Modulation mode identification method, device, medium and apparatus of signal

Abstract

The application discloses a modulation mode identification method and device of a signal, a medium and equipment, and belongs to the technical field of signal processing. The method comprises the following steps: identifying whether a wireless signal is a low signal-to-noise ratio signal by using a low signal-to-noise ratio signal identification module in a modulation mode identification model; if yes, performing Markov conversion on the low signal-to-noise ratio signal by using a Markov conversion module to obtain a heat map; processing the heat map by using a threshold threshold mechanism and a coordination attention mechanism in a noise reduction autoencoder module to obtain time domain characteristic information; and performing feature extraction and classification processing on the time domain characteristic information by using a feature extraction and classification module to obtain the modulation mode of the wireless signal. The application can identify the modulation mode after noise reduction processing of the low signal-to-noise ratio signal, and the accuracy is improved.

Description

Technical Field

[0001] The present application relates to the field of signal processing technology, and in particular to a method, device, medium and equipment for identifying a modulation mode of a signal. Background Art

[0002] Automatic Modulation Recognition (AMR) plays a vital role in wireless communications. It is a technology that can identify a signal's modulation type without knowing its source or modulation information. AMR helps avoid radio interference and optimize spectrum allocation, and is widely used in military and civilian applications such as signal authentication, interference detection, communication countermeasures, spectrum monitoring, and information security.

[0003] Modulated signals can include high signal-to-noise ratio signals and low signal-to-noise ratio signals. The noise of high signal-to-noise ratio signals is small, and their modulation mode can be identified more accurately; while the noise of low signal-to-noise ratio signals is more serious. Accurately identifying the modulation mode of low signal-to-noise ratio signals becomes a challenging research. Summary of the Invention

[0004] This application provides a method, apparatus, medium, and device for identifying the modulation mode of a signal, which is used to solve the problem that low signal-to-noise ratio signals have severe noise and are difficult to identify their modulation mode. The technical solution is as follows:

[0005] According to a first aspect of the present application, a method for identifying a modulation mode of a signal is provided, the method comprising:

[0006] Obtaining a wireless signal to be identified and a trained modulation mode recognition model, wherein the wireless signal is a low signal-to-noise ratio signal or a high signal-to-noise ratio signal, and the modulation mode recognition model includes a low signal-to-noise ratio signal recognition module, a Markov transformation module, a denoising autoencoder module, and a feature extraction and classification module;

[0007] Using the low signal-to-noise ratio signal identification module to identify whether the wireless signal is a low signal-to-noise ratio signal;

[0008] If the wireless signal is a low signal-to-noise ratio signal, performing a Markov transformation on the low signal-to-noise ratio signal using the Markov transformation module to obtain a heat map;

[0009] Processing the heat map using the threshold mechanism and the coordinated attention mechanism in the denoising autoencoder module to obtain time domain feature information;

[0010] The feature extraction and classification module is used to perform feature extraction and classification processing on the time domain feature information to obtain the modulation mode of the wireless signal.

[0011] In one possible implementation, the low signal-to-noise ratio signal identification module includes a first convolutional layer, a second convolutional layer, and a fully connected layer. Using the low signal-to-noise ratio signal identification module to identify whether the wireless signal is a low signal-to-noise ratio signal includes:

[0012] Performing convolution processing on the wireless signal using the first convolution layer to obtain first convolution feature information; performing convolution processing on the first convolution feature information using the second convolution layer to obtain second convolution feature information; and processing the second convolution feature information using the fully connected layer to obtain a classification result of the wireless signal, where the classification result is a low signal-to-noise ratio signal or a high signal-to-noise ratio signal;

[0013] Determine whether the wireless signal is a low signal-to-noise ratio signal according to the classification result.

[0014] In a possible implementation, performing Markov transformation on the low signal-to-noise ratio signal using the Markov transformation module to obtain a heat map includes:

[0015] Dividing Q discrete distribution quantile units according to the distribution of the low signal-to-noise ratio signal;

[0016] Splitting the real component and the imaginary component in the low signal-to-noise ratio signal, and mapping the low signal-to-noise ratio signal to the Q quantile units according to the real component and the imaginary component;

[0017] Generate a Q×Q weighted adjacency matrix for each low SNR signal through migration probability;

[0018] The weighted adjacency matrix is ​​Gaussian blurred using a Gaussian blur algorithm to obtain a heat map.

[0019] In one possible implementation, the denoising autoencoder module includes an encoder, a denoising unit, and a decoder. The process of processing the heat map using the threshold mechanism and the coordinated attention mechanism in the denoising autoencoder module to obtain time domain feature information includes:

[0020] Encoding the heat map using the encoder to obtain an encoding result;

[0021] Performing denoising on the encoding result using the coordinated attention mechanism in the denoising unit to obtain denoising features;

[0022] Decoding the noise reduction feature using the decoder to obtain a decoding result;

[0023] The heat map and the decoding result are multiplied based on a threshold mechanism to obtain time domain feature information.

[0024] In a possible implementation, the feature extraction and classification module comprises a feature extraction network and a feature classification network, the feature extraction and classification of the time domain feature information by using the feature extraction and classification module comprises:

[0025] extracting features of the time domain feature information by using the feature extraction network to obtain intermediate feature information;

[0026] classifying the intermediate feature information by using the feature classification network to obtain the modulation mode of the wireless signal.

[0027] In a possible implementation, the feature extraction network comprises a third convolutional layer, a fourth convolutional layer, a first pooling layer, a second pooling layer, a third pooling layer, and three convolutional blocks, each convolutional block comprising a convolutional unit, a batch normalization unit, and an activation function, the feature extraction of the time domain feature information by using the feature extraction network to obtain intermediate feature information comprises:

[0028] performing convolutional processing on the time domain feature information by using the third convolutional layer to obtain third convolutional feature information;

[0029] performing pooling processing on the third convolutional feature information by using the first pooling layer to obtain first pooling feature information;

[0030] performing convolutional processing on the first pooling feature information by using the fourth convolutional layer to obtain fourth convolutional feature information;

[0031] performing pooling processing on the fourth convolutional feature information by using the second pooling layer to obtain second pooling feature information;

[0032] processing the second pooling feature information by using the three convolutional blocks in sequence to obtain fifth convolutional feature information;

[0033] performing pooling processing on the fifth convolutional feature information by using the third pooling layer to obtain intermediate feature information.

