A method and system for modulated signal identification

Abstract

This invention relates to a modulation signal recognition method. The modulation signal recognition method includes: acquiring I / Q data of a modulation signal; constructing a lightweight complex residual attention neural network; and inputting the I / Q data into the lightweight complex residual attention neural network to obtain a modulation signal recognition result. The constructed lightweight complex residual attention neural network includes multiple complex convolutional modules, a residual attention module, and a classifier. Based on this modulation signal recognition scheme, the recognition rate of modulation signals can be improved, and the resource consumption rate can be reduced.

Description

Technical Field

[0001] This invention relates to the field of signal recognition technology, and in particular to a method and system for identifying modulated signals. Background Technology

[0002] Automatic Modulation Recognition (AMR) technology is widely used in civilian and military fields such as spectrum monitoring and electronic reconnaissance. Traditional modulation recognition methods extract instantaneous and statistical features of signals from the time and frequency domains for identification. Feature extraction is easily affected by noise, and the recognition performance of the algorithm is poor at low signal-to-noise ratios.

[0003] In recent years, deep learning has achieved great success in computer vision and natural language processing thanks to its powerful feature extraction capabilities. Some researchers are focusing on applying deep learning methods to AMR (Automatic Mapping). Current methods generally extract features from the time-frequency diagram and constellation diagram of the signal, transforming the problem of modulated signal recognition into an image recognition problem. This improves the recognition rate to some extent under low signal-to-noise ratio conditions. However, the process of converting the modulated signal into an image and the image recognition process both consume a lot of computing resources, limiting the deployment of classification models on resource-constrained devices and reducing the real-time performance of communication.

[0004] In recent years, some researchers have adopted end-to-end modulation signal recognition methods to avoid the uncertainty of manual feature extraction. They utilize the in-phase and quadrature (I / Q) components of the modulation signal obtained from GNU Radio as inputs into deep learning models for recognition, achieving better classification of modulation signals. For example, for the recognition of radio modulation signals, a two-dimensional I / Q signal is used to represent the complex-valued modulation signal. In a real-valued model, the real and imaginary parts of the signal are independent. To fully utilize the correlation between the real and imaginary parts, Tu et al. constructed a complex neural network model to classify modulation signals. Compared with the real-valued model, the complex-valued model achieved higher recognition performance. Furthermore, to enhance important information and reduce interference from irrelevant information, attention mechanisms are widely used in deep learning. An SE attention mechanism was proposed, which uses two-dimensional global pooling to compute channel attention, improving performance with lower computational cost. Woo et al. proposed a channel and spatial attention mechanism, mixing channel and spatial information to enhance input features. Lin et al. proposed a time-frequency attention mechanism, learning useful channel, frequency, and time features from spectral images to improve the recognition performance of existing CNNs. However, the performance of real-valued deep learning models in modulation recognition applications still cannot achieve the desired results. Summary of the Invention

[0005] The purpose of this invention is to provide a modulation signal recognition method and system that can improve the recognition rate of modulation signals and reduce resource consumption.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A modulation signal identification method, comprising:

[0008] Acquire the I / Q data of the modulated signal;

[0009] A lightweight complex residual attention neural network is constructed. This network includes multiple complex convolutional modules, a residual attention module, and a classifier. The complex convolutional modules extract complex domain features from the input data. The residual attention module extracts residual features from the complex domain features. The classifier obtains a classification result based on the extracted residual features.

[0010] The I / Q data is input into the lightweight complex residual attention neural network to obtain the modulation signal recognition result.

[0011] Preferably, the complex convolution module includes a real-valued convolution module and a linear transformation module; the linear transformation module is disposed in the real-valued convolution module.

[0012] Preferably, the real-valued convolution module includes a convolutional layer, a batch normalization layer, and a ReLU activation layer cascaded in sequence; the linear transformation module includes a linear matrix;

[0013] The linear matrix is ​​positioned between the convolutional layer and the batch normalization layer.

[0014] Preferably, the linear matrix is ​​a 3×2 linear matrix.

