A modulation mode intelligent identification method based on lightweight network
By introducing reverse residual structure and linear bottleneck structure in intelligent identification of modulation mode, combining the average pooling layer and width factor hyperparameters, the contradiction between network accuracy and lightweight requirements is solved, and efficient modulation mode identification is achieved. for mobile devices and wireless communication networks.
Patent Information
- Application Number
- CN202210503877.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-10
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2042-05-10
AI Technical Summary
The existing intelligent identification technology of modulation methods has a contradiction between the accuracy of network models and the need for lightweight scale, making it difficult to weigh the accuracy and complexity, especially in mobile applications where computational complexity and accuracy requirements are high. In the following, it is difficult to meet the needs of wireless communication networks.
Using a modulation method based on lightweight network intelligent identification method, by introducing reverse residual structure and linear bottleneck structure, combining the average pooling layer, the amount of network parameters is reduced, and the network input and output dimensions are adjusted through the width factor hyperparameters to reduce the computational complexity and ensure prediction accuracy.
While reducing the network scale, it ensures the performance of modulation method identification, significantly reduces the computational complexity and parameter volume, and improves the network convergence speed. It is suitable for mobile devices and lightweight wireless communication networks.
Smart Images

Figure CN114912486B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of wireless communications, and in particular, relates to a modulation mode intelligent identification method based on a lightweight network. Background Art
[0002] Intelligent recognition of modulation mode is a vital part of wireless communication networks and one of the main methods of cognitive radio. Intelligent recognition of modulation mode is based on signals received from different devices, identifying their different modulation types, which helps to recover the transmitted information and determine the appropriate demodulation method. At the same time, it can improve the efficiency of spectrum utilization when spectrum resources are scarce, and is an indispensable intelligent communication technology in wireless communication. Existing intelligent recognition technologies for modulation mode can be divided into two types: model-driven methods and data-driven methods. Model-based methods are further divided into likelihood function-based methods and feature-based methods. Likelihood function-based methods treat modulation mode intelligent recognition as a hypothesis testing problem. Feature-based methods aim to find better features of received signals. However, these two methods have high computational complexity and require complete prior knowledge.
[0003] In recent years, data-driven methods can directly capture information related to the difference between categories from the differences between raw data to perform modulation classification. For example, the automatic modulation classification scheme based on deep learning for I / Q samples has attracted widespread attention due to its superior accuracy. In their paper "Over-the-air deep learning based radio signal classification" (IEEE Journal of Selected Topics in Signal Processing., vol. 12, no. 1, pp. 168-179, 2018), O'Shea TJ, Roy T, and Clancy TC proposed two customized network architectures, VGG and ResNet, based on the RadioML2018.01A dataset. It uses the in-phase part and the orthogonal part (I / Q) of the received signal as input for feature extraction, and experimentally proves that the accuracy is significantly improved under this data-driven method. However, these networks are large in scale, and the number of parameters and computational complexity have high requirements on computer equipment, which exceeds the capabilities of many mobile and portable devices. In the paper "Sparsely connected CNN for efficient automatic modulation recognition" (IEEE Transactions on Vehicular Technology, vol. 69, no. 12, pp. 15557-15568, 2020.), Tunze GB, Huynh-The T, Lee JM et al. proposed a network architecture, SCGNet, that uses deep convolution and point convolution in lightweight network MobilNets to extract deep features, ensuring that data details can be retained as the network model deepens, while also reducing computational complexity. However, the accuracy of the network has decreased.
[0004] The above methods have the contradiction between network accuracy and complexity, which makes them difficult to use in practice. How to strike a balance between the two, while ensuring the deployment of efficient deep learning architecture, reduce the complexity (i.e., training time, computing cost, and memory cost) in mobile applications, and meet the lightweight requirements of future wireless communication networks, urgently requires the development of new modulation mode intelligent recognition solutions.