[0034] In a possible implementation, the method further comprises:

[0035] obtaining a training data set, each training sample in the training data set comprising a wireless signal and a label, the wireless signal being a low signal-to-noise ratio signal or a high signal-to-noise ratio signal, and the label being used to mark an actual modulation mode of the wireless signal;

[0036] creating a modulation mode identification model, the modulation mode identification model comprising a low signal-to-noise ratio signal identification module, a Markov conversion module, a noise reduction autoencoder module, and a feature extraction and classification module;

[0037] Using the low signal-to-noise ratio signal identification module to screen training samples with low signal-to-noise ratio signals from the training data set;

[0038] For each screened training sample, the Markov transformation module is used to perform a Markov transformation on the low signal-to-noise ratio signal to obtain a heat map; the threshold mechanism and the coordinated attention mechanism in the denoising autoencoder module are used to process the heat map to obtain time domain feature information; the feature extraction and classification module is used to perform feature extraction and classification processing on the time domain feature information to obtain a predicted modulation mode;

[0039] The loss function of the modulation mode recognition model is used to calculate the loss of the predicted modulation mode and the actual modulation mode, and the modulation mode recognition model is trained according to the calculation result.

[0040] According to a second aspect of the present application, a device for identifying a modulation mode of a signal is provided, the device comprising:

[0041] An acquisition module, configured to acquire a wireless signal to be identified and a trained modulation mode recognition model, wherein the wireless signal is a low signal-to-noise ratio signal or a high signal-to-noise ratio signal, and the modulation mode recognition model includes a low signal-to-noise ratio signal recognition module, a Markov transition module, a denoising autoencoder module, and a feature extraction and classification module;

[0042] an identification module, configured to identify whether the wireless signal is a low signal-to-noise ratio signal using the low signal-to-noise ratio signal identification module;

[0043] The identification module is further configured to, if the wireless signal is a low signal-to-noise ratio signal, perform a Markov transformation on the low signal-to-noise ratio signal using the Markov transformation module to obtain a heat map;

[0044] The recognition module is further configured to process the heat map using the threshold mechanism and the coordinated attention mechanism in the denoising autoencoder module to obtain time domain feature information;

[0045] The identification module is further configured to utilize the feature extraction and classification module to perform feature extraction and classification processing on the time domain feature information to obtain the modulation mode of the wireless signal.

[0046] According to a third aspect of the present application, a computer-readable storage medium is provided, wherein the storage medium stores at least one instruction, and the at least one instruction is loaded and executed by a processor to implement the signal modulation mode identification method as described above.

[0047] According to a fourth aspect of the present application, a computer device is provided, comprising the above-mentioned signal modulation mode identification device.

[0048] The beneficial effects of the technical solution provided by this application include at least:

[0049] Customized optimization is performed based on the signal characteristics under different signal-to-noise ratio conditions. For low signal-to-noise ratio signals, the low signal-to-noise ratio signals are first enhanced using the data enhancement optimization method of Markov transform to enhance the robustness of the modulation recognition model for low signal-to-noise ratio signals; then, the data is denoised by the denoising autoencoder based on the threshold gate mechanism and the coordinated attention mechanism, which effectively reduces the interference of noise on the recognition results. Finally, the obtained time domain feature information is extracted and classified, thereby improving the accuracy of modulation recognition of low signal-to-noise ratio signals.

[0050] By adopting a network structure based on denoising autoencoder and an asymmetric convolutional neural network structure to construct a modulation recognition model, the modulation recognition model can use the lowest possible number of parameters while fully completing feature extraction.

[0051] By introducing a coordinated attention mechanism, feature information is multiplied and weighted in different dimensions to complete the recalibration of the original features. The obtained feature information is more advantageous than the original image. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0053] Figure 1 This is a flowchart of a method for training a signal modulation mode recognition model provided by one embodiment of the present application;

[0054] Figure 2 This is a schematic diagram of the structure of a low signal-to-noise ratio signal recognition module provided by an embodiment of the present application;

[0055] Figure 3 This is a schematic diagram of the structure of a Markov transformation module provided by an embodiment of the present application;

[0056] Figure 4 This is a schematic diagram of the structure of a noise reduction autoencoder module provided by an embodiment of the present application;

[0057] Figure 5 This is a schematic diagram of the structure of a feature extraction network provided by an embodiment of the present application;

[0058] Figure 6 This is a flowchart of a method for identifying a modulation mode of a signal provided by an embodiment of the present application;

[0059] Figure 7 This is a structural block diagram of a signal modulation mode identification device provided by an embodiment of the present application;

[0060] Figure 8 This is a structural block diagram of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0061] In order to make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0062] The modulation mode recognition model includes a low signal-to-noise ratio signal recognition module, a Markov transformation module, a denoising autoencoder module, and a feature extraction and classification module. The low signal-to-noise ratio signal recognition module is used to classify wireless signals into high signal-to-noise ratio signals and low signal-to-noise ratio signals, and output the low signal-to-noise ratio signals to the Markov transformation module. The Markov transformation module is used to perform a Markov transformation on the low signal-to-noise ratio signal to obtain a heat map, and output the heat map to the denoising autoencoder module. The denoising autoencoder module is used to reduce the noise on the heat map to obtain time-domain feature information, and output the time-domain feature information to the feature extraction and classification module. The feature extraction and classification module is used to extract and classify the time-domain feature information to obtain the debugging method of the wireless signal. The structure and specific processing methods of each module are described in detail below and are not repeated here.

[0063] The following describes the training process of the modulation recognition model.

[0064] like Figure 1 As shown, it shows a method flow chart of a method for training a signal modulation mode recognition model provided by an embodiment of the present application. The training method of the signal modulation mode recognition model can be applied to a computer device. The training method of the signal modulation mode recognition model may include:

[0065] Step 101: Obtain a training data set. Each training sample in the training data set includes a wireless signal and a label. The wireless signal is a low signal-to-noise ratio signal or a high signal-to-noise ratio signal. The label is used to mark the actual modulation mode of the wireless signal.

[0066] In this example, data from the public dataset RML2016.10B is selected as wireless signals, and in-phase quadrature (IQ) data is generated based on each wireless signal. The dataset contains 10 standard modulation methods, including digital modulation and analog modulation. The specific types are as follows:

[0067] (1) Digital modulation: Phase Shift Keying (PSK), Quadrature Phase Shift Keying (QPSK), 8 Phase Shift Keying (8PSK), 16 Quadrature Amplitude Modulation (QAM), 64QAM, Gaussian Frequency Shift Keying (GFSK), Continuous-Phase Frequency Shift Keying (CPFSK), 4 Pulse Amplitude Modulation (PAM);

[0068] (2) Analog modulation: Wide-Band Frequency Modulation (WBFM) and Amplitude Modulation Double Sideband (AM-DSB).

[0069] Under 20 signal-to-noise ratio (SNR) conditions (ranging from -20dB to 18dB, with 2dB intervals), each modulation mode contains 128 dual-channel IQ data points at each SNR value. Then, eight adjacent groups of IQ data for the same modulation mode are merged, converting the original 100,000 IQ data points of 2*128 in length into 12,500 IQ data points of 2*1024 in length. Each converted IQ data point and label form a training sample. The label is used to annotate a modulation mode.

