Modulation mode identification method used under negative signal-to-noise ratio condition

Abstract

The invention provides a modulation mode identification method used under a negative signal-to-noise ratio condition, and relates to the technical field of noise reduction technology, modulation mode identification and deep neural network. Modulation mode identification is realized through signal noise reduction and feature extraction. The method comprises the following steps: constructing a double-current time-frequency noise reduction network; constructing a multi-scale convolutional neural network; respectively training the double-flow time-frequency noise reduction network and the multi-scale convolutional neural network to obtain the double-flow time-frequency noise reduction network with the best noise reduction effect and the multi-scale convolutional neural network with the best recognition effect; and obtaining an input signal, and performing modulation mode identification on the input signal based on the double-current time-frequency noise reduction network with the best noise reduction effect and the multi-scale convolutional neural network with the best identification effect to obtain a modulation mode identification result of the input signal. The method can improve the signal noise reduction effect and the modulation mode identification precision in a complex signal environment, has high robustness and identification precision under the condition of a negative signal-to-noise ratio, and is suitable for a complex and changeable communication environment.

Description

Technical Field

[0001] The present invention relates to the fields of noise reduction technology, modulation mode recognition and deep neural network technology, and in particular to a modulation mode recognition method for negative signal-to-noise ratio conditions. Background Art

[0002] Modulation mode recognition is an intermediate process between signal detection and demodulation. Accurately determining the modulation mode used by a signal plays a crucial role in spectrum management, system optimization, and improving anti-interference capabilities. With the rapid development of communication technology, the complexity and variability of communication environments, and the continuous emergence of new modulation modes, conventional recognition methods and theories often struggle to adapt to practical needs and are unable to effectively identify communication signals. This makes modulation mode recognition under negative signal-to-noise ratio conditions even more challenging.

[0003] Deep learning methods, with their powerful feature learning capabilities, high flexibility, and ability to process complex data, have demonstrated revolutionary performance improvements in multiple fields, including communications. In the field of modulation recognition, deep learning can automatically extract high-level abstract features from raw signals. These features often better reflect the signal's essential properties than manually designed features, greatly improving recognition accuracy and robustness. Especially under negative signal-to-noise ratio conditions, deep learning models can learn effective strategies for distinguishing different modulation modes under varying noise levels through training with large amounts of data. This is of great significance for improving the adaptability and intelligence of communication systems.

[0004] Although existing deep learning-based modulation pattern recognition methods perform well in ideal scenarios, they still have the following key defects, including: insufficient environmental robustness. When faced with complex electromagnetic environments such as nonlinear distortion, multipath fading, and strong noise interference, the model's ability to extract time-frequency features is easily degraded, resulting in a significant decrease in recognition accuracy; excessive data dependence, with a high degree of reliance on large-scale, diverse, and accurately labeled training data sets, severely limiting the model's generalization ability; weak interpretability. Deep learning models (such as CNN and Transformer) usually run in a "black box" form, making it difficult to clearly reveal their specific mechanism for extracting modulation features, which is not conducive to model debugging and optimization. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned prior art and provide a method for modulation mode identification under negative signal-to-noise ratio conditions. The method realizes modulation mode identification by signal noise reduction and feature extraction, and comprises the following steps:

[0006] Construct a dual-stream time-frequency denoising network to denoise the input signal and obtain a denoised signal;

[0007] Construct a multi-scale convolutional neural network to identify the modulation mode of the noise reduction signal and obtain the modulation mode recognition result;

[0008] The dual-stream time-frequency denoising network and the multi-scale convolutional neural network are trained separately to obtain the dual-stream time-frequency denoising network with the best denoising effect and the multi-scale convolutional neural network with the best recognition effect;

[0009] The input signal is obtained, and the modulation mode of the input signal is identified based on the dual-stream time-frequency denoising network with the best denoising effect and the multi-scale convolutional neural network with the best recognition effect, to obtain the modulation mode recognition result of the input signal.