[0005] Therefore, it is necessary to improve the problems existing in the above-mentioned prior art to solve the shortcomings of the prior art. Summary of the invention
[0006] In view of the contradiction between the accuracy of the network model and the demand for lightweight network scale faced by the existing modulation mode intelligent identification technology in actual communication scenarios, the present invention proposes a modulation mode intelligent identification method based on a lightweight network. While greatly reducing the network scale parameters, the prediction accuracy of the network is guaranteed and the network convergence speed is fast.
[0007] To achieve the above object, the modulation mode intelligent identification method based on lightweight network of the present invention comprises the following steps:
[0008] Step 1: signal data acquisition, performing vector conversion processing on the in-phase part I and the orthogonal part Q of the received signal to form a modulation identification signal;
[0009] Step 2: Create a category label, and create a corresponding category label file based on the collected modulation identification signal;
[0010] Step 3: Preprocess I / Q data;
[0011] Step 4: Build the network model;
[0012] Step 5: Use training data to train the network model;
[0013] Step 6: Determine whether the network training is finished. If so, execute step 7. If not, increase the number of training iterations by one and continue to train the network in step 5.
[0014] Step 7: Input the test set data into the network;
[0015] Step 8: Output the test results.
[0016] Furthermore, in step 1, the modulation classification of the received signal can be determined as a K-type hypothesis test, assuming that the received signal H under the k-th modulation hypothesis k for
[0017] H k :x k (n) = s k (n)+ω k (n),n=1,2,...,N
[0018] Among them, s k (n) represents the transmission signal, x k (n) represents the received signal, N is the number of signal symbols, ω k (n) has a mean of zero and a variance of σ 2 Additive Gaussian white noise,
[0019] Will receive signal x k The in-phase part I and the orthogonal part Q of (n) are directly input into the neural network without normalization.k (n) converted to vector x k , representing the I / Q signal samples as a vector, given as
[0020]
[0021] Among them, I k and Q k They represent the in-phase and quadrature parts of the received signal respectively.
[0022] Furthermore, in step 3, after receiving the signal sample, features are extracted from the original data using a deep learning solution, which is input into a fully connected layer to convert the feature dimension. Feature learning is represented by converting the original data into The process of mapping to the L-dimensional vector y is given by
[0023]
[0024] Among them, the mapping function f represents the feature learning model with a fully connected layer, Represents x k In the vector space, y represents the feature vector output by the fully connected layer. Represents the vector space where y is located.
[0025] Furthermore, step 4 includes the following steps:
[0026] Step 4a, in the input stage, the raw data of the signal samples are converted into tensors to match the requirements of the deep learning framework;
[0027] Step 4b, build a network using a linear bottleneck structure and an inverse residual structure to extract features from the input signal, and use an average pooling layer to adapt to the input length;
[0028] Step 4c, set the width factor hyperparameter to adjust the input and output dimensions, thereby further adjusting the network parameters;
[0029] Further, in step 4b,
[0030] The network structure adopts a linear bottleneck structure and an inverse residual structure. The inverse residual structure constructed by deep convolution and point-by-point convolution fully extracts data features, reduces the number of parameters and ensures its accuracy. Batch normalization is used after each convolution layer to speed up the training convergence speed, and the ReLU6 nonlinear function is used for activation. The network structure includes one convolution layer, 14 linear bottleneck structures, a convolution kernel size of 3×3, and batch normalization is added during the training process.
[0031] First, a 3×3 convolutional layer with a stride of 2 is used to increase the feature dimension from 2 to 32 through depthwise convolution.
[0032] Then, 14 linear bottleneck structures are passed in turn. Assume that the expansion factor of each linear bottleneck structure is t, the input dimension is M, and point-by-point convolution is used to increase the dimension to t×M. Then, the linear bottleneck structure is repeatedly trained. During repetition, only the step size in the first linear bottleneck structure is s, and the others are defaulted to 1. After point-by-point convolution compresses the output dimension, the linear bottleneck structure increases the dimension to 128, and then passes through a 1×1 convolution layer to increase the dimension to 512.
