Radar signal modulation identification method and device for self-supervised contrast mask reconstruction

Abstract

The invention discloses a radar signal modulation identification method and device for self-supervised contrast mask reconstruction, and belongs to the technical field of radar signal modulation identification, and the method comprises the following steps: obtaining a radar signal, constructing a radar modulation signal data set containing label data and label-free data, and converting the radar modulation signal data set into a time-frequency image; building a self-supervised contrast mask image reconstruction model comprising an online branch and a target branch; the method comprises the following steps: pre-training a self-supervised contrast mask image reconstruction model by using label-free data, performing data enhancement and random mask operation on a time-frequency image to generate double views, respectively inputting an online branch and a target branch, updating model parameters by jointly optimizing reconstruction loss and contrast loss, and obtaining a pre-training weight; and migrating the pre-training weight to a downstream identification network, freezing part of encoder parameters, and performing fine tuning by using a small amount of labeled data to obtain a radar signal modulation identification model. According to the invention, the radar signal modulation identification precision in a complex electromagnetic environment is improved.

Description

Technical Field

[0001] The present invention relates to a radar signal modulation recognition method and device based on self-supervised contrast mask reconstruction, belonging to the technical field of radar signal modulation recognition. Background Art

[0002] Radar signal modulation recognition (RSMR) is a key technology in radar signal processing and wireless communication systems, with widespread applications in electronic countermeasures, spectrum monitoring, and intelligent communications. Its core goal is to accurately identify the modulation mode of a received signal without prior information. In practical applications, modulation recognition still faces multiple challenges. First, the complexity of the electromagnetic environment often subjects signals to strong interference and background noise, significantly increasing the difficulty of recognition. Second, with the increasing sophistication of radar system architectures and signal designs, modulation modes have become highly diverse and complex. Traditional recognition algorithms exhibit significant adaptability issues when dealing with new modulation modes, especially under low signal-to-noise ratio (SNR) conditions. Furthermore, many practical scenarios place stringent requirements on system response speed, placing higher demands on modulation recognition methods to balance accuracy and computational efficiency.

[0003] Traditional modulation recognition methods primarily rely on manually designed feature extraction algorithms, such as cyclic spectrum, instantaneous frequency, and time-frequency diagram analysis. While these methods can achieve reasonable recognition performance under ideal conditions, they rely on the expertise of signal processing experts and are less robust in non-ideal environments. Furthermore, these methods struggle to model complex modulation structures and mixed signals, and their recognition accuracy falls short of meeting the requirements of modern systems. Therefore, machine learning methods have been introduced into RSMR research to automatically extract more discriminative features.

[0004] With the development of machine learning, algorithms such as support vector machines (SVMs) have begun to be applied to modulation recognition tasks. These methods have achieved some performance improvements by performing classification training on raw signals or manually extracted features. However, they still rely on strong feature engineering capabilities, limiting their generalization. In recent years, deep learning technology has achieved breakthroughs in image, speech, and signal processing. By constructing deep neural networks, deep learning methods can learn effective representations from large amounts of data, significantly improving recognition accuracy. In particular, architectures such as convolutional neural networks (CNNs) and self-attention can automatically extract deep semantic features from raw I / Q signals, spectrograms, or time-frequency plots, significantly improving recognition performance. Currently, a growing number of researchers are considering applying these methods to radar signal modulation recognition.

[0005] While deep learning methods have demonstrated powerful modeling capabilities in the field of RSMR, their widespread application still faces two core challenges: First, deep models typically require large amounts of high-quality annotated data for training. However, radar signal data acquisition and annotation are extremely expensive, making it difficult to collect large amounts of high-quality, reliable radar data in practice. Second, under extreme conditions such as low signal-to-noise ratios, the model's feature extraction and discrimination performance remain limited. Training modulation recognition models with strong generalization and robustness under low-resource and weak-supervision conditions has become a hot topic and a key challenge in current research.

[0006] To address the reliance of supervised learning on labeled data, self-supervised learning (SSL) offers a solution. In recent years, it has made significant progress in fields such as computer vision and speech processing, and has gradually been introduced into modulation recognition. Self-supervised learning employs pre-training tasks (such as predicting occluded regions, sorting transformations, and discriminative contrast samples) to learn universal representations that can be transferred to downstream tasks, thereby reducing reliance on labeled data. Current mainstream self-supervised methods can be broadly categorized into two types: discriminative learning and generative learning. The former constructs positive and negative sample pairs, enabling the model to learn a feature space structure that remains consistent across different transformations for the same instance and distinguishes between different instances. The latter primarily employs a mask modeling strategy, randomly masking portions of the input and guiding the model to learn local and global structural relationships by predicting the masked content. Typical contrastive and generative learning frameworks have been applied to modulation recognition, with promising results. However, single-strategy learning models have their own limitations: contrastive learning focuses on global feature alignment and ignores local semantic integrity; while generative methods can capture detailed structures, they are prone to fuzzy expressions in global discrimination tasks, affecting classification performance. Summary of the Invention

[0007] In order to solve the above problems, the present invention proposes a radar signal modulation recognition method and device based on self-supervised contrast mask reconstruction, which can improve the radar signal modulation recognition performance in low-resource scenarios.

[0008] The technical solution adopted by the present invention to solve its technical problems is: In a first aspect, an embodiment of the present invention provides a radar signal modulation recognition method based on self-supervised contrast mask reconstruction, comprising the following steps: Step 1: Acquire radar signals and construct a radar modulation signal dataset containing labeled data and unlabeled data, perform improved multiple synchronous compression transform (IMSST) time-frequency analysis on the radar modulation signal, and convert it into a time-frequency image; Step 2: Build a self-supervised contrast mask image reconstruction model. The self-supervised contrast mask image reconstruction model includes an online branch and a target branch. The online branch includes an improved ResNet encoder with an integrated cross-time-frequency pooling (CTFP) module, a reconstruction decoder, and a projection head. The target branch includes a target encoder and a projection head whose parameters are updated via an exponential moving average (EMA) mechanism. Step 3: Pre-train the self-supervised contrast mask image reconstruction model using the unlabeled data, apply data augmentation and random masking operations to the time-frequency image to generate a dual view, input the dual view into the online branch and the target branch respectively, and update the model parameters by jointly optimizing the reconstruction loss and the contrast loss to obtain the pre-trained weights; Step 4: Migrate the pre-trained weights to the downstream recognition network, freeze some encoder parameters, and then use a small amount of labeled data for fine-tuning to obtain a radar signal modulation recognition model.

[0009] As a possible implementation of this embodiment, step 1 includes: Step 11, generating a plurality of radar modulation signals according to the radar signal; Step 12: Perform time-frequency transformation on the radar modulated signal using the IMSST method. The IMSST method decomposes the signal through multi-scale analysis and then adjusts the time-frequency energy distribution through synchronous squeezing to concentrate the energy on the actual time-frequency trajectory. Step 13: After time-frequency transformation, a time-frequency image with a size of 1×64×64 is obtained; Step 14: Integrate all the obtained time-frequency images to form a radar modulation signal time-frequency image dataset containing labeled data and unlabeled data.

[0010] As a possible implementation of this embodiment, step 2 includes: Step 21: Design the overall framework of the self-supervised contrast mask image reconstruction model, and clarify the composition and connection relationship between the online branch and the target branch; Step 22, constructing an improved ResNet encoder integrated with a cross-time-frequency pooling (CTFP) module, wherein the improved ResNet encoder includes five consecutive stages (layer1 to layer5); Step 23: Design an online branch reconstruction decoder. The reconstruction decoder consists of a five-layer deconvolution module. Each layer includes a transposed convolution, batch normalization, and a ReLU activation function. It restores the feature space size layer by layer and ultimately outputs a reconstructed image with the same size as the input time-frequency map (1×64×64). Step 24: Design the projection heads of the online branch and the target branch. The projection heads are both two-layer fully connected networks. The first layer maps the high-dimensional features (512 dimensions) output by the encoder to 256 dimensions, and the second layer maps them to 128 dimensions, which are used to project the features into the contrast space. Step 25, set the parameter update mechanism of the target encoder: the target encoder parameters are synchronously updated from the improved ResNet encoder of the online branch through the EMA mechanism; Step 26, construct the joint loss function: , in and are the weight coefficients of reconstruction loss and contrast loss, and They are reconstruction loss and contrast loss respectively.

