Radar signal modulation recognition method and device based on self-supervised contrastive mask reconstruction

Abstract

The application discloses a radar signal modulation identification method and device based on self-supervised contrast mask reconstruction, and belongs to the technical field of radar signal modulation identification. The method comprises the following steps: acquiring a radar signal and constructing a radar modulation signal dataset containing labeled data and unlabeled data, and converting the radar modulation signal dataset into a time-frequency image; a self-supervised contrast mask image reconstruction model comprising an online branch and a target branch is built; the self-supervised contrast mask image reconstruction model is pre-trained by using the unlabeled data, data enhancement and random mask operation are performed on the time-frequency image to generate double views, the double views are respectively input into the online branch and the target branch, model parameters are updated by jointly optimizing reconstruction loss and contrast loss, and pre-training weights are obtained; the pre-training weights are migrated to a downstream identification network, a part of encoder parameters is frozen, a small amount of labeled data is used for fine-tuning, and a radar signal modulation identification model is obtained. The application improves the radar signal modulation identification precision in a complex electromagnetic environment.

Description

TECHNICAL FIELD

[0001] The application relates to a radar signal modulation recognition method and device based on self-supervised contrastive mask reconstruction, and belongs to the technical field of radar signal modulation recognition. BACKGROUND

[0002] Radar signal modulation recognition (RSMR) is a key technology in radar signal processing and wireless communication systems, and has wide applications in electronic countermeasures, spectrum monitoring and intelligent communication. The core goal is to accurately identify the modulation mode from the received signal without prior information. In practical applications, the modulation recognition task still faces many challenges. On the one hand, the complexity of the electromagnetic environment makes the signal often affected by strong interference and background noise, significantly increasing the difficulty of recognition. On the other hand, with the increasing complexity of radar system and signal design, the modulation mode presents high diversity and complexity, and traditional recognition algorithms show obvious lack of adaptability in the face of new modulation modes, especially under low signal-to-noise ratio (SNR) conditions. In addition, many actual scenarios have strict requirements for the response speed of the system, which also puts higher standards on the balance between precision and operation efficiency of the modulation recognition method.

[0003] Traditional modulation recognition methods mainly rely on manually designed feature extraction algorithms, such as cyclic spectrum, instantaneous frequency and time-frequency diagram analysis. Although these methods can achieve certain recognition performance under ideal conditions, they rely on the experience of signal processing experts and have poor robustness in non-ideal environments. In addition, it is difficult for this kind of method to model complex modulation structures and mixed signals, and the recognition accuracy is difficult to meet the needs of modern systems. Therefore, machine learning methods are introduced into RSMR research to automatically extract more discriminative features.

[0004] With the development of machine learning, algorithms such as support vector machine (SVM) have been applied to modulation recognition tasks. Through classification training on original signals or manually extracted features, these methods have improved performance to some extent. However, they still rely on strong feature engineering capabilities and have limited generalization. In recent years, deep learning technology has made breakthrough progress in image, speech and signal processing fields. Deep learning methods can learn effective representations from a large amount of data through the construction of deep neural networks, which can greatly improve recognition accuracy. In particular, architectures such as convolutional neural network (CNN) and self-attention can automatically extract deep semantic features from original I / Q signals, spectrum diagrams or time-frequency diagrams, significantly improving recognition performance. Now, more and more researchers are considering applying it to radar signal modulation recognition.

[0005] Although deep learning methods have shown strong modeling capabilities in the field of RSMR, two core problems still exist in its wide application: first, deep models usually require a large amount of high-quality labeled data for training, but the cost of radar signal data collection and labeling is extremely high, and it is difficult to collect large-scale high-quality reliable radar data in practice; second, the feature extraction ability and discrimination performance of the model are still limited under extreme conditions such as low signal-to-noise ratio. How to train a modulation recognition model with strong generalization and high robustness under low resource and weak supervision has become a hot and difficult point of current research.

[0006] To solve the dependence of supervised learning on labeled data, self-supervised learning (SSL) provides a solution for such problems, which has made significant progress in computer vision, speech processing and other fields in recent years, and has been gradually introduced into the modulation recognition task. Self-supervised learning learns general representations by designing pre-training tasks (such as predicting occluded regions, sorting transformations, and distinguishing contrastive samples), and migrates to downstream tasks, thereby reducing the dependence on labeled data. Current mainstream self-supervised methods can be roughly divided into two categories: discriminative learning and generative learning. The former constructs positive and negative sample pairs, so that the model learns a feature space structure that remains consistent for the same instance under different transformations and distinguishes between different instances; the latter mainly adopts a mask modeling strategy, randomly masks part of the input, and predicts the masked content to guide the model to learn the structural relationship between local and global. Typical contrastive learning and generative learning frameworks have been tried for modulation recognition tasks and have achieved certain results. However, the learning mode of a single strategy has its own limitations: contrastive learning focuses on global feature alignment, ignoring local semantic integrity; while generative methods can capture detailed structures, but are prone to ambiguous expressions in global discrimination tasks, affecting classification performance. SUMMARY

[0007] To solve the above problems, the present application provides a radar signal modulation recognition method and device based on self-supervised contrastive mask reconstruction, which can improve the performance of radar signal modulation recognition in low resource scenarios.

