A micro intelligent modulation recognition method based on resampling mechanism

Abstract

The present application provides a kind of micro intelligent modulation identification method based on resampling mechanism, specific steps include: generating multiple signal multiple signal-to-noise ratio modulation identification dataset, construct micro intelligent modulation identification network based on resampling mechanism, train micro intelligent modulation identification network based on resampling mechanism and identify the modulation type of signal.The present application introduces resampling mechanism, uniformly samples long signal data input to reduce data length, converts from bar data to stick data, greatly reduces the parameter amount of model.The present application replaces residual network with dilated residual network at the tail of network, further enhances the information extraction capability of network in processing long bar data, ensures that signal characteristics can be effectively retained in model propagation process.The present application adds self-attention mechanism to the back end of dilated residual network, utilizes its dynamic learning ability, effectively integrates and filters the signal characteristics extracted by front part network, so as to obtain the overall cognition of signal.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, and specifically to a micro-intelligent modulation recognition method based on a resampling mechanism. Background Technology

[0002] Communication signals, as a commonly used signal system, are widely applied in commerce, industry, and electromagnetic countermeasures, showing great application potential. However, with the continuous evolution of communication signal systems, the complexity of signals is increasing, and the continuous iteration of communication equipment has provided more possibilities for the transmission methods of communication signals. This poses a significant challenge to receiving or reconnaissance equipment. In addition to collecting signals according to frequency bands or known parameters, receiving equipment also needs to understand the signal modulation method to obtain corresponding information. Currently, communication signal modulation identification technology is widely used in many fields, mainly including communications, broadcasting, radar, and electronic measurement. To adapt to rapidly changing modulation methods, more adaptable modulation identification methods are needed.

[0003] Traditional modulation recognition methods rely on probability theory or known parameters for identification, or on salient signal features such as instantaneous features and cyclic stationary spectrum features. However, these methods are often limited to specific signals and lack applicability. With the advent of machine learning, SVM, decision trees, and BP neural networks began to be applied to this work, but these methods were gradually replaced by later-developed neural networks due to their weak fitting capabilities. As deep learning technology matures and its feature extraction capabilities improve, researching rapid identification and classification techniques for communication signals is one of the directions in the development of artificial intelligence.

[0004] Since CNNs were first applied to modulation recognition technology, various deep learning methods have been applied to signal modulation recognition. Subsequent temporal neural networks, primarily LSTM and GRU, surpassed the original CNN networks due to their temporal memory and temporal correlation. To reduce the task difficulty, some researchers transformed time-series data into graphs, introducing time-domain and frequency-domain graphs, and using image classification methods to classify signals. However, both LSTM and GRU, which are based on temporal networks, and image classification methods all have a large number of parameters and computational costs, making them difficult to apply in real-world scenarios. Summary of the Invention

[0005] This invention addresses the problem that existing networks have high hardware resource requirements in terms of parameter quantity and computational load, and the large weight files also prevent the networks from being directly applied to embedded boards. It provides a micro intelligent modulation recognition method based on a resampling mechanism. By utilizing signal characteristics and introducing a resampling mechanism, the method greatly reduces the data length of the signal input. Furthermore, it employs a more applicable expansion mechanism and self-attention mechanism in the backend, which significantly reduces the number of model parameters and computational load while ensuring the accuracy of modulation recognition.

[0006] This invention provides a miniature intelligent modulation recognition method based on a resampling mechanism, comprising the following steps: S1. Divide the collected communication signals into a detection dataset and a training dataset; S2. Construct a miniature intelligent modulation recognition network based on a resampling mechanism; The miniature intelligent modulation recognition network includes a resampling module, a residual module, and an expanded residual module connected in sequence, and a self-attention module and a fully connected layer connected in sequence to the outputs of the residual module and the expanded residual module; The resampling module samples the input data at equal intervals and superimposes the sampled data on the channel layer to change the size of the input data and obtain compressed data. The residual module mines the relationships between different dimensions of the compressed data and preserves the original signal information during forward propagation; The dilated residual module includes residual convolution and dilated convolution. The outputs of the residual module and the dilated residual module are summed to obtain the signal after fusing deep and shallow features. The self-attention module extracts and integrates features from the signal after fusing features from deep and shallow layers; The fully connected layer converges the output features of the self-attention module to obtain the modulation feature determination result of the signal; S3. Train the micro-intelligent modulation recognition network. Input the training dataset of multi-type and multi-signal-noise ratio modulation signals into the micro-intelligent modulation recognition network. Use the Adam optimizer and gradient descent method for iterative training. After reaching the target, obtain the weight file of the micro-modulation recognition network model for multi-type and multi-signal-noise ratio modulation recognition. S4. Use a miniature intelligent modulation recognition network based on a resampling mechanism to identify signal categories.