[0011] As a possible implementation of this embodiment, in step 21, the online branch includes an improved ResNet encoder with an integrated Cross Time-Frequency Pooling (CTFP) module, a reconstruction decoder, and a projection head. The target branch includes a target encoder and a projection head. The target encoder has the same structure as the improved ResNet encoder, and its parameters are synchronously updated by the parameters of the improved ResNet encoder in the online branch via an exponential moving average (EMA) mechanism. The radar signal is transformed into a 1×64×64 time-frequency map through IMSST. Two views are generated through two types of enhancement operations. The view with a random mask (masking ratio 75%) is input into the online branch, and the unmasked view that has undergone conventional data augmentation (cropping, scaling, and noise perturbation) is input into the target branch.

[0012] As a possible implementation of this embodiment, the specific process of step 22 is as follows: Based on the ResNet-18 architecture, five consecutive stages are set up: layer1-layer5. In the layer1 stage, the initial feature extraction of the input time-frequency map is performed through the convolution layer and the maximum pooling layer. From layer2 to layer5, deep feature modeling is performed through residual blocks. The number of channels increases from 64 to 128 to 256 to 512, and the feature map resolution decreases. A CTFP module is inserted after the 3×3 convolution of the last residual unit in each stage from layer 2 to layer 5, and after all residual blocks in the layer 5 stage. The CTFP module performs one-dimensional strip pooling in the time dimension and frequency dimension respectively, and enhances the time-frequency correlation of features by fusing the pooling results. Specifically, it includes: Perform horizontal strip pooling on the input feature map and output for: , where x is the input feature map and W is the feature map width; Perform vertical strip pooling on the input feature map and output for: , where H is the feature map height; Pooling the horizontal strips Compared with the vertical strip pooling result After 1×1 convolution fusion and sigmoid activation, it is multiplied element by element with the original feature map to obtain the enhanced features: , , in, represents element-wise multiplication, is the sigmoid function, is a 1×1 convolution, is the output of cross time-frequency pooling (CTFP), and C is the channel dimension of the feature map.

[0013] As a possible implementation of this embodiment, constructing the joint loss function includes: Define the reconstruction loss that only calculates the pixel error in the mask area: , in Indicates the combination of the obscured positions, and Represent the pixel values ​​of the original image and the model reconstructed image at the i-th position respectively; InfoNCE loss is used as the comparison loss: , in , It is the representation vector corresponding to the same sample after different data enhancements. is the cosine similarity, is the temperature coefficient, is the total number of comparison samples; Two learnable weight parameters are introduced for contrast loss and reconstruction loss respectively and , construct the final joint loss function.

[0014] As a possible implementation of this embodiment, step 3 includes: Step 31: Prepare an unlabeled time-frequency image dataset: Use the improved multiple synchronous compression transform (IMSST) to obtain 1×64×64 time-frequency images. The images in the unlabeled time-frequency image dataset are derived from ten radar modulation signals (LFM, SFM, BPSK, LFM-BPSK, SFM-BPSK, EQFM, FSK, 4FSK, NS, and Frank), and cover samples with a signal-to-noise ratio (SNR) range of -16 dB to 4 dB (at 2 dB intervals). Step 32, generating dual-view input: performing two-class processing on each unlabeled time-frequency image in the dataset to construct a first view and a second view; Step 33, dual-branch feature extraction and processing: The first view is fed into an improved ResNet encoder with an integrated cross-time-frequency pooling (CTFP) module for online branch processing. After extracting high-dimensional features, the feature is split into two paths: one path is fed into a reconstruction decoder to reconstruct the complete time-frequency image, and the other path is fed into a projection head after global average pooling and mapped to the contrast space. The second view is fed into a target encoder (with the same structure as the improved ResNet encoder, and its parameters are synchronously updated from the online branch encoder via an exponential moving average (EMA) mechanism) for target branch processing. After feature extraction, it is mapped to the contrast space through global average pooling and the projection head. Step 34, joint loss function calculation; Step 35, model parameter update and pre-training weight acquisition: optimize the total loss through backpropagation, update the parameters of the improved ResNet encoder, reconstruction decoder and projection head; update the target encoder and projection head parameters through the EMA mechanism; jointly optimize the reconstruction loss and contrast loss to complete pre-training, and finally save the parameters of the improved ResNet encoder as pre-training weights.

[0015] As a possible implementation of this embodiment, the first view is used as the input of the online branch: conventional data augmentation (including random cropping, scale transformation, and noise perturbation) is first applied, and then a random masking operation is performed to divide the image into 8×8 patches (a total of 64 blocks), and 75% (48 blocks) of the patches are randomly masked; The second view is the target branch input: only the same regular data augmentation as the first view is applied, without masking.

[0016] As a possible implementation of this embodiment, step 4 includes: Step 41: Construct a downstream recognition network structure. The downstream recognition network uses an improved ResNet encoder with an integrated cross-time-frequency pooling (CTFP) module as the feature extraction backbone, and a lightweight classification head is connected to the encoder output. The classification head includes two fully connected layers. The first layer maps the 512-dimensional features output by the encoder to 256 dimensions, and the second layer maps them to an output dimension that matches the number of modulation types (10 categories) to output the probability distribution of each category. Step 42, transfer pre-trained weights: reconstruct the contrast mask image from the improved ResNet encoder parameters obtained by pre-training the supervised model, and transfer them as a whole to the corresponding encoder of the downstream recognition network as the initial weights; Step 43, freeze some encoder parameters: freeze all parameters of the first three stages (layer1 to layer3) of the improved ResNet encoder, and only keep the parameters of the last two stages (layer4 to layer5) and the classification head parameters trainable; Step 44, prepare a fine-tuning dataset: Use a small amount of labeled time-frequency image data. The data of the fine-tuning dataset comes from IMSST time-frequency graphs of ten radar modulation signals (LFM, SFM, BPSK, LFM-BPSK, SFM-BPSK, EQFM, FSK, 4FSK, NS, Frank), covering the signal-to-noise ratio range of -16 dB to 4 dB (at 2 dB intervals). The number of labeled samples is 50-200 for each signal type and each signal-to-noise ratio. Step 45, fine-tuning training process: take the fine-tuning dataset as input, output the category prediction results through the downstream recognition network, and use the cross-entropy loss function to calculate the error between the prediction results and the labels; use the Adam optimizer to optimize the cross-entropy loss, update the parameters of the last two stages of the trainable encoder and the classification head parameters, set the initial learning rate to 0.0001, the batch size to 64, the training rounds to 200 rounds, and use the cosine annealing mechanism to decay the learning rate from the initial value to 1×10⁻ 6 ; Step 46, obtaining a radar signal modulation recognition model: After the fine-tuning training is completed, all parameters of the downstream recognition network are saved to form the final radar signal modulation recognition model.

[0017] In a second aspect, an embodiment of the present invention provides a radar signal modulation identification device based on self-supervised contrast mask reconstruction, comprising: A dataset construction module is used to acquire radar signals and construct a radar modulation signal dataset containing labeled data and unlabeled data, perform improved multiple synchronous compression transform (IMSST) time-frequency analysis on the radar modulation signal, and convert it into a time-frequency image; A model building module for building a self-supervised contrastive mask image reconstruction model, comprising an online branch and a target branch. The online branch comprises an improved ResNet encoder with an integrated cross-time-frequency pooling (CTFP) module, a reconstruction decoder, and a projection head. The target branch comprises a target encoder and a projection head whose parameters are updated via an exponential moving average (EMA) mechanism. a pre-training module for pre-training the self-supervised contrastive mask image reconstruction model using the unlabeled data, applying data augmentation and random masking operations to the time-frequency image to generate a dual view, inputting the dual view into the online branch and the target branch respectively, and updating the model parameters by jointly optimizing the reconstruction loss and the contrast loss to obtain the pre-trained weights; The fine-tuning module is used to transfer the pre-trained weights to the downstream recognition network, freeze some encoder parameters, and then use a small amount of labeled data for fine-tuning to obtain a radar signal modulation recognition model.

[0018] The beneficial effects of the technical solutions of the embodiments of the present invention are as follows: The present invention utilizes unlabeled data to learn features that have both structural integrity and discriminability through the joint optimization of contrastive learning and mask image reconstruction; designs a cross-time-frequency pooling module to enhance the time-frequency correlation modeling capability, and ultimately achieves high-precision recognition with a small amount of labeled data. This solves the problems of traditional methods such as strong dependence on labeled data and poor performance in low signal-to-noise ratio scenarios, and significantly improves recognition accuracy in complex electromagnetic environments.

[0019] The present invention fully utilizes unlabeled data through self-supervised pre-training, and only requires 10% labeled data to achieve recognition performance close to that of full supervision, reducing dependence on labeled data and lowering data labeling costs.

[0020] The mask reconstruction task of the present invention forces the model to learn noise-resistant time-frequency structural features, and contrastive learning enhances category discrimination. It maintains a recognition rate of more than 43% under an extremely low signal-to-noise ratio of -16dB, thereby improving the robustness to low signal-to-noise ratios.