[0008] The technical scheme adopted by the present application to solve its technical problems is:

[0009] In a first aspect, the present application provides a radar signal modulation recognition method based on self-supervised contrastive mask reconstruction, comprising the following steps:

[0010] Step 1, acquiring radar signals and constructing a radar modulation signal dataset containing labeled data and unlabeled data, performing improved multiple simultaneous synchronization compression transformation (IMSST) time-frequency analysis processing on the radar modulation signal, and converting it into a time-frequency image;

[0011] Step 2, a self-supervised contrastive mask image reconstruction model is built, the self-supervised contrastive mask image reconstruction model comprises an online branch and a target branch, the online branch comprises an improved ResNet encoder integrating a cross-time-frequency pooling (CTFP) module, a reconstruction decoder and a projection head, and the target branch comprises a target encoder and a projection head whose parameters are updated through an exponential moving average (EMA) mechanism;

[0012] Step 3, the self-supervised contrastive mask image reconstruction model is pre-trained using the unlabeled data, data enhancement and random mask operation are performed on the time-frequency image to generate double views, the online branch and the target branch are respectively input, the model parameters are updated through joint optimization of reconstruction loss and contrast loss, and pre-training weights are obtained;

[0013] Step 4, the pre-training weights are migrated to a downstream recognition network, after freezing part of the encoder parameters, a small amount of labeled data is used for fine-tuning to obtain a radar signal modulation recognition model.

[0014] As a possible implementation manner of the embodiment, the step 1 comprises:

[0015] Step 11, a plurality of radar modulation signals are generated according to the radar signal;

[0016] Step 12, an IMSST method is used to perform time-frequency transformation on the radar modulation signal, the IMSST decomposes the signal through multi-scale analysis, and then adjusts the time-frequency energy distribution through synchronous squeezing to make the energy gather on the actual time-frequency track;

[0017] Step 13, after the time-frequency transformation processing, a time-frequency image with a size of 1×64×64 is obtained;

[0018] Step 14, all obtained time-frequency images are integrated to form a radar modulation signal time-frequency image dataset containing labeled data and unlabeled data.

[0019] As a possible implementation manner of the embodiment, the step 2 comprises:

[0020] Step 21, the overall framework of the self-supervised contrastive mask image reconstruction model is designed, and the composition and connection relationship of the online branch and the target branch are determined;

[0021] Step 22, an improved ResNet encoder integrating a cross-time-frequency pooling (CTFP) module is constructed, the improved ResNet encoder comprises five consecutive stages (layer1 to layer5);

[0022] Step 23, design the reconstruction decoder of the online branch, which is composed of five deconvolution modules, each layer contains transpose convolution, batch normalization and ReLU activation function, and the feature space size is recovered layer by layer, and finally the reconstructed image with the same size as the input time-frequency map (1×64×64) is output;

[0023] Step 24, design the projection head of the online branch and the target branch, which 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 to 128 dimensions, which is used to project the features to the contrast space;

[0024] Step 25, set the parameter update mechanism of the target encoder: the parameters of the target encoder are updated synchronously from the improved ResNet encoder of the online branch through the EMA mechanism;

[0025] Step 26, construct the joint loss function:

[0026] ,

[0027] Wherein and are the weight coefficients of the reconstruction loss and the contrast loss respectively, and are the reconstruction loss and the contrast loss respectively.

[0028] As a possible implementation manner of the embodiment, in step 21, the online branch includes an improved ResNet encoder integrated with a 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 the parameters of the target encoder are updated synchronously from the improved ResNet encoder of the online branch through the exponential moving average (EMA) mechanism; the 1×64×64 time-frequency map obtained by transforming the radar signal through IMSST is generated into two views through two types of enhancement operations, wherein the view with random mask (masking ratio 75%) is input into the online branch, and the unmasked view with regular data enhancement (cropping, scale transformation, noise disturbance) is input into the target branch.

[0029] As a possible implementation manner of the embodiment, the specific process of step 22 is as follows:

[0030] Taking ResNet-18 as the basic architecture, five consecutive stages layer1-layer5 are set, the layer1 stage performs initial feature extraction on the input time-frequency map through convolution layers and maximum pooling layers, and the layer2 to layer5 stages perform deep feature modeling through residual blocks, accompanied by increasing the channel number in the mode of 64→128→256→512 and decreasing the feature map resolution;

[0031] After the 3x3 convolution of the last residual unit in each stage of layer2 to layer5 stage, and after all residual blocks of layer5 stage, the CTFP module is inserted; the CTFP module respectively performs one-dimensional strip pooling in time dimension and frequency dimension, and enhances the time-frequency correlation of features by fusing the pooling results, specifically including:

[0032] Performing horizontal strip pooling on the input feature map, output is: wherein x is the input feature map, and W is the width of the feature map;

[0033] Performing vertical strip pooling on the input feature map, output is:

[0034] wherein H is the height of the feature map;

[0035] The horizontal strip pooling result and the vertical strip pooling result are fused by 1x1 convolution, and then are activated by sigmoid, and then are multiplied with the original feature map element by element to obtain the enhanced feature:

[0036] ,

[0037] ,

[0038] wherein, represents element-wise multiplication, is a sigmoid function, is 1x1 convolution, is the output of cross-time-frequency pooling (CTFP), and C is the channel dimension of the feature map.

[0039] As a possible implementation manner of the embodiment, the joint loss function is constructed, including:

[0040] Defining a reconstruction loss that only calculates the pixel error of the mask area:

[0041] ,

[0042] wherein represents the position of the mask, and respectively represent the pixel value of the original image and the model reconstructed image at the i-th position;

[0043] Using InfoNCE loss as the contrast loss:

[0044] ,

[0045] wherein , is the representation vector corresponding to the same sample after different data augmentation, is the cosine similarity, is the temperature coefficient, is the total number of contrast samples;

[0046] The corresponding contrast loss and reconstruction loss respectively introduce two learnable weight parameters and to construct the final joint loss function.