[0033] Finally, the average pooling layer is used to perform pooling operation to reduce the feature dimension and generate classification results. The output dimension generated by the previous linear bottleneck structure is input into the average pooling layer, and each output dimension of each corresponding channel in the last convolutional layer is averaged. The output dimension of the average pooling layer is C×1×1, where C is the number of output channels.
[0034] Furthermore, in step 4c, the width factor hyperparameter α is set to evenly refine each convolutional layer. The value range of α is 0.4 to 0.75, and the computational complexity and parameter amount of the overall network are reduced to the original α. 2 .
[0035] Furthermore, in step 4c, the optimal value of α is 0.5.
[0036] Furthermore, in step five,
[0037] First, the trainable parameters of the network are randomly initialized. Since the sequence is always most correlated with the data point closest to the current time, the farther the distance, the lower the correlation. Therefore, the initial trainable parameters of the network are randomly initialized according to the monotonically decreasing law.
[0038] Secondly, the training data is input into the network training in batches. The batch size can be adjusted. The training error of each batch is back-propagated to optimize the network parameters. The error loss function uses the cross entropy loss function to measure the gap between the model output and the true modulation mode label, which can be written as
[0039]
[0040] Among them, K represents the number of classes, N represents the number of samples, and p j,k is the output of the Softmax function, indicating the probability that sample i belongs to the kth class, q j,k is an indicator variable, given by
[0041]
[0042] When all training data are trained once, it is considered as one training iteration.
[0043] The beneficial effects of the present invention are:
[0044] First, since the present invention introduces an inverse residual structure and a linear bottleneck structure, and uses an average pooling layer instead of a traditional fully connected layer, the network scale is greatly reduced and the number of parameters is greatly reduced compared to the traditional deep learning framework.
[0045] Second, since the linear bottleneck structure in the modulation mode intelligent identification method based on a lightweight network designed by the present invention is composed of deep convolution and point-by-point convolution, the loss of data details can be avoided as the network goes deeper, which greatly reduces the network parameters while ensuring the performance of modulation mode recognition.
[0046] Third, the width factor hyperparameter is introduced to adjust the dimension of network input and output, thereby further adjusting the quantity and accuracy of network parameters and improving the flexibility of network scale. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] Figure 1 is a flow chart of the present invention;
[0048] Figure 2 This is a diagram of a modulation mode intelligent identification network framework of the present invention;
[0049] Figure 3 It is a comparison chart of network scale parameters using the present invention and other existing technologies;
[0050] Figure 4 It is a comparison chart of prediction accuracy of the present invention and other existing technologies under different signal-to-noise ratio conditions;
[0051] Figure 5 This is a comparison chart of the training convergence speed using the present invention and other existing technologies. DETAILED DESCRIPTION
[0052] The present invention will be further described below in conjunction with the accompanying drawings.
[0053] Combined with Figure 1 The specific steps of the method of the present invention are described as follows:
[0054] Step 1: Signal data acquisition and processing.
[0055] Modulation classification can be determined as a K-type hypothesis test, where K represents the number of modulation schemes. The received signal H under the k-th modulation hypothesis is k for
[0056] H k :x k (n) = s k (n)+ω k (n),n=1,2,...,N
[0057] Among them, s k(n) and x k (n) represents the transmission signal and the received signal respectively. N is the number of signal symbols. ω k (n) has a mean of zero and a variance of σ 2 Additive white Gaussian noise.
[0058] The in-phase part and the quadrature part (I / Q) of the received signal are used simultaneously. These two parts obey the same independent distribution and can be directly input into the network model in step 4 without normalization. k (n) converted to vector x k , the I / Q signal samples can be represented as a vector, given as
[0059]
[0060] In English, the first two letters of Real are "Re" to represent the real part of a complex number, and the first two letters of Imaginary are "Im" to represent the imaginary part of a complex number. k and Q k Respectively represent the in-phase part and the orthogonal part of the received signal, the original data x k It can be specifically expressed in the form of a matrix, given as
[0061]
[0062] Wherein, N is the number of signal symbols.
[0063] Step 2: Create a corresponding category label file based on the collected modulation identification signal.