[0021] The CTFP module of the present invention enhances the time-frequency correlation modeling through time-frequency dual-dimensional global pooling and attention fusion, effectively capturing the non-stationary modulation characteristics of radar signals (such as the time-frequency trajectory of frequency hopping and linear frequency modulation).

[0022] The improved ResNet encoder in this paper has approximately 11M parameters and a computational complexity of 274.85M FLOPs, taking into account both accuracy and efficiency and is suitable for deployment in resource-constrained scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a flow chart of a radar signal modulation recognition method based on self-supervised contrast mask reconstruction according to an exemplary embodiment; Figure 2 1 is a schematic structural diagram of a radar signal modulation identification device based on self-supervised contrast mask reconstruction according to an exemplary embodiment; Figure 3 1 is a diagram showing a framework of a contrast mask-pushing reconstruction model according to an exemplary embodiment; Figure 4 is a schematic diagram of a cross-time-frequency pooling mechanism according to an exemplary embodiment; Figure 5 1 is a structural diagram of an encoder improved based on ResNet according to an exemplary embodiment; Figure 6 is a schematic diagram showing the recognition accuracy of each signal at each signal-to-noise ratio according to an exemplary embodiment of the method of the present invention; Figure 7 FIG. 4 is a schematic diagram showing the influence of different encoder pairs on recognition accuracy in the proposed framework according to an exemplary embodiment. DETAILED DESCRIPTION

[0024] In order to more clearly illustrate the technical features of the present invention, the present invention is described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0025] like Figure 1 As shown, an embodiment of the present invention provides a radar signal modulation recognition method based on self-supervised contrast mask reconstruction, comprising the following steps: Step 1: Acquire radar signals and construct a radar modulation signal dataset containing labeled data and unlabeled data, perform improved multiple synchronous compression transform (IMSST) time-frequency analysis on the radar modulation signal, and convert it into a time-frequency image; Step 2: Build a self-supervised contrast mask image reconstruction model. The self-supervised contrast mask image reconstruction model includes an online branch and a target branch. The online branch includes an improved ResNet encoder with an integrated cross-time-frequency pooling (CTFP) module, a reconstruction decoder, and a projection head. The target branch includes a target encoder and a projection head whose parameters are updated via an exponential moving average (EMA) mechanism. Step 3: Pre-train the self-supervised contrast mask image reconstruction model using the unlabeled data, apply data augmentation and random masking operations to the time-frequency image to generate a dual view, input the dual view into the online branch and the target branch respectively, and update the model parameters by jointly optimizing the reconstruction loss and the contrast loss to obtain the pre-trained weights; Step 4: Migrate the pre-trained weights to the downstream recognition network, freeze some encoder parameters, and then use a small amount of labeled data for fine-tuning to obtain a radar signal modulation recognition model.

[0026] As a possible implementation of this embodiment, step 1 includes: Step 11, generating a plurality of radar modulation signals according to the radar signal; Step 12: Perform time-frequency transformation on the radar modulated signal using the IMSST method. The IMSST method decomposes the signal through multi-scale analysis and then adjusts the time-frequency energy distribution through synchronous squeezing to concentrate the energy on the actual time-frequency trajectory. Step 13: After time-frequency transformation, a time-frequency image with a size of 1×64×64 is obtained; Step 14: Integrate all the obtained time-frequency images to form a radar modulation signal time-frequency image dataset containing labeled data and unlabeled data.

[0027] As a possible implementation of this embodiment, step 2 includes: Step 21: Design the overall framework of the self-supervised contrast mask image reconstruction model, and clarify the composition and connection relationship between the online branch and the target branch; Step 22, constructing an improved ResNet encoder integrated with a cross-time-frequency pooling (CTFP) module, wherein the improved ResNet encoder includes five consecutive stages (layer1 to layer5); Step 23: Design an online branch reconstruction decoder. The reconstruction decoder consists of a five-layer deconvolution module. Each layer includes a transposed convolution, batch normalization, and a ReLU activation function. It restores the feature space size layer by layer and ultimately outputs a reconstructed image with the same size as the input time-frequency map (1×64×64). Step 24: Design the projection heads of the online branch and the target branch. The projection heads are both two-layer fully connected networks. The first layer maps the high-dimensional features (512 dimensions) output by the encoder to 256 dimensions, and the second layer maps them to 128 dimensions, which are used to project the features into the contrast space. Step 25, set the parameter update mechanism of the target encoder: the target encoder parameters are synchronously updated from the improved ResNet encoder of the online branch through the EMA mechanism; Step 26, construct the joint loss function: , in and are the weight coefficients of reconstruction loss and contrast loss, and They are reconstruction loss and contrast loss respectively.

[0028] As a possible implementation of this embodiment, in step 21, the online branch includes an improved ResNet encoder with an integrated cross time-frequency pooling (CTFP) module, a reconstruction decoder, and a projection head. The target branch includes a target encoder and a projection head. The target encoder has the same structure as the improved ResNet encoder, and its parameters are synchronously updated by the parameters of the improved ResNet encoder in the online branch via an exponential moving average (EMA) mechanism. The radar signal is transformed using an IMSST transform to obtain a 1×64×64 time-frequency map. Two views are generated through two types of enhancement operations. The view with a random mask (masking ratio 75%) is input to the online branch, and the unmasked view that has undergone conventional data augmentation (cropping, scaling, and noise perturbation) is input to the target branch.

[0029] As a possible implementation of this embodiment, the specific process of step 22 is as follows: Based on the ResNet-18 architecture, five consecutive stages are set up: layer1-layer5. In the layer1 stage, the initial feature extraction of the input time-frequency map is performed through the convolution layer and the maximum pooling layer. From layer2 to layer5, deep feature modeling is performed through residual blocks. The number of channels increases from 64 to 128 to 256 to 512, and the feature map resolution decreases. A CTFP module is inserted after the 3×3 convolution of the last residual unit in each stage from layer 2 to layer 5, and after all residual blocks in the layer 5 stage. The CTFP module performs one-dimensional strip pooling in the time dimension and frequency dimension respectively, and enhances the time-frequency correlation of features by fusing the pooling results. Specifically, it includes: Perform horizontal strip pooling on the input feature map and output for: , where x is the input feature map and W is the feature map width; Perform vertical strip pooling on the input feature map and output for: , where H is the feature map height; Pooling the horizontal strips Compared with the vertical strip pooling result After 1×1 convolution fusion and sigmoid activation, it is multiplied element by element with the original feature map to obtain the enhanced features: , , in, represents element-wise multiplication, is the sigmoid function, is a 1×1 convolution, is the output of cross time-frequency pooling (CTFP), and C is the channel dimension of the feature map.

[0030] As a possible implementation of this embodiment, constructing the joint loss function includes: Define the reconstruction loss that only calculates the pixel error in the mask area: , in Indicates the combination of the obscured positions, and Represent the pixel values ​​of the original image and the model reconstructed image at the i-th position respectively; InfoNCE loss is used as the comparison loss: , in , It is the representation vector corresponding to the same sample after different data enhancements. is the cosine similarity, is the temperature coefficient, is the total number of comparison samples; Two learnable weight parameters are introduced for contrast loss and reconstruction loss respectively and , construct the final joint loss function.

[0031] As a possible implementation of this embodiment, step 3 includes: Step 31: Prepare an unlabeled time-frequency image dataset: Use the improved multiple synchronous compression transform (IMSST) to obtain 1×64×64 time-frequency images. The images in the unlabeled time-frequency image dataset are derived from ten radar modulation signals (LFM, SFM, BPSK, LFM-BPSK, SFM-BPSK, EQFM, FSK, 4FSK, NS, and Frank), and cover samples with a signal-to-noise ratio (SNR) range of -16 dB to 4 dB (at 2 dB intervals). Step 32, generating dual-view input: performing two-class processing on each unlabeled time-frequency image in the dataset to construct a first view and a second view; Step 33, dual-branch feature extraction and processing: The first view is fed into an improved ResNet encoder with an integrated cross-time-frequency pooling (CTFP) module for online branch processing. After extracting high-dimensional features, the feature is split into two paths: one path is fed into a reconstruction decoder to reconstruct the complete time-frequency image, and the other path is fed into a projection head after global average pooling and mapped to the contrast space. The second view is fed into a target encoder (with the same structure as the improved ResNet encoder, and its parameters are synchronously updated from the online branch encoder via an exponential moving average (EMA) mechanism) for target branch processing. After feature extraction, it is mapped to the contrast space through global average pooling and the projection head. Step 34, joint loss function calculation; Step 35, model parameter update and pre-training weight acquisition: optimize the total loss through backpropagation, update the parameters of the improved ResNet encoder, reconstruction decoder and projection head; update the target encoder and projection head parameters through the EMA mechanism; jointly optimize the reconstruction loss and contrast loss to complete pre-training, and finally save the parameters of the improved ResNet encoder as pre-training weights.

