selection and hyperparameter tuning are carried out; iterating in this way until the number of training times or the degree of convergence meets the preset conditions, a trained deep learning network is obtained. For example, the preset verification interval value is 50 iteration times.

[0037] The samples in the test set are input into the trained deep learning network, and the predicted results are obtained. According to the predicted results and the true labels of the samples in the test set, the recognition accuracy is calculated to judge the performance of the trained deep learning network; optionally, the test set samples of 11 modulation modes at 20 signal-to-noise ratios are input into the trained deep learning network, and the recognition accuracy is obtained. After calculation, the average recognition accuracy of the trained deep learning network at each signal-to-noise ratio can be obtained.

[0038] In this embodiment, the loss of the nth training process is the cross-entropy loss function.

[0039] In this embodiment, when training the model, the maximum number of iterations is 200, the initial learning rate is 0.001, the decay factor is 0.5, the learning rate starting condition is that the loss has not decreased within five iterations, the batch size is set to 512, and the waiting rounds for early stopping are 20.

[0040] In this embodiment, during the test process, the test set samples have 20 signal-to-noise ratio levels from -20dB to 18dB, and each level contains 4000 samples (20 classes × 200 samples / class), with a total of 44000 test samples.

[0041] The evaluation metrics are the single signal-to-noise ratio accuracy and the average accuracy. The single signal-to-noise ratio accuracy refers to the ratio of the number of correct predictions to the total number of test samples at a single signal-to-noise ratio level; the average accuracy is the arithmetic mean of the accuracies at 20 signal-to-noise ratio levels.

[0042] In summary, the present invention proposes a signal modulation recognition method based on deep learning. First, a Gaussian filter is used to denoise the input signal to reduce the interference of noise on subsequent processing. Secondly, the denoised signal is subjected to short-time Fourier transform (STFT) and mapped to the color space to generate a color time-frequency diagram, which is used as the input of the deep learning network for subsequent modulation recognition tasks. Thirdly, the adopted deep learning network integrates a feature pyramid structure, a SimAM parameter-free attention mechanism, and a residual idea. Among them, the feature pyramid structure can achieve multi-scale feature fusion of the signal time-frequency diagram, effectively capture the global and local information of the signal, and the SimAM parameter-free attention mechanism can make the network focus on the key features of the time-frequency diagram, thereby improving the network's recognition ability of signal modulation modes. The introduction of the residual idea effectively solves the problem of gradient disappearance that may occur in the training process of the deep learning network, further improving the training efficiency and stability of the network. Through comparative experiments with current advanced deep learning networks, the network proposed by the present invention shows significant advantages in the accuracy of signal modulation recognition, can more accurately identify signals of different modulation methods, and has high practical value and broad application prospects.

[0043] Based on the same inventive concept, the present invention also provides a signal modulation recognition device based on deep learning for implementing the signal modulation recognition method provided in the above embodiments of the present invention. For the embodiments of the method, please refer to the above, and details will not be described herein again; the device includes: A data acquisition module for acquiring the modulation signal to be recognized; A data processing module for removing noise from the modulation signal to be recognized by using a Gaussian filter, performing short-time Fourier transform on the noise-removed modulation signal to be recognized to obtain a grayscale time-frequency diagram; performing normalization processing on the grayscale time-frequency diagram, and converting the normalized grayscale time-frequency diagram into a color time-frequency diagram by using a color mapping function; A data recognition module for processing the color time-frequency diagram by using a trained deep learning network to obtain a classification result of the modulation signal to be recognized; Among them, the trained deep learning network is obtained by training the initial deep learning network with data of a preset category as the training data set.

[0044] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variant are intended to cover non-exclusive inclusion, so that an article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the article or device including the said element. Words such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The orientation or positional relationship indicated by "upper", "lower", "left", "right", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as a limitation on the present invention.

[0045] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification.

[0046] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A signal modulation recognition method based on deep learning, characterized in that Including: Obtain the modulation signal to be recognized; Use a Gaussian filter to remove noise from the modulation signal to be recognized, perform short-time Fourier transform on the noise-removed modulation signal to be recognized to obtain a grayscale time-frequency diagram; perform normalization processing on the grayscale time-frequency diagram, and use a color mapping function to convert the normalized grayscale time-frequency diagram into a color time-frequency diagram; Use the trained deep learning network to process the color time-frequency diagram to obtain the classification result of the modulation signal to be recognized; Among them, the trained deep learning network is obtained by training the initial deep learning network with data of a preset category as the training data set.

2. The signal modulation recognition method based on deep learning according to claim 1, characterized in that, The modulation signal to be recognized after noise removal has the following expression: Among them, σ represents the standard deviation of the Gaussian filter, r represents the radius of the Gaussian kernel, x(n) represents the modulation signal to be recognized, and G(r) represents the Gaussian filter.

3. The signal modulation recognition method based on deep learning according to claim 1, characterized in that The trained deep learning network includes a trained first convolutional module, a trained first residual block, a trained second residual block, a trained first channel adjustment module, a trained second channel adjustment module, a trained upsampling module, a trained second convolutional module, a trained attention mechanism module, a trained global average pooling layer, and a trained fully connected layer; The step of using the trained deep learning network to process the color time-frequency diagram to obtain the classification result of the modulation signal to be recognized includes: Use the trained first convolutional module to extract features from the color time-frequency diagram to obtain a first feature; use the trained first channel adjustment module to adjust the channels of the first feature to obtain a processed first feature; Use the trained first residual block to extract features from the first feature to obtain a second feature; use the trained second channel adjustment module to adjust the channels of the second feature to obtain a processed second feature; Use the trained second residual block to extract features from the second feature to obtain a third feature; use the trained upsampling module to perform an upsampling operation on the third feature to obtain a processed third feature; Perform multi-scale fusion on the processed first feature, the processed second feature, and the processed third feature to obtain a fused feature; Use the trained second convolutional module to extract features from the fused feature to obtain a fourth feature; Use the trained attention mechanism module to process the fourth feature, fuse spatial information and channel information to obtain a fifth feature; Use the trained global average pooling layer to process the fifth feature to obtain a sixth feature; Use the trained fully connected layer to process the sixth feature to obtain the classification result of the modulation signal to be recognized.

