A signal modulation recognition method and system based on double-branch VGG16 network feature-level fusion

Abstract

The application discloses a signal modulation recognition method and system based on double-branch VGG16 network feature-level fusion, and belongs to the technical field of radar signal intelligent recognition. The method first extracts intra-pulse features of typical radar signals, specifically adopts two time-frequency feature analysis methods of short-time Fourier transform and continuous wavelet transform to generate corresponding feature images. Then, the images are input into two VGG16 main neural networks which are the same in structure but independent of each other to capture local texture features in the time-frequency graph. Through fusion of secondary features output by each branch network, accurate recognition of six types of modulation signals is finally realized. In order to verify the performance of the method, the system evaluates the recognition accuracy change under different signal-to-noise ratios. The experimental results show that the method based on feature-level fusion can effectively extract multi-dimensional features of radar signals, maintain high recognition performance and has good practical value.

Description

Technical Field

[0001] This invention relates to the field of intelligent radar signal recognition technology, specifically to a signal modulation recognition method and system based on feature-level fusion of a dual-branch VGG16 network. Background Technology

[0002] Radar signal modulation identification is one of the core technologies in fields such as electronic countermeasures and radar reconnaissance. Its purpose is to accurately identify the modulation type of the signal by analyzing the intra-pulse characteristics of the radar signal, so as to provide a basis for subsequent target location and threat assessment.

[0003] Traditional radar signal identification methods primarily rely on pulse descriptor words (PDWs), which contain parameters such as pulse amplitude, pulse width, and pulse repetition frequency. Signal identification is achieved through statistical analysis of these parameters. However, with the widespread application of new radar systems (such as phased array radar and cognitive radar), the battlefield electromagnetic environment has become increasingly complex. Radar signals exhibit dense overlap and varied characteristics, making PDWs alone insufficient to fully characterize the essential features of the signals, leading to a significant decrease in identification accuracy. Therefore, extracting intra-pulse features of radar signals for modulation type identification has become a key research focus.

[0004] Currently, intra-pulse feature extraction mainly includes methods such as time-frequency analysis and higher-order statistical analysis. Among them, time-frequency analysis methods can simultaneously reflect the time and frequency domain characteristics of a signal and are commonly used for intra-pulse feature extraction. Typical time-frequency analysis methods include Short-Time Fourier Transform (STFT) and Continuous Wavelet Transform (CWT). STFT has good time-frequency convergence and is suitable for characterizing stationary or slowly varying signal components, but its ability to capture transient abrupt changes in signal characteristics is limited. CWT has multi-resolution analysis capabilities and can effectively extract transient details of signals, but its time-frequency profile stability is insufficient. A single time-frequency feature extraction method cannot fully cover the complex features of a signal, resulting in limited generalization ability and difficulty in dealing with complex scenarios such as multi-signal aliasing and low signal-to-noise ratio.

[0005] In recent years, deep learning technology has been widely used in the field of radar signal recognition due to its powerful adaptive feature extraction capabilities. However, existing deep learning-based radar signal recognition methods still have shortcomings: on the one hand, most methods use a single network structure to process a single feature, failing to fully utilize the complementary information between different features, resulting in insufficient feature representation; on the other hand, deep learning models usually require a large amount of training data to achieve ideal recognition results, while obtaining radar signal samples in practical applications is quite difficult, limiting the practicality of the models.

[0006] To address the aforementioned issues, some studies have attempted to improve recognition performance using feature fusion techniques. Existing fusion methods are mostly data-level fusions (such as directly concatenating different feature maps and inputting them into the network). This approach lacks in-depth feature processing, easily introducing a large amount of redundant information, leading to excessively high feature dimensionality, increased computational complexity, and failure to fully leverage the complementary advantages of different features, resulting in poor fusion performance. Therefore, there is an urgent need for a signal modulation recognition method that can effectively mine the complementarity of multi-domain features, suppress redundant information, and improve recognition performance. Summary of the Invention

[0007] The technical problem this invention aims to solve is the poor modulation recognition performance of current radar signal modulation recognition technologies with small sample data. It proposes a signal modulation recognition method and system based on feature-level fusion using a dual-branch VGG16 network. This invention extracts and fuses multi-domain deep features of the signal, enriching the feature representation and enhancing the separability between categories. This enables high-accuracy and robust automatic identification of intra-pulse modulation types of radar signals in complex electromagnetic environments.

[0008] The technical solution adopted by this invention to solve the above problems is as follows: a signal modulation recognition method based on feature-level fusion of a dual-branch VGG16 network, comprising three core parts: radar signal feature data construction, recognition network framework design, and feature-level fusion technology implementation, as detailed below: (1) Construction of radar signal feature data Selection of identification targets: Six typical radar signal modulation types were selected as identification targets, including binary phase coding (BPSK), quadrature phase coding (QPSK), single carrier frequency (CW), two-frequency coding (BFSK), linear frequency modulation (LFM), and nonlinear frequency modulation (NLFM), covering common radar signal modulation modes and having wide applicability.