[0032] As a possible implementation of this embodiment, the first view is used as the input of the online branch: conventional data augmentation (including random cropping, scale transformation, and noise perturbation) is first applied, and then a random masking operation is performed to divide the image into 8×8 patches (a total of 64 blocks), and 75% (48 blocks) of the patches are randomly masked; The second view is the target branch input: only the same regular data augmentation as the first view is applied, without masking.

[0033] As a possible implementation of this embodiment, step 4 includes: Step 41: Construct a downstream recognition network structure. The downstream recognition network uses an improved ResNet encoder with an integrated cross-time-frequency pooling (CTFP) module as the feature extraction backbone, and a lightweight classification head is connected to the encoder output. The classification head includes two fully connected layers. The first layer maps the 512-dimensional features output by the encoder to 256 dimensions, and the second layer maps them to an output dimension that matches the number of modulation types (10 categories) to output the probability distribution of each category. Step 42, transfer pre-trained weights: reconstruct the contrast mask image from the improved ResNet encoder parameters obtained by pre-training the supervised model, and transfer them as a whole to the corresponding encoder of the downstream recognition network as the initial weights; Step 43, freeze some encoder parameters: freeze all parameters of the first three stages (layer1 to layer3) of the improved ResNet encoder, and only keep the parameters of the last two stages (layer4 to layer5) and the classification head parameters trainable; Step 44, prepare a fine-tuning dataset: Use a small amount of labeled time-frequency image data. The data of the fine-tuning dataset comes from IMSST time-frequency graphs of ten radar modulation signals (LFM, SFM, BPSK, LFM-BPSK, SFM-BPSK, EQFM, FSK, 4FSK, NS, Frank), covering the signal-to-noise ratio range of -16 dB to 4 dB (at 2 dB intervals). The number of labeled samples is 50-200 for each signal type and each signal-to-noise ratio. Step 45, fine-tuning training process: take the fine-tuning dataset as input, output the category prediction results through the downstream recognition network, and use the cross-entropy loss function to calculate the error between the prediction results and the labels; use the Adam optimizer to optimize the cross-entropy loss, update the parameters of the last two stages of the trainable encoder and the classification head parameters, set the initial learning rate to 0.0001, the batch size to 64, the training rounds to 200 rounds, and use the cosine annealing mechanism to decay the learning rate from the initial value to 1×10⁻ 6 ; Step 46, obtaining a radar signal modulation recognition model: After the fine-tuning training is completed, all parameters of the downstream recognition network are saved to form the final radar signal modulation recognition model.

[0034] like Figure 2 As shown, an embodiment of the present invention provides a radar signal modulation recognition device based on self-supervised contrast mask reconstruction, comprising: A dataset construction module is used to acquire radar signals and construct a radar modulation signal dataset containing labeled data and unlabeled data, perform improved multiple synchronous compression transform (IMSST) time-frequency analysis on the radar modulation signal, and convert it into a time-frequency image; A model building module for building a self-supervised contrastive mask image reconstruction model, comprising an online branch and a target branch. The online branch comprises an improved ResNet encoder with an integrated cross-time-frequency pooling (CTFP) module, a reconstruction decoder, and a projection head. The target branch comprises a target encoder and a projection head whose parameters are updated via an exponential moving average (EMA) mechanism. a pre-training module for pre-training the self-supervised contrastive mask image reconstruction model using the unlabeled data, applying data augmentation and random masking operations to the time-frequency image to generate a dual view, inputting the dual view into the online branch and the target branch respectively, and updating the model parameters by jointly optimizing the reconstruction loss and the contrast loss to obtain the pre-trained weights; The fine-tuning module is used to transfer the pre-trained weights to the downstream recognition network, freeze some encoder parameters, and then use a small amount of labeled data for fine-tuning to obtain a radar signal modulation recognition model.

[0035] The specific implementation of the present invention can be divided into the following steps.

[0036] Step 1: Construct a data set of ten types of radar modulation signals and convert them into time-frequency images using the improved multiple synchronous compression transform (IMSST) time-frequency analysis process.

[0037] t The radar signal received at each moment can usually be expressed as a complex signal in the following form: , in, represents the amplitude envelope of the signal, represents the instantaneous frequency, represents the instantaneous phase, is additive noise, usually modeled as zero-mean Gaussian white noise.

[0038] Ten radar modulation signals, including LFM, SFM, BPSK, LFM-BPSK, SFM-BPSK, EQFM, FSK, 4FSK, NS, and Frank, are generated using MATLAB 2021b as simulation data to verify the effectiveness of the model. Here, [·] represents a random parameter set, f s 、 f c 、 N s 、 B 、 T p 、 N bc 、 cpp 、 M and N p represents the sampling frequency, carrier frequency, number of samples, bandwidth, symbol width, Barker code length, number of codes per cycle, frequency step size, and code period, respectively. All signals were simulated with varying signal-to-noise ratios (SNRs). The SNR range was set from -16 dB to 4 dB, with 2 dB intervals, to ensure data diversity and fairness in the simulation. The specific signal parameters are shown in Table 1.

[0039] Table 1 Signal parameters

[0040] To accurately extract the time-frequency characteristics of radar signals, the improved multiscale time-frequency analysis method (IMSST) is used to perform a time-frequency transform on the original signal. The IMSST method effectively improves the resolution and clarity of the time-frequency representation through multiscale transformation and an improved synchronous squeezing transform. Its core concept is to decompose the signal using multiscale analysis and concentrate the time-frequency energy on the actual time-frequency trajectory through synchronous squeezing. IMSST can be described by the following formula: , Among them, the variable and represents discrete frequencies, Indicates time n The discrete short-time Fourier transform at is the Kronecker delta function. represents two rounding operations on the instantaneous frequency of the multi-synchronous compression transform, Indicates the number of iterations.

[0041] Step 2: Build a self-supervised training model for contrast mask reconstruction based on the CTFP enhancement module.

[0042] 2.1 Overall design of the contrast mask image reconstruction model: like Figure 3 , the overall framework structure of the contrast mask image reconstruction model proposed in the present invention. Aiming at the task of identifying the modulation type of radar signals, the present invention constructs a self-supervised feature modeling framework that combines the contrast learning strategy with the mask image reconstruction mechanism, aiming to effectively utilize the time-frequency structural features in unlabeled data. The method consists of two core stages: a self-supervised pre-training stage and a supervised downstream recognition stage. In the pre-training stage, the encoder learns robust and discriminative latent feature representations under the synergy of contrast constraints and reconstruction goals; in the downstream stage, by transferring the trained encoder weights, a linear classifier is trained on a small number of labeled samples to achieve modulation category discrimination.

[0043] Specifically, the raw radar signal is first converted into a time-frequency map. Two types of augmentation operations are then applied: one involves conventional data augmentation (such as cropping, rescaling, and noise perturbation), while the other involves introducing random masks (with a masking ratio of 75%) to construct scenes with missing structural information. The two views are then fed into an online encoder and a target encoder for feature extraction. The online encoder is updated via backpropagation, while the parameters of the target encoder are synchronously updated by the online encoder using an exponential moving average to ensure training stability. After feature extraction, the high-dimensional features generated by the masked view are fed into a decoder to reconstruct the corresponding time-frequency map. The output is compared with the unmasked image at the pixel level to form a reconstruction loss. Simultaneously, the feature representations of the two enhanced images are mapped to a contrast space via global pooling and projection. A loss function such as InfoNCE is used to enhance the similarity between positive pairs and distance negative samples, forming a contrast loss. By jointly optimizing these two losses, the model learns features that are both discriminative and capable of expressing details. After pre-training, the encoder is fixed and a lightweight classification head is trained under small-sample supervision to effectively recognize radar modulation signals. This approach, centered on the collaborative design of structural modeling and discriminative learning, significantly improves the model's generalization and recognition capabilities for complex modulation patterns.

