Automatic modulation recognition method for large dynamic signal-to-noise ratio

Abstract

This invention discloses an automatic modulation identification method for high dynamic signal-to-noise ratio (SNR) conditions, belonging to the field of modulation detection technology. The method includes the following steps: acquiring modulation signal data under different SNR conditions; dataset partitioning and preprocessing; constructing a dual-channel fusion neural network; and identifying the modulation type to be classified. This invention improves the representation ability of modulation signal data under low SNR conditions by fusing the SNR category of the received signal with the signal convolutional features, thereby achieving accurate identification of modulation signal data under different SNR conditions. This invention effectively improves the adaptability of modulation identification algorithms under high dynamic SNR conditions, ensuring both accuracy and efficiency, thus providing a guarantee for intelligent identification of modulation signals.

Description

Technical Field

[0001] This invention belongs to the field of modulation detection technology, specifically relating to an automatic modulation identification method for large dynamic signal-to-noise ratio conditions. Background Technology

[0002] Automatic modulation identification technology is an important aspect of cognitive radio. By identifying and analyzing received signals, it obtains information such as the modulation type and modulation parameters, providing more information for subsequent signal analysis and processing. It has significant and far-reaching implications for electronic warfare, radio management, and military reconnaissance.

[0003] Traditional modulation identification methods are mostly tested based on high signal-to-noise ratio (SNR) data under ideal conditions. However, in practical wireless communication applications, current communication systems are characterized by diverse application scenarios, complex propagation environments, and dense communication equipment. This causes noise to affect the signal during transmission and reception, resulting in large SNR fluctuations and low SNR in the received signal. Consequently, traditional identification methods struggle to accurately identify the modulation type under these conditions. Therefore, accurately identifying the modulation type of a signal under large dynamic SNR has become a hot topic and a challenge in automatic modulation identification technology, possessing significant theoretical and engineering application value for national development, people's livelihood, and national defense.

[0004] The patent "Modulation Signal Recognition Method Based on Curriculum Learning" (CN 110300078 A) provides a method for recognizing modulation signals based on curriculum learning. This method involves acquiring a training modulation signal sampling sequence and corresponding labeled data, and preprocessing the sampling sequence. A deep residual network is constructed, with the preprocessed sampling sequence used as input and the labeled data of the sampling sequence serving as the modulation type corresponding to the largest component in the deep residual network's output vector. The constructed deep residual network is trained using a curriculum learning training strategy, resulting in a trained network. The grayscale image of the modulation signal to be recognized is then used as input to the trained network, and the modulation type corresponding to the largest component in the network's output vector is the recognized modulation type. By using a curriculum learning training strategy to guide the training of the deep residual network, overfitting to data noise is avoided, allowing the network to learn a more robust model more quickly. This improves recognition performance in noisy environments and avoids the problem of low recognition rates caused by signal noise in existing networks.

[0005] The patent "Method for Modulation Signal Recognition Based on Complexity Features under Low Signal-to-Noise Ratio" (CN 104796365 A) provides a method for identifying modulation signals based on complexity features under low signal-to-noise ratio. This method involves discretizing the received unknown signal to obtain a discrete signal sequence, then recombining this sequence to obtain a reconstructed signal sequence. Multifractal dimension operations are then performed on the reconstructed signal sequence to obtain the multifractal dimension features of the modulation signal. The extracted multifractal dimension features of the unknown signal are then compared with the multifractal dimension features of known communication modulation signals calculated in a database using grey relational analysis. The modulation type of the signal with the highest correlation is selected as the modulation type of the unknown signal, thus achieving modulation type classification and identification. This invention does not require long-term signal observation or a large number of signal samples. The calculation method is simple, and different signal sequence grouping methods and different fractal dimension selection methods can be used to highlight different signal features, achieving the goal of identifying different communication signal modulation types under low signal-to-noise ratio.