[0047] As a possible implementation manner of the embodiment, the step 3 comprises:

[0048] Step 31, preparing an unlabeled time-frequency image dataset: obtaining a 1×64×64 size time-frequency image by using an improved multiple synchronous synchronous compression transformation (IMSST), and the images of 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 signal-to-noise ratios (SNRs) of-16 dB to 4 dB (with an interval of 2 dB);

[0049] Step 32, generating a double-view input: performing two types of processing on each unlabeled time-frequency image in the dataset to construct a first view and a second view;

[0050] Step 33, double-branch feature extraction and processing: the first view input is processed by an improved ResNet encoder integrated with a cross-time-frequency pooling (CTFP) module in an online branch to extract high-dimensional features, which are then divided into two paths: one path inputs a reconstruction decoder to reconstruct a complete time-frequency image, and the other path inputs a projection head after global average pooling to map to a contrast space; the second view input is processed by a target encoder (with the same structure as the improved ResNet encoder, and the parameters are updated from the online branch encoder by an exponential moving average (EMA) mechanism) in a target branch to extract features, which are then mapped to the contrast space by global average pooling and the projection head;

[0051] Step 34, joint loss function calculation;

[0052] Step 35, model parameter updating and pre-training weight acquisition: the parameters of the improved ResNet encoder, the reconstruction decoder, and the projection head are updated by back propagation optimization of the total loss; the parameters of the target encoder and the projection head are updated by the EMA mechanism; the reconstruction loss and the contrast loss are jointly optimized to complete pre-training, and finally the parameters of the improved ResNet encoder are saved as pre-training weights.

[0053] As a possible implementation manner of the embodiment, the first view is an online branch input: first, conventional data enhancement (including random cropping, scale transformation, noise disturbance) is applied, and then random mask operation is performed to divide the image into 8x8 patches (a total of 64 patches), and 75% (48 patches) of the regions are randomly masked;

[0054] The second view is a target branch input: only the same conventional data enhancement as the first view is applied, and no mask is performed.

[0055] As a possible implementation manner of the embodiment, the step 4 comprises:

[0056] Step 41, constructing a downstream recognition network structure: the downstream recognition network takes an improved ResNet encoder with an integrated cross-time-frequency pooling (CTFP) module as a feature extraction backbone, and connects a lightweight classification head at the output end of the encoder; 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 matching the number of modulation types (10 categories) for outputting the probability distribution of each category;

[0057] Step 42, migrating pre-training weights: the improved ResNet encoder parameters obtained by pre-training the contrast mask image reconstruction self-supervised model are migrated to the corresponding encoder of the downstream recognition network as initial weights;

[0058] Step 43, freezing part of the encoder parameters: all parameters in the first three stages (layer1 to layer3) of the improved ResNet encoder are frozen, and only the parameters of the last two stages (layer4 to layer5) and the classification head parameters are trainable;

[0059] Step 44, preparing a fine-tuning dataset: a small amount of labeled time-frequency image data is used, and the data of the fine-tuning dataset is derived from IMSST time-frequency graphs of ten radar modulation signals (LFM, SFM, BPSK, LFM-BPSK, SFM-BPSK, EQFM, FSK, 4FSK, NS, and Frank), covering signal-to-noise ratios of -16dB to 4dB (with an interval of 2dB), and the number of labeled samples is 50-200 for each signal per signal-to-noise ratio;

[0060] Step 45, fine-tuning training process: taking the fine-tuning dataset as input, the downstream recognition network outputs a category prediction result, and a cross-entropy loss function is used to calculate the error between the prediction result and the label; the Adam optimizer is used to optimize the cross-entropy loss to update the trainable encoder parameters of the last two stages and the classification head parameters, the initial learning rate is set to 0.0001, the batch size is 64, the training rounds are 200, and the learning rate is decayed from the initial value to 1x10⁻ 6 ;

[0061] Step 46, obtaining a radar signal modulation identification model: after fine-tuning is completed, all parameters of the downstream identification network are saved to form a final radar signal modulation identification model.

[0062] In a second aspect, the embodiment of the present application provides a radar signal modulation identification device based on self-supervised contrastive mask reconstruction, comprising:

[0063] A data set construction module is configured to obtain radar signals and construct a radar modulation signal data set containing labeled data and unlabeled data, and perform improved multiple simultaneous synchronization compression transformation (IMSST) time-frequency analysis processing on the radar modulation signal to convert it into a time-frequency image.

[0064] A model building module is configured to build a self-supervised contrastive mask image reconstruction model, wherein the self-supervised contrastive mask image reconstruction model comprises an online branch and a target branch, the online branch comprises an improved ResNet encoder integrated with a cross-time-frequency pooling (CTFP) module, a reconstruction decoder, and a projection head, and the target branch comprises a target encoder and a projection head whose parameters are updated through an exponential moving average (EMA) mechanism.

[0065] A pre-training module is configured to pre-train the self-supervised contrastive mask image reconstruction model using the unlabeled data, apply data enhancement and random mask operation to the time-frequency image to generate double views, input the double views into the online branch and the target branch respectively, update the model parameters through joint optimization of reconstruction loss and contrast loss, and obtain pre-training weights.

[0066] A fine-tuning module is configured to migrate the pre-training weights to a downstream identification network, freeze part of the encoder parameters, and fine-tune using a small amount of labeled data to obtain a radar signal modulation identification model.