[0044] 2.2 Cross-time-frequency pooling structure: This paper proposes a context modeling mechanism suitable for time-frequency graph structure: cross-time-frequency pooling, such as Figure 4 As shown in Figure 2. This structure applies one-dimensional pooling operations on the time dimension and frequency dimension respectively, models the global structural information along each direction, and enhances the original feature representation through lightweight fusion. Unlike two-dimensional average pooling, the proposed strip pooling averages all feature values ​​in rows or columns. Therefore, the output after horizontal strip merging is It can be written as: , Similarly, the output after vertical strip pooling is It can be written as: , Given the horizontal and vertical strip pooling layers, due to the long and narrow kernel shape, it is easy to establish long-range dependencies between discretely distributed regions and encode regions with banded structures. At the same time, due to the use of narrower kernels in the vertical or horizontal direction, it is also able to focus on capturing local details in the other dimension. These characteristics make the proposed strip pooling different from traditional spatial pooling that relies on square kernels. In order to obtain outputs that contain more useful global priors The present invention first and Combining the following together, we get : , , in, represents element-wise multiplication, is the sigmoid function, It is a 1×1 convolution.

[0045] The design of CTFP fully considers the structural characteristics of time-frequency graphs: the time dimension models sequentiality, and the frequency dimension models spectral patterns. Compared to standard pooling methods, this module can perceive global structure along both time and frequency, significantly improving the model's representation and recognition capabilities for complex radar signals such as frequency hopping, linear frequency modulation, and dual-band modulation. This invention will systematically integrate CTFP into the encoder module as its core context enhancement component, thereby improving overall representation capabilities and pre-training performance.

[0046] 2.3 Encoder Design: In the present invention, the classic model ResNet in the CNN architecture is selected as the backbone encoder. Among them, ResNet effectively alleviates the gradient vanishing and performance degradation problems faced by deep networks by introducing the residual connection mechanism, and improves the network's expression ability and training stability. The standard ResNet structure is usually composed of a plurality of residual units (Bottleneck modules) stacked together. Each unit contains several convolutional layers, and realizes the direct transmission of information flow between layers through the identity mapping channel. The method of the present invention adopts ResNet-18 as the basic encoder. Its network structure consists of five consecutive stages (layer1 to layer5). The first stage performs preliminary feature extraction on the input time-frequency map. The subsequent stages complete deeper feature modeling through residual blocks, and the number of channels gradually increases and the resolution of the feature map decreases.

[0047] Unlike standard ResNets, to better adapt to the structural characteristics of time-frequency graph data, the proposed method inserts a custom-designed Cross Time-Frequency Pooling (CTFP) module after the 3×3 convolution in the last residual unit of each stage, and after all residual blocks in the final stage. This module is designed to improve the model's ability to model multi-scale contextual information, thereby enhancing its sensitivity and ability to represent complex modulation features.

[0048] Specifically, let the input time-frequency graph be First, the initial feature extraction is performed through the convolution layer and the maximum pooling layer of Layer1 to obtain the feature map The feature map is then processed through four stages of residual modules: , After each stage of processing, a CTFP module is inserted to enhance the context of its output, namely: , The advantage of this structural design lies in the complementary global modeling capabilities provided by the CTFP module and the local convolution mechanism of ResNet: the convolution module excels at extracting local edges, textures, and short-range dependency features, while the CTFP module can capture global structural information across long distances, time, and frequency, effectively compensating for the modeling shortcomings of conventional CNN architectures with asymmetric input structures. Furthermore, because the CTFP module is lightweight, its introduction does not significantly increase model complexity or training time, making it highly practical and scalable.

[0049] 2.4 Decoder Design: In self-supervised reconstruction tasks, the decoder is responsible for restoring the high-dimensional features output by the encoder into a two-dimensional image representation corresponding to the original time-frequency map. While using more complex decoding structures can generally achieve lower reconstruction errors, related research indicates that overly powerful decoders may overburden the modeling task during pre-training, causing the encoder to rely on the decoder for feature restoration, which is detrimental to the encoder's ability to learn universal representations. Therefore, recent self-supervised learning methods have generally tended to adopt simple yet powerful lightweight decoder structures to avoid this problem and encourage the encoder to focus on learning key discriminative features. The decoder designed in this invention consists of a five-level deconvolution module. Each layer consists of a transposed convolution, batch normalization, and a ReLU activation function, which gradually restores the spatial dimensions and restores the image structure through layer-by-layer operations. It is important to note that the design of this decoder architecture does not aim for the ultimate image reconstruction quality, but rather intentionally maintains the simplicity of the network, making the decoding process merely an auxiliary tool for pre-training, thereby ensuring that the encoder can lead feature learning and extract effective information with discriminative capabilities.

[0050] 2.5 Loss Design: To leverage the complementary advantages of different pre-training tasks in self-supervised learning, this paper introduces a reconstruction loss and a contrastive loss into the overall framework, corresponding to the mask image reconstruction branch and the contrastive learning branch, respectively. This design aims to guide the model through multi-task collaborative training to achieve a balance between local detail modeling and global discriminative feature extraction, thereby improving the versatility of the representation and adaptability to downstream tasks.

[0051] In the reconstruction branch, this paper adopts a mechanism similar to that of the Masked Autoencoder, randomly masking parts of the input time-frequency map and performing feature extraction and image restoration only on the remaining visible areas. To ensure that the reconstruction task focuses on the information-missing areas, only the reconstruction error of the masked areas is measured, and the reconstruction loss is defined as: , in Indicates the combination of the obscured positions, and Represents the pixel value at position i in the original image and the model-reconstructed image, respectively. This loss encourages the model to learn how to recover missing information from incomplete input, thereby strengthening the encoder's ability to model fine-grained local features and structures.

[0052] However, relying solely on the reconstruction objective may result in the model learning information that is primarily focused on low-level texture features and lacks the ability to capture abstract semantics. Therefore, we further introduce contrastive loss as supplementary supervision to enhance the model's discriminative ability and the structure of the feature space. Specifically, different enhancements of the same image are considered as positive sample pairs, and enhancements from different images are considered as negative sample pairs. The contrastive learning loss is used to define the contrastive objective: , in , It is the representation vector corresponding to the same sample after different data enhancements. is the cosine similarity, is the temperature coefficient, is the total number of comparison samples. This loss guides the model to learn to aggregate similar samples and separate heterogeneous samples, thereby improving the discriminative and generalization capabilities of features in downstream classification tasks.

[0053] In order to dynamically balance the optimization objectives of contrastive learning and reconstruction tasks during training, the present invention introduces two learnable weight parameters into the final loss function: and , corresponding to contrast loss and reconstruction loss respectively. The present invention adopts the uncertainty weighting method. The total loss is defined as follows: , in and These are the weight coefficients for the reconstruction loss and contrastive loss, respectively. These two weights, as trainable parameters of the model, are automatically updated during the optimization process, eliminating the need for manual hyperparameter setting and enabling dynamic adjustment of task importance.

[0054] The entire encoder structure is as follows Figure 5As shown in the figure, the CTFP module is integrated into the Bottleneck structure of each stage of ResNet to improve the time-frequency modeling capability; at the same time, combined with channel expansion and cross-dimensional feature fusion mechanism, the network's representation ability and time-frequency discrimination ability are significantly enhanced.

[0055] In step 3, the dataset in step 1 is used to train the constructed contrast mask image reconstruction model, and finally the modulation recognition of the radar signal is realized.

[0056] First, unlabeled samples are selected from the radar signal time-frequency image constructed in step 1. Random augmentation and masking operations are applied to them, generating two input images. The masked image serves as the input to the online branch, while the unmasked, enhanced image serves as the input to the target branch. After feature extraction from the online branch's input image, the encoder extracts features and feeds them into two modules: one is fed into a reconstruction decoder to reconstruct the original image, where a reconstruction loss is calculated based on the difference between the original and reconstructed images; the other is fed into a contrastive learning space after global average pooling and projection. The target branch, using a target encoder maintained by a momentum update mechanism, extracts corresponding features and similarly passes through the projection head into the contrastive space, forming positive pairs with the online branch. Feature consistency is enforced using a contrastive loss. These two tasks are jointly optimized, synergistically driving the model to learn discriminative and structure-aware time-frequency embedding features. After pre-training, the model parameters are transferred and some encoder modules are frozen. Fine-tuning is performed using a small amount of labeled data, ultimately achieving effective recognition of radar modulation types.