[0015] Preferably, the residual attention module includes a convolutional layer, a first attention channel, and a second attention channel;

[0016] The input to the convolutional layer is the output of the complex convolution module; the output of the convolutional layer is used as the input to the first attention channel and the second attention channel, respectively; the weighted output of the first attention channel and the output of the second attention channel are then input to the classifier.

[0017] Preferably, the size of the convolution kernel of the convolutional layer is 1×1.

[0018] Preferably, the classifier includes multiple fully connected layers.

[0019] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0020] The present invention provides a modulation signal recognition method. After constructing a lightweight complex residual attention neural network including multiple complex convolution modules, residual attention modules and a classifier, the modulation signal recognition result is obtained by using this lightweight complex residual attention neural network based on the I / Q data of the modulation signal. This method can improve the modulation signal recognition rate and reduce the resource consumption rate.

[0021] Corresponding to the modulation signal identification method provided above, the present invention also provides a modulation signal identification system, the system comprising:

[0022] The I / Q data acquisition unit is used to acquire the I / Q data of the modulated signal;

[0023] A neural network construction unit is used to construct a lightweight complex residual attention neural network. The lightweight complex residual attention neural network includes multiple complex convolution modules, a residual attention module, and a classifier. The complex convolution modules are used to extract complex domain features from the input data. The residual attention modules are used to extract residual features from the complex domain features. The classifier is used to obtain a classification result based on the extracted residual features.

[0024] The signal recognition unit is used to input the I / Q data into the lightweight complex residual attention neural network to obtain the modulation signal recognition result.

[0025] Since the technical effects achieved by the modulation signal recognition system provided by this invention are the same as those achieved by the modulation signal recognition method provided above, they will not be described again here. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 A flowchart of the modulation signal identification method provided by the present invention;

[0028] Figure 2 This is a schematic diagram of the structure of a lightweight complex residual attention neural network provided in an embodiment of the present invention;

[0029] Figure 3 This is a schematic diagram of the complex convolution module provided in an embodiment of the present invention;

[0030] Figure 4 This is a schematic diagram of the structure of the residual attention module provided in an embodiment of the present invention;

[0031] Figure 5 This is an overall structure diagram of a lightweight complex residual attention neural network provided in an embodiment of the present invention;

[0032] Figure 6 This is a block diagram illustrating the implementation of the complex convolution module provided in an embodiment of the present invention.

[0033] Figure 7 This is a block diagram of the residual attention module implementation provided in an embodiment of the present invention;

[0034] Figure 8 The recognition results of the compensation model provided in the embodiments of the present invention at various signal-to-noise ratios are shown in the figure.

[0035] Figure 9 The confusion matrix diagram of the CVRANN model provided in this embodiment of the invention at a signal-to-noise ratio of 10dB;

[0036] Figure 10 The comparison model provided in this embodiment of the invention has recognition results at various signal-to-noise ratios.

[0037] Figure 11 This is a schematic diagram illustrating the recognition rate of actual collected data under various models in embodiments of the present invention;

[0038] Figure 12 This is a schematic diagram of the modulation signal recognition system provided by the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] The purpose of this invention is to provide a modulation signal recognition method and system that can improve the recognition rate of modulation signals and reduce resource consumption.

[0041] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] like Figure 1 As shown, the modulation signal identification method provided by the present invention includes:

[0043] Step 100: Acquire the I / Q data of the modulation signal.

[0044] Step 101: Construct a lightweight complex residual attention neural network. For example... Figure 2 As shown, the constructed lightweight complex residual attention neural network includes multiple complex convolutional modules, a residual attention module, and a classifier. The complex convolutional modules extract complex domain features from the input data. The residual attention module extracts residual features from the complex domain features. The classifier obtains the classification result based on the extracted residual features. The classifier consists of multiple fully connected layers.

[0045] Each complex convolution module includes a real-valued convolution module and a linear transformation module. The linear transformation module is located within the real-valued convolution module. The real-valued convolution module consists of cascaded convolutional layers, batch normalization layers, and ReLU activation layers. The linear transformation module includes a linear matrix. This linear matrix is ​​positioned between the convolutional layers and the batch normalization layers, such as... Figure 3 As shown. A 3×2 linear matrix is ​​preferred.