[0010] Furthermore, the input signal is a two-dimensional matrix containing time series data and frequency components.

[0011] Furthermore, the dual-stream time-frequency denoising network includes: a frequency division module, a temporal information extraction module and an adaptive attention mechanism SE module; the frequency division module includes an octave convolution operation OctConv, and the temporal information extraction module includes a gated recurrent unit GRU.

[0012] Furthermore, the dual-stream time-frequency denoising network denoises the input signal to obtain a denoised signal. The specific method is as follows:

[0013] S1 obtains the sample input signal X IQ ∈R i*m , where i is the IQ branch of the sample input signal, and m is the number of sampling points of the sample input signal;

[0014] S2 samples the input signal X IQ Input the dual-stream time-frequency denoising network and use the frequency division module to sample the input signal X IQ Decompose and input the sample into signal X IQ Decompose into high-frequency signal H and low-frequency signal L, perform feature extraction on high-frequency signal H and low-frequency signal L respectively, and obtain high-frequency feature h and low-frequency feature l;

[0015] S3 uses the time series information extraction module to extract time series features from the high-frequency feature h and the low-frequency feature l, respectively, to obtain the high-frequency time series feature h′ and the low-frequency time series feature l′;

[0016] S4 sends the high-frequency temporal feature h′ and the low-frequency temporal feature l′ to the adaptive attention mechanism module for feature recalibration, and obtains the recalibrated low-frequency temporal feature h″ and the recalibrated high-frequency temporal feature l″;

[0017] S5 upsamples the recalibrated low-frequency time series feature h″, and the upsampled low-frequency time series feature dimension h″ upsanmple The upsampled low-frequency time series feature h″ is the same as the recalibrated high-frequency time series feature l″ in dimension.upsanmple It is fused with the recalibrated high-frequency time series feature l″ to obtain the noise reduction signal F.

[0018] Furthermore, the dual-stream time-frequency denoising network is trained to obtain the dual-stream time-frequency denoising network with the best denoising effect. The specific method is as follows:

[0019] The mean square error (MSE) loss function is used to optimize the network parameters of the dual-stream time-frequency denoising network, and the dual-stream time-frequency denoising network model with the best denoising effect is obtained.

[0020] Furthermore, a multi-scale convolutional neural network is constructed, including: a hybrid attention mechanism module, a BiLSTM-CNN module and a fully connected layer. The BiLSTM-CNN module includes a bidirectional long short-term memory network layer BiLSTM and a convolutional neural network layer CNN.

[0021] Furthermore, a multi-scale convolutional neural network is used to identify the modulation mode of the noise reduction signal to obtain the modulation mode recognition result. The specific method is as follows:

[0022] Step 1: Preprocess the noise reduction signal F to obtain a preprocessed noise reduction signal R;

[0023] The pre-processing includes a dimension adjustment operation;

[0024] Step 2: Input the preprocessed denoised signal R into the hybrid attention mechanism module to perform channel attention extraction and spatial attention extraction to obtain the channel attention feature map A and the spatial attention feature map B;

[0025] Step 3: Input the preprocessed denoised signal R into the BiLSTM-CNN module, use the bidirectional long short-term memory network layer BiLSTM to extract time series features, and use the convolutional neural network layer CNN to extract time-frequency features, and obtain the time series features T and time-frequency features F′ respectively;

[0026] Step 4: Fuse the channel attention feature map A, spatial attention feature map B, temporal feature T and time-frequency feature F′ of different scales to obtain the fused feature G;

[0027] Step 5: Input the fusion feature G into the fully connected layer and use the cross entropy loss function for classification to obtain the modulation mode recognition result;

[0028] Cross Entropy Loss Function As shown in the following formula:

[0029]

[0030] Among them, N is the number of categories, n is the element index, X (n) is the true category of the nth element, is the category of the predicted n-th element.