[0057] The effect of the present invention is further illustrated by the following comparative experiments: 1 Experimental environment configuration: All experiments conducted in this paper were conducted on a Linux computing platform with the following configuration: an Intel Xeon Gold 6122 processor with 64GB of RAM and an NVIDIA RTX 3090 graphics card (with 24GB of video memory). The software environment included Python 3.8, the PyTorch deep learning framework 1.12.1, and GPU driver and computing support provided by CUDA 11.6 and CuDNN 8.4. During the self-supervised pre-training phase, the total number of training epochs was set to 400, with a batch size of 4096. Initially, a low warm-start learning rate (1e-6) was used, gradually increasing to a base learning rate of 0.00001 to mitigate the risk of instability caused by rapid initial parameter updates. Later, the learning rate was gradually decayed to a minimum learning rate (1e-6), and the learning rate was dynamically adjusted using a cosine annealing mechanism to improve overall convergence performance. During the fine-tuning phase, the number of training epochs was set to 200, with a batch size of 64. The base learning rate was increased to 0.0001, and the warm-up and decay mechanisms were still used, maintaining a learning rate range of 1e-6 to 0.0001. Throughout training, the momentum parameter was set to 0.99, the temperature parameter τ was set to 0.2, and the optimizer selected was AdamW, which has weight decay capabilities. For input processing, a random masking strategy was used as the model's masking mechanism. Specifically, the time-frequency map was divided into 8×8 patches (a total of 64 blocks), 75% (i.e., 48 blocks) were randomly selected for masking, and only the remaining area was retained for reconstruction to enhance the model's robustness to incomplete input. Furthermore, to improve the sample diversity and representation capabilities of self-supervised learning, the method of the present invention introduces multiple enhancement techniques, including random cropping, horizontal flipping, color jitter, and the mixed sample enhancement methods MIXUP and CUTMIX, enabling the model to better learn semantic associations between samples.

[0058] 2 Experimental results analysis and evaluation: 2.1 Ablation Experiment (1) Structural level: In order to comprehensively evaluate the role of each module in the modulation recognition performance of the framework of the present invention, the present invention carried out a systematic ablation experimental analysis from the perspective of structural composition. As shown in Table 2, a baseline model (Supervised) without any self-supervisory mechanism or structure enhancement module was first constructed. This model is completely based on the traditional supervised training method, with a test accuracy of 83.18%, which serves as the basis for subsequent comparison. On this basis, after the cross-time-frequency pooling module (CTFP) designed in the method of the present invention is integrated into each stage of ResNet-50, the accuracy is improved to 85.07%. This module captures remote context information on the time axis and frequency axis respectively, and fuses the two types of information into a model, which makes up for the problem of conventional two-dimensional pooling ignoring the axial structure in the process of time-frequency image processing, and effectively improves the network's sensitivity and capture ability to modulation patterns. Furthermore, under the premise of introducing the structure enhancement module, the influence of different pre-training objectives on the model performance is explored. Based on the structure containing the CTFP module, contrastive learning loss and mask reconstruction loss are introduced for self-supervised training. Results show that when the contrastive loss alone is applied, the model achieves a recognition accuracy of 85.62%, demonstrating that by building a contrastive relationship between the time-frequency views, the model learns more discriminative feature representations. When pre-trained using only the reconstruction loss, the accuracy also reaches 85.47%, also outperforming the model without pre-training. This demonstrates that the mask-based local restoration task can guide the model to learn structural details of the time-frequency graph and strengthen local feature modeling. These results demonstrate that even using only a single self-supervised objective can lead to significant performance improvements. Finally, a comprehensive CMIR training framework based on the proposed method is constructed, integrating contrastive learning and reconstruction mechanisms into a unified self-supervised strategy, further improving the accuracy of the trained model to 86.35%. Compared to the schemes using the contrastive loss and reconstruction loss separately, the combined scheme achieves performance gains of 0.73% and 0.88%, respectively, validating their complementary nature in feature learning: the contrastive objective helps learn global discriminative structure, while the reconstruction objective enhances the model's ability to model fine-grained structure. The two mechanisms work synergistically, significantly enhancing the encoder's representational generalization capabilities under unlabeled data conditions.

[0059] Table 2 Ablation experiments at the structural level

[0060] (2) Masking level: Based on the above, in order to further verify the role and stability of the designed masking mechanism in self-supervised training, the impact of different masking ratios on model performance was deeply analyzed. Five different masking ratios were designed for comparative experiments, namely: 0 (no masking), 0.25, 0.5, 0.65 and 0.75, as shown in Table 3. The experimental results show that a reasonable setting of the masking ratio can significantly improve the learning effect of the model. When no masking is used (masking ratio = 0), the recognition accuracy of the model is 85.21%, indicating that even without occlusion, the image reconstruction task itself can have a regularization effect on the encoder to a certain extent, promoting its learning of potential features. However, with the increase of the masking ratio, the model performance showed slight fluctuations and reached the highest accuracy of 86.35% when the masking ratio was 0.75. This experimental phenomenon shows that in this task scenario, a higher occlusion ratio actually enhances the feature learning ability of the model. This result can be attributed to the fact that high-scale masking significantly increases the complexity of the reconstruction process, forcing the model to rely more on the global structural information of the input data to complete the reconstruction, thereby learning a more abstract and generalizable latent representation.

[0061] Table 3 Ablation experiments at the mask level

[0062] 2.2 Signal-to-Noise Ratio and Category Accuracy Analysis: In order to comprehensively examine the recognition robustness and generalization performance of the method of the present invention in complex noise environments, the classification performance of the model under different sample conditions was deeply analyzed from the two dimensions of signal-to-noise ratio and modulation signal type. Figure 6The model demonstrates classification accuracy for 10 types of modulation signals under 11 different SNR settings (ranging from −16dB to 4dB). Overall, the model demonstrates excellent recognition capabilities in the medium-to-high SNR range (i.e., SNR ≥ −6dB), with classification accuracy exceeding 98% for most modulation types. This demonstrates the method's strong interference immunity and adaptability to complex channel conditions found in real-world environments. Furthermore, in low SNR environments (such as −16dB and −14dB), despite noise significantly affecting the clarity of the modulation structure, the model maintains a certain level of discrimination, demonstrating that the self-supervised pre-training strategy effectively enhances the model's robustness in noisy environments. The performance differences among modulation types show that some modulation schemes maintain high recognition rates across the entire SNR range. For example, eqfm, lfm, ns, and frank modulation schemes demonstrate stability across all SNR conditions. eqfm achieves 100% accuracy in nearly all test scenarios, demonstrating its significant time-frequency characteristics, making it easily modeled and recognized. In contrast, the accuracy of modulation schemes such as BPSK, 4FSK, FSK, and LFM-BPSK drops significantly at low signal-to-noise ratios (SNRs), particularly between -16dB and -12dB. At -14dB, 4FSK and BPSK only reach approximately 30.00% and 18.75%, respectively. This result indicates that these schemes exhibit more blurred time-frequency plots under severe noise interference, posing significant challenges for feature extraction. However, as the SNR improves, recognition performance for these schemes rapidly recovers, returning to near-100% accuracy at levels of -6dB and above. Furthermore, for combined modulation schemes (such as LFM-BPSK and SFM-BPSK), although recognition performance is slightly lower than that of single modulation signals at extremely low SNRs, accuracy rapidly increases to approximately 90% when the SNR rises to around -6dB, demonstrating the model's strong modeling and discrimination capabilities for analyzing mixed modulation structures.

[0063] 2.3 The impact of different encoders on model performance: In order to further evaluate the versatility and scalability of the self-supervised framework proposed in this paper, different mainstream encoder structures are selected for comparative experiments under the same training configuration and loss mechanism to examine their modulation recognition performance under different signal-to-noise ratio conditions. Figure 7The selected encoders include the classic convolutional ResNet18, the standard Vision Transformer (ViT), the hierarchical modeling-based Swin Transformer, and the proposed ResNet18 with the CTFP module (ResNet18+CTFP). These four architectures represent the current mainstream architectures for image modeling: the traditional CNN, the global attention Transformer, the local sliding window Transformer, and the structured enhanced convolutional network with a frequency-time perception mechanism. The experimental results, shown in the table, show that under a fully self-supervised training mechanism, the four encoder types all demonstrate excellent recognition capabilities under high signal-to-noise ratio (SNR ≥ 0 dB). The classification accuracy of ResNet18, ViT, and SwinT all stabilizes at around 99%, demonstrating that the proposed method has good downstream adaptability and can be combined with various backbone networks to effectively classify modulated signals. However, in low signal-to-noise ratio environments (SNR ≤ −10 dB), the performance gap between the different network architectures becomes apparent, reflecting their structural differences in noise immunity and feature robustness. Specifically, in the two extreme channel conditions of −14dB and −16dB, ViT achieved accuracies of 46.75% and 32.75%, respectively. This is slightly better than SwinT (49.60% and 35.20%), but both are lower than the traditional convolutional architecture ResNet-18 (54.50% and 39.25%). This demonstrates that while the Transformer architecture possesses strong global modeling capabilities, its sensitivity to local details under strong noise interference limits its ability to extract key information. It is worth noting that integrating the Cross Time-Frequency Pooling (CTFP) module into ResNet-18 further improves the model's performance across the full SNR range. In particular, in the three extreme channel conditions of −16dB, −14dB, and −12dB, the accuracies reached 43.00%, 57.50%, and 76.00%, respectively, representing improvements of approximately 3% to 4% compared to the base version. This performance gain clearly shows that the CTFP module enhances the model's ability to express effective features in non-stationary interference scenarios by cross-modeling structural information in the time and frequency dimensions, thereby improving the overall discrimination effect.