[0007] The miniature intelligent modulation recognition method based on a resampling mechanism described in this invention, in a preferred embodiment, includes the following steps in step S1: S11. Classify and integrate the collected communication signal data according to different signal-to-noise ratios and different modulation methods; S12. Each signal is 1024 in length, with a corresponding label ranging from 0 to the number of categories in the dataset. Additional labels are added to the SNR to obtain a multi-signal-to-noise ratio modulation recognition dataset. S13. Select 30% of the data as the detection dataset and 70% as the training dataset; The present invention discloses a micro-intelligent modulation recognition method based on a resampling mechanism. In a preferred embodiment, the resampling module converts the communication signal data (M1, L1, N1) into (M1', L1', N1), where M1 is the number of communication signal channels acquired, M1 is 1, L1 is the length of the acquired communication signal, N1 is the width of the acquired communication signal, M1' is the number of signal channels output by the resampling module, M1'≥2, and L1' is the length of the signal output by the resampling module, L1'=L1 / M1'. The resampling module includes at least two layers of network modules with downsampling capabilities.

[0008] The present invention provides a micro-intelligent modulation recognition method based on a resampling mechanism. In a preferred embodiment, the number of residual modules is five. Each residual module includes five convolutional layers and one max pooling layer. The output of the first convolutional layer is added to the output of the third convolutional layer, and one branch of the result is input to the fourth convolutional layer. The other branch is added to the output of the fifth convolutional layer, and the output of the fifth convolutional layer is sent to the max pooling layer. The kernel size of the first convolutional layer is... ; The last four convolutional kernels of the first residual module are all of size 1. The number of convolutional kernels is 16; the last four convolutional kernels of the second, third, fourth, and fifth residual modules are all of size 16. Both have 16 channels.

[0009] The present invention discloses a micro-intelligent modulation recognition method based on a resampling mechanism. In a preferred embodiment, the structure of the dilated residual module is the same as that of the residual module, except that the second, third, fourth, and fifth network modules inside the dilated residual module are replaced with dilated convolutional layers.

[0010] The micro-intelligent modulation recognition method based on resampling mechanism described in this invention, as a preferred embodiment, includes a self-attention module comprising: a Q branch, a K branch, and a V branch; The Q branch is a single convolutional kernel. The convolutional layers are used to learn the channel-level features of the signal after the fusion of deep and shallow features. The K branch and V branch are two consecutive convolutional kernels with a size of [missing value]. The convolutional layer is used to calculate the positional and layer features of the signal after the fusion of deep and shallow features; The Q branch, K branch, and V branch share input features. The Q branch result is element-wise multiplied with the K branch result and then passed through the Softmax function before being element-wise multiplied with the V branch result. The multiplied result is then fused with the features from the deep and shallow layers through the ReLU activation layer, and the signals are added together to obtain the output of the self-attention module.

[0011] The miniature intelligent modulation recognition method based on a resampling mechanism described in this invention, as a preferred embodiment, has the following output from the self-attention module: ; in, , and The features extracted from the Q branch, K branch, and V branch are respectively. For the number of channels, This is an element-wise multiplication operation.

[0012] The miniature intelligent modulation recognition method based on a resampling mechanism described in this invention, as a preferred embodiment, uses the following weight gradient update method for the V branch: ; in, branch for V The weight parameters, , and For constants during parameter update, by and control; Branches and The weight gradient update method for the branch is the same as that for the V branch.