[0031] Furthermore, the multi-scale convolutional neural network is trained to obtain the multi-scale convolutional neural network with the best recognition effect. The specific method is as follows:

[0032] The multi-scale convolutional neural network is optimized based on the stochastic gradient descent method to obtain a multi-scale convolutional neural network model with the best recognition effect.

[0033] Furthermore, the modulation mode of the input signal is identified based on the dual-stream time-frequency denoising network with the best denoising effect and the multi-scale convolutional neural network with the best recognition effect, and the modulation mode recognition result of the input signal is obtained, including:

[0034] Use the dual-stream time-frequency denoising network model with the best denoising effect to denoise the input signal Z to obtain the denoised signal Z′, and transmit the denoised signal Z′ to the multi-scale convolutional neural network with the best recognition effect;

[0035] The multi-scale convolutional neural network model with the best recognition effect is used to identify the modulation mode of the denoised signal Z′, and the modulation mode recognition result P of the input signal Z is obtained.

[0036] The beneficial effects of adopting the above technical solution are as follows: the modulation mode recognition method for negative signal-to-noise ratio conditions provided by the present invention is based on a dual-stream time-frequency denoising network, which performs signal denoising and feature extraction through a frequency division module, a timing information extraction module, and an adaptive attention mechanism module; based on a multi-scale convolutional neural network, it further extracts time-frequency features and performs modulation mode recognition through a hybrid attention mechanism module, a BiLSTM-CNN module, and a fully connected layer. The modulation mode recognition method for negative signal-to-noise ratio conditions provided by the present invention can improve the signal denoising effect and modulation mode recognition accuracy in complex signal environments, has strong robustness and recognition accuracy under negative signal-to-noise ratio conditions, and is suitable for complex and changing communication environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A diagram showing the dual-stream time-frequency noise reduction network structure provided by an embodiment of the present invention;

[0038] Figure 2 A diagram of the multi-scale convolutional neural network structure provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0039] The following embodiments of the present invention are described in further detail with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.

[0040] A modulation mode identification method for negative signal-to-noise ratio conditions in this embodiment includes a dual-stream time-frequency noise reduction network and a multi-scale convolutional neural network. This method achieves accurate modulation mode identification through signal noise reduction and feature extraction, and includes the following steps:

[0041] Construct a dual-stream time-frequency denoising network to denoise the input signal and obtain a denoised signal;

[0042] Construct a multi-scale convolutional neural network to identify the modulation mode of the noise reduction signal and obtain the modulation mode recognition result;

[0043] The dual-stream time-frequency denoising network and the multi-scale convolutional neural network are trained separately to obtain the dual-stream time-frequency denoising network with the best denoising effect and the multi-scale convolutional neural network with the best recognition effect;

[0044] The input signal is acquired and the modulation mode of the input signal is identified based on the dual-stream time-frequency denoising network with the best denoising effect and the multi-scale convolutional neural network with the best recognition effect, thereby obtaining the modulation mode recognition result of the input signal.

[0045] The input signal is a two-dimensional matrix containing time series data and frequency components;

[0046] In this embodiment, a dual-stream time-frequency noise reduction network is constructed, such as Figure 1 As shown, it includes: a frequency division module, a temporal information extraction module and an adaptive attention mechanism SE module; the frequency division module includes an octave convolution operation OctConv, and the temporal information extraction module includes a gated recurrent unit GRU;

[0047] The dual-stream time-frequency denoising network denoises the input signal to obtain a denoised signal. The specific method is as follows:

[0048] S1 obtains the sample input signal X IQ ∈R i*m , where i is the IQ branch of the sample input signal, and m is the number of sampling points of the sample input signal;

[0049] S2 samples the input signal X IQ Input the dual-stream time-frequency denoising network and use the frequency division module to sample the input signal X IQ Decompose and input the sample into signal X IQ Decompose into high-frequency signal H and low-frequency signal L, perform feature extraction on high-frequency signal H and low-frequency signal L respectively, and obtain high-frequency feature h and low-frequency feature l;