[0067] The technical scheme of the embodiment of the present application has the following beneficial effects:

[0068] The present application uses unlabeled data to learn features with complete structure and discriminability through joint optimization of contrastive learning and mask image reconstruction, designs a cross-time-frequency pooling module to enhance time-frequency correlation modeling capability, and finally realizes high-precision identification under a small amount of labeled data, solving the problem of strong dependence on labeled data and poor performance in low signal-to-noise ratio scenarios of traditional methods, and significantly improving the identification accuracy in complex electromagnetic environments.

[0069] The present application fully utilizes unlabeled data through self-supervised pre-training, and only 10% of labeled data is needed to achieve identification performance close to full supervision, reducing the dependence on labeled data and reducing the cost of data labeling.

[0070] The mask reconstruction task of the application forces the model to learn the time-frequency structure features resistant to noise, and the contrast learning enhances the category distinguishability, and still maintains an identification rate of more than 43% under an extreme low signal-to-noise ratio of-16dB, thereby improving the low signal-to-noise ratio robustness.

[0071] The CTFP module of the application enhances the time-frequency correlation modeling by time-frequency two-dimensional global pooling and attention fusion, and effectively captures the non-stationary modulation features (such as the time-frequency trajectory of frequency hopping and linear frequency modulation) of the radar signal.

[0072] The application improves the ResNet encoder parameter quantity by about 11M, and the calculation complexity is 274.85M FLOPs, which balances the accuracy and efficiency and is suitable for deployment in resource-limited scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0073] Figure 1 It is a radar signal modulation identification method flow chart of self-supervised contrast mask reconstruction according to an exemplary embodiment;

[0074] Figure 2 It is a structural schematic diagram of a radar signal modulation identification device of self-supervised contrast mask reconstruction according to an exemplary embodiment;

[0075] Figure 3 It is a contrast mask push-to-reconstruction model framework diagram according to an exemplary embodiment;

[0076] Figure 4 It is a cross-time-frequency pooling mechanism schematic diagram according to an exemplary embodiment;

[0077] Figure 5 It is a ResNet improved encoder structure diagram according to an exemplary embodiment;

[0078] Figure 6 It is an identification accuracy diagram of the application method for each signal at each signal-to-noise ratio according to an exemplary embodiment;

[0079] Figure 7 It is a diagram showing the influence of different encoders on the identification accuracy in the proposed framework according to an exemplary embodiment. DETAILED DESCRIPTION

[0080] In order to more clearly illustrate the technical features of the application scheme, the application will be described in detail below through specific embodiments, and combined with the accompanying drawings.

[0081] As shown in Figure 1 , the radar signal modulation identification method of self-supervised contrast mask reconstruction provided by the embodiment of the application comprises the following steps:

[0082] Step 1, acquiring radar signals and constructing a radar modulation signal dataset containing labeled data and unlabeled data, performing improved multiple synchronous synchrosqueezing transform (IMSST) time-frequency analysis processing on the radar modulation signals, and converting them into time-frequency images;

[0083] Step 2, building a self-supervised contrastive mask image reconstruction model, which includes an online branch and a target branch, the online branch containing an improved ResNet encoder integrated with a cross-time-frequency pooling (CTFP) module, a reconstruction decoder, and a projection head, and the target branch containing a target encoder and a projection head whose parameters are updated through an exponential moving average (EMA) mechanism;

[0084] Step 3, pre-training the self-supervised contrastive mask image reconstruction model using the unlabeled data, applying data augmentation and random mask operations to the time-frequency images to generate double views, inputting them into the online branch and the target branch respectively, updating the model parameters through joint optimization of reconstruction loss and contrastive loss, and obtaining pre-training weights;

[0085] Step 4, migrating the pre-training weights to a downstream recognition network, fine-tuning it with a small amount of labeled data after freezing part of the encoder parameters, and obtaining a radar signal modulation recognition model.

[0086] As a possible implementation manner of the embodiment, the step 1 comprises:

[0087] Step 11, generating multiple radar modulation signals according to radar signals;

[0088] Step 12, performing time-frequency transformation on the radar modulation signals using the IMSST method, which decomposes signals through multi-scale analysis and adjusts time-frequency energy distribution through synchronous squeezing to concentrate energy on actual time-frequency tracks;

[0089] Step 13, obtaining time-frequency images with a size of 1×64×64 after time-frequency transformation processing;

[0090] Step 14, integrating all obtained time-frequency images to form a radar modulation signal time-frequency image dataset containing labeled data and unlabeled data.

[0091] As a possible implementation manner of the embodiment, the step 2 comprises:

[0092] Step 21, designing the overall framework of the self-supervised contrastive mask image reconstruction model, and specifying the composition and connection relationship of the online branch and the target branch;

[0093] Step 22, constructing an improved ResNet encoder integrated with cross-time-frequency pooling (CTFP) module, which includes five consecutive stages (layer1 to layer5);

[0094] Step 23, designing a reconstruction decoder for online branch, which is composed of five deconvolution modules, each containing transpose convolution, batch normalization and ReLU activation function, recovering feature space size layer by layer, and finally outputting a reconstructed image consistent with the input time-frequency map size (1×64×64);

[0095] Step 24, designing projection heads for online branch and target branch, both of which are two-layer fully connected networks, the first layer maps high-dimensional features (512 dimensions) output by the encoder to 256 dimensions, and the second layer maps to 128 dimensions, used to project features to contrast space;

[0096] Step 25, setting the parameter update mechanism of the target encoder: the parameters of the target encoder are updated synchronously from the improved ResNet encoder of the online branch through the EMA mechanism;

[0097] Step 26, constructing a joint loss function:

[0098] ,

[0099] wherein and are the weight coefficients of reconstruction loss and contrast loss respectively, and are the reconstruction loss and contrast loss respectively.