[0013] The present invention provides a micro-intelligent modulation recognition method based on a resampling mechanism. In a preferred embodiment, the fully connected layer includes a first fully connected layer and a second fully connected layer connected in sequence. The first fully connected layer flattens the output of the self-attention module and outputs it to the second fully connected layer to obtain the number of signal categories. The final category recognition probability is obtained by the Softmax function, and the highest probability is taken as the final signal category recognition result. The number of neurons in the second fully connected layer is the same as the total number of categories of the communication signals to be identified.

[0014] In the present invention, a miniature intelligent modulation recognition method based on a resampling mechanism is preferred in which N1 is 2 when the acquired signal includes I signal and Q signal. The resampling module converts the (1,1024,2) I / Q signal data into (4,256,2). The first residual module outputs (16,128,1) data, the second residual module outputs (16,64,1) data, the third residual module outputs (16,32,1) data, the fourth residual module outputs (16,16,1) data, and the fifth residual module and the dilated residual module both output (16,8,1) data. The input length of the first fully connected layer is 128 and the output length is 64, and the output length of the second fully connected layer is 24.

[0015] This invention relates to the fields of communication signal recognition and classification and deep learning in communication signal processing, specifically to a miniature intelligent modulation recognition method based on a resampling mechanism applicable to multiple types of communication signals.

[0016] This invention proposes a micro-signal intelligent modulation recognition method based on a resampling mechanism, which reduces the number of network parameters and the network weight file obtained through training. The size of the weight file is close to that of most quantized model weight files, making it easy to port and conveniently applied to various platforms, thus enhancing the practicality of intelligent signal modulation recognition.

[0017] The present invention has the following advantages: This invention utilizes residual networks from deep learning as the basic network architecture for accurate and rapid identification, inheriting their high accuracy and versatility in signal modulation identification. This invention achieves intelligent modulation identification of multiple types of communication signals using a miniature model, facilitating practical applications. A resampling mechanism is introduced to uniformly sample long signal data inputs, reducing data length and transforming bar-shaped data into rod-shaped data, significantly reducing the number of model parameters. This invention replaces the residual network with an dilated residual network at the network tail, further enhancing the network's information extraction capability in processing long bar-shaped data and ensuring that signal features are effectively preserved during model propagation. This invention adds a self-attention mechanism at the back end of the dilated residual network, utilizing its dynamic learning capability to effectively integrate and filter the signal features extracted by the preceding network, thereby obtaining an overall understanding of the signal. This invention can achieve stable signal identification at most signal-to-noise ratios (>6dB). This invention is applicable to various modulation identification methods and has strong applicability. The final unquantized weight file generated by this invention can reach a size of 672KB, achieving the level of conventional network pruning and quantization, and has greater practical value. Attached Figure Description

[0018] Figure 1 This is a flowchart of a miniature intelligent modulation recognition method based on a resampling mechanism; Figure 2 This is a schematic diagram of a micro-modulation recognition network structure for a micro-intelligent modulation recognition method based on a resampling mechanism; Figure 3 This is a schematic diagram of the structure of a self-attention module in a miniature intelligent modulation recognition method based on a resampling mechanism; Figure 4 This is a comparison of the recognition accuracy of a micro-intelligent modulation recognition method based on a resampling mechanism with other models. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Example 1

[0020] like Figure 1 As shown, a micro-intelligent modulation recognition method based on a resampling mechanism includes the following steps: S1. Generate a multi-signal, multi-signal-to-noise ratio modulation recognition dataset: S1.1 Classify and integrate the collected communication signal data according to different signal-to-noise ratios and different modulation methods; S1.2. Each signal is 1024 in length, and its corresponding label ranges from 0 to the number of categories in the dataset. Additional labels are added for the SNR. S1.3 Select 30% of the data as the detection dataset and 70% as the training dataset; S2. Construct a miniature intelligent modulation recognition network based on a resampling mechanism: The network structure of this invention is based on the ResNet network. See the overall network structure diagram below. Figure 2 .

[0021] S2.1, Resampling module; The resampling module is similar to uniform sampling. It samples the data at equal intervals and superimposes the sampled data on the channel layer, converting the (1, 1024, 2) I / Q signal data into (4, 256, 2). Uniform sampling minimizes the loss of signal features. Even if the frequency of some signals is affected, this portion of the signal is not lost after sampling but is superimposed on the channel layer. Therefore, the network can still notice this information during the learning process, and it will not significantly affect the modulation category evaluation result.