[0050] S3 uses the time series information extraction module to extract time series features from the high-frequency feature h and the low-frequency feature l, respectively, to obtain the high-frequency time series feature h′ and the low-frequency time series feature l′;

[0051] S4 sends the high-frequency temporal feature h′ and the low-frequency temporal feature l′ to the adaptive attention mechanism module for feature recalibration, and obtains the recalibrated low-frequency temporal feature h″ and the recalibrated high-frequency temporal feature l″;

[0052] S5 upsamples the recalibrated low-frequency time series feature h″, and the upsampled low-frequency time series feature dimension h″ upsanmple The upsampled low-frequency time series feature h″ is the same as the recalibrated high-frequency time series feature l″ in dimension. upsanmple It is fused with the recalibrated high-frequency time series feature l″ to obtain the noise reduction signal F;

[0053] The dual-stream time-frequency denoising network is trained to obtain the dual-stream time-frequency denoising network with the best denoising effect. The specific method is as follows:

[0054] The mean square error (MSE) loss function is used to optimize the network parameters of the dual-stream time-frequency denoising network, and the dual-stream time-frequency denoising network model with the best denoising effect is obtained.

[0055] In this embodiment, the frequency division module uses octave convolution to decompose the input signal into a high-frequency signal and a low-frequency signal, and uses the high-frequency channel and the low-frequency channel to extract features of the high-frequency signal and the low-frequency signal respectively to obtain high-frequency features and low-frequency features; the high-frequency channel focuses on the details of the signal, while the low-frequency channel focuses on the global structure of the signal. The high-frequency channel outputs signal detail information, and the low-frequency channel outputs signal global information;

[0056] The timing information extraction module uses the gated recurrent unit (GRU) to process the signal details output by the high-frequency channel and the global signal information output by the low-frequency channel in the frequency division module. The gated recurrent unit (GRU) processes timing data through a gating mechanism, learns the temporal dependency of the signal, selectively retains key information and discards redundant information, removes time-related noise in the signal, retains the key timing information of the signal, and outputs the timing information of each time point.

[0057] The adaptive attention mechanism module performs global average pooling on the output features of each channel to obtain the global description information of each channel. It then calculates the weight of each channel through a fully connected network, normalizes the weight of each channel through the Sigmoid activation function, and multiplies the normalized channel weight with each channel of the input feature map channel by channel to obtain the noise reduction signal.

[0058] In this embodiment, a multi-scale convolutional neural network is constructed, such as Figure 2 As shown, it includes a hybrid attention mechanism module, a BiLSTM-CNN module and a fully connected layer. The BiLSTM-CNN module includes a bidirectional long short-term memory network layer BiLSTM and a convolutional neural network layer CNN.

[0059] The multi-scale convolutional neural network recognizes the modulation mode of the noise reduction signal and obtains the modulation mode recognition result. The specific method is as follows:

[0060] Step 1: Preprocess the noise reduction signal F to obtain a preprocessed noise reduction signal R;

[0061] The preprocessing includes a dimensionality modulation operation;

[0062] Step 2: Input the preprocessed denoised signal R into the hybrid attention mechanism module to perform channel attention extraction and spatial attention extraction to obtain the channel attention feature map A and the spatial attention feature map B;

[0063] Step 3: Input the preprocessed denoised signal R into the BiLSTM-CNN module, use the bidirectional long short-term memory network layer BiLSTM to extract time series features, and use the convolutional neural network layer CNN to extract time-frequency features, and obtain the time series features T and time-frequency features F′ respectively;

[0064] Step 4: Fuse the channel attention feature map A, spatial attention feature map B, temporal feature T and time-frequency feature F′ of different scales to obtain the fused feature G;

[0065] Step 5: Input the fusion feature G into the fully connected neural network and use the cross entropy loss function for classification to obtain the modulation mode recognition result;

[0066] Cross Entropy Loss Function As shown in the following formula:

[0067]