[0070] Step 102: Create a modulation mode recognition model.

[0071] The modulation mode recognition model includes a low signal-to-noise ratio signal recognition module, a Markov conversion module, a denoising autoencoder module and a feature extraction and classification module.

[0072] Step 103: Filter training samples with low signal-to-noise ratio signals from the training data set using a low signal-to-noise ratio signal identification module.

[0073] The low SNR signal recognition module includes two convolutional neural networks and one fully connected layer, primarily used to classify low SNR signals into two categories: high SNR signals (high SNR IQ data) and low SNR signals (low SNR IQ data). For high SNR signals, their modulation scheme is identified based on other techniques; for low SNR signals, further enhancement is performed before their modulation scheme is identified, i.e., step 104 is executed.

[0074] Specifically, the low signal-to-noise ratio signal recognition module includes a first convolutional layer, a second convolutional layer, and a fully connected layer. The low signal-to-noise ratio signal recognition module is used to screen training samples with low signal-to-noise ratio signals from the training data set, which can include: for each training sample, using the first convolutional layer to perform convolution processing on the wireless signal in the training sample to obtain first convolution feature information; using the second convolutional layer to perform convolution processing on the first convolution feature information to obtain second convolution feature information; using the fully connected layer to process the second convolution feature information to obtain a classification result of the wireless signal, where the classification result is a low signal-to-noise ratio signal or a high signal-to-noise ratio signal; and retaining each training sample whose classification result is a low signal-to-noise ratio signal.

[0075] like Figure 2 As shown in the figure, the input data is Input[2, 1024], the first convolutional feature information obtained after processing by the first convolutional layer is Conv[1024, 64, 3], the second convolutional feature information obtained after processing by the second convolutional layer is Conv[64, 32, 3], and the information obtained after processing by the fully connected layer is Fc[128, 10]. The final classification result is low SNR signal or high SNR signal. The low SNR signal recognition module uses convolutional neural networks (CNNs) to extract signal features. This method achieves the best classification accuracy at a SNR threshold of -4dB, with an accuracy rate of up to 97%.

[0076] In step 104, for each screened training sample, the Markov transformation module is used to perform Markov transformation on the low signal-to-noise ratio signal to obtain a heat map; the threshold mechanism and coordinated attention mechanism in the denoising autoencoder module are used to process the heat map to obtain time domain feature information; the feature extraction and classification module is used to extract and classify the time domain feature information to obtain a predicted modulation method.

[0077] The Markov transformation module uses the Markov matrix to transform the one-dimensional IQ data into a 64*64 matrix, and based on this, generates the corresponding heat map.

[0078] Specifically, using a Markov transformation module to perform Markov transformation on a low signal-to-noise ratio signal to obtain a heat map can include: dividing the low signal-to-noise ratio signal into Q discrete distribution quantile units according to the distribution of the low signal-to-noise ratio signal; splitting the real component and the imaginary component in the low signal-to-noise ratio signal, and mapping the low signal-to-noise ratio signal to Q quantile units according to the real component and the imaginary component; generating a Q×Q weighted adjacency matrix for each low signal-to-noise ratio signal through migration probability; and performing Gaussian fuzzy processing on the weighted adjacency matrix using a Gaussian fuzzy algorithm to obtain a heat map.

[0079] Since the low signal-to-noise ratio signal is a time domain signal sequence, Q discrete distribution quantile units can be divided according to the distribution of the sequence value. Then, the MarkovTransitionField function in the Pyts.image library 7 is used in Python to split the real and imaginary components of the low signal-to-noise ratio signal, and the low signal-to-noise ratio signal is mapped to Q quantile units according to the real and imaginary components. A weighted adjacency matrix M of size Q×Q is generated by the migration probability. The flow of time series data in different quantile intervals is revealed by counting the number of transitions in the quantile interval. The weighted adjacency matrix M can also eliminate the influence of data distribution and time step on time series dependence, has stronger robustness to data distribution and time series dependence, and can accurately describe the dynamic changes of data:

[0080]

[0081] Among them, W ij / x i Indicates q i and q j The transition probability of the corresponding quantile relationship on the weighted adjacency matrix.

[0082] However, the size of the heat map generated by Markov (that is, the weighted adjacency matrix M) is 256×256, which covers a lot of redundant information. Before applying the weighted adjacency matrix M to the deep learning model, the complexity and computational efficiency of the model need to be considered. To this end, in this embodiment, Gaussian blur technology is used to reduce the dimension of the weighted adjacency matrix M. The gaussian_filter function provided by the Python library scipy.ndimage is used. The filter center value sigma is 2.0 and the mode is constant. The original weighted adjacency matrix M dimension is reduced to 64×64 through the Gaussian blur algorithm. The total number of vectors is 4096, and the size of the heat map can be reshaped to 4×1024. Figure 3 shown.

[0083] The denoising autoencoder module processes signal data through unsupervised learning of the intrinsic structure and distribution of the input data. To reduce the computational complexity of the denoising autoencoder module, this embodiment introduces a thresholding mechanism and a coordinated attention (CA) mechanism to remove noise and unimportant features, thereby enhancing the signal's characteristics. Furthermore, the thresholding and CA mechanisms can further learn and extract temporal feature information.

[0084] The threshold mechanism is used to further filter noise in wireless signals. By adaptively learning and setting an appropriate threshold, this mechanism can suppress signal components below the threshold and reduce the impact of noise on signal quality.

[0085] The CA mechanism helps enhance the extraction of time-domain feature information across channels and space, while also accounting for long-range dependencies. By introducing the CA mechanism, the modulation recognition model can be made more lightweight while focusing on important feature information. The denoising autoencoder module can effectively reduce signal interference and noise, improving recognition accuracy. With a parameter count of approximately 13k, it has no impact on the model parameter magnitude. Its introduction not only maintains the low-complexity design of the overall network model, but also enhances signal robustness and anti-interference capabilities, providing an effective denoising and optimization solution for modulation recognition.

[0086] like Figure 4 As shown, the denoising autoencoder module includes an encoder, a denoising unit and a decoder. The threshold threshold mechanism and the coordinated attention mechanism in the denoising autoencoder module are used to process the heat map to obtain time domain feature information, which can include: encoding the heat map using the encoder to obtain an encoding result; performing denoising processing on the encoding result using the coordinated attention mechanism in the denoising unit to obtain denoising features; decoding the denoising features using the decoder to obtain a decoding result; and multiplying the heat map and the decoding result based on the threshold threshold mechanism to obtain time domain feature information.