[0009] Time-frequency feature extraction: Two complementary time-frequency analysis methods, Short-Time Fourier Transform (STFT) and Continuous Wavelet Transform (CWT), are used to extract intra-pulse features from radar signals. STFT Extraction: Set the window function of STFT to Hanning window, with a window length of 256 sampling points and an overlap rate of 50%. Convert the time-domain signal into an STFT time-frequency graph through STFT. This time-frequency graph can clearly characterize the stationary time-frequency profile and energy distribution law of the signal. CWT Extraction: Morlet wavelet is selected as the mother wavelet, and the wavelet scale range is set to 1-256. The time domain signal is converted into a CWT time-frequency plot through CWT. This time-frequency plot can accurately capture the transient change characteristics and high-frequency detail information of the signal. Feature map standardization: The STFT and CWT time-frequency maps are normalized, with their dimensions uniformly constrained to 256×256×2 (height×width×number of channels) to ensure consistency of the feature maps and meet the input requirements of subsequent networks. Through the above processing, a comprehensive and highly discriminative radar signal feature database is constructed.

[0010] (2) Identify network framework design This invention employs a symmetrical dual-branch parallel VGG16 network structure. The two branches have identical structures but independent parameters, processing the STFT time-frequency feature map and the CWT time-frequency feature map respectively, thus achieving parallel deep learning across multiple features. The specific structure of each branch is as follows: Input layer: Receives a standardized 256×256×2 two-dimensional time-frequency feature map; Convolutional module group: Each branch contains 6 consecutive convolutional modules (Block 1-Block 6), and the structural design of each convolutional module is as follows: Block 1 consists of two stacked 3×3 convolutional layers, each with 64 output channels. Batch normalization and ReLU activation functions are then applied sequentially after each convolutional layer. Block2 consists of two stacked 3×3 convolutional layers, each with 128 output channels. Batch normalization and ReLU activation functions are then applied sequentially after each convolutional layer. Block3 consists of three stacked 3×3 convolutional layers, each with 256 output channels. Batch normalization and ReLU activation functions are then applied sequentially after each convolutional layer. Block4 consists of three stacked 3×3 convolutional layers, each with 512 output channels. Batch normalization and ReLU activation functions are then applied sequentially after each convolutional layer. Block5 consists of three stacked 3×3 convolutional layers, each with 512 output channels. Batch normalization and ReLU activation functions are then applied sequentially after each convolutional layer. Block 6 consists of three stacked 3×3 convolutional layers, each with 512 output channels. Batch normalization and ReLU activation functions are then applied sequentially after each convolutional layer. Pooling layers: A 2×2 max pooling layer is inserted between Block1 and Block2, between Block2 and Block3, between Block3 and Block4, between Block4 and Block5, and between Block5 and Block6. The pooling stride is 2. Max pooling is used to achieve downsampling, expand the receptive field and compress the feature map space dimension, and realize hierarchical abstraction of features. Block6 finally outputs a fixed-length one-dimensional high-level feature vector of 1×1×512. The two branches output the STFT depth feature vector and the CWT depth feature vector, respectively.

[0011] (3) Feature-level fusion technology implementation Feature-level fusion technology achieves deep fusion and optimization of multiple features through four key steps, as follows: Step 1: Multi-feature extraction. For the input radar time-domain signal, STFT and CWT time-frequency feature extraction are performed in parallel to generate 256×256×2 STFT time-frequency maps and CWT time-frequency maps, comprehensively describing the signal features from two physical dimensions: stationary time-frequency profile and transient change details. Step 2: Secondary Feature Learning. The STFT time-frequency plot and CWT time-frequency plot are input into two independent dual-branch VGG16 networks respectively. Through the network's convolutional modules and pooling layers, deep abstract learning is performed on the time-frequency features to extract highly discriminative STFT high-level feature vectors (dimension 1×512) and CWT high-level feature vectors (dimension 1×512). Step 3: Feature-level fusion. The high-level feature vectors of STFT and CWT are concatenated and fused using a channel-dimensional stacking approach to obtain a fused feature vector (1×1024 dimensions). This fusion method maps heterogeneous features to a unified high-dimensional discriminative space, preserving the unique information of each modality while strengthening the deep interaction and complementarity of cross-modal features and adaptively suppressing redundant information. Step 4: Classification and Model Solidification. The fused feature vectors are input into a fully connected network for training. The fully connected network consists of two hidden layers and one output layer. First hidden layer: input dimension is 1024, output dimension is 512, activation function is ReLU; Second hidden layer: input dimension is 512, output dimension is 256, activation function is ReLU; Output layer: Input dimension is 256, output dimension is 6 (corresponding to 6 modulation signal types), activation function is Softmax, output probability value for each modulation type; During network training, the cross-entropy loss function is used as the optimization objective, and the Adam optimizer is used for parameter updates. The initial learning rate is set to 0.001, and it decays to 0.9 times its original value every 10 epochs. The number of training iterations is set to 100 epochs. After training, the optimal model parameters and structure are saved to form a deployable recognition module. During the testing phase, the fixed model is loaded, and test samples that were not used in the training (including radar signals with different signal-to-noise ratios and modulation types) are input. The signal modulation type is determined by the probability value output by the model, and the recognition task is completed.