[0006] The main drawbacks of existing technologies are as follows: First, most existing modulation type identification methods are tested based on high signal-to-noise ratio (SNR) data under ideal conditions. When the received signal has noise, resulting in a low SNR, the modulation identification method cannot effectively extract the signal features during the feature extraction stage due to noise interference, leading to poor identification accuracy. Second, when the SNR fluctuates significantly, existing identification algorithms have poor adaptability and low model robustness. Third, traditional modulation identification methods require a large amount of complex computation and prior knowledge, resulting in high algorithm complexity and failing to meet the requirements of real-time, high-efficiency detection. Summary of the Invention

[0007] In view of the above-mentioned technical problems in the prior art, the present invention proposes an automatic modulation identification method for large dynamic signal-to-noise ratio. The method is reasonably designed, overcomes the shortcomings of the prior art, and has good results.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] An automatic modulation identification method for high dynamic signal-to-noise ratio conditions includes the following steps:

[0010] Step 1: Acquire modulation signal data under different signal-to-noise ratios;

[0011] Step 2: Dataset partitioning and preprocessing;

[0012] Step 3: Construct a dual-channel fusion neural network;

[0013] Step 4: Identify the modulation type to be classified.

[0014] Preferably, in step 1, 14 different types of modulation signals are generated through simulation, including 2FSK, 4FSK, 8FSK, 16FSK, BPSK, QPSK, 8PSK, 16PSK, 2ASK, 4ASK, 8ASK, 16ASK, 16QAM, and 64QAM. Each signal has 10,000 samples, totaling 140,000 sample data. Each signal data consists of two signals, I and Q, with 1,024 sampling points per signal. The data format is 140,000 * 2 * 1024. The signal-to-noise ratio (SNR) of each signal ranges from -20dB to 20dB, increasing in 2dB increments. The number of modulation signals is the same for each SNR.

[0015] Preferably, in step 2, the simulated modulation signal data is evenly divided into a training set and a test set in a 7:3 ratio. The training set and the test set contain modulation signal data with 14 different modulation types and 20 signal-to-noise ratios. The training set contains a total of 98,000 data samples, and the test set contains a total of 42,000 data samples.

[0016] Preferably, in step 2, normalization is used to preprocess the input data. After normalization, the preprocessed data is limited to the range [-1, 1], and all indicators are on the same order of magnitude. The normalization transformation function is:

[0017]

[0018] Where max(x) is the maximum value in the sample data x, and min(x) is the minimum value in the sample data x.

[0019] Preferably, in step 3, the dual-channel fusion neural network is a neural network composed of two sub-networks; the network model includes an input layer, a signal-to-noise ratio classification sub-network, a feature extraction sub-network, a feature fusion layer, a fully connected layer, and an output layer;

[0020] The specific construction steps are as follows:

[0021] Step 3.1: Input the received IQ data of the modulated signal into two sub-networks respectively. The two sub-networks process in parallel and output the corresponding feature maps.

[0022] Step 3.2: Perform feature fusion of the two sub-networks along the channel dimension;

[0023] Channel-dimensional connections require the two feature maps to be the same size. Therefore, the signal-to-noise ratio category output by the signal-to-noise ratio classification sub-network is copied and expanded to the same size as the feature map output by the feature extraction sub-network. This allows the feature maps of the two sub-networks to be connected in the channel dimension to form a fused feature layer.

[0024] Step 3.3: Input the fused feature layer into the fully connected layer for modulation category determination;

[0025] Step 3.4: Finally, the output layer outputs the final modulation category;

[0026] The dual-channel fusion neural network model is trained using the training set, and the optimal training model is obtained by continuously optimizing the loss function.

[0027] Preferably, in step 3.1, the signal-to-noise ratio (SNR) classification subnetwork is used to determine the SNR of the input signal and output the SNR category; it includes two convolutional neural network layers and a Softmax layer; the convolutional neural network layers extract SNR feature information from the input data, and the Softmax layer is an effective multivariate classifier that uses the subtle SNR features extracted by the convolutional neural network layers to calculate the probability value of each SNR category and selects the largest component as the final output component, thereby completing the identification of the SNR category.