[0064] Overall, while the Transformer architecture demonstrates superior performance in natural image tasks, its performance in radar time-frequency image modeling is somewhat unstable. This phenomenon suggests that the native architecture is not yet fully adapted to the high non-stationarity and low redundancy inherent in radar data, and still faces certain generalization bottlenecks. However, the ResNet architecture combined with the CTFP module demonstrates greater adaptability in high-noise and complex modulation environments, further validating the practicality and effectiveness of the proposed architecture.

[0065] 2.4 Comparison with other methods: To demonstrate the performance advantages of the proposed model in radar signal modulation recognition, we selected several representative neural network architectures for comparative evaluation. These included SVM, ResNet, MobileNetV2, ConvNeXt, MobileViT, VAN, LPINe, and MAPNet. These methods are widely used in computer vision and radar signal processing, each with its own unique characteristics, representing different model design approaches and features.

[0066] Under fully supervised conditions (100% labeled data), a systematic evaluation of the modulation recognition capabilities of various representative methods at varying signal-to-noise ratio (SNR) levels was conducted. As shown in Table 4, the results demonstrate that traditional methods, such as support vector machines, are nearly ineffective at low SNRs, achieving accuracy rates of only 25.12% and less than 35% at −16dB and −10dB, respectively. This is primarily due to their heavy reliance on artificial static features, making them incapable of modeling dynamic time-frequency patterns in the presence of noise. In comparison, deep learning models significantly outperform SVMs overall. For example, ResNet and ConvNeXt, these residual-based networks, maintain stable performance at medium and high SNRs, but exhibit significant performance degradation at low SNRs, with accuracy rates of 54.5% and 48%, respectively, at −14dB. While MobileNetV2 performs reasonably well at high SNRs, its accuracy drops rapidly at low SNRs, demonstrating the limited robustness of lightweight architectures to noise. Attention-based models such as VAN and MobileViT demonstrate moderate recognition capabilities within the SNR range of −6dB to +4dB, but their performance fluctuates significantly in extreme noise scenarios. For example, MobileViT achieves only 42.13% accuracy at −16dB, indicating that its cross-channel modeling capabilities are insufficient to effectively mitigate performance losses caused by feature degradation. Domain-specific architectures such as MAPNet and LPINet achieve good results at SNRs of −4dB and above, but also experience performance degradation at low SNRs, suggesting that their embedded prior mechanisms struggle to maintain stability in extreme environments. LSYNet, an existing self-supervised method, demonstrates relatively good robustness at lower SNRs, but its overall performance still falls short of the CMIR proposed in this paper. Overall, CMIR achieves optimal or near-optimal classification performance across the entire SNR range, particularly exceeding 90% accuracy in scenarios of −10dB and above, and reaching 100% at 0dB. At extremely low SNR (−16dB, −14dB), the accuracy rates reached 43.00% and 57.50% respectively, significantly outperforming other comparison models.

[0067] Table 4 Recognition accuracy of different methods under different labeled training data

[0068] Under weak supervision, using only 10% labeled samples, all models generally experience performance degradation, but the robustness differences between different architectures are even more pronounced, reflecting their generalization capabilities under limited supervision. Fully supervised networks such as ResNet, ConvNeXt, and MobileNetV2 experience a significant decline in recognition accuracy when training data is scarce. For example, at −10 dB, the accuracy drops to 71.13%, 33.63%, and 66.88%, respectively, demonstrating their high reliance on labeled data. Attention-driven models such as MobileViT perform more erratically, prone to overfitting when the training set is reduced, with accuracy dropping to 47.88% at −12 dB. Although MAPNet and LPINet incorporate domain priors, their structural advantages are limited when training data is limited, resulting in accuracies of only 26.13% and 31.38%, respectively, at −16 dB. While LSYNet utilizes a contrastive learning mechanism and performs relatively robustly at −10dB and −8dB noise levels (with accuracy rates of 75.23% and 84.14%, respectively), it lacks a reconstruction objective and thus suffers from the dual effects of extreme noise and insufficient data. In contrast, CMIR, with its self-supervised training strategy combining contrast and reconstruction, effectively mitigates the model degradation caused by weak supervision. Accuracies reached 76.38%, 86.38%, and 92.75% at −10dB, −8dB, and −6dB, respectively; even at −16dB noise, accuracy remained at 34.75%. In the medium-to-high signal-to-noise ratio range (0dB and above), accuracy remained consistently above 97%, surpassing all compared methods.

[0069] 2.5 Complexity Analysis: While ensuring excellent recognition performance, the model's computational overhead and parameter size are also key considerations during practical deployment. To evaluate the resource feasibility of the proposed method, a quantitative comparison of the computational complexity (measured in FLOPs) and parameter count of various comparison models at a standard input dimension (1×64×64) is presented, as shown in Table 5. The evaluation covers a wide range of categories, including typical convolutional architectures, lightweight networks, Transformer-based models, specialized structures for modulation recognition, and self-supervised learning methods. The relevant statistical results are detailed in the table. The proposed CMIR method has approximately 274.85M FLOPs and 18.31M parameters, placing it in the middle to high range among the compared models. In comparison, MobileNetV2 and MobileViT have significantly lower computational costs, with 26.06M and 21.43M FLOPs, respectively, and parameters under 3M, demonstrating their advantages for deployment on edge devices and mobile platforms. However, these lightweight models significantly lag behind performance-oriented architectures in terms of accuracy. Among high-complexity models, ConvNeXt-tiny has the highest computational requirements, reaching 362.88M FLOPs, while LSYNet and LPINet have 183.54M and 232.59M FLOPs, respectively, also at relatively high levels. In terms of parameter size, ConvNeXt-tiny and LSYNet have 27.8M and 14.75M parameters, respectively, significantly exceeding the proposed architecture. Notably, despite their extremely compact parameter size (less than 0.3M), LPINet and MAPNet exhibited significant performance disadvantages in previous recognition experiments, suggesting that their lightweight designs present accuracy bottlenecks. A comprehensive evaluation demonstrates that the proposed approach maintains leading recognition accuracy without employing a highly complex or bloated network structure, demonstrating an overall good balance between performance and resource consumption. While its resource usage is slightly increased compared to ultra-lightweight models, its structure maintains a modular, clear, and easy-to-understand design, facilitating project implementation and maintenance.

[0070] Table 5 Comparison of FLOPs and Params of different models

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A radar signal modulation recognition method based on self-supervised contrast mask reconstruction, characterized in that: The steps include: Step 1: Acquire radar signals and construct a radar modulation signal dataset containing labeled data and unlabeled data, perform improved multiple synchronous compression transform time-frequency analysis on the radar modulation signal, and convert it into a time-frequency image; Step 2: Build a self-supervised contrast mask image reconstruction model, which includes an online branch and a target branch. The online branch includes an improved ResNet encoder with an integrated cross-time-frequency pooling module, a reconstruction decoder, and a projection head. The target branch includes a target encoder and a projection head whose parameters are updated by an exponential sliding average mechanism. Step 3: Pre-train the self-supervised contrast mask image reconstruction model using the unlabeled data, apply data augmentation and random masking operations to the time-frequency image to generate a dual view, input the dual view into the online branch and the target branch respectively, and update the model parameters by jointly optimizing the reconstruction loss and the contrast loss to obtain the pre-trained weights; Step 4: Migrate the pre-trained weights to the downstream recognition network, freeze some encoder parameters, and then use a small amount of labeled data for fine-tuning to obtain a radar signal modulation recognition model.

2. The radar signal modulation recognition method based on self-supervised contrast mask reconstruction according to claim 1 is characterized in that: The step 1 comprises: Step 11, generating a plurality of radar modulation signals according to the radar signal; Step 12: Perform time-frequency transformation on the radar modulated signal using the IMSST method. The IMSST method decomposes the signal through multi-scale analysis and then adjusts the time-frequency energy distribution through synchronous squeezing to concentrate the energy on the actual time-frequency trajectory. Step 13: After time-frequency transformation, a time-frequency image with a size of 1×64×64 is obtained; Step 14: Integrate all the obtained time-frequency images to form a radar modulation signal time-frequency image dataset containing labeled data and unlabeled data.