[0068] Among them, N is the number of categories, n is the element index, X (n) is the true category of the nth element, is the category of the predicted nth element;

[0069] The multi-scale convolutional neural network is trained to obtain the multi-scale convolutional neural network with the best recognition effect. The specific method is as follows:

[0070] The multi-scale convolutional neural network is optimized based on the stochastic gradient descent method to obtain the multi-scale convolutional neural network model with the best recognition effect;

[0071] In this embodiment, the hybrid attention mechanism module CBAM enhances the network's ability to focus on important features by weighting the features of the channel dimension and the spatial dimension. In the channel attention part, global average pooling and global maximum pooling are first used to aggregate the input features to obtain the importance weight of each channel. Then, the dependencies between channels are learned through a small fully connected layer or convolutional layer to generate the weight of each channel. In the spatial attention part, the convolution operation is used to weight the spatial dimension of the input feature map to emphasize the key spatial areas in the signal. The weighted results of the channel attention and spatial attention are combined to form the final weighted feature representation. The hybrid attention mechanism module CBAM can ensure that the network focuses on the key information in the signal and enhance the accuracy of modulation mode recognition.

[0072] The BiLSTM-CNN module comprises a bidirectional long short-term memory (BiLSTM) network layer and a convolutional neural network (CNN) layer, which extracts signal time-frequency features from different dimensions (time and frequency domains). The BiLSTM layer learns the temporal dependencies of signals from preceding and following temporal information, capturing long-term dependencies and enhancing the ability to extract temporal features. The convolutional neural network (CNN) layer extracts local features from the time and frequency domains through convolution operations, fusing multi-scale information layer by layer to enhance the understanding of complex signals. The combination of the BiLSTM layer and the convolutional neural network (CNN) enables the network to effectively extract features in both the time and frequency domains, thereby enhancing the ability to recognize modulation modes.

[0073] The hybrid attention mechanism (CBAM) and BiLSTM-CNN modules extract the signal's time-frequency features. These features are input into a fully connected layer for classification, and the resulting signal modulation is identified. Based on the extracted features, the fully connected neural network accurately identifies the signal's modulation mode, demonstrating strong robustness and accuracy, especially under negative signal-to-noise ratio conditions.

[0074] The modulation mode of the input signal is identified based on the dual-stream time-frequency denoising network with the best denoising effect and the multi-scale convolutional neural network with the best recognition effect. The modulation mode recognition results of the input signal are obtained, including:

[0075] Use the dual-stream time-frequency denoising network model with the best denoising effect to denoise the input signal Z to obtain the denoised signal Z′, and transmit the denoised signal Z′ to the multi-scale convolutional neural network with the best recognition effect;

[0076] Use the multi-scale convolutional neural network model with the best recognition effect to identify the modulation mode of the denoised signal Z′, and obtain the modulation mode recognition result P of the input signal Z;

[0077] In this embodiment, a sample X in the simulated communication signal data set is collected. IQ∈R 2*1024 As the sample input signal, 2 represents the IQ split of the signal, and 1024 represents the number of signal sampling points. The dimensions of the high-frequency signal and low-frequency signal output by the frequency division module in the dual-stream time-frequency denoising network are 32×1024 and 32×512, respectively. The dimension of the denoised signal output by the dual-stream time-frequency denoising network is 2×1024, which is the same as the dimension of the sample input signal. In the multi-scale modulation mode recognition network, the denoised signal is further processed and the dimension of the denoised signal is adjusted to 64×1024. The dimension of the modulation mode recognition result output by the multi-scale modulation mode recognition network is 8, where the element of each dimension represents the probability or confidence that the corresponding sample belongs to a certain modulation mode.