[0046] like Figure 4 As shown, the residual attention module includes convolutional layers, a first attention channel, and a second attention channel. The input to the convolutional layer is the output of the complex convolutional module. The output of the convolutional layer serves as the input to both the first and second attention channels. The weighted sum of the outputs of the first and second attention channels is then input to the classifier. The preferred size of the convolutional kernel is 1×1.

[0047] Step 102: Input the I / Q data into a lightweight complex residual attention neural network to obtain the modulation signal recognition result.

[0048] Based on the above description, this invention fuses a complex convolution module and a residual attention module to form a complex residual attention mechanism, and constructs a complex residual attention neural network, such as... Figure 5 As shown, the middle section represents the overall structure of the complex dense residual neural network. During construction, firstly, a base network (BAN) consisting of four convolutional layers and three fully connected layers is used. Based on the BAN, a linear transformation is added to each convolutional layer to convert real-valued convolutional modules into complex convolutional modules, enabling the extraction of complex-domain features from the modulated signal. Next, a residual attention module is added after the last convolutional layer to extract residual features from the signal. Finally, the residual features are input into the fully connected layers for recognition. Figure 5In the diagram, the left side represents the Complex-Valued Convolution (CvConv) module, which includes convolution (Conv) layers, LT layers, batch normalization (BN) layers, and ReLU activation layers. The right side represents the Residual Attention (RA) module, which includes a convolutional layer with a kernel size of 1x1, a max pooling (MaxP) layer, and a mean pooling (MeanP) layer.

[0049] Because the complex space is more robust to noise than the real space, and different modulation types result in specific structural features of I / Q signals in the complex plane, the complex structure neural network model constructed in this invention can extract richer signal features and obtain better noise immunity.

[0050] This invention uses the implementation of complex convolution operations within a real-valued deep learning framework through linear transformation as an example to illustrate the implementation process of the modulation signal recognition method provided above. For example... Figure 6 As shown, N×2 represents the dimension of the data features, and s′ and s represent the number of feature channels. Matrix K is the linear transformation matrix.

[0051] In deep learning frameworks, the real-valued convolution operation process is as follows: by zero-padding the input data and convolving it with a convolution kernel, N×2 input features are transformed into N×3 output features:

[0052]

[0053] Assume the input is a two-dimensional I / Q complex signal and the convolution kernel complex weight parameters are X. n (n∈1,2,...,N) and W m (m∈1,2,...M), the complex number representation is shown in equation (2), where, I N Q N They are X n The in-phase and quadrature components, R M V M They are W m The real and imaginary parts of the integer, in the complex field, are given by input X. n With W m The one-dimensional convolution operation is shown in equation (3).

[0054]

[0055] C 1D =X n *W m =(I N *R M -QN *V M )+i(I N *V M +Q N *R M (3)

[0056] Comparing the output of two-dimensional real-valued convolution with the calculation results of one-dimensional complex convolution, such as... Figure 7 As shown, through linear transformation, the first column of the matrix on the right side of equation (1) is subtracted from the third column to obtain the real part of the one-dimensional complex convolution operation, and the second column is directly used as the imaginary part to achieve the result of the one-dimensional complex convolution. That is, by multiplying the output result of equation (1) with the linear transformation matrix K, the complex convolution operation can be implemented in the real-valued deep learning model.

[0057] Attention mechanisms, derived from the field of image recognition, mimic human visual attention by focusing on key target regions in an image and suppressing other useless information. Based on this, to extract more effective recognition features in modulation recognition, attention mechanisms are applied to the field of modulation recognition, further extracting key features in the complex domain of the signal for classification and recognition. The specific structure is as follows... Figure 7 As shown, the residual attention module consists of a single convolutional layer and two attention channels.