[0064] Eight different digital modulation modes are used for training and testing in the patent of this invention, namely 8PSK, BPSK, CPFSK, GFSK, PAM4, 16QAM, 64QAM, and QPSK, and the signal noise is additive white Gaussian noise (AWGN).
[0065] Step 3: I / Q data preprocessing.
[0066] After receiving a large number of signal samples, CNN and RNN solutions based on deep learning are used to extract features from the original data. Then, the extracted features are input into the fully connected layer (FC) for information integration, and then the feature dimension is converted. Feature learning can be expressed as converting the original data The process of mapping to the L-dimensional vector y is given by
[0067]
[0068] Among them, the mapping function f represents the feature learning model with a fully connected layer, Represents x k In the vector space, y represents the feature vector output by the fully connected layer. Represents the vector space where y is located.
[0069] Step 4: Build the network model.
[0070] In the input stage, the raw data of signal samples are converted into tensors, and the visual feature space is used as the embedding space to match the requirements of the deep learning framework.
[0071] Both the linear bottleneck structure and the reverse residual network use depthwise separable convolution to reduce the amount of computation. After using depthwise separable convolution, the interaction between the number of output channels and the size of the convolution kernel is broken. Depthwise separable convolution consists of two parts: depthwise convolution and pointwise convolution. Depthwise convolution applies a single filter on each input channel, and then uses pointwise convolution to create a linear combination of depthwise outputs. If the number of input channels is M, the number of output channels is C, and the convolution kernel size is D k ×D k and feature map size is D F ×D F , the computational cost of depth-wise separable convolution is that of standard convolution:
[0072]
[0073] The linear bottleneck structure evolved from the bottleneck structure and was first used in MobileNetV2. Assuming the number of input channels is M and the expansion factor is t, the first layer of this structure uses point-by-point convolution for dimensionality increase; the second layer uses a 3×3 convolution kernel for depth convolution; the third layer uses point-by-point convolution again, associates the features after depth convolution and outputs the specified number of channels C. There are two cases for using the linear bottleneck structure. The first is to use the residual structure when the step size is 1, and the second is not to use the residual structure when the step size is 2. Through the linear bottleneck structure, the number of parameters and the amount of convolution calculation are effectively reduced relative to the standard convolution. The network is optimized in terms of space and time.
[0074] The reverse residual structure is optimized based on the residual structure (Residuals) in ResNet. In fact, it adds residual propagation on the basis of the linear bottleneck structure. The reverse residual structure uses the first layer of point-by-point convolution to increase the dimension, then uses depth convolution, and then uses point-by-point convolution to reduce the dimension. This is exactly the opposite of the residual structure in ResNet. Such convolution operations are more conducive to reducing the number of parameters and the amount of calculation.
[0075] Based on the above two structures, combined with the attached Figure 2 , we built a network structure for automatic modulation recognition. The network architecture consists of a convolutional layer followed by Figure 2The 14 linear bottleneck structures described in . We use ReLU6 as the nonlinear activation function. All the above convolution kernels have a size of 3×3 and batch normalization is added during training.
[0076] First, a 3×3 convolution layer with a stride of 2 is used to increase the feature dimension from 2 to 32. Then, 14 linear bottleneck structures are passed in sequence. Assume that the expansion factor of each linear bottleneck structure is t, which means that point-by-point convolution is used to increase the dimension to t×M first, and then the linear bottleneck structure is repeated. When repeated, only the step size in the first linear bottleneck structure is s, and the rest are 1 by default. After using deep convolution, it is compressed to the output dimension through point-by-point convolution. The linear bottleneck structure increases the number of dimensions to 128. In this process, the inverse residual structure constructed by deep convolution and point-by-point convolution fully extracts data features, reduces the number of parameters and ensures its accuracy. Batch normalization (BN) is used after each convolution layer to speed up the training convergence speed, and the ReLU6 nonlinear function is used for activation. The convolution operation in the linear bottleneck structure is as follows:
[0077]
[0078] For a linear bottleneck structure with size h×w, expansion factor t, kernel size k, input channels M, and output channels C, the required computational effort is:
[0079] h·w·M·t(M+k 2 +C)
[0080] Compared with the traditional fully connected layer, we use the average pooling layer (GAP) for pooling operation to reduce the feature dimension and generate classification results. The feature map generated by the previous linear bottleneck structure is input into the average pooling layer, and each feature map of each corresponding channel in the last convolutional layer is averaged. The output dimension of the average pooling layer is C×1×1. This method can effectively alleviate the occurrence of overfitting problems.