[0087] The feature extraction and classification module extracts the time domain feature information through three groups of convolution blocks with different kernels. Finally, the fully connected layer maps the outputs of multiple neurons to probabilities in [0, 1] through the SoftMax activation function to achieve classification of various modulation modes.

[0088] Specifically, the feature extraction network includes a third convolutional layer, a fourth convolutional layer, a first pooling layer, a second pooling layer, a third pooling layer and three convolutional blocks. Each convolutional block includes a convolution unit, a batch normalization unit and an activation function. The feature extraction network is used to extract features from the time domain feature information to obtain intermediate feature information, which may include: using the third convolutional layer to convolve the time domain feature information to obtain third convolutional feature information; using the first pooling layer to pool the third convolutional feature information to obtain first pooled feature information; using the fourth convolutional layer to convolve the first pooled feature information to obtain fourth convolutional feature information; using the second pooling layer to pool the fourth convolutional feature information to obtain second pooled feature information; using the three convolutional blocks in sequence to process the second pooled feature information to obtain fifth convolutional feature information; using the third pooling layer to pool the fifth convolutional feature information to obtain intermediate feature information.

[0089] like Figure 5As shown in the figure, the feature extraction network consists of 14 convolutional layers and 3 pooling layers. The first convolutional layer, Conv1 (also known as the third convolutional layer), processes temporal feature information with an input dimension of 4×1024. The convolution kernel size is 3*3, and the padding size is (3, 1). To optimize signal feature extraction, the first pooling layer, Maxpool1 (also known as the first pooling layer), reduces the size of the feature map. The second convolutional layer, Conv2 (also known as the fourth convolutional layer), is the same as the first convolutional layer, Conv1, and the second pooling layer, Maxpool2 (also known as the second pooling layer), is the same as the first pooling layer, Maxpool1. Subsequently, the output feature map passes through the third convolutional block Conv-Block1, Conv-Block2 and Conv-Block3 in sequence. Each convolutional block Conv-Block consists of 4 convolution units, 4 batch normalization units (Batchnorm2d) and the activation function ReLU. Among them, the first and last convolution units in each convolutional block Conv-Block use a 1×1 kernel matrix to reduce the dimension of the feature map channel, which affects the computational complexity; the second and third convolution units use asymmetric kernel matrices with kernel sizes of 3×1 and 1×3, respectively, instead of a 3×3 kernel matrix, to reduce the number of trainable parameters. The third pooling layer Maxpool3 (that is, the third pooling layer) uses a 3*16 convolution kernel to match the subsequent feature fusion operation.

[0090] The obtained intermediate feature information is input into a fully connected (FC) layer with 128 units, a ReLU activation layer, and a dropout layer that randomly discards 70% of the neurons. Finally, the recognition result is output through a fully connected layer with softmax activation. The dropout layer is used to prevent overfitting.

[0091] The ReLU activation function is:

[0092]

[0093] Where: z[i] represents the feature map value.

[0094] Step 105 , performing loss calculation on the predicted modulation mode and the actual modulation mode according to the loss function of the modulation mode recognition model, and training the modulation mode recognition model according to the calculation result.

[0095] The modulation recognition model combines the cross-entropy loss function with the SoftMax (regression) function classifier to more accurately measure the difference between the predicted modulation mode and the actual modulation mode, thereby reconstructing more accurate signal values ​​and improving the accuracy of model training.

[0096] The cross entropy loss function is calculated as follows: for a set of data points, we calculate the average information entropy between the probability distribution predicted by the model and the actual label distribution. Its mathematical expression is as follows:

[0097]

[0098] Where m is the number of labels, t[i] is the actual label, and p[i] is the predicted probability of the label, which is calculated as:

[0099]

[0100] Among them, z[i] is the feature map after Relu activation, and s is the number of categories of the classification task.

[0101] During the training phase, based on the loss calculation results, the adaptive moment estimation (Adam) optimization algorithm and backpropagation mechanism are used to update the values ​​of the trainable parameters in the modulation recognition model. Multiple iterations are performed and the learning rate is adjusted until the modulation recognition model reaches stability, and the model parameters are saved. These model parameters are then reloaded into the modulation recognition model. The trained modulation recognition model is then used to identify modulation modes on a test dataset. The loss and accuracy values ​​for channels ranging from -10dB to 20dB are then verified against a validation dataset. This allows the model performance to be evaluated and the model parameters to be adjusted until the modulation recognition model reaches stability.

[0102] In summary, the training method of the signal modulation mode recognition model provided by the embodiment of the present application distinguishes high and low signal-to-noise ratio signals through a low signal-to-noise ratio signal recognition module, thereby realizing segmented training. Customized optimization is performed based on the signal characteristics under different signal-to-noise ratio conditions. For low signal-to-noise ratio signals, the low signal-to-noise ratio signals are first enhanced using the data enhancement optimization method of Markov transform to enhance the robustness of the modulation mode recognition model for low signal-to-noise ratio signals; then, the data is denoised by a denoising autoencoder based on a threshold threshold mechanism and a coordinated attention mechanism, effectively reducing the interference of noise on the recognition results. Finally, the obtained time domain feature information is feature extracted and classified, thereby improving the accuracy of modulation mode recognition of low signal-to-noise ratio signals.

[0103] By adopting a network structure based on denoising autoencoder and an asymmetric convolutional neural network structure to construct a modulation recognition model, the modulation recognition model can use the lowest possible number of parameters while fully completing feature extraction.

[0104] By introducing a coordinated attention mechanism, feature information is multiplied and weighted in different dimensions to complete the recalibration of the original features. The obtained feature information is more advantageous than the original image.

[0105] like Figure 6FIG2 shows a flow chart of a method for identifying a modulation mode of a signal provided by an embodiment of the present application. The method for identifying a modulation mode of a signal can be applied to a computer device. The method for identifying a modulation mode of a signal can include:

[0106] Step 601: Obtain a wireless signal to be identified and a trained modulation mode recognition model. The wireless signal is a low signal-to-noise ratio signal or a high signal-to-noise ratio signal. The modulation mode recognition model includes a low signal-to-noise ratio signal recognition module, a Markov transformation module, a denoising autoencoder module, and a feature extraction and classification module.

[0107] Among them, the modulation mode recognition model is based on Figure 1 The model obtained by training with the training method shown in the figure has a detailed structure and data processing flow. Figure 2-4 .

[0108] Step 602: Use a low signal-to-noise ratio signal identification module to identify whether the wireless signal is a low signal-to-noise ratio signal.

[0109] The low SNR signal recognition module includes two convolutional neural networks and one fully connected layer. It primarily classifies low SNR signals into high and low SNR signals. For high SNR signals, the modulation scheme is identified based on other techniques. For low SNR signals, further enhancement is performed before the modulation scheme is identified, which is performed in step 504.