[0100] As a possible implementation manner of the embodiment, in step 21, the online branch includes an improved ResNet encoder integrated with a 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 updated synchronously from the parameters of the improved ResNet encoder of the online branch through the exponential moving average (EMA) mechanism; the radar signal is transformed into a 1×64×64 time-frequency map through IMSST, and two views are generated through two types of enhancement operations, wherein the view with random mask (masking ratio 75%) is input into the online branch, and the unmasked view with regular data enhancement (clipping, scale transformation, noise disturbance) is input into the target branch.

[0101] As a possible implementation manner of the embodiment, the specific process of step 22 is as follows:

[0102] ResNet-18 as the basic architecture, set five consecutive stages layer1-layer5, layer1 stage through the convolution layer and the maximum pooling layer to the initial feature extraction of input time-frequency graph, layer2 to layer5 stage through the residual block deep feature modeling, accompanied by the number of channels with 64→128→256→512 way of increasing and feature map resolution decreases;

[0103] In layer2 to layer5 stage in each stage after the last residual unit of 3×3 convolution, and layer5 stage all residual block after inserting CTFP module; the CTFP module respectively in time dimension and frequency dimension on the execution of one-dimensional strip pool, through the fusion of the pool result enhances the time-frequency correlation of the feature, specifically including:

[0104] The input feature map is executed horizontally strip pool, and the output For: , wherein x is the input feature map, and W is the width of the feature map;

[0105] The input feature map is executed vertically strip pool, and the output For:

[0106] , wherein H is the height of the feature map;

[0107] The horizontal strip pool result And the vertical strip pool result Through 1×1 convolution fusion after sigmoid activation, and then multiplied by the original feature map element by element, the enhanced feature is obtained:

[0108] ,

[0109] ,

[0110] Wherein, Indicates element by element multiplication, Sigmoid function, 1×1 convolution, The output of cross-time-frequency pool (CTFP) is C, the channel dimension of the feature map.

[0111] As a possible implementation manner of the embodiment, the construction of the joint loss function includes:

[0112] Define the reconstruction loss only calculating the mask area pixel error:

[0113] ,

[0114] Wherein Indicates the position of the shielded combination, with respectively denote the pixel value of the original image and the model reconstructed image at the i-th position;

[0115] The InfoNCE loss is used as the contrastive loss:

[0116] ,

[0117] where , is the corresponding representation vector of the same sample after different data augmentation, is the cosine similarity, is the temperature coefficient, is the total number of contrastive samples;

[0118] The contrastive loss and the reconstruction loss are respectively introduced into two learnable weight parameters and to construct the final joint loss function.

[0119] As a possible implementation manner of the embodiment, the step 3 comprises:

[0120] Step 31, preparing an unlabeled time-frequency image dataset: obtaining a 1×64×64 size time-frequency image by using an improved multiple synchronous synchronous compression transformation (IMSST), and the images of the unlabeled time-frequency image dataset are derived from ten radar modulation signals (LFM, SFM, BPSK, LFM-BPSK, SFM-BPSK, EQFM, FSK, 4FSK, NS, Frank), and cover samples with signal-to-noise ratios (SNRs) of-16 dB to 4 dB (with an interval of 2 dB);

[0121] Step 32, generating a double-view input: performing two types of processing on each unlabeled time-frequency image in the dataset to construct a first view and a second view;

[0122] Step 33, double-branch feature extraction and processing: the first view input is processed by an improved ResNet encoder integrated with a cross-time-frequency pooling (CTFP) module in an online branch to extract high-dimensional features, which are then divided into two paths: one path inputs a reconstruction decoder to reconstruct a complete time-frequency image, and the other path inputs a projection head to map to a contrastive space after global average pooling; the second view input is processed by a target encoder (with the same structure as the improved ResNet encoder, and the parameters are updated from the online branch encoder by an exponential moving average (EMA) mechanism) in a target branch to extract features, which are then mapped to the contrastive space by global average pooling and the projection head;

[0123] Step 34, calculating a joint loss function;

[0124] Step 35, model parameter updating and pre-training weight acquisition: update and improve the parameters of the ResNet encoder, reconstruction decoder and projection head by back propagation optimization of the total loss; the target encoder and projection head parameters are updated by the EMA mechanism; the pre-training is completed by jointly optimizing the reconstruction loss and contrast loss, and finally the parameters of the improved ResNet encoder are saved as pre-training weights.

[0125] As a possible implementation manner of the embodiment, the first view is an online branch input: first, apply conventional data enhancement (including random cropping, scale transformation, noise disturbance), and then perform random mask operation to divide the image into 8x8 patches (a total of 64 patches), and randomly mask 75% (48 patches) of the area;

[0126] The second view is a target branch input: only the same conventional data enhancement as the first view is applied, and no mask is performed.

[0127] As a possible implementation manner of the embodiment, the step 4 includes:

[0128] Step 41, constructing a downstream recognition network structure: the downstream recognition network takes the improved ResNet encoder with an integrated cross-time-frequency pooling (CTFP) module as a feature extraction backbone, and connects a lightweight classification head at the output end of the encoder; 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 matching the number of modulation types (10 categories) for outputting the probability distribution of each category;

[0129] Step 42, migrating pre-training weights: the improved ResNet encoder parameters obtained by pre-training the contrast mask image reconstruction self-supervised model are migrated to the corresponding encoder of the downstream recognition network as initial weights;

[0130] Step 43, freezing part of the encoder parameters: freeze all parameters in 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 parameters of the classification head trainable;

[0131] Step 44, preparing a fine-tuning data set: a small amount of labeled time-frequency image data is used, and the data of the fine-tuning data set is derived from IMSST time-frequency graphs of ten radar modulation signals (LFM, SFM, BPSK, LFM-BPSK, SFM-BPSK, EQFM, FSK, 4FSK, NS, and Frank), covering signal-to-noise ratios of -16dB to 4dB (with an interval of 2dB), and the number of labeled samples is 50-200 for each signal per signal-to-noise ratio.