[0022] Taking (1, 1024, 2) data input as an example, in order to achieve a feature size of (4, 256, 2), it is necessary to go through two layers of network modules with downsampling function. The resampling module changes the data size by superimposing the sampled data, thereby quickly achieving data compression and greatly reducing the computational load and network parameters of the network.

[0023] Meanwhile, because the resampling module reshapes the signal data, with the same convolution kernel parameters, its convolution operation count is 1 / n of that without the resampling module, where n is the resampling rate. This allows the computational cost of the network to be reduced by a factor of several.

[0024] S2.2, Construct the residual module; Its front end is a convolutional kernel of size . The convolutional layer has 16 kernels, followed by four consecutive kernels of size [size missing]. The network uses 16 convolutional layers. The output of the first convolutional layer is added to the output of the third convolutional layer. One branch of the result is fed into the fourth convolutional layer, and the other branch is added to the output of the fifth convolutional layer. Finally, the output of the fifth convolutional layer is fed into a max-pooling layer. Each network layer has a stride of 1. The residual module consists of five layers, used in the second, third, fourth, fifth, and sixth network modules of the entire model. The difference is that, because the third dimension in the data input size (4, 256, 2) is 2, to explore the relationships between different dimensions of the data, the size of the last four convolutional kernels in residual module 1 (the second layer of the model's network) needs to be set to... .

[0025] Because the resampling module reshapes the data, the required network computation operations are greatly reduced. To prevent network overfitting, the number of convolution kernels in subsequent networks also needs to be reduced accordingly. Therefore, 16 is used as the number of convolution kernels in each network layer.

[0026] Meanwhile, the residual module can retain the original signal information during the forward propagation of the model, avoiding information loss caused by a large number of model layers.

[0027] S2.3, Expansion Residual Module; The dilated residual module is similar to the residual module, but the convolutional layers at the second, third, fourth, and fifth positions (within the module) are replaced with dilated convolutions. The kernel size and stride of the convolutions are the same as those of the residual module, and the dilation rate is set to 2.

[0028] It is used in parallel with a residual network, and the results of the two modules are element-wise summed and input into the self-attention module.

[0029] Compared to ordinary convolution, dilated convolution has a larger receptive field, which can effectively handle long data types. It has a stronger feature integration capability and is more suitable for processing signal-type data.

[0030] assumed It is a convolution kernel. It is the convolution stride. The formula for calculating the receptive field of a convolutional layer is: ; in This refers to the receptive field of the current convolutional layer. It is the receptive field of the previous level network. If the convolution stride of the convolutional layer is... s and convolution kernel k With sizes set to 1 and 3 respectively, the receptive field size of a single layer is only 3, requiring a large number of network layers to cover 1024 data points. However, due to practical limitations, modulation recognition tasks do not employ excessive convolutional networks.

[0031] Dilated convolution increases the efficiency of the convolution by padding the kernel with zeros. The receptive field is larger than that of a single ordinary convolution. ; With the convolution stride and kernel size set to 1 and 3 respectively, and the dilation rate r to 2, the receptive field of a single layer is 5, while the receptive field of its second layer will become 9, which is much larger than that of ordinary convolution.

[0032] At this point, compared to the fully residual network ResNet, the dilated residual convolution has a receptive field that is twice that of the ordinary residual convolutional network, thus possessing a deeper signal feature extraction capability.

[0033] To effectively utilize the feature information of different depths of signals extracted from different receptive fields, this layer of the network uses the sum of a residual network and an dilated residual network as the output.

[0034] S2.4 Construct a self-attention module; The self-attention module has three branches: Q, K, and V.

[0035] The Q branch is a single convolutional kernel. The convolutional layer.

[0036] The K and V branches are two consecutive convolutional kernels with a size of [missing value]. The convolutional layer.

[0037] It shares input features. The Q-branch result is element-wise multiplied with the K-branch result, then multiplied with the V-branch by the Softmax function, and finally added to the input of the self-attention module through the ReLU activation layer to obtain the output of the self-attention module.

[0038] Its structural diagram is as follows Figure 3 As shown.