[0110] Specifically, the low signal-to-noise ratio signal identification module includes a first convolutional layer, a second convolutional layer, and a fully connected layer. Using the low signal-to-noise ratio signal identification module to identify whether the wireless signal is a low signal-to-noise ratio signal can include: using the first convolutional layer to convolve the wireless signal to obtain first convolution feature information; using the second convolutional layer to convolve the first convolution feature information to obtain second convolution feature information; using the fully connected layer to process the second convolution feature information to obtain a classification result of the wireless signal, the classification result being a low signal-to-noise ratio signal or a high signal-to-noise ratio signal; and determining whether the wireless signal is a low signal-to-noise ratio signal based on the classification result.

[0111] like Figure 2As shown in the figure, the input data is Input[2, 1024], the first convolutional feature information obtained after processing by the first convolutional layer is Conv[1024, 64, 3], the second convolutional feature information obtained after processing by the second convolutional layer is Conv[64, 32, 3], and the information obtained after processing by the fully connected layer is Fc[128, 10]. The final classification result is low SNR signal or high SNR signal. The low SNR signal recognition module uses convolutional neural networks (CNN) to extract signal features. This method achieves the best classification accuracy at a SNR threshold of -4dB, with an accuracy rate of up to 97%.

[0112] Step 603: If the wireless signal is a low signal-to-noise ratio signal, a Markov transformation module is used to perform Markov transformation on the low signal-to-noise ratio signal to obtain a heat map.

[0113] The Markov transformation module uses the Markov matrix to transform the one-dimensional IQ data into a 64*64 matrix, and based on this, generates the corresponding heat map.

[0114] Specifically, using a Markov transformation module to perform Markov transformation on a low signal-to-noise ratio signal to obtain a heat map can include: dividing the low signal-to-noise ratio signal into Q discrete distribution quantile units according to the distribution of the low signal-to-noise ratio signal; splitting the real component and the imaginary component in the low signal-to-noise ratio signal, and mapping the low signal-to-noise ratio signal to Q quantile units according to the real component and the imaginary component; generating a Q×Q weighted adjacency matrix for each low signal-to-noise ratio signal through migration probability; and performing Gaussian fuzzy processing on the weighted adjacency matrix using a Gaussian fuzzy algorithm to obtain a heat map.

[0115] Since the low signal-to-noise ratio signal is a time domain signal sequence, Q discrete distribution quantile units can be divided according to the distribution of the sequence value. Then, the MarkovTransitionField function in the Pyts.image library 7 is used in Python to split the real and imaginary components of the low signal-to-noise ratio signal, and the low signal-to-noise ratio signal is mapped to Q quantile units according to the real and imaginary components. A weighted adjacency matrix M of size Q×Q is generated by the migration probability. The flow of time series data in different quantile intervals is revealed by counting the number of transitions in the quantile interval. The weighted adjacency matrix M can also eliminate the influence of data distribution and time step on time series dependence, has stronger robustness to data distribution and time series dependence, and can accurately describe the dynamic changes of data:

[0116]

[0117] Among them, W ij / x iIndicates q i and q j The transition probability of the corresponding quantile relationship on the weighted adjacency matrix.

[0118] However, the size of the heat map generated by Markov (that is, the weighted adjacency matrix M) is 256×256, which covers a lot of redundant information. Before applying the weighted adjacency matrix M to the deep learning model, the complexity and computational efficiency of the model need to be considered. To this end, in this embodiment, Gaussian blur technology is used to reduce the dimension of the weighted adjacency matrix M. The gaussian_filter function provided by the Python library scipy.ndimage is used. The filter center value sigma is 2.0 and the mode is constant. The original weighted adjacency matrix M dimension is reduced to 64×64 through the Gaussian blur algorithm. The total number of vectors is 4096, and the size of the heat map can be reshaped to 4×1024. Figure 3 shown.

[0119] Step 604: Process the heat map using the threshold mechanism and coordinated attention mechanism in the denoising autoencoder module to obtain time domain feature information.

[0120] The denoising autoencoder module processes signal data through unsupervised learning of the intrinsic structure and distribution of the input data. To reduce the computational complexity of the denoising autoencoder module, this embodiment introduces a thresholding mechanism and a call-conversion (CA) mechanism to remove noise and unimportant features, thereby enhancing the signal's characteristics. Furthermore, the thresholding and CA mechanisms can further learn and extract temporal feature information.

[0121] The threshold mechanism is used to further filter noise in wireless signals. By adaptively learning and setting an appropriate threshold, this mechanism can suppress signal components below the threshold and reduce the impact of noise on signal quality.

[0122] The CA mechanism helps enhance the extraction of time-domain feature information across channels and space, while also accounting for long-range dependencies. By introducing the CA mechanism, the modulation recognition model can be made more lightweight while focusing on important feature information. The denoising autoencoder module can effectively reduce signal interference and noise, improving recognition accuracy. With a parameter count of approximately 13k, it has no impact on the model parameter magnitude. Its introduction not only maintains the low-complexity design of the overall network model, but also enhances signal robustness and anti-interference capabilities, providing an effective denoising and optimization solution for modulation recognition.

[0123] like Figure 4As shown, the denoising autoencoder module includes an encoder, a denoising unit and a decoder. The threshold threshold mechanism and the coordinated attention mechanism in the denoising autoencoder module are used to process the heat map to obtain time domain feature information, which can include: encoding the heat map using the encoder to obtain an encoding result; performing denoising processing on the encoding result using the coordinated attention mechanism in the denoising unit to obtain denoising features; decoding the denoising features using the decoder to obtain a decoding result; and multiplying the heat map and the decoding result based on the threshold threshold mechanism to obtain time domain feature information.

[0124] Step 605: When the feature extraction and classification module includes a feature extraction network and a feature classification network, the feature extraction network is used to extract features from the time domain feature information to obtain intermediate feature information; the feature classification network is used to classify the intermediate feature information to obtain the modulation mode of the wireless signal.

[0125] The feature extraction and classification module extracts the time domain feature information through three groups of convolution blocks with different kernels. Finally, the fully connected layer maps the outputs of multiple neurons to probabilities in [0, 1] through the SoftMax activation function to achieve classification of various modulation modes.