[0132] Step 45, fine-tuning training process: taking the fine-tuning dataset as input, outputting the category prediction result through the downstream recognition network, calculating the error of the prediction result and the label by using the cross-entropy loss function; using the Adam optimizer to optimize the cross-entropy loss, updating the parameters of the last two stages of the trainable encoder and the classification head, setting the initial learning rate to 0.0001, the batch size to 64, the training rounds to 200 rounds, and the learning rate to decay from the initial value to 1x10 6 ;

[0133] Step 46, obtaining a radar signal modulation recognition model: after the fine-tuning training is completed, saving all parameters of the downstream recognition network to form a final radar signal modulation recognition model.

[0134] As shown in Figure 2 , the radar signal modulation recognition device provided by the embodiment of the application comprises:

[0135] A dataset construction module is configured to acquire radar signals and construct a radar modulation signal dataset containing labeled data and unlabeled data, perform improved multiple simultaneous synchronization compression transformation (IMSST) time-frequency analysis processing on the radar modulation signals, and convert the radar modulation signals into time-frequency images.

[0136] A model construction module is configured to construct a self-supervised contrastive mask image reconstruction model, wherein the self-supervised contrastive mask image reconstruction model comprises an online branch and a target branch, the online branch comprises an improved ResNet encoder integrated with a cross-time-frequency pooling (CTFP) module, a reconstruction decoder, and a projection head, and the target branch comprises a target encoder and a projection head whose parameters are updated through an exponential moving average (EMA) mechanism.

[0137] A pre-training module is configured to pre-train the self-supervised contrastive mask image reconstruction model by using the unlabeled data, apply data enhancement and random mask operation to the time-frequency images to generate double views, input the double views into the online branch and the target branch respectively, update the model parameters by jointly optimizing the reconstruction loss and the contrast loss, and obtain pre-training weights.

[0138] A fine-tuning module is configured to migrate the pre-training weights to a downstream recognition network, freeze part of the encoder parameters, fine-tune the downstream recognition network by using a small amount of labeled data, and obtain a radar signal modulation recognition model.

[0139] The specific implementation of the application can be divided into the following steps.

[0140] Step 1, constructing a radar modulation signal dataset of ten kinds of modulation signals and performing improved multiple simultaneous synchronization compression transformation IMSST time-frequency analysis processing to convert the radar modulation signal dataset into time-frequency images.

[0141] tThe radar signal received at time t can be generally represented as a complex signal of the form:

[0142] ,

[0143] where, denotes the amplitude envelope of the signal, denotes the instantaneous frequency, denotes the instantaneous phase, is additive noise, usually modeled as a zero-mean Gaussian white noise.

[0144] 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 denote the sampling frequency, carrier frequency, sample number, bandwidth, symbol width, Barker code length, code number per period, frequency step, and code period, respectively. All signals are simulated according to different signal-to-noise ratios (SNRs), with an SNR range of -16 dB to 4 dB and an interval of 2 dB to ensure data diversity and fairness of the simulation experiment. The specific parameters of the signals are shown in Table 1.

[0145] Table 1 Signal parameters

[0146]

[0147] In order to accurately extract the time-frequency features of the radar signal, an improved multi-scale time-frequency analysis method (IMSST) is used to perform time-frequency transformation on the original signal. The IMSST method effectively improves the resolution and clarity of the time-frequency representation through multi-scale transformation and improved synchronous squeezing transformation. The core idea is to use multi-scale analysis to decompose the signal and concentrate the time-frequency energy on the actual time-frequency track through synchronous squeezing. IMSST can be described by the following formula:

[0148] ,

[0149] where the variables and denotes a discrete frequency, denotes time n a discrete short-time Fourier transform at time is the Kronecker delta function. denotes two rounding operations on the instantaneous frequency of the multi-sync compression transform, denotes the number of iterations.

[0150] Step 2, build a contrast mask reconstruction self-supervised training model based on the CTFP enhancement module.

[0151] 2.1 Overall design of the contrast mask image reconstruction model:

[0152] As Figure 3 , the overall framework structure of the contrast mask image reconstruction model proposed in the application. The application faces the recognition task of radar signal modulation type, and constructs a self-supervised feature modeling framework combining contrast learning strategy and mask image reconstruction mechanism, aiming to effectively utilize the time-frequency structure features in unlabeled data. The method includes two core stages: a self-supervised pre-training stage and a supervised downstream recognition stage. In the pre-training stage, the encoder learns a robust and discriminative latent feature representation under the joint action of contrast constraints and reconstruction objectives; while in the downstream stage, the trained encoder weights are transferred to train a linear classifier on a small number of labeled samples to realize the discrimination of modulation categories.