3. The radar signal modulation recognition method based on self-supervised contrast mask reconstruction according to claim 1 is characterized in that: The step 2 comprises: Step 21: Design the overall framework of the self-supervised contrast mask image reconstruction model, and clarify the composition and connection relationship between the online branch and the target branch; Step 22, constructing an improved ResNet encoder integrating a cross-time-frequency pooling module, wherein the improved ResNet encoder includes five consecutive stages; Step 23: Design an online branch reconstruction decoder. The reconstruction decoder consists of a five-layer deconvolution module. Each layer includes a transposed convolution, batch normalization, and a ReLU activation function. The feature space size is restored layer by layer, and the final output is a reconstructed image with the same size as the input time-frequency map. Step 24: Design the projection heads of the online branch and the target branch. The projection heads are both two-layer fully connected networks. The first layer maps the 512-dimensional high-dimensional features output by the encoder to 256 dimensions, and the second layer maps them to 128 dimensions, which are used to project the features into the contrast space. Step 25, set the parameter update mechanism of the target encoder: the target encoder parameters are synchronously updated from the improved ResNet encoder of the online branch through the EMA mechanism; Step 26, construct the joint loss function: ， in and are the weight coefficients of reconstruction loss and contrast loss, and They are reconstruction loss and contrast loss respectively.

4. The radar signal modulation recognition method based on self-supervised contrast mask reconstruction according to claim 3 is characterized in that: In step 21, the online branch includes an improved ResNet encoder with an integrated cross-time-frequency pooling module, a reconstruction decoder, and a projection head; the target branch includes a target encoder and a projection head. The target encoder has the same structure as the improved ResNet encoder, and its parameters are synchronously updated by the improved ResNet encoder parameters of the online branch through an exponential sliding average mechanism. The radar signal is transformed into a 1×64×64 time-frequency map through IMSST. Two views are generated through two types of enhancement operations. The view with random mask is input into the online branch, and the unmasked view after data enhancement is input into the target branch.

5. The radar signal modulation recognition method based on self-supervised contrast mask reconstruction according to claim 3 is characterized in that: The specific process of step 22 is as follows: Based on the ResNet-18 architecture, five consecutive stages are set up: layer1-layer5. In the layer1 stage, the initial feature extraction of the input time-frequency map is performed through the convolution layer and the maximum pooling layer. From layer2 to layer5, deep feature modeling is performed through residual blocks. The number of channels increases from 64 to 128 to 256 to 512, and the feature map resolution decreases. A CTFP module is inserted after the 3×3 convolution of the last residual unit in each stage from layer 2 to layer 5, and after all residual blocks in the layer 5 stage. The CTFP module performs one-dimensional strip pooling in the time dimension and frequency dimension respectively, and enhances the time-frequency correlation of features by fusing the pooling results. Specifically, it includes: Perform horizontal strip pooling on the input feature map and output for: , where x is the input feature map and W is the feature map width; Perform vertical strip pooling on the input feature map and output for: , where H is the feature map height; Pooling the horizontal strips Compared with the vertical strip pooling result After 1×1 convolution fusion and sigmoid activation, it is multiplied element by element with the original feature map to obtain the enhanced features: ， ， in, represents element-wise multiplication, is the sigmoid function, is a 1×1 convolution, is the output of cross-time-frequency pooling, and C is the channel dimension of the feature map.

6. The radar signal modulation recognition method based on self-supervised contrast mask reconstruction according to claim 3 is characterized in that: The constructing of the joint loss function includes: Define the reconstruction loss that only calculates the pixel error in the mask area: ， in Indicates the combination of the obscured positions, and Represent the pixel values ​​of the original image and the model reconstructed image at the i-th position respectively; InfoNCE loss is used as the comparison loss: ， in , It is the representation vector corresponding to the same sample after different data enhancements. is the cosine similarity, is the temperature coefficient, is the total number of comparison samples; Two learnable weight parameters are introduced for contrast loss and reconstruction loss respectively and , construct the final joint loss function.

7. The radar signal modulation recognition method based on self-supervised contrast mask reconstruction according to claim 1, characterized in that: The step 3 comprises: Step 31: Prepare an unlabeled time-frequency image dataset: Use an improved multiple synchronous compression transform to obtain 1×64×64 size time-frequency images. The images in the unlabeled time-frequency image dataset are derived from ten radar modulation signals and cover samples with a signal-to-noise ratio of -16 dB to 4 dB. Step 32, generating dual-view input: performing two-class processing on each unlabeled time-frequency image in the dataset to construct a first view and a second view; Step 33, dual-branch feature extraction and processing: The first view is input into the improved ResNet encoder with integrated cross-time-frequency pooling module for online branch processing. After extracting high-dimensional features, it is divided into two paths: one path is input into the reconstruction decoder to reconstruct the complete time-frequency image, and the other path is input into the projection head after global average pooling and mapped to the contrast space; the second view is input into the target encoder for target branch processing, and after extracting features, it is mapped to the contrast space through global average pooling and the projection head; Step 34, joint loss function calculation; Step 35, model parameter update and pre-training weight acquisition: optimize the total loss through backpropagation, update the parameters of the improved ResNet encoder, reconstruction decoder and projection head; update the target encoder and projection head parameters through the EMA mechanism; jointly optimize the reconstruction loss and contrast loss to complete pre-training, and finally save the parameters of the improved ResNet encoder as pre-training weights.

8. The radar signal modulation recognition method based on self-supervised contrast mask reconstruction according to claim 7, characterized in that: The first view is the input of the online branch: conventional data augmentation is first applied, followed by a random masking operation, which divides the image into 8×8 patches and randomly masks 75% of the area; The second view is the target branch input: only the same regular data augmentation as the first view is applied, without masking.

9. The radar signal modulation recognition method based on self-supervised contrast mask reconstruction according to any one of claims 1 to 8, characterized in that: The step 4 comprises: Step 41: Construct a downstream recognition network structure: the downstream recognition network uses an improved ResNet encoder with an integrated cross-time-frequency pooling module as the feature extraction backbone, and a lightweight classification head is connected to the encoder output. The classification head includes two fully connected layers. The first layer maps the 512-dimensional features output by the encoder to 256 dimensions, and the second layer maps to an output dimension that matches the number of modulation types, and is used to output the probability distribution of each category. Step 42, transfer pre-trained weights: reconstruct the contrast mask image from the improved ResNet encoder parameters obtained by pre-training the supervised model, and transfer them as a whole to the corresponding encoder of the downstream recognition network as the initial weights; Step 43, freeze some encoder parameters: freeze all parameters of the first three stages of the improved ResNet encoder, and only keep the parameters of the last two stages and the classification head parameters trainable; Step 44, prepare a fine-tuning dataset: use a small amount of labeled time-frequency image data. The data of the fine-tuning dataset is derived from IMSST time-frequency graphs of ten radar modulation signals, covering signal-to-noise ratios from -16dB to 4dB. The number of labeled samples is 50-200 for each type of signal at each signal-to-noise ratio. Step 45, fine-tuning training process: take the fine-tuning dataset as input, output the category prediction results through the downstream recognition network, and use the cross-entropy loss function to calculate the error between the prediction results and the labels; use the Adam optimizer to optimize the cross-entropy loss, update the parameters of the last two stages of the trainable encoder and the classification head parameters, set the initial learning rate to 0.0001, the batch size to 64, the training rounds to 200 rounds, and use the cosine annealing mechanism to decay the learning rate from the initial value to 1×10⁻ 6 ; Step 46, obtaining a radar signal modulation recognition model: After the fine-tuning training is completed, all parameters of the downstream recognition network are saved to form the final radar signal modulation recognition model.

10. A radar signal modulation recognition device based on self-supervised contrast mask reconstruction, characterized in that: include: A data set construction module is used to acquire radar signals and construct a radar modulation signal data set containing labeled data and unlabeled data, perform improved multiple synchronous compression transform time-frequency analysis on the radar modulation signal, and convert it into a time-frequency image; A model building module for building a self-supervised contrast mask image reconstruction model, wherein the self-supervised contrast mask image reconstruction model includes an online branch and a target branch. The online branch includes an improved ResNet encoder with an integrated cross-time-frequency pooling module, a reconstruction decoder, and a projection head. The target branch includes a target encoder and a projection head whose parameters are updated via an exponential sliding average mechanism. a pre-training module for pre-training the self-supervised contrastive mask image reconstruction model using the unlabeled data, applying data augmentation and random masking operations to the time-frequency image to generate a dual view, inputting the dual view into the online branch and the target branch respectively, and updating the model parameters by jointly optimizing the reconstruction loss and the contrast loss to obtain the pre-trained weights; The fine-tuning module is used to transfer the pre-trained weights to the downstream recognition network, freeze some encoder parameters, and then use a small amount of labeled data for fine-tuning to obtain a radar signal modulation recognition model.

Images