[0126] Specifically, the feature extraction network includes a third convolutional layer, a fourth convolutional layer, a first pooling layer, a second pooling layer, a third pooling layer and three convolutional blocks. Each convolutional block includes a convolution unit, a batch normalization unit and an activation function. The feature extraction network is used to extract features from the time domain feature information to obtain intermediate feature information, which may include: using the third convolutional layer to convolve the time domain feature information to obtain third convolutional feature information; using the first pooling layer to pool the third convolutional feature information to obtain first pooled feature information; using the fourth convolutional layer to convolve the first pooled feature information to obtain fourth convolutional feature information; using the second pooling layer to pool the fourth convolutional feature information to obtain second pooled feature information; using the three convolutional blocks in sequence to process the second pooled feature information to obtain fifth convolutional feature information; using the third pooling layer to pool the fifth convolutional feature information to obtain intermediate feature information.

[0127] like Figure 5As shown, the feature extraction network is composed of 14 convolutional layers and 3 pooling layers. The first convolutional layer Conv1 (i.e., the third convolutional layer) processes time-domain feature information with an input dimension of 4*1024, a convolution kernel size of 3*3, and a padding size of (3, 1). To optimize the extraction of signal features, the first pooling layer Maxpool1 (i.e., the first pooling layer) reduces the size of the feature map, and the second convolutional layer Conv2 (i.e., the fourth convolutional layer) is the same as the first convolutional layer Conv1. The second pooling layer Maxpool2 (i.e., the second pooling layer) is the same as the first pooling layer Maxpool1. Subsequently, the output feature map sequentially passes through the third convolutional block Conv-Block1, Conv-Block2, and Conv-Block3. Each convolutional block Conv-Block is composed of four convolutional units, four batch normalization units (Batchnorm2d), and an activation function ReLU. Among them, the first and last convolutional units in each convolutional block Conv-Block use a 1*1 kernel matrix to reduce the dimension of the feature map channel, which affects the complexity of the calculation. The second and third convolutional units use asymmetric kernel matrices with kernel sizes of 3*1 and 1*3, respectively, instead of a 3*3 kernel matrix, to reduce the number of trainable parameters. The third pooling layer Maxpool3 (i.e., the third pooling layer) uses a 3*16 convolution kernel to match the subsequent feature fusion operation.

[0128] The obtained intermediate feature information is input into a fully connected (FC) layer with 128 units, a ReLU activation layer, and a dropout layer that randomly discards 70% of the neurons. Finally, a fully connected layer with a softmax activation outputs the recognition result, where the dropout layer is used to prevent overfitting.

[0129] The ReLU activation function is:

[0130]

[0131] where z[i] represents the feature mapping value.

[0132] In summary, the signal modulation mode recognition method provided by the embodiments of the present application is customized and optimized for signal characteristics under different signal-to-noise ratio conditions. For low signal-to-noise ratio signals, the data enhancement optimization method of Markov transformation is used to enhance the low signal-to-noise ratio signals, to enhance the robustness of the modulation mode recognition model to low signal-to-noise ratio signals. Then, the denoising autoencoder based on the threshold threshold mechanism and the coordination attention mechanism is used to denoise the data, effectively reducing the interference of noise on the recognition result. Finally, the time-domain feature information obtained is subjected to feature extraction and classification, thereby improving the accuracy of the modulation mode recognition of low signal-to-noise ratio signals.

[0133] The modulation mode recognition model is constructed by adopting a network structure based on a noise-reducing auto-encoder and a neural network structure of asymmetric convolution, so that the modulation mode recognition model can use as few parameters as possible while sufficiently completing feature extraction.

[0134] By introducing a coordinated attention mechanism, the feature information is multiplied and weighted in different dimensions to complete the recalibration of the original features, and the obtained feature information is more advantageous than the original image.

[0135] As shown in Figure 7 The signal modulation mode recognition device provided by the embodiment of the present application is shown in a structural block diagram, and the signal modulation mode recognition device can be applied to a computer device. The signal modulation mode recognition device can include:

[0136] The acquisition module 710 is configured to acquire a wireless signal to be identified and a trained modulation mode recognition model. The wireless signal is a low signal-to-noise ratio signal or a high signal-to-noise ratio signal, and the modulation mode recognition model includes a low signal-to-noise ratio signal recognition module, a Markov conversion module, a noise-reducing auto-encoder module, and a feature extraction and classification module.

[0137] The recognition module 720 is configured to identify whether the wireless signal is a low signal-to-noise ratio signal by using the low signal-to-noise ratio signal recognition module.

[0138] The recognition module 720 is further configured to perform Markov conversion on the low signal-to-noise ratio signal by using the Markov conversion module to obtain a heat map if the wireless signal is a low signal-to-noise ratio signal.

[0139] The recognition module 720 is further configured to process the heat map by using a threshold mechanism and a coordinated attention mechanism in the noise-reducing auto-encoder module to obtain time-domain feature information.

[0140] The recognition module 720 is further configured to perform feature extraction and classification processing on the time-domain feature information by using the feature extraction and classification module to obtain the modulation mode of the wireless signal.

[0141] In an optional embodiment, the low signal-to-noise ratio signal recognition module includes a first convolutional layer, a second convolutional layer, and a fully connected layer, and the recognition module 720 is further configured to:

[0142] perform convolution processing on the wireless signal by using the first convolutional layer to obtain first convolutional feature information, perform convolution processing on the first convolutional feature information by using the second convolutional layer to obtain second convolutional feature information, and perform processing on the second convolutional feature information by using the fully connected layer to obtain a classification result of the wireless signal, the classification result being a low signal-to-noise ratio signal or a high signal-to-noise ratio signal.

[0143] The classification result is used to determine whether the wireless signal is a low signal-to-noise ratio signal.

[0144] In an optional embodiment, the identification module 720 is further configured to:

[0145] Divide Q discrete distribution quantile units according to the distribution of low signal-to-noise ratio signals;

[0146] Splitting the real component and the imaginary component in the low signal-to-noise ratio signal, and mapping the low signal-to-noise ratio signal to Q quantile units according to the real component and the imaginary component;

[0147] Generate a Q×Q weighted adjacency matrix for each low SNR signal through migration probability;

[0148] The weighted adjacency matrix is ​​Gaussian blurred using the Gaussian blur algorithm to obtain a heat map.

[0149] In an optional embodiment, the denoising autoencoder module includes an encoder, a denoising unit and a decoder, and the identification module 720 is further configured to:

[0150] Encode the heat map using an encoder to obtain an encoding result;

[0151] The denoising unit uses the coordinated attention mechanism to denoise the encoded result and obtain the denoising feature.

[0152] Decoding the noise reduction features using a decoder to obtain a decoding result;

[0153] The heat map and the decoding result are multiplied based on the threshold mechanism to obtain the time domain feature information.

[0154] In an optional embodiment, the feature extraction and classification module includes a feature extraction network and a feature classification network, and the recognition module 720 is further configured to:

[0155] Use the feature extraction network to extract the time domain feature information and obtain the intermediate feature information;

[0156] The feature classification network is used to classify the intermediate feature information and obtain the modulation mode of the wireless signal.