[0081] Finally, the width factor hyperparameter α is introduced to evenly refine each network at each layer. At this time, the number of input channels is αM, the number of output channels is αC, and the computational complexity of the depthwise separable convolution is:
[0082] D k ·D k αM D F ·D F +αM·αC·D F ·D F
[0083] α is generally set to 0.4 to 0.75, which can reduce the amount of computation and parameters of the overall network by α 2, which allows for more flexibility in building smaller and faster models. Our network finally chose 0.5 as the width factor, which significantly reduced the number of parameters while maintaining high accuracy.
[0084] Step 5: Use the training data to train the network model.
[0085] In the first step, the trainable parameters of the network are randomly initialized. Since the sequence is always most correlated with the data point closest to the current time, the farther the distance, the lower the correlation. Therefore, the initial trainable parameters of the network are randomly initialized according to the monotonically decreasing law. The number of epochs for initial network training is equal to 1, the maximum number of iterations is 40, and the learning rate is 0.01.
[0086] In the second step, the training data is fed into the network for training in batches. The batch size can be adjusted, and the training error of each batch is back-propagated to optimize the network parameters. The error loss function uses the cross entropy loss function to measure the gap between the model output and the true modulation mode label, which can be written as
[0087]
[0088] Where K represents the number of classes and N represents the number of samples. i,k is the output of the Softmax function, indicating the probability that data sample i belongs to category k. i,k is an indicator variable, given by
[0089]
[0090] When all training data are trained once, it is called an epoch.
[0091] Step 6: Determine whether the network training is complete.
[0092] Determine whether the current epoch has reached the maximum epoch. If so, proceed to step 7. If not, increase the epoch by one and continue to step 5 to train the network.
[0093] Step 7: Input the test set data into the network.
[0094] Step 8: Output the test results.
[0095] The effect of the present invention is further described below in conjunction with simulation experiments.
[0096] 1. Simulation conditions and parameter settings:
[0097] The simulation experiment of the present invention is carried out on the simulation platform of Python3.6 and Pytorch1.5.0. The computer CPU model is Intel Core i5, equipped with an independent graphics card of NVIDIA GeForceGTX 1650. In the embodiment of the present invention, we used the public data set RadioML2016.10A to evaluate the results. The data is generated by GNU Radio and consists of 11 adjustment methods (8 digital modulation methods and 3 analog modulation methods), with a signal-to-noise ratio range of -20dB to 20dB and a step size of 2dB. At each different signal-to-noise ratio, each modulation type has 1000 samples, and each sample has 128 points. The entire data set consists of a total of 220,000 I / Q vectors.
[0098] Eight digital adjustment methods in the data set were selected, with the signal-to-noise ratio ranging from -6dB to 14dB. A total of 88,000 vectors were used, of which 80% were training sets and 20% were test sets. The model was trained using SGD, with an initial learning rate of 0.01, a weight decay of 5×10-4, and a maximum number of iterations of 40.
[0099] 2. Simulation content:
[0100] Attached Figure 3 This is a comparison chart of network scale parameters and computational complexity of the present invention and the prior art. Paramssize (parameter quantity), Flops (computational complexity), and Madd computational complexity were selected for comparison. It can be seen that in terms of parameter quantity, the present invention is much smaller than the convolutional neural network, and also has advantages over the residual network (ResNet) and the visual geometric neural network (VGG), which is only 0.4Mb; the Madd computational complexity is 2.22Mb, and the number of floating-point operations per second is 1.15M, which is 0.01% of the convolutional neural network, 18% of the residual network, and 30% of the visual geometric neural network. It can be seen that the network is lightweight compared to other modulation recognition schemes, and both network parameters and computational complexity have been greatly reduced.