[0078] During the training phase of the dual-stream time-frequency denoising network and the multi-scale modulation recognition network, Adam was used to optimize network parameters. The training batch size was 64, and the learning rate was initialized to 0.0001. The dual-stream time-frequency denoising network used the mean square error (MSE) as the loss function, while the multi-scale modulation recognition network used the cross-entropy loss function. All experiments were conducted on an NVIDIA RTX 4090 with the PyTorch toolbox installed. This method uses two different models to construct a joint network. The specific model parameters are shown in Table 1, which also describes the entire network process using IQ data input.

[0079] Table 1 Parameters of the dual-stream time-frequency denoising network model

[0080]

[0081] Table 2 Multi-scale modulation recognition network model parameters

[0082]

[0083] This embodiment provides a modulation mode recognition method for negative signal-to-noise ratio conditions. Based on the existing deep learning model, an innovative neural network architecture is designed, including a dual-stream time-frequency denoising network and a multi-scale convolutional neural network. This architecture combines the advantages of the attention mechanism, the bidirectional long short-term memory network layer BiLSTM, and the convolutional neural network layer CNN, aiming to enhance the model's ability to capture time series signal features, its robustness to noise interference, and its ability to distinguish complex modulation patterns. The introduction of the attention mechanism enables the model to focus on the most discriminative parts of the signal and ignore irrelevant or noisy information, thereby improving recognition accuracy and efficiency; the bidirectional long short-term memory network layer BiLSTM is used to process the signal's temporal dependence and effectively capture the dynamic characteristics of the modulated signal; and the convolutional neural network layer CNN is responsible for extracting the signal's local frequency and time domain features, providing a rich information basis for the final classification decision.

[0084] The dual-stream time-frequency noise reduction network and multi-scale modulation recognition network provided in this embodiment significantly improve modulation recognition performance under negative signal-to-noise ratio conditions. This not only increases recognition accuracy but also enhances the model's generalization capabilities, providing strong technical support for spectrum management, system optimization, and the formulation of anti-interference strategies in complex and changing communication environments. This research achievement not only promotes the development of modulation recognition technology but also paves a new path for the intelligent and adaptive design of future communication systems.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A modulation mode identification method for use under negative signal-to-noise ratio conditions, characterized by: The following steps are involved: Construct a dual-stream time-frequency denoising network to denoise the input signal and obtain a denoised signal; Construct a multi-scale convolutional neural network to identify the modulation mode of the noise reduction signal and obtain the modulation mode recognition result; The dual-stream time-frequency denoising network and the multi-scale convolutional neural network are trained separately to obtain the dual-stream time-frequency denoising network with the best denoising effect and the multi-scale convolutional neural network with the best recognition effect; The input signal is obtained, and the modulation mode of the input signal is identified based on the dual-stream time-frequency denoising network with the best denoising effect and the multi-scale convolutional neural network with the best recognition effect, to obtain the modulation mode recognition result of the input signal.

2. The modulation mode identification method for negative signal-to-noise ratio conditions according to claim 1, characterized in that: The input signal is a two-dimensional matrix containing time series data and frequency components.

3. The modulation mode identification method for negative signal-to-noise ratio conditions according to claim 1, characterized in that: A dual-stream time-frequency denoising network is constructed, including: a frequency division module, a temporal information extraction module and an adaptive attention mechanism SE module; the frequency division module includes an octave convolution operation OctConv, and the temporal information extraction module includes a gated recurrent unit GRU.