[0039] The formula for this self-attention module can be written as: ; in , and This refers to the features extracted from the three branches. It refers to the number of channels. This represents element-wise multiplication. The function calculates the probability of the element at each position. Then, The weights generated by the function and The features in the samples are used to extract features from each other to obtain the final output.

[0040] Because the signal data is elongated and rectangular, after multiple dimensionality reductions, the pairwise matrix multiplication of Q, K, and V will cause the features to change in both length and width dimensions. This leads to information failure, so it is necessary to change the matrix multiplication in the typical self-attention mechanism to element-wise multiplication.

[0041] The branch is primarily used to learn channel-level features. Location-level features are calculated by other branches. The function assigns weights to each feature. This invention employs a method following the self-attention module. The layer performs a non-linear fitting to the final result, therefore the formula for the self-attention module can be changed to: ; , Medium weight parameters The gradient formula is: ; in , and It is a constant in the parameter update process, which can be obtained from... and Control. With The branches are the same. , The parameters in the model also follow a similar gradient update rule. According to this rule, the weights of the three branches can be dynamically updated, demonstrating dynamic learning capabilities, and they can influence each other. Furthermore, dynamic parameters can filter noisy data from the signal and increase the weight of effective features in the back-parameter update, enhancing the model's recognition performance. Building upon this, the invention also adds a skip connection structure to avoid overfitting.

[0042] Since the resampling module causes a certain degree of loss of temporal logic in the signal data due to the reshaping of the data, resulting in a decrease in the overall recognition accuracy of the network, in order to compensate for this loss of accuracy, this invention uses a self-attention module with stronger information integration capabilities to enhance the signal feature extraction capability of the network, and places it after the sixth network layer (the sum of elements of the residual module and the dilated residual module) that integrates the deep and shallow features of the signal, so as to effectively utilize the information integration capability of the self-attention module.

[0043] S2.5, construct a fully connected layer; The output of the self-attention module is flattened, with an output length of 128 or 64, and then input into the second fully connected layer. The final output is the number of categories, which is then processed by the Softmax function to obtain the final category recognition probability. The highest probability is taken as the final recognition result.

[0044] In this embodiment, there are a total of 24 signal categories, therefore the number of neurons in the second fully connected layer is 24.

[0045] The fully connected layer is used to receive the signal information features integrated by the self-attention module, and to converge the features to obtain the modulation feature determination result of the signal.

[0046] S3. Train a miniature intelligent modulation recognition network based on a resampling mechanism; The training set of multi-type multi-signal-to-noise ratio modulated signals is input into a miniature intelligent modulation recognition network based on a resampling mechanism. The Adam optimizer is used with a learning rate of 0.001, and gradient descent is used for iterative training to obtain the weight file of the miniature modulation recognition network model for multi-type multi-signal-to-noise ratio modulation recognition. S4. Identify the modulation type of the signal: The input signal is clipped to a length of 1024 and input into the proposed micro-modulation recognition network model to obtain the final recognition result.

[0047] The input and output parameters of each module of the network are shown in Table 1: Table 1 Model Input / Output Parameters

[0048] The use of a resampling module can significantly reduce the data size and the number of subsequent network operations without increasing computational cost (the cost of a single-layer convolutional network is reduced to 1 / n of the original computational cost, where n is the resampling rate).

[0049] The effects of this invention can be further illustrated by the following simulation experiments.

[0050] Simulation experimental conditions: The hardware platform for the simulation experiment of this invention is as follows: the processor is an Intel i9-10940X with a main frequency of 3.30GHz, the memory is 64G, the graphics card is an NVIDIA GeForce RTX 2080 Ti with 12GB of video memory.

[0051] The software platform for the simulation experiment of this invention is: Windows 10 operating system and Python 3.6.

[0052] The simulation experiments in this invention use 30% of the data from the RML2018.01a dataset, a widely used dataset in the fields of wireless communication and radio frequency machine learning, primarily used for developing and evaluating wireless signal classification algorithms. This dataset contains signal samples with various modulation styles, generated using software-defined radio (SDR) technology, covering different signal-to-noise ratio (SNR) ranges, totaling 2,555,904 data points.

[0053] Simulation experiment content and result analysis: The simulation experiment of this invention is carried out according to the following steps.