[0153] Specifically, first, the radar original signal is converted into a time-frequency graph, and two types of enhancement operations are applied to it: one is conventional data enhancement (such as cropping, scale transformation, and noise disturbance, etc.), and the other is to introduce a random mask (masking ratio is 75%) on this basis to construct a scenario with missing structure information. Two views are sent into the online encoder and the target encoder for feature extraction. Among them, the online encoder is updated by backpropagation, while the parameters of the target encoder are updated by the online encoder through exponential moving average to ensure training stability. After feature extraction, the high-dimensional features generated by the masked view are input into the decoder to reconstruct the corresponding time-frequency graph, and the output results are compared with the unmasked image at the pixel level to form the reconstruction loss. At the same time, the feature representations of the two enhanced graphs are mapped to the contrast space through global pooling and projection transformation, and the loss function such as InfoNCE is used to strengthen the similarity between positive sample pairs and pull away negative samples, forming the contrast loss. By jointly optimizing the above two types of losses, the model is guided to learn features with discriminative and detailed expression ability. After pre-training, the encoder is fixed and a lightweight classification head is trained under the condition of small sample supervision, so as to realize the effective recognition of radar modulation signals. The method takes the collaborative design of structure modeling and discriminative learning as the core, which significantly improves the generalization and recognition ability of the model for complex modulation styles.

[0154] 2.2 Cross-time-frequency pooling structure:

[0155] The present application proposes a context modeling mechanism suitable for time-frequency map structure: cross-time-frequency pooling, as shown in Figure 4 . The structure respectively applies one-dimensional pooling operation in time dimension and frequency dimension, models the global structure 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 a row or column. Therefore, the output of horizontal strip merging can be written as:

[0156] ,

[0157] Similarly, the output after vertical strip pooling can be written as:

[0158] ,

[0159] 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 discrete distributed areas and encode areas with strip structure. At the same time, due to the use of a narrower kernel 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 pool different from traditional spatial pool which relies on square kernel. In order to obtain the output containing more useful global priors, the present application first combines and together as follows, obtaining :

[0160] ,

[0161] ,

[0162] where, denotes element-wise multiplication, is a sigmoid function, is a 1x1 convolution.

[0163] The design of CTFP fully considers the structural characteristics of time-frequency map: the time dimension models the sequence, and the frequency dimension models the spectral pattern. Compared with the standard pooling method, this module can respectively perceive the global structure along the time and frequency, significantly improving the model's expression and recognition ability for complex radar signals such as frequency hopping, linear frequency modulation, and dual-band modulation. Subsequently, the present application will systematically integrate CTFP in the encoder module as its core context enhancement component, thereby improving the overall representation ability and pre-training performance.

[0164] 2.3 Encoder design:

[0165] In the present application, 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 a residual connection mechanism, improving the network's expression ability and training stability. The standard ResNet structure is usually composed of multiple residual units (Bottleneck modules) stacked together, each unit contains several convolutional layers, and the direct transmission of inter-layer information flow is realized through the identity mapping channel. The method of the present application uses ResNet-18 as the basic encoder, and its network structure is composed of five consecutive stages (layer1 to layer5). The first stage performs preliminary feature extraction on the input time-frequency graph, and the subsequent stages complete deeper feature modeling through residual blocks, accompanied by a gradual increase in the number of channels and a decrease in the resolution of the feature map.

[0166] Unlike the standard ResNet, to better adapt to the structural characteristics of time-frequency graph data, the method of the present application inserts a custom-designed cross-time-frequency pooling module (CTFP) after the 3x3 convolution in the last residual unit of each stage, and after all residual blocks in the last stage. The module aims to improve the model's ability to model multi-scale context information, thereby enhancing its sensitivity and representation ability for complex modulation features.

[0167] Specifically, let the input time-frequency graph be First, the initial feature map is obtained by performing initial feature extraction on the input time-frequency graph through the convolutional layers and max-pooling layers of Layer1.

[0168] ,

[0169] After the processing of each stage is completed, a CTFP module is inserted to enhance the context of its output, i.e.

[0170] ,

[0171] The advantage of this structure design is that the global modeling ability provided by the CTFP module and the local convolution mechanism of ResNet form a complementary relationship: the convolution module is good at extracting local edge, texture, and short-range dependency features, while the CTFP module can capture long-distance global structure information across time and frequency, effectively making up for the modeling shortcoming of conventional CNN architecture under asymmetric structure input. In addition, since the CTFP module is a lightweight module, its introduction will not significantly increase the model complexity and training time, and has good practicality and scalability.

[0172] 2.4 Decoder design:

[0173] In the self-supervised reconstruction task, 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. Although using a more complex decoding structure can generally obtain lower reconstruction error, related studies have pointed out that the decoder's ability is too strong, which may share too much modeling task in the pre-training process, so that the encoder tends to rely on the decoder to restore the features, thereby not conducive to the encoder itself to learn the representation with universality. Therefore, in recent years, self-supervised learning methods have generally tended to use simple but functional light decoder structures to avoid this problem and encourage the encoder to focus on learning key features with discriminability. The decoder designed by the invention is composed of five deconvolution modules. Each layer is composed of transpose convolution, batch normalization and ReLU activation function, which gradually restores the spatial size to restore the image structure through layer-by-layer operation. It should be pointed out that the design of this decoder architecture does not aim to pursue the ultimate image reconstruction quality, but consciously maintains the simplicity of the network, so that the decoding process is only used as an auxiliary tool for pre-training, thereby ensuring that the encoder can dominate feature learning and extract effective information with recognition ability.

[0174] 2.5 Loss design:

[0175] In order to fully exert the complementary advantages of different pre-training tasks in self-supervised learning, the invention jointly introduces reconstruction loss and contrast loss in the overall framework, corresponding to the mask image reconstruction branch and the contrast learning branch respectively. This design aims to guide the model to balance between local detail modeling and global discriminative feature extraction through multi-task collaborative training, thereby improving the universality of representation and the adaptability of downstream tasks.