4. The modulation mode identification method for negative signal-to-noise ratio conditions according to claim 3, characterized in that: The dual-stream time-frequency denoising network denoises the input signal to obtain a denoised signal. The specific method is as follows: S1 obtains the sample input signal X IQ ∈R i*m , where i is the IQ branch of the sample input signal, and m is the number of sampling points of the sample input signal; S2 samples the input signal X IQ Input the dual-stream time-frequency denoising network and use the frequency division module to sample the input signal X IQ Decompose and input the sample into signal X IQ Decompose into high-frequency signal H and low-frequency signal L, perform feature extraction on high-frequency signal H and low-frequency signal L respectively, and obtain high-frequency feature h and low-frequency feature l; S3 uses the time series information extraction module to extract time series features from the high-frequency feature h and the low-frequency feature l, respectively, to obtain the high-frequency time series feature h′ and the low-frequency time series feature l′; S4 sends the high-frequency temporal feature h′ and the low-frequency temporal feature l′ to the adaptive attention mechanism module for feature recalibration, and obtains the recalibrated low-frequency temporal feature h″ and the recalibrated high-frequency temporal feature l″; S5 upsamples the recalibrated low-frequency time series feature h″, and the upsampled low-frequency time series feature dimension h″ upsanmple The upsampled low-frequency time series feature h″ is the same as the recalibrated high-frequency time series feature l″ in dimension. upsanmple It is fused with the recalibrated high-frequency time series feature l″ to obtain the noise reduction signal F.

5. The modulation mode identification method for negative signal-to-noise ratio conditions according to claim 4, characterized in that: The dual-stream time-frequency denoising network is trained to obtain the dual-stream time-frequency denoising network with the best denoising effect. The specific method is as follows: The mean square error (MSE) loss function is used to learn and optimize the network parameters of the dual-stream time-frequency denoising network, and a dual-stream time-frequency denoising network model with the best denoising effect is obtained.

6. The modulation mode identification method for negative signal-to-noise ratio conditions according to claim 1, characterized in that: Construct a multi-scale convolutional neural network, including: a hybrid attention mechanism module, a BiLSTM-CNN module and a fully connected layer. The BiLSTM-CNN module includes a bidirectional long short-term memory network layer BiLSTM and a convolutional neural network layer CNN.

7. The modulation mode identification method for negative signal-to-noise ratio conditions according to claim 6, characterized in that: The multi-scale convolutional neural network recognizes the modulation mode of the noise reduction signal and obtains the modulation mode recognition result. The specific method is as follows: Step 1: Preprocess the noise reduction signal F to obtain a preprocessed noise reduction signal R; The pre-processing includes a dimension adjustment operation; Step 2: Input the preprocessed denoised signal R into the hybrid attention mechanism module to perform channel attention extraction and spatial attention extraction to obtain the channel attention feature map A and the spatial attention feature map B; Step 3: Input the preprocessed denoised signal R into the BiLSTM-CNN module, use the bidirectional long short-term memory network layer BiLSTM to extract time series features, and use the convolutional neural network layer CNN to extract time-frequency features, and obtain the time series features T and time-frequency features F′ respectively; Step 4: Fuse the channel attention feature map A, spatial attention feature map B, temporal feature T and time-frequency feature F′ of different scales to obtain the fused feature G; Step 5: Input the fusion feature G into the fully connected layer and use the cross entropy loss function for classification to obtain the modulation mode recognition result; Cross Entropy Loss Function As shown in the following formula: Among them, N is the number of categories, n is the element index, X (n) is the true category of the nth element, is the category of the predicted n-th element.

8. The modulation mode identification method for negative signal-to-noise ratio conditions according to claim 7, characterized in that: The multi-scale convolutional neural network is trained to obtain the multi-scale convolutional neural network with the best recognition effect. The specific method is as follows: The multi-scale convolutional neural network is optimized based on the stochastic gradient descent method to obtain a multi-scale convolutional neural network model with the best recognition effect.

9. The modulation mode identification method for negative signal-to-noise ratio conditions according to claim 1, characterized in that: The modulation mode of the input signal is identified based on the dual-stream time-frequency denoising network with the best denoising effect and the multi-scale convolutional neural network with the best recognition effect. The modulation mode recognition results of the input signal are obtained, including: Use the dual-stream time-frequency denoising network model with the best denoising effect to denoise the input signal Z to obtain the denoised signal Z′, and transmit the denoised signal Z′ to the multi-scale convolutional neural network with the best recognition effect; The multi-scale convolutional neural network model with the best recognition effect is used to identify the modulation mode of the denoised signal Z′, and the modulation mode recognition result P of the input signal Z is obtained.

Images