[0101] Attached Figure 4 It is a comparison chart of modulation recognition accuracy under different signal-to-noise ratio conditions using the present invention and the prior art. Figure 4The horizontal axis in the figure represents different signal-to-noise ratios (dB), and the vertical axis represents the accuracy of modulation mode recognition. The broken line marked with a star represents the modulation recognition accuracy curve of the method of the present invention, the broken line marked with a circle represents the modulation recognition accuracy curve of the visual geometric neural network under different signal-to-noise ratios, the broken line marked with a cross represents the modulation recognition accuracy curve of the residual network under different signal-to-noise ratios, and the broken line marked with a square represents the modulation recognition accuracy curve of the convolutional neural network under different signal-to-noise ratios. The signal-to-noise ratio ranges from -6dB to 14dB, and simulation experiments are carried out using the methods of the present invention and the prior art. By comparing the accuracy of the modulation recognition methods obtained by the four methods, it can be seen that the prediction accuracy of the present invention is significantly higher than that of the traditional method. From the overall trend, the prediction accuracy increases with the increase of the signal-to-noise ratio. In a low signal-to-noise ratio environment, the present method is about 2dB better than other technical methods. When the signal-to-noise ratio is -6dB, the prediction accuracy of the present invention reaches about 55%, which exceeds the accuracy of other methods by about 15%; as the signal-to-noise ratio increases, the accuracy also increases significantly. When the signal-to-noise ratio is 0dB, the prediction accuracy of the present invention has reached the maximum value, reaching about 93%, which exceeds the accuracy of the method based on visual geometric neural network and convolutional neural network by about 15%, and exceeds the accuracy of the residual network by about 5%. In a high signal-to-noise ratio environment, the accuracy of the present invention can reach up to 95%, which exceeds the accuracy of the visual geometric neural network and convolutional neural network by about 10%, and is higher than the accuracy of the residual network by about 2%.
[0102] Attached Figure 5 This is a comparison chart of network training speed using the present invention and the prior art. Figure 5 The horizontal axis represents the number of training times, and the vertical axis represents the loss function value. The broken line marked with a star represents the training speed convergence curve of the method of the present invention, the broken line marked with a circle represents the training speed convergence curve of the visual geometric neural network, the broken line marked with a cross represents the training speed convergence curve of the residual network, and the broken line marked with a square represents the training speed convergence curve of the convolutional neural network. By comparing the training speed convergence curves obtained by these four methods, it can be seen that the training speed of the method of the present invention is significantly faster than the existing method for intelligent recognition of modulation modes. The method of the present invention completes convergence in only about 10 epochs, while the methods based on the visual geometric neural network and the convolutional neural network complete convergence in about 20 epochs.
[0103] Based on the above simulation results and analysis, the intelligent modulation mode identification method based on a lightweight network proposed in the present invention can significantly reduce the network scale and computational complexity compared with the existing methods while ensuring the performance of modulation mode identification, and the convergence speed is faster, which enables the present invention to be better applied in actual communication scenarios.
[0104] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.