[0012] Compared with the prior art, the advantages of the present invention are as follows: (1) Richer feature representation: By extracting dual time-frequency features of STFT and CWT and combining deep learning of dual-branch VGG16 network, the stationary time-frequency profile and transient change details of the signal are fully explored, and a multi-dimensional complementary feature representation is constructed, which overcomes the defect of incomplete description by a single feature. (2) Superior recognition performance: The feature-level fusion strategy realizes high-level interaction of heterogeneous features and suppression of redundant information, which improves the discriminativeness of features. Experiments have verified that the recognition accuracy of the six types of modulated signals exceeds 89% in the signal-to-noise ratio range of [-10,20] dB, and the recognition accuracy of CW signal reaches 91%. It still maintains high recognition robustness in low signal-to-noise ratio and multi-signal aliasing scenarios. (3) More efficient computation: The dual-branch VGG16 network adopts the design of "small-sized convolutional kernels and deep stacking", which effectively controls the number of network parameters while ensuring feature extraction capabilities; feature-level fusion reduces the dimension of fused features, reduces redundant information, and improves computational efficiency and model training speed. (4) Stronger generalization ability: The design of multi-feature complementary extraction and deep fusion enables the model to adapt to the radar signal recognition needs of different signal-to-noise ratios and different modulation types. It can achieve ideal results without a large amount of training data and has high practical application value. Attached Figure Description

[0013] Figure 1 Flowchart of feature-level fusion algorithm.

[0014] Figure 2 VGG16 network architecture.

[0015] Figure 3 STFT time-frequency diagrams of six radar signals.

[0016] Figure 4 CWT time-frequency diagrams of six radar signals.

[0017] Figure 5 The training set features of the algorithm model are represented by the t-SNE dimensionality reduction distribution.

[0018] Figure 6 Test set classification confusion matrix based on feature-level fusion algorithm. Detailed Implementation

[0019] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] 1. Experimental Environment Setup Hardware environment: CPU is Intel Core i7-12700K, GPU is NVIDIA RTX 3090 (24GB VRAM), memory is 64GB, and storage capacity is 2TB SSD; Software environment: The operating system is Ubuntu 20.04 LTS, the deep learning framework is PyTorch 1.12.0, the Python version is 3.8.10, and other dependent libraries include NumPy 1.21.5, Matplotlib 3.5.2, and Scikit-learn 1.0.2.

[0021] 2. Data Preparation Signal generation: MATLAB 2022b was used to generate 6 types of radar modulation signals (BPSK, QPSK, CW, BFSK, LFM, NLFM). The signal sampling frequency was 100MHz, the sampling duration was 10μs, and the number of sampling points for each signal was 1000. Signal-to-noise ratio (SNR) settings: The SNR range is [-10, -8, -6, -4, -2, 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20] dB, with a total of 16 SNR levels; Sample size: For each modulation signal, 100 training samples and 50 test samples are generated under each signal-to-noise ratio condition. The total number of training samples is 6×16×100=9600 and the total number of test samples is 6×16×50=4800. Feature map generation: Perform STFT and CWT processing on each sample to generate STFT time-frequency map and CWT time-frequency map. Normalize the two time-frequency maps to a size of 256×256×2 to construct the feature databases of training set and test set.

[0022] 3. Network training and model solidification Network initialization: Based on the above-described network identification framework, construct a dual-branch VGG16 network and randomly initialize the network parameters; Training parameter settings: Batch size is 32, cross-entropy loss function is used as the basis for loss calculation, the initial learning rate of Adam optimizer is 0.001, the learning rate decay coefficient is 0.9, and the number of training iterations is 100 epochs; Model training: Input the feature data of the training set into the dual-branch VGG16 network for model training. Calculate the accuracy and loss value of the training set after each epoch of training. Save the model parameters and structure with the lowest loss value and the highest accuracy to form a fixed model.