[0058] in, The input data for the residual attention module is given by c′, h, and w, which represent the number of channels, height, and width of the input data features, respectively. The convolutional kernel size is c×1×1, where c is the number of output channels, and 1×1 represents the height and width of the kernel. Max Pooling and Avg Pooling are used for max pooling and average pooling, respectively, for the dual attention channels. When processing the input data, the residual attention module extracts features from the input data x through the convolutional layer and outputs data... To further eliminate interference from useless information and extract key information, a dual-channel attention approach is used to pool the data to obtain pooled features m. i and p i Then, the two pooled data are weighted and summed to obtain the final classification feature f. i , i∈{1,2,...,c}. Where m i The weighted cumulative value of the i-channel data after spatial pooling is obtained by the spatial attention layer in the dual attention channel, as shown in equation (4), where r i (j) represents the weight of the j-th feature element in the feature space of channel i, and α is a coefficient parameter used to control r i The magnitude of (j) as α approaches infinity, r i (j) becomes a Dirac delta function for the independent variable y. i(j) Only the largest element has a value, and all other elements are 0. That is, when α approaches infinity, the spatial pooling of the input data in equation (4) becomes the max pooling, as shown in equation (5).

[0059]

[0060] m i =max(y i (1),y i (2),…,y i (h×w)) (5)

[0061] Similarly, the other pooling operation in the dual attention channel is average pooling, as shown below:

[0062]

[0063] Finally, as shown in equation (7), the max-pooling feature m output from the dual attention channels is... i With p i The residual structure is formed by weighted summation to obtain the residual features of the classifier input.

[0064] f i =p i +λm i (7)

[0065] The advantages of the modulation signal recognition method described above will be illustrated through experiments below.

[0066] Experiments were conducted using the open-source dataset RadioML2016.10a to evaluate the proposed modulation recognition model. RadioML2016.10a includes common modulation schemes in communication systems: three analog modulations and eight digital modulations, namely 8PSK, AM-DSB, AM-SSB, BPSK, CPFSK, GFSK, PAM4, QAM16, QAM64, QPSK, and WBFM. This dataset has 20 different signal-to-noise ratios (SNR) ranging from -20dB to 18dB with a stride of 2. Each modulation scheme has 1000 signals per SNR, and each signal consists of 2x128 IQ signals, for a total of 220,000 modulated signals. The dataset was divided into training and testing sets in a 7:3 ratio. A complex residual attention recognition model was built based on the PyTorch deep learning framework, with Adam selected as the optimizer, for both training and testing on the dataset. To analyze the effectiveness of the proposed complex residual attention mechanism, a compensation experiment was conducted using a BAN as the base network. Complex convolutional modules, residual attention modules, and complex residual attention modules were added to identify the modulated signal. Furthermore, the proposed AMR model was compared with CNN2 and CLDNN models, and the recognition performance of the proposed AMR model was verified by identifying a dataset of modulated signals actually acquired using a software radio platform. Finally, the resource consumption of each model was analyzed, using the number of model parameters as the evaluation metric for model size and floating-point operations per second as the evaluation metric for model running speed.

[0067] To verify the effectiveness of the complex residual attention mechanism proposed in this invention, compensation experiments were conducted. First, a complex-valued convolutional neural network (CVCNN) was formed by adding a complex convolutional module. Second, a residual attention neural network (RANN) was formed by adding a residual attention module. Third, a complex convolutional module plus a residual attention module was added, resulting in the proposed complex residual attention neural network (CVRANN). The compensation experiments involved inputting pre-defined test sets into the compensation models to evaluate the recognition performance of each compensation model, such as... Figure 8 As shown.

[0068] Depend on Figure 8As can be seen, the recognition rate of each model gradually increases with the increase of the signal-to-noise ratio (SNR). At all SNR levels, the compensated models CVCNN and RANN, which incorporate complex convolutional modules and residual attention modules, outperform BAN in recognition performance. Similarly, at all SNR levels, the CVRANN model proposed in this invention achieves a higher recognition rate than the other models, reaching 90% at high SNR levels, a 10% improvement compared to BAN. Compared to applying only a single complex convolutional module and residual attention module, the recognition rate is improved by 5% and 8%, respectively.

[0069] Further analysis based on the compensation experiment results shows that, compared with the real-valued model, the CVRANN model proposed in this invention can extract more effective features in the complex domain of the signal by utilizing the complex convolution module. Compared with the model lacking the residual attention mechanism, the residual attention mechanism can more effectively aggregate the extracted complex domain features and eliminate the influence of irrelevant information.