[0054] The simulation experiment covers the entire modulation recognition process, from signal acquisition to input network and output results.

[0055] Step A involves randomly selecting data from the RML2018.01a dataset and truncating it to 1024.

[0056] Step B involves dividing the data into training and testing sets at a ratio of 7:3 under different signal-to-noise ratios and categories. The final training set contains 1,789,132 sets of data, and the testing set contains 766,771 sets of data.

[0057] Step C involves inputting the signal data and labels from the training set into the micro-modulation recognition network. The data is shuffled before training, but random sorting is not used during the training process. The model is saved every 100 epochs and tested on the validation set.

[0058] Step D involves inputting the test set signal data into the trained micro-modulation recognition network to obtain the detection results.

[0059] The following is combined Figure 4 The resulting figures further illustrate the effects of the present invention.

[0060] Figure 4 In this embodiment, S_AMR_128 is a network with a first fully connected layer of 128 neurons, and S_AMR_64 is a network with a first fully connected layer of 64 neurons. The other methods, LSTM, ResNet, and CNN, are for comparison. Figure Four The curves represent the recognition accuracy of different networks under different signal-to-noise ratios. It can be seen that the proposed network is not much different from other mainstream networks in terms of accuracy.

[0061] The recognition accuracy Acc is calculated using the following formula and used as an evaluation metric to assess the recognition results. All calculation results are plotted in Table 1.

[0062] Let TP be the number of samples that are actually positive and whose predicted results are also positive, and FN be the number of samples that are actually positive and whose predicted results are negative.

[0063] The formula for accuracy (Acc) is as follows: ; Wherein, FP is the number of samples that are actually negative but are predicted to be positive, and TN is the number of samples that are actually negative and are also predicted to be negative.

[0064] Table 1. Quantitative analysis of the method of the present invention under multiple modulation recognition and multiple signal-to-noise ratios in simulation experiments.

[0065] In Table 1, a higher accuracy value indicates a more precise recognition result. As shown in Table 1, the network proposed in this invention achieves an 80% recognition accuracy at signal-to-noise ratios (SNR) above 8dB, and can perform modulation recognition tasks across multiple classes and SNRs. Although the network's accuracy is slightly lower, it significantly surpasses other networks in terms of parameter count and computation time, making it more suitable for deployment scenarios. Ultimately, its weight file size is only 672KB, consistent with the quantized results of most networks. This experiment demonstrates the feasibility of this invention in modulation recognition tasks, effectively reducing hardware and storage resource burdens.

[0066] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A micro-intelligent modulation recognition method based on a resampling mechanism, characterized in that: Includes the following steps: S1. Divide the collected communication signals into a detection dataset and a training dataset; S2. Construct a miniature intelligent modulation recognition network based on a resampling mechanism; The micro-intelligent modulation recognition network includes a resampling module, a residual module, and an expanded residual module connected in sequence, and a self-attention module and a fully connected layer connected in sequence to the outputs of the residual module and the expanded residual module; The resampling module samples the input data at equal intervals and superimposes the sampled data on the channel layer to change the size of the input data and obtain compressed data. The residual module mines the relationships between different dimensions of the compressed data and retains the original signal information during forward propagation. The dilated residual module includes residual convolution and dilated convolution. The outputs of the residual module and the dilated residual module are summed to obtain the signal after deep and shallow feature fusion. The self-attention module extracts and integrates features from the signal after the fusion of deep and shallow layer features; The fully connected layer converges the output features of the self-attention module to obtain the modulation feature determination result of the signal; S3. Train the micro-intelligent modulation recognition network by inputting the multi-type multi-signal-noise ratio modulation signal training dataset into the micro-intelligent modulation recognition network, using the Adam optimizer and gradient descent method for iterative training, and obtaining the micro-modulation recognition network model weight file for multi-type multi-signal-noise ratio modulation recognition after reaching the target. S4. Use the micro-intelligent modulation recognition network based on the resampling mechanism to identify the signal category.