[0028] Preferably, in step 3.1, the feature extraction sub-network is used to perform global feature extraction on the input signal to obtain convolutional features hidden deep within the signal; it includes four structural residual layers of different sizes; the structural residual layer includes two sequential convolution operations, and after each convolution, a ReLU activation function is used, and a short-circuit connection is added to add the input feature map x to the output feature map F(x) after the convolution operation, thereby obtaining the total feature map output H(x), which forces the network to learn the difference between the input and the output during the training process; the four residual layers extract features from the input data and output a convolutional feature map containing signal features.

[0029] Preferably, based on the optimally trained dual-channel fusion neural network, the modulation signal data to be classified is input into the dual-channel fusion neural network, and the modulation type corresponding to the largest component in the output vector of the dual-channel fusion neural network is the identified modulation type.

[0030] The beneficial technical effects of this invention are as follows:

[0031] This invention abandons the previous method of using a single neural network for modulation signal type identification. Instead, it designs a dual-channel fusion neural network to quickly and accurately identify modulation signal types under high dynamic signal-to-noise ratios (SNR). By fusing the SNR category of the received signal with the signal convolutional features, it improves the representation ability of modulation signal data under low SNR, thereby achieving accurate identification of modulation signal data under different SNRs. This invention effectively improves the adaptability of modulation recognition algorithms under high dynamic SNR, ensuring both accuracy and efficiency, thus providing a guarantee for intelligent identification of modulation signals. Attached Figure Description

[0032] Figure 1 This is a flowchart of an automatic modulation recognition algorithm for high dynamic signal-to-noise ratio.

[0033] Figure 2 This is a diagram of a dual-channel fused neural network model.

[0034] Figure 3 This is a schematic diagram of a structured residual layer. Detailed Implementation

[0035] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0036] To overcome the shortcomings of existing technologies, such as low recognition rate and poor robustness under high dynamic signal-to-noise ratios, this invention aims to propose a stable and efficient modulation signal recognition method. A dual-channel fusion neural network recognition model is proposed. After receiving signal data, the network first inputs the data into two parallel sub-networks. One sub-network is a signal-to-noise ratio classification sub-network, used to determine the signal-to-noise ratio of the input signal and output the signal-to-noise ratio category. The other sub-network is a feature extraction sub-network, whose main function is to perform global feature extraction on the input signal, obtaining convolutional features hidden deep within the signal. The two sub-networks process in parallel, ensuring efficiency while extracting different features from the signal. Then, the feature maps extracted by the two sub-networks are fused along the channel dimension. The fused features are used for final modulation recognition and classification.

[0037] By fusing convolutional features and signal-to-noise ratio (SNR) features, the network can easily distinguish the modulation types of high and low SNR data. This allows for increased weighting of low SNR data in the loss function during model training, making the network focus more on low SNR data and enhancing its feature learning. This improves the network's fault tolerance when learning signal features, thereby enhancing its adaptability to different SNR levels and addressing the problem of low recognition rates caused by SNR fluctuations. Furthermore, parallel processing by two sub-networks improves algorithm efficiency, enabling real-time and efficient modulation type processing.

[0038] An automatic modulation identification method for high dynamic signal-to-noise ratio conditions, the process of which is as follows: Figure 1 As shown, the specific steps include the following:

[0039] (1) Acquire modulation signal data under different signal-to-noise ratios

[0040] Fourteen different types of modulation signals were generated through simulation, including 2FSK, 4FSK, 8FSK, 16FSK, BPSK, QPSK, 8PSK, 16PSK, 2ASK, 4ASK, 8ASK, 16ASK, 16QAM, and 64QAM. Each signal had 10,000 samples, totaling 140,000 data samples. Each signal consisted of two signals, I and Q, with 1024 sampling points per signal, resulting in a data format of 140,000 * 2 * 1024. Furthermore, the signal-to-noise ratio (SNR) ranged from -20 dB to 20 dB, increasing in 2 dB increments, with the same number of modulation signals for each SNR. To more closely resemble the real channel environment, multipath fading, additive white Gaussian noise, and impulse noise were randomly added to the simulated signal data.

[0041] (2) Dataset partitioning and preprocessing

[0042] The simulated modulation signal data is evenly divided into a training set and a test set in a 7:3 ratio. The training set and the test set contain modulation signal data of 14 different modulation types and 20 signal-to-noise ratios. Therefore, the training set of this invention contains a total of 98,000 data samples, and the test set contains a total of 42,000 data samples.