[0070] To further analyze the recognition performance of each modulated signal, a confusion matrix was plotted based on the recognition results of the CVRANN model, as shown in the following figure. Figure 9 As shown.

[0071] Figure 9 The horizontal and vertical axes represent the true and predicted labels, respectively. The intersection of the true and predicted labels represents the confusion area, the diagonal intersection represents the correctly identified region, and the remaining intersections represent the confusion area. Figure 9 As can be seen, due to the similarity between signal modulation methods, 16QAM is mainly confused with 64QAM, WBFM is mainly confused with AM-DSB, and for the other modulation signals, the CVRANN model proposed in this invention can basically achieve 100% recognition.

[0072] Comparative experiment

[0073] The RadioML2016.10a open-source dataset was used to perform recognition tests on CNN2, CLDNN, and the proposed CVRANN network, respectively. To verify the effectiveness of the proposed classification model, the dataset was input into CNN2, CLDNN, and the proposed CVRANN network for recognition, and comparative analysis was conducted. The recognition results are as follows: Figure 10 As shown.

[0074] Depend on Figure 10 It is evident that the proposed model outperforms existing models in terms of recognition performance.

[0075] To analyze the differences in signal feature extraction from the above models, 1024 samples were randomly selected from the test set and input into the models. The TSNE dimensionality reduction algorithm was used to visualize the output features of the first fully connected layer of the model. It was found that the signal features extracted by the CNN2 model exhibited inter-class overlap, while the signal features extracted by the CLDNN model showed less overlap but smaller inter-class distances. In contrast, the visualized feature maps of the CVRANN model proposed in this invention showed significant differences between classes, resulting in a higher final recognition rate. Figure 9 The results shown are consistent.

[0076] To verify the robustness of the proposed model, 10 types of modulated signals were acquired using software-defined radio transceivers (SAM-60MK2 and VSG60A) and a software platform (LABVIEW): AM, FM, 2ASK, BPSK, QPSK, 8PSK, 2FSK, 4FSK, 16QAM, and 64QAM. During actual acquisition, 5000 samples were collected for each type of modulated signal, with each sample consisting of 2x1024 IQ signals. The signal-to-noise ratio at the transmitting end was set to 0dB. The acquired data was input into CNN2, BAN, CLDNN, and CVRANN models for training and recognition. The results are as follows: Figure 11 As shown, compared with other models that do not apply complex attention mechanisms, the CVRANN model still exhibits superior recognition performance, verifying the recognition feasibility of the complex residual attention neural network proposed in this invention.

[0077] Resource consumption analysis

[0078] Recognition performance and resource consumption are two important indicators for evaluating the practical application of modulation recognition models. The experimental results above show that the CVRANN model outperforms the basic model and other compensation models at various signal-to-noise ratios. Based on this, the resource consumption of the aforementioned compensation and comparison models is further examined. The resource consumption of each model is evaluated using the number of model parameters as the evaluation metric for model size and the number of floating-point operations per second as the evaluation metric for model running speed. The results are shown in Table 1.

[0079] As shown in Table 1, the experimental results indicate that, compared to the basic BAN model, the models with complex convolution modules and residual convolution modules exhibit reduced resource consumption to some extent. Further analysis reveals the reasons for this. Building upon the BAN model, firstly, the application of linear combinations to perform complex convolution operations in real-valued deep learning ensures that the signal feature dimension after each convolution operation remains consistently 2×128, avoiding the increased resource consumption caused by the increased feature dimension after multiple convolution operations. Secondly, the residual attention module focuses on the key features of the data, eliminating irrelevant information and further reducing the feature dimension to a certain extent, thus lowering resource consumption. In contrast, the CVRANN model increases model complexity to improve recognition performance, resulting in slightly higher resource consumption than CNN2 and CLDNN. However, compared to the improved recognition rate, the increased resource consumption is within a manageable range.

[0080] Table 1. Resource Consumption of the Basic Model and the Compensation Model

[0081]

[0082] The experimental results and analysis above show that the CVRANN model proposed in this invention uses a complex convolution module to extract effective features in the complex domain of the modulated signal and further uses a residual attention module to extract key features. This not only improves the recognition performance of the model, but also reduces the resource consumption of the model to a certain extent. Compared with the basic model BAN, the number of model parameters is reduced dramatically and the floating-point operation is reduced by nearly 3 / 5.