4. The method for signal modulation recognition based on deep learning according to claim 3, characterized in that The trained first residual block includes a first convolutional layer, a first batch normalization layer, a first activation layer, a second convolutional layer, a second batch normalization layer, and a second activation layer; the step of using the trained first residual block to extract features from the first feature to obtain a second feature includes: The first convolution layer is used to extract features from the first feature to obtain a first output. The first batch normalization layer is used to process the first output to obtain a second output. The first activation layer is used to process the second output to obtain a third output. The second convolution layer is used to extract features from the third output to obtain a fourth output. The second batch normalization layer is used to process the fourth output to obtain a fifth output; The first feature and the fifth output are concatenated to obtain a sixth output; The second activation layer is used to process the sixth output to obtain a second feature.

5. The signal modulation recognition method based on deep learning according to claim 3, wherein The trained second residual block includes a first branch, a second branch, and a third activation layer. The first branch includes a third convolution layer, a third batch normalization layer, a fourth activation layer, a fourth convolution layer, and a fourth batch normalization layer. The second branch includes a fifth convolution layer and a fifth batch normalization layer. Using the trained second residual block to extract features from the second feature to obtain a third feature includes: The third convolution layer is used to extract features from the second feature to obtain a seventh output. The third batch normalization layer is used to process the seventh output to obtain an eighth output. The fourth activation layer is used to process the eighth output to obtain a ninth output. The fourth convolution layer is used to extract features from the ninth output to obtain a tenth output. The fourth batch normalization layer is used to process the tenth output to obtain an eleventh output; The fifth convolution layer is used to extract features from the second feature to obtain a twelfth output. The fifth batch normalization layer is used to process the twelfth output to obtain a thirteenth output; The eleventh output and the thirteenth output are concatenated to obtain a fourteenth output; The third activation layer is used to process the fourteenth output to obtain a third feature.

6. The signal modulation recognition method based on deep learning according to claim 3, wherein The minimum energy function of the trained attention mechanism module has the following expression: Among them, λ represents the regularization parameter, t represents the target neuron of a single channel in the input features, M represents the number of energy functions, q represents the serial number of the target neuron, and xq represents the estimated value of the target neuron.

7. The method for signal modulation recognition based on deep learning according to claim 3, wherein The trained first channel adjustment module, the trained second channel adjustment module, and the trained upsampling module constitute a feature pyramid multi-scale fusion module.

8. The method for signal modulation recognition based on deep learning according to claim 1, wherein The training process of the trained deep learning network includes: Obtain data of multiple preset categories. The data of the preset categories are modulation signals, and the modulation methods of the modulation signals include analog modulation and digital modulation; Denoise the data of multiple preset categories, perform short-time Fourier transform on the denoised data of multiple preset categories to obtain multiple grayscale time-frequency map data samples. Perform normalization processing on the multiple grayscale time-frequency map data samples, and use a color mapping function to convert the normalized multiple grayscale time-frequency map data samples into multiple color time-frequency map data samples; Divide the multiple color time-frequency map data samples into a training set, a validation set, and a test set according to a preset ratio, and label the samples in the training set, the validation set, and the test set to obtain the true labels of the samples in the training set, the validation set, and the test set; Input some samples in the training set into the j-th deep learning network to be trained for training to obtain the prediction results output during the j-th training process; Calculate the loss based on the predicted results output during the j-th training process and the true labels of the samples for training the deep learning network to be trained in the j-th time, and use it as the loss of the j-th training process; Perform backpropagation based on the loss of the j-th training process to update the network parameters of the deep learning network to be trained in the j-th time, and obtain the deep learning network to be trained in the (j + 1)-th time; at the same time, every time the preset verification interval value is reached, input the samples in the verification set into the deep learning network to be trained, obtain the predicted results output, and combine the true labels of the samples in the verification set to perform model selection and hyperparameter tuning; iterate in this way until the number of training times or the degree of convergence meets the preset conditions, and obtain the trained deep learning network; Input the samples in the test set into the trained deep learning network, obtain the predicted results output, and calculate the recognition accuracy according to the predicted results and the true labels of the samples in the test set to judge the performance of the trained deep learning network.

9. The signal modulation recognition method based on deep learning according to claim 8, characterized in that, The loss of the j-th training process is the cross-entropy loss function.

10. A signal modulation recognition device based on deep learning, characterized in that, It includes: A data acquisition module for acquiring the modulation signal to be recognized; A data processing module for removing noise from the modulation signal to be recognized by using a Gaussian filter, performing short-time Fourier transform on the noise-removed modulation signal to be recognized to obtain a gray time-frequency diagram; performing normalization processing on the gray time-frequency diagram, and converting the normalized gray time-frequency diagram into a color time-frequency diagram by using a color mapping function; A data recognition module for processing the color time-frequency diagram by using the trained deep learning network to obtain the classification result of the modulation signal to be recognized; Among them, the trained deep learning network is obtained by training the initial deep learning network with data of a preset category as the training data set.

Images