[0043] Before inputting the data into the neural network, the input data is first preprocessed; this invention uses normalization. After normalization, the preprocessed data is confined to a certain range ([-1,1]), ensuring that all indicators are on the same order of magnitude. This eliminates the adverse effects of outlier data, making it suitable for comprehensive comparative evaluation. The normalization transformation function is:

[0044]

[0045] Where max(x) is the maximum value in the sample data x, and min(x) is the minimum value in the sample data x.

[0046] (3) Constructing a dual-channel fusion neural network

[0047] The automatic modulation recognition algorithm designed in this invention for high dynamic signal-to-noise ratio conditions is a neural network consisting of two sub-networks. The network model is as follows: Figure 2 As shown, the network model consists of an input layer, a signal-to-noise ratio classification subnetwork, a feature extraction subnetwork, a fusion layer, a fully connected layer, and an output layer.

[0048] First, the IQ data of the received modulated signal are input into two sub-networks respectively. The two sub-networks process the data in parallel and output the corresponding feature maps.

[0049] The signal-to-noise ratio (SNR) classification subnetwork is a conventional classification network, consisting of two convolutional neural network layers and a Softmax layer. The convolutional neural network layers extract SNR feature information from the input data, while the Softmax layer is an effective multivariate classifier that uses the subtle SNR features extracted by the convolutional layers to calculate the probability value of each SNR category and selects the component with the largest value as the final output component, thereby completing the identification of the SNR category.

[0050] The feature extraction subnetwork consists of four structural residual layers of different sizes, such as... Figure 3 As shown, the structural residual layer consists of two sequential convolutional operations, followed by a ReLU activation function after each convolution. Additionally, a shortcut connection is added to add the input feature map x to the output feature map F(x) after the convolution operation, resulting in the total feature map output H(x), which forces the network to learn the difference between the input and output during training. The four residual layers extract features from the input data and output convolutional feature maps containing signal features.

[0051] Then, feature fusion is performed between the two sub-networks along the channel dimension. Channel-dimensional connection requires the two feature maps to be of the same size. Therefore, the signal-to-noise ratio (SNR) category output from the SNR classification sub-network is copied and expanded to the same size as the feature map output from the feature extraction sub-network. This allows the feature maps of the two sub-networks to be connected along the channel dimension, forming a fused feature layer. The fused feature layer is then input into a fully connected layer for modulation category determination, and finally, the output layer outputs the final modulation category.

[0052] The dual-channel fusion neural network model is trained using the training set, and the optimal training model is obtained by continuously optimizing the loss function.

[0053] (4) Identify the modulation type to be classified

[0054] Based on the optimally trained dual-channel fusion neural network, the modulation signal data to be classified is input into the neural network, and the modulation type corresponding to the largest component in the network's output vector is the identified modulation type.

[0055] The key points and protection points of this invention include:

[0056] (1) The test data of this invention comes from modulation signal samples under different signal-to-noise ratios.

[0057] (2) In view of the complex data features under different signal-to-noise ratios, the present invention designs and uses a dual-channel fusion neural network to mine the deep-seated stable features contained in the modulation signal, so as to achieve accurate classification of the modulation signal category.

[0058] (3) The present invention calculates the signal-to-noise ratio category of the received signal and uses fused features for model training and learning. During the training process, the weight of the loss function for low signal-to-noise ratio data is increased to strengthen the feature learning of low signal-to-noise ratio data and improve the model's adaptability under different signal-to-noise ratios.

[0059] (4) The present invention designs a dual-channel feature fusion method to extract different features of signal data in parallel, thereby achieving rapid identification of modulated signals.