[0083] Based on the above description, compared with real-valued models such as BAN, CNN2, and CLDNN, the complex residual attention neural network model proposed in this invention has better recognition performance and lower resource consumption.

[0084] In addition, corresponding to the modulation signal identification method provided above, the present invention also provides a modulation signal identification system, such as... Figure 12 As shown, the system includes:

[0085] I / Q data acquisition unit 1 is used to acquire the I / Q data of the modulated signal.

[0086] Neural network building unit 2 is used to construct a lightweight complex residual attention neural network. This lightweight complex residual attention neural network includes multiple complex convolutional modules, a residual attention module, and a classifier. The complex convolutional modules are used to extract complex domain features from the input data. The residual attention modules are used to extract residual features from the complex domain features. The classifier is used to obtain the classification result based on the extracted residual features.

[0087] Signal recognition unit 3 is used to input I / Q data into a lightweight complex residual attention neural network to obtain the modulation signal recognition result.

[0088] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0089] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for identifying modulated signals, characterized in that, include: Acquire the I / Q data of the modulated signal; A lightweight complex residual attention neural network is constructed. This network includes multiple complex convolutional modules, a residual attention module, and a classifier. The complex convolutional modules extract complex domain features from the input data. The residual attention module extracts residual features from the complex domain features. The classifier obtains a classification result based on the extracted residual features. Each complex convolutional module includes a real-valued convolutional module and a linear transformation module. The linear transformation module is nested within the real-valued convolutional modules. The real-valued convolutional module comprises cascaded convolutional layers, batch normalization layers, and ReLU activation layers. The linear transformation module includes a linear matrix, which is positioned between the convolutional layers and the batch normalization layers. The I / Q data is input into the lightweight complex residual attention neural network to obtain the modulation signal recognition result; In deep learning frameworks, the real-valued convolution operation process is as follows: The input data is zero-paddinged and convolved with a convolution kernel to... Transform input features into The dimensional output features are: (1) The input two-dimensional I / Q complex signal and the convolution kernel complex weight parameters are respectively and The complex number representation is shown in equation (2), where, They are In-phase and quadrature components, They are The real and imaginary parts of the integers, in the complex field, are input. and The one-dimensional convolution operation is shown in equation (3); (2) (3) By comparing the output of two-dimensional real-valued convolution with the calculation result of one-dimensional complex convolution, through linear transformation, the first column of the matrix on the right side of equation (1) is subtracted from the third column as the real part of the one-dimensional complex convolution operation, and the second column is directly used as the imaginary part to realize the result of one-dimensional complex convolution.

2. The modulation signal identification method according to claim 1, characterized in that, The linear matrix is ​​a 3×2 linear matrix.

3. The modulation signal identification method according to claim 1, characterized in that, The residual attention module includes a convolutional layer, a first attention channel, and a second attention channel; The input to the convolutional layer is the output of the complex convolution module; the output of the convolutional layer is used as the input to the first attention channel and the second attention channel, respectively; the weighted output of the first attention channel and the output of the second attention channel are then input to the classifier.

4. The modulation signal identification method according to claim 3, characterized in that, The kernel size of the convolutional layer is 1×1.

5. The modulation signal identification method according to claim 1, characterized in that, The classifier comprises multiple fully connected layers.

6. A modulation signal identification system, characterized in that, The modulation signal identification system is used to implement the modulation signal identification method as described in any one of claims 1-5; the system includes: The I / Q data acquisition unit is used to acquire the I / Q data of the modulated signal; A neural network construction unit is used to construct a lightweight complex residual attention neural network. The lightweight complex residual attention neural network includes multiple complex convolution modules, a residual attention module, and a classifier. The complex convolution modules are used to extract complex domain features from the input data. The residual attention modules are used to extract residual features from the complex domain features. The classifier is used to obtain a classification result based on the extracted residual features. The signal recognition unit is used to input the I / Q data into the lightweight complex residual attention neural network to obtain the modulation signal recognition result.

Images