[0157] In an optional embodiment, the feature extraction network includes a third convolutional layer, a fourth convolutional layer, a first pooling layer, a second pooling layer, a third pooling layer, and three convolutional blocks, each of which includes a convolution unit, a batch normalization unit, and an activation function. The identification module 720 is further configured to:

[0158] Using the third convolutional layer to perform convolution processing on the time domain feature information to obtain third convolution feature information;

[0159] Using the first pooling layer to perform pooling processing on the third convolution feature information to obtain first pooled feature information;

[0160] Using the fourth convolutional layer to perform convolution processing on the first pooled feature information to obtain fourth convolution feature information;

[0161] Using the second pooling layer to perform pooling processing on the fourth convolution feature information to obtain second pooled feature information;

[0162] The second pooled feature information is processed in sequence using three convolution blocks to obtain the fifth convolution feature information;

[0163] The third pooling layer is used to perform pooling on the fifth convolution feature information to obtain intermediate feature information.

[0164] In an optional embodiment, the acquisition module 710 is further configured to acquire a training data set, where each training sample in the training data set includes a wireless signal and a label, where the wireless signal is a low signal-to-noise ratio signal or a high signal-to-noise ratio signal, and the label is used to mark an actual modulation mode of the wireless signal;

[0165] Creating a module for creating a modulation recognition model, the modulation recognition model includes a low signal-to-noise ratio signal recognition module, a Markov conversion module, a noise reduction autoencoder module, and a feature extraction and classification module;

[0166] The identification module 720 is further configured to filter training samples having low signal-to-noise ratio signals from the training data set using the low signal-to-noise ratio signal identification module;

[0167] The recognition module 720 is further configured to perform a Markov transformation on the low signal-to-noise ratio signal using the Markov transformation module for each screened training sample to obtain a heat map; process the heat map using the threshold mechanism and the coordinated attention mechanism in the denoising autoencoder module to obtain time domain feature information; and perform feature extraction and classification processing on the time domain feature information using the feature extraction and classification module to obtain a predicted modulation mode;

[0168] The training module is used to calculate the loss of the predicted modulation mode and the actual modulation mode according to the loss function of the modulation mode recognition model, and train the modulation mode recognition model according to the calculation results.

[0169] In summary, the modulation mode identification device for a signal provided in an embodiment of the present application is customized and optimized according to the signal characteristics under different signal-to-noise ratio conditions. For low signal-to-noise ratio signals, the low signal-to-noise ratio signal is first enhanced using the data enhancement optimization method of Markov transform to enhance the robustness of the modulation mode identification model for low signal-to-noise ratio signals; then, the data is denoised by a denoising autoencoder based on a threshold threshold mechanism and a coordinated attention mechanism, which effectively reduces the interference of noise on the recognition results; finally, the obtained time domain feature information is feature extracted and classified, thereby improving the accuracy of modulation mode recognition of low signal-to-noise ratio signals.

[0170] By adopting a network structure based on denoising autoencoder and an asymmetric convolutional neural network structure to construct a modulation recognition model, the modulation recognition model can use the lowest possible number of parameters while fully completing feature extraction.

[0171] By introducing a coordinated attention mechanism, feature information is multiplied and weighted in different dimensions to complete the recalibration of the original features. The obtained feature information is more advantageous than the original image.

[0172] like Figure 8 , which shows a schematic structural diagram of a computer device 800 suitable for implementing an embodiment of the present invention. Figure 8 The computer device 800 shown is only an example and should not limit the functions and scope of use of the embodiments of the present invention.

[0173] like Figure 7 As shown, the computer device 800 includes a central processing unit (CPU) 801, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 802 or a program loaded from a storage portion 808 into a random access memory (RAM) 803. Various programs and data required for the operation of the computer device 800 are also stored in the RAM 803. The CPU 801, ROM 802, and RAM 803 are connected to each other via a bus 804. An input / output (I / O) interface is also connected to the bus 804.

[0174] The following components are connected to the I / O interface 805: an input section 806 including a keyboard, a mouse, and the like; an output section 807 including devices such as a cathode ray tube (CRT), a liquid crystal display (LCD), and a speaker; a storage section 808 including a hard disk; and a communication section 809 including a network interface card such as a LAN card or a modem. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to the I / O interface 805 as needed. A removable medium 811, such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory, is installed in the drive 810 as needed, so that computer programs read therefrom can be installed into the storage section 808 as needed.

[0175] In particular, according to the embodiments disclosed in the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, the embodiments disclosed in the present invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program comprising program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 809, and / or installed from a removable medium 811. When the computer program is executed by the central processing unit (CPU) 801, the above-mentioned functions defined in the system of the present invention are performed.

[0176] It should be noted that the computer-readable medium described in the present invention can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media can include, but are not limited to, an electrical connection having one or more conductors, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In the present invention, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. This propagated data signal can take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transport a program for use by or in conjunction with an instruction execution system, apparatus, or device. Program code embodied on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wireline, optical fiber cable, RF, or any suitable combination thereof.

[0177] The flowcharts and block diagrams in the accompanying drawings illustrate the possible implementation architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, program segment, or a part of code, and the above-mentioned module, program segment, or a part of code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flowchart, and the combination of boxes in the block diagram or flowchart, can be implemented with a dedicated hardware-based system that performs the specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0178] The modules involved in the embodiments of the present invention may be implemented in software or hardware. The modules described may also be provided in a processor. For example, they may be described as follows: a processor includes a sending module, an acquisition module, a determination module, and a first processing module. The names of these modules do not, in some cases, limit the modules themselves. For example, the sending module may also be described as a "module for sending a picture acquisition request to the connected server."

[0179] It should be noted that the signal modulation mode identification device provided in the above embodiment only uses the division of the above functional modules as an example to illustrate when performing signal modulation mode identification. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the signal modulation mode identification device can be divided into different functional modules to complete all or part of the functions described above. In addition, the signal modulation mode identification device provided in the above embodiment and the signal modulation mode identification method embodiment are of the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0180] Those skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware, or by a program to instruct the relevant hardware, and the program may be stored in a computer-readable storage medium, which may be a read-only memory, a disk, or an optical disk, etc.

[0181] The above description is not intended to limit the embodiments of the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the embodiments of the present application should be included in the scope of protection of the embodiments of the present application.