[0060] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. An automatic modulation recognition method for large dynamic signal-to-noise ratio, characterized in that: Comprising the following steps: Step 1: Obtain modulation signal data under different signal-to-noise ratios; Step 2: Data set division and preprocessing; Step 3: Construct a dual-channel fusion neural network; the dual-channel fusion neural network is a neural network composed of two sub-networks; the network model comprises an input layer, a signal-to-noise ratio classification sub-network, a feature extraction sub-network, a feature fusion layer, a full connection layer and an output layer; The specific construction steps are as follows: Step 3.1: input the IQ data of the received modulation signal into the two sub-networks respectively, and the two sub-networks are processed in parallel and output corresponding feature maps; Step 3.2: perform feature fusion of the two sub-networks in the channel dimension; The channel dimension connection requires that the two feature maps have the same size, so the signal-to-noise ratio categories output by the signal-to-noise ratio classification sub-network are copied and expanded to the same size as the feature maps output by the feature extraction sub-network, so that the feature maps of the two sub-networks can be connected in the channel dimension to form a fused feature layer; Step 3.3: input the fused feature layer into the full connection layer to determine the modulation category; Step 3.4: finally output the final modulation category by the output layer; Use the training set to train the constructed dual-channel fusion neural network model, and obtain the optimal training model by continuously optimizing the loss function; The signal-to-noise ratio classification sub-network is used to judge the signal-to-noise ratio of the input signal and output the signal-to-noise ratio category; it comprises two convolutional neural network layers and a Softmax layer; the convolutional neural network layer extracts signal-to-noise ratio feature information from the input data, the Softmax layer is an effective multivariate classifier, which uses the signal-to-noise ratio fine features extracted by the convolutional neural network layer to calculate the probability values of each signal-to-noise ratio category, and selects the maximum component as the final output component, thereby completing the recognition of the signal-to-noise ratio category; Step 4: identify the modulation type to be classified.

2. The method of automatic modulation recognition for large dynamic signal-to-noise ratio according to claim 1, characterized in that: In step 1, 14 different types of modulation signals including 2FSK, 4FSK, 8FSK, 16FSK, BPSK, QPSK, 8PSK, 16PSK, 2ASK, 4ASK, 8ASK, 16ASK, 16QAM and 64QAM are generated by simulation, each signal has 10000 samples, a total of 140000 sample data, each signal data is composed of IQ two signals, each signal has 1024 sampling points, the data form is 140000*2*1024, and the signal-to-noise ratio of each signal ranges from -20dB to 20dB, with a progression of every 2dB, and the number of modulation signals at each signal-to-noise ratio is the same.

3. The method of claim 1, wherein the method is for automatic modulation recognition in a large dynamic signal-to-noise ratio. In step 2, the simulation generated modulation signal data is uniformly divided into training set and test set according to the ratio of 7:3, the modulation signal data in the training set and the test set contains 14 different modulation types and 20 signal-to-noise ratios, the training set contains a total of 98000 data samples, and the test set contains a total of 42000 data samples.

4. The method of automatic modulation recognition for large dynamic signal-to-noise ratio according to claim 1, characterized in that: In step 2, the input data is preprocessed by normalization, and the original data is normalized to limit the preprocessed data in the range of [-1, 1] and make each index in the same order of magnitude; the conversion function of normalization is: ； wherein is the maximum value in the sampled data is the minimum value in the sampled data is the minimum value in the sampled data is the minimum value in the sampled data 5. The method of automatic modulation recognition for large dynamic signal-to-noise ratio according to claim 1, characterized in that: In step 3.1, the feature extraction sub-network is used for global feature extraction of the input signal to obtain the deep convolution features contained in the signal; it includes four layers of residual layers with different sizes; the residual layer includes two sequential convolution operations, and uses the ReLu activation function after each convolution, and adds a short circuit connection to add the input feature map x and the output feature map F(x) after convolution operation, so as to obtain the total feature map output H(x), so as to force the network to learn the difference between input and output in the training process; the four-layer residual layer extracts the features of the input data and outputs the convolution feature map containing the signal features.

6. The method of automatic modulation recognition for large dynamic signal-to-noise ratio according to claim 1, characterized in that: In step 4, based on the trained optimal dual-channel fusion neural network, the modulation signal data to be classified is input into the dual-channel fusion neural network, and the modulation type corresponding to the maximum component in the output vector of the dual-channel fusion neural network is the recognized modulation type.

Images