[0023] 4. Model Testing and Performance Evaluation Testing process: Input the test set feature data into the solidified model, and the model outputs the probability values ​​of 6 types of modulation signals corresponding to each test sample. Select the category with the highest probability value as the recognition result. Performance metrics: Recognition accuracy (overall accuracy and accuracy of various signals) and confusion matrix are used as performance evaluation metrics; Test results: Overall accuracy: Within a signal-to-noise ratio range of [-10, 20] dB, the overall recognition accuracy of the model exceeds 89%; Accuracy rates for various signals: CW signal has the highest recognition accuracy, reaching 91%; the recognition accuracy rates for BPSK, QPSK, BFSK, LFM, and NLFM signals are all between 89% and 90%. Confusion matrix analysis: As can be seen from the confusion matrix in Figure 6, the intra-class aggregation of various signals is strong and the inter-class interference is small. Only a small number of misclassifications exist under low signal-to-noise ratio conditions, which proves that the model has good recognition performance.

[0024] 5. Results Analysis Experimental results show that the signal modulation recognition method based on feature-level fusion of dual-branch VGG16 network proposed in this invention effectively improves the accuracy and robustness of radar signal modulation recognition through complementary extraction of STFT and CWT dual time-frequency features, deep learning of dual-branch network, and feature-level fusion optimization. It can adapt to the recognition needs in complex electromagnetic environments and has important practical application value.

[0025] Industrial applicability of the invention The signal modulation recognition method and system of the present invention can be directly applied to fields such as radar reconnaissance, electronic countermeasures, and communication signal processing. Specific application scenarios include: Military field: Used for battlefield radar signal reconnaissance, quickly identifying the modulation type of enemy radar, and providing a basis for electronic countermeasures decision-making; Civilian applications: Used for signal monitoring and identification in civil aviation radar and meteorological radar to ensure the normal operation of radar systems; In the field of communications: it is used for modulation type identification of communication signals to achieve adaptive reception and processing of communication signals.

[0026] The technical solution of this invention has been verified by experiments. The model training and deployment process is simple, the hardware requirements are in line with the existing industrial-grade equipment configuration, and it can be quickly implemented through software programming, which has significant industrial applicability.

[0027] In addition to the above embodiments, the present invention also includes other embodiments. All technical solutions formed by equivalent transformation or equivalent substitution should fall within the protection scope of the claims of the present invention.

Claims

1. A signal modulation recognition method based on double-branch VGG16 network feature level fusion, characterized in that, Includes the following steps: (1) Perform short-time Fourier transform (STFT) and continuous wavelet transform (CWT) on the radar signal to generate the corresponding STFT time-frequency diagram and CWT time-frequency diagram; (2) Input the STFT time-frequency diagram and CWT time-frequency diagram into two VGG16 branch networks with the same structure and independent parameters, respectively, and obtain the corresponding high-dimensional feature vectors through feature extraction; (3) The two high-dimensional feature vectors are concatenated and fused along the channel dimension to obtain a fused feature vector; (4) Input the fused feature vector into the fully connected layer network for training, and output the signal modulation type identification result through the trained model.

2. The method of claim 1, wherein, In step (1), the radar signal includes binary phase coded (BPSK), quadrature phase coded (QPSK), single carrier frequency (CW), two-frequency coded (BFSK), linear frequency modulation (LFM), and nonlinear frequency modulation (NLFM) signals.

3. The method of claim 1, wherein, In step (1), the STFT time-frequency diagram and CWT time-frequency diagram are subjected to dimension reduction or sampling processing, and the dimension is uniformly constrained to 256×256×2.

4. The method of claim 1, wherein, In step (2), each VGG16 branch network includes multiple convolutional module groups, each convolutional module group consists of 2-3 stacked 3×3 convolutional layers, each convolutional layer is followed by batch normalization and ReLU activation function, and 2×2 max pooling layers are inserted between adjacent convolutional module groups; the end of the VGG16 branch network outputs the high-dimensional feature vector through a global average pooling layer.

5. The method of claim 1, wherein, In step (4), the output of the fully connected layer network uses the Softmax activation function to output the recognition probability of different modulation types.

6. The method of claim 1, wherein, When training the fully connected layer network, the cross-entropy loss function is used as the optimization objective, and the optimal model parameters and structure are saved after training.

7. A signal modulation recognition system based on double-branch VGG16 network feature-level fusion, characterized in that, include: Feature extraction module: used to perform short-time Fourier transform (STFT) and continuous wavelet transform (CWT) on radar signals to generate STFT time-frequency plots and CWT time-frequency plots; Dual-branch feature learning module: includes two VGG16 branch networks with identical structures and independent parameters, used to extract features from the STFT time-frequency plot and CWT time-frequency plot respectively, and output the corresponding high-dimensional feature vectors; Feature fusion module: used to concatenate and fuse the two high-dimensional feature vectors along the channel dimension to obtain a fused feature vector; Classification and recognition module: includes a fully connected layer network, used to train the fused feature vector and output the signal modulation type recognition result.

8. The system according to claim 7, characterized in that, It also includes a model storage module, which is used to save the optimal model parameters and structure after training.

Images