Claims

1. A method for identifying a modulation mode of a signal, characterized in that: The method comprises: Obtaining a wireless signal to be identified and a trained modulation mode recognition model, wherein the wireless signal is a low signal-to-noise ratio signal or a high signal-to-noise ratio signal, and the modulation mode recognition model includes a low signal-to-noise ratio signal recognition module, a Markov transformation module, a denoising autoencoder module, and a feature extraction and classification module; Using the low signal-to-noise ratio signal identification module to identify whether the wireless signal is a low signal-to-noise ratio signal; If the wireless signal is a low signal-to-noise ratio signal, performing a Markov transformation on the low signal-to-noise ratio signal using the Markov transformation module to obtain a heat map; Processing the heat map using the threshold mechanism and the coordinated attention mechanism in the denoising autoencoder module to obtain time domain feature information; Using the feature extraction and classification module to perform feature extraction and classification processing on the time domain feature information to obtain the modulation mode of the wireless signal; The denoising autoencoder module includes an encoder, a denoising unit and a decoder. The threshold threshold mechanism and the coordinated attention mechanism in the denoising autoencoder module are used to process the heat map to obtain time domain feature information, including: using the encoder to encode the heat map to obtain an encoding result; using the coordinated attention mechanism in the denoising unit to perform denoising on the encoding result to obtain denoising features; using the decoder to decode the denoising features to obtain a decoding result; and multiplying the heat map and the decoding result based on the threshold threshold mechanism to obtain time domain feature information.

2. The method for identifying the modulation mode of a signal according to claim 1, wherein: The low signal-to-noise ratio signal identification module includes a first convolutional layer, a second convolutional layer, and a fully connected layer. The low signal-to-noise ratio signal identification module is used to identify whether the wireless signal is a low signal-to-noise ratio signal, including: Performing convolution processing on the wireless signal using the first convolution layer to obtain first convolution feature information; performing convolution processing on the first convolution feature information using the second convolution layer to obtain second convolution feature information; and processing the second convolution feature information using the fully connected layer to obtain a classification result of the wireless signal, where the classification result is a low signal-to-noise ratio signal or a high signal-to-noise ratio signal; Determine whether the wireless signal is a low signal-to-noise ratio signal according to the classification result.

3. The method for identifying the modulation mode of a signal according to claim 1, wherein: The step of performing Markov transformation on the low signal-to-noise ratio signal using the Markov transformation module to obtain a heat map includes: Dividing Q discrete distribution quantile units according to the distribution of the low signal-to-noise ratio signal; Splitting the real component and the imaginary component in the low signal-to-noise ratio signal, and mapping the low signal-to-noise ratio signal to the Q quantile units according to the real component and the imaginary component; Generate a Q×Q weighted adjacency matrix for each low SNR signal through migration probability; The weighted adjacency matrix is ​​Gaussian blurred using a Gaussian blur algorithm to obtain a heat map.

4. The method for identifying the modulation mode of a signal according to claim 1, wherein: The feature extraction and classification module includes a feature extraction network and a feature classification network. The feature extraction and classification module is used to perform feature extraction and classification processing on the time domain feature information to obtain the modulation mode of the wireless signal, including: Using the feature extraction network to extract features from the time domain feature information to obtain intermediate feature information; The intermediate feature information is classified and processed using the feature classification network to obtain a modulation mode of the wireless signal.

5. The method for identifying the modulation mode of a signal according to claim 4, wherein: The feature extraction network includes a third convolutional layer, a fourth convolutional layer, a first pooling layer, a second pooling layer, a third pooling layer and three convolutional blocks, each convolutional block includes a convolution unit, a batch normalization unit and an activation function, and the feature extraction network is used to extract the time domain feature information to obtain intermediate feature information, including: Performing convolution processing on the time domain feature information using the third convolution layer to obtain third convolution feature information; Performing pooling processing on the third convolutional feature information using the first pooling layer to obtain first pooled feature information; Performing convolution processing on the first pooled feature information using the fourth convolution layer to obtain fourth convolution feature information; Performing pooling processing on the fourth convolutional feature information using the second pooling layer to obtain second pooled feature information; Sequentially processing the second pooled feature information using three convolution blocks to obtain fifth convolution feature information; The third pooling layer is used to perform pooling processing on the fifth convolutional feature information to obtain intermediate feature information.

6. The method for identifying the modulation mode of a signal according to any one of claims 1 to 5, characterized in that: The method further comprises: Acquire a training data set, where each training sample in the training data set includes a wireless signal and a label, the wireless signal being a low signal-to-noise ratio signal or a high signal-to-noise ratio signal, and the label being used to mark an actual modulation mode of the wireless signal; Creating a modulation recognition model, the modulation recognition model comprising a low signal-to-noise ratio signal recognition module, a Markov transformation module, a noise reduction autoencoder module, and a feature extraction and classification module; Using the low signal-to-noise ratio signal identification module to screen training samples with low signal-to-noise ratio signals from the training data set; For each screened training sample, the Markov transformation module is used to perform a Markov transformation on the low signal-to-noise ratio signal to obtain a heat map; the threshold mechanism and the coordinated attention mechanism in the denoising autoencoder module are used to process the heat map to obtain time domain feature information; the feature extraction and classification module is used to perform feature extraction and classification processing on the time domain feature information to obtain a predicted modulation mode; The loss function of the modulation mode recognition model is used to calculate the loss of the predicted modulation mode and the actual modulation mode, and the modulation mode recognition model is trained according to the calculation result.

7. A signal modulation mode identification device, characterized in that: The device comprises: An acquisition module, configured to acquire a wireless signal to be identified and a trained modulation mode recognition model, wherein the wireless signal is a low signal-to-noise ratio signal or a high signal-to-noise ratio signal, and the modulation mode recognition model includes a low signal-to-noise ratio signal recognition module, a Markov transition module, a denoising autoencoder module, and a feature extraction and classification module; an identification module, configured to identify whether the wireless signal is a low signal-to-noise ratio signal using the low signal-to-noise ratio signal identification module; The identification module is further configured to, if the wireless signal is a low signal-to-noise ratio signal, perform a Markov transformation on the low signal-to-noise ratio signal using the Markov transformation module to obtain a heat map; The recognition module is further configured to process the heat map using the threshold mechanism and the coordinated attention mechanism in the denoising autoencoder module to obtain time domain feature information; The identification module is further configured to perform feature extraction and classification processing on the time domain feature information using the feature extraction and classification module to obtain a modulation mode of the wireless signal; The denoising autoencoder module includes an encoder, a denoising unit and a decoder. The recognition module is further used to: encode the heat map using the encoder to obtain an encoding result; perform denoising processing on the encoding result using the coordinated attention mechanism in the denoising unit to obtain denoising features; decode the denoising features using the decoder to obtain a decoding result; and multiply the heat map and the decoding result based on a threshold mechanism to obtain time domain feature information.

8. A computer-readable storage medium, characterized in that The storage medium stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the signal modulation mode identification method according to any one of claims 1 to 6.

9. A computer device, characterized in that: The computer device includes: the signal modulation mode identification device according to claim 7.

Images