Claims
1. A modulation mode intelligent identification method based on a lightweight network, characterized in that: The steps include: Step 1: signal data acquisition, performing vector conversion processing on the in-phase part I and the orthogonal part Q of the received signal to form a modulation identification signal; Step 2: Create a category label, and create a corresponding category label file based on the collected modulation identification signal; Step 3: Preprocess I / Q data; Step 4: Build the network model; Step 5: Use training data to train the network model; Step 6: Determine whether the network training is finished. If so, execute step 7. If not, increase the number of training iterations by one and continue to train the network in step 5. Step 7: Input the test set data into the network; Step 8: Output the test results; Step 4 includes the following steps: Step 4a, in the input stage, the raw data of the signal sample is converted into a tensor to match the requirements of the deep learning framework; Step 4b, build a network using a linear bottleneck structure and an inverse residual structure to extract features from the input signal, and use an average pooling layer to adapt to the input length; Step 4c, set the width factor hyperparameter to adjust the input and output dimensions, thereby further adjusting the network parameters; In step 4b, The network structure adopts a linear bottleneck structure and an inverse residual structure. The inverse residual structure constructed by deep convolution and point-by-point convolution fully extracts data features, reduces the number of parameters and ensures its accuracy. Batch normalization is used after each convolution layer to speed up the training convergence speed, and the ReLU6 nonlinear function is used for activation. The network structure includes one convolution layer, 14 linear bottleneck structures, a convolution kernel size of 3×3, and batch normalization is added during the training process. First, a 3×3 convolutional layer with a stride of 2 is used to increase the feature dimension from 2 to 32 through depthwise convolution. Then, 14 linear bottleneck structures are passed in turn. Assume that the expansion factor of each linear bottleneck structure is t, the input dimension is M, and point-by-point convolution is used to increase the dimension to t×M. Then, the linear bottleneck structure is repeatedly trained. During repetition, only the step size in the first linear bottleneck structure is s, and the others are defaulted to 1. After point-by-point convolution compresses the output dimension, the linear bottleneck structure increases the dimension to 128, and then passes through a 1×1 convolution layer to increase the dimension to 512. Finally, the average pooling layer is used for pooling operation to reduce the feature dimension and generate classification results. The output dimension generated by the previous linear bottleneck structure is input into the average pooling layer, and each output dimension of each corresponding channel in the last convolutional layer is averaged. The output dimension of the average pooling layer is C×1×1, where C is the number of output channels. In step 4c, the width factor hyperparameter α is set to evenly refine each convolutional layer. The value range of α is 0.4~0.
75. The computational complexity and parameter amount of the overall network are reduced to the original α. 2 ; In step 4c, the value of α is 0.
5.
2. According to claim 1, a modulation mode intelligent identification method based on a lightweight network is characterized in that: In step 1, the modulation classification of the received signal can be determined as a K-type hypothesis test, assuming that the received signal H under the k-th modulation hypothesis k for H k :x k (n)=s k (n)+ω k (n),n=1,2,...,N Among them, s k (n) represents the transmission signal, x k (n) represents the received signal, N is the number of signal symbols, ω k (n) has a mean of zero and a variance of σ 2 Additive Gaussian white noise, Will receive signal x k The in-phase part I and the orthogonal part Q of (n) are directly input into the neural network without normalization. k (n) is converted to vector x k , the I / Q signal samples are represented as a vector, given as Among them, I k and Q k They represent the in-phase and quadrature parts of the received signal respectively.
3. According to the method of intelligent recognition of modulation mode based on lightweight network in claim 1, it is characterized in that: In step 3, after receiving the signal sample, the feature is extracted from the original data using a deep learning solution, and the feature is input into the fully connected layer to convert the feature dimension. The feature learning is represented by converting the original data into The process of mapping to the L-dimensional vector y is given by Among them, the mapping function f represents the feature learning model with a fully connected layer, Represents x k In the vector space, y represents the feature vector output by the fully connected layer. Represents the vector space where y is located.
4. According to the method of intelligent recognition of modulation mode based on lightweight network in claim 1, it is characterized in that: In step five, First, the trainable parameters of the network are randomly initialized. Since the sequence is always most correlated with the data point closest to the current time, the farther the distance, the lower the correlation. Therefore, the initial trainable parameters of the network are randomly initialized according to the monotonically decreasing law. Secondly, the training data is input into the network training in batches. The batch size can be adjusted. The training error of each batch is back-propagated to optimize the network parameters. The error loss function uses the cross entropy loss function to measure the gap between the model output and the true modulation mode label, which can be written as Among them, K represents the number of classes, N represents the number of samples, and p j,k is the output of the Softmax function, indicating the probability that sample i belongs to the kth class, q j,k is an indicator variable, given by When all training data are trained once, it is considered as one training iteration.