2. The micro-intelligent modulation recognition method based on a resampling mechanism according to claim 1, characterized in that: Step S1 includes the following steps: S11. Classify and integrate the collected communication signal data according to different signal-to-noise ratios and different modulation methods; S12. Each signal is 1024 in length, with a corresponding label ranging from 0 to the number of categories in the dataset. Additional labels are added to the SNR to obtain a multi-signal-to-noise ratio modulation recognition dataset. S13. Select 30% of the data as the detection dataset and 70% as the training dataset.

3. The micro-intelligent modulation recognition method based on a resampling mechanism according to claim 1, characterized in that: The resampling module converts the communication signal data (M1, L1, N1) into (M1', L1', N1), where M1 is the number of communication signal channels acquired, M1 is 1, L1 is the length of the acquired communication signal, N1 is the width of the acquired communication signal, M1' is the number of signal channels output by the resampling module, M1'≥2, and L1' is the length of the signal output by the resampling module, L1'=L1 / M1'. The resampling module includes at least two layers of network modules with downsampling capabilities.

4. The micro-intelligent modulation recognition method based on a resampling mechanism according to claim 1, characterized in that: The number of residual modules is five. Each residual module includes five convolutional layers and one max pooling layer. The output of the first convolutional layer is added to the output of the third convolutional layer and one branch of the result is input to the fourth convolutional layer. The other branch is added to the output of the fifth convolutional layer. The output of the fifth convolutional layer is sent to the max pooling layer. The kernel size of the first convolutional layer is 1. ; The last four convolutional kernels of the first residual module are all of the same size. The number of convolutional kernels is 16; the last four convolutional kernels of the second, third, fourth, and fifth residual modules are all [missing information - likely related to kernel size]. Both have 16 channels.

5. The micro-intelligent modulation recognition method based on a resampling mechanism according to claim 4, characterized in that: The structure of the dilated residual module is the same as that of the residual module, except that the second, third, fourth, and fifth network modules inside the dilated residual module are replaced with dilated convolutional layers.

6. The micro-intelligent modulation recognition method based on a resampling mechanism according to claim 1, characterized in that: The self-attention module includes: Q branch, K branch, and V branch; The Q branch is a single convolutional kernel. The convolutional layers are used to learn the channel-level features of the signal after the fusion of deep and shallow features. The K branch and the V branch are two consecutive convolutional kernels with a size of [missing value]. The convolutional layer is used to calculate the positional and planar features of the signal after the fusion of deep and shallow features; The Q branch, K branch, and V branch share input features. The Q branch result is element-wise multiplied with the K branch result and then multiplied with the V branch result after passing through the Softmax function. The result after multiplication is finally fused with the deep and shallow layer features through the ReLU activation layer and then added to obtain the output result of the self-attention module.

7. The micro-intelligent modulation recognition method based on a resampling mechanism according to claim 6, characterized in that: The output of the self-attention module is: ； in, , and The features extracted from the Q branch, K branch, and V branch are respectively. For the number of channels, This is an element-wise multiplication operation.

8. The micro-intelligent modulation recognition method based on a resampling mechanism according to claim 7, characterized in that: The weight gradient update method for branch V is as follows: ； in, branch for V The weight parameters, , and For constants during parameter update, by and control; Branches and The weight gradient update method for the branch is the same as that for the V branch.

9. The micro-intelligent modulation recognition method based on a resampling mechanism according to claim 1, characterized in that: The fully connected layer includes a first fully connected layer and a second fully connected layer connected in sequence. The first fully connected layer flattens the output of the self-attention module and outputs it to the second fully connected layer to obtain the number of signal categories. The final category recognition probability is obtained by the Softmax function, and the highest probability is taken as the final signal category recognition result. The number of neurons in the second fully connected layer is the same as the total number of categories of the communication signals to be identified.

10. A micro-intelligent modulation recognition method based on a resampling mechanism according to claim 3, characterized in that: When the acquired signal includes both I and Q signals, N1 is 2; The resampling module converts the (1,1024,2) I / Q signal data into (4,256,2). The first residual module outputs (16,128,1) data, the second residual module outputs (16,64,1) data, the third residual module outputs (16,32,1) data, the fourth residual module outputs (16,16,1) data, and the fifth residual module and the dilated residual module both output (16,8,1) data. The input length of the first fully connected layer is 128 and the output length is 64, and the output length of the second fully connected layer is 24.

Images