[0176] In the reconstruction branch, the invention uses a mechanism similar to Masked Autoencoder, which randomly masks part of the input time-frequency map area, and only extracts features and restores images through the remaining visible area. In order to ensure that the reconstruction task focuses on the information missing area, only the reconstruction error of the masked area is measured, and the reconstruction loss is defined as:

[0177]

[0178] wherein represents the position of the masked combination, and represent the pixel values of the original image and the model reconstruction image at the i-th position. This loss encourages the model to learn how to recover the missing information from the incomplete input, thereby strengthening the encoder's ability to model fine-grained local features and structures.

[0179] ​However, relying only on reconstruction targets may cause the model to learn mainly low-level texture features, lacking the ability to capture abstract semantics. Therefore, a contrastive loss is further introduced as a supplement to enhance the discriminative ability of the model and the structural nature of the feature space. Specifically, different augmentations of the same image are regarded as positive sample pairs, and augmentations from different images are regarded as negative sample pairs. The contrastive learning loss is defined as:

[0180] ,

[0181] where , is the representation vector corresponding to the same sample after different data augmentation, is the cosine similarity, is the temperature coefficient, is the total number of contrastive samples. This loss guides the model to learn to aggregate samples of the same class and pull away samples of different classes, thereby improving the discriminative ability and generalization ability of the features in the downstream classification task.

[0182] To dynamically balance the optimization objectives of contrastive learning and reconstruction tasks during training, the present application introduces two learnable weight parameters and in the final loss function, corresponding to the contrastive loss and the reconstruction loss, respectively. The present application adopts the uncertainty weighting method. The total loss is defined as follows:

[0183] ,

[0184] where and are the weight coefficients of the reconstruction loss and the contrastive loss, respectively. The two weights are automatically updated as trainable parameters of the model during the optimization process, without the need for manual setting of hyperparameters, thereby achieving dynamic adjustment of the importance of the tasks.

[0185] The entire encoder structure is shown in Figure 5 , integrating the CTFP module in the Bottleneck structure of each stage of ResNet to improve the time-frequency modeling capability. At the same time, combining channel expansion and cross-dimensional feature fusion mechanism, the representation ability and time-frequency discriminative ability of the network are significantly enhanced.

[0186] Step 3: Train the contrastive mask image reconstruction model constructed in step 1 using the dataset in step 1, and finally realize the modulation recognition of the radar signal.

[0187] First, select the unlabeled samples from the radar signal time-frequency image constructed in step 1, and apply random enhancement and mask operation to them respectively to generate two input images; among them, the mask image is used as the input of the online branch, and the enhanced image without mask is used as the input of the target branch. The input image of the online branch is sent to two modules after feature extraction by the encoder: one is sent to the reconstruction decoder to reconstruct the original complete image, and the reconstruction loss is calculated based on the difference between the original image and the reconstructed image; the other is sent to the contrast learning space through global average pooling and projection head. The target branch uses the target encoder maintained by the momentum update mechanism to extract the corresponding features, which are also sent to the contrast space through the projection head to form a positive sample pair with the online branch, and the consistency of the features is strengthened through the contrast loss. The above two tasks are jointly optimized to cooperatively drive the model to learn time-frequency embedding features with discriminative and structural perception ability. After pre-training, the model parameters are migrated and some encoder modules are frozen, and a small amount of labeled data is used for fine-tuning to finally realize effective recognition of radar modulation types.

[0188] The effects of the present application are further illustrated by the following comparative experiments:

[0189] 1. Experimental environment configuration:

[0190] All experiments involved in the present application are completed on a Linux computing platform configured as follows: Intel Xeon Gold 6122 processor, 64 GB of memory, NVIDIA RTX 3090 graphics card (24 GB of video memory). The software environment includes Python 3.8 version, the deep learning framework is PyTorch 1.12.1, and the GPU driver and computing support are provided by CUDA11.6 and CuDNN 8.4. In the self-supervised pre-training stage, the total number of training rounds is set to 400 rounds, and the batch size (BatchSize) is 4096. A lower warm-up learning rate (1e-6) is used at the beginning of training, which gradually rises to the base learning rate 0.00001 to alleviate the instability risk caused by the initial parameter update; In the later period, it gradually decays to the minimum learning rate (1e-6), and combines the cosine annealing (cosine annealing) mechanism to dynamically adjust the learning rate to improve the overall convergence performance. In the fine-tuning stage, the training rounds are set to 200 rounds, and the Batch Size is 64. The base learning rate is raised to 0.0001, and the warm-up and decay mechanisms are still used, and the learning rate range is kept between 1e-6 and 0.0001. During the entire training process, the momentum parameter is set to 0.99, the temperature parameter τ is 0.2, and the optimizer selects AdamW with weight decay capability. In terms of input processing, a random masking strategy is used as the mask mechanism of the model. The specific operation is as follows: divide the time-frequency graph into 8x8 patches (a total of 64 blocks), randomly select 75% (i.e. 48 blocks) of them for masking, and only keep the remaining area for performing the reconstruction task to enhance the robustness of the model to incomplete input. In addition, in order to improve the sample diversity and representation ability of self-supervised learning, the present application method also introduces a variety of enhancement techniques, including: random cropping, horizontal flipping, color jitter (Color Jitter), and mixed sample enhancement methods MIXUP and CUTMIX, so that the model can better learn the semantic association between samples.

[0191] 2 Experimental result analysis and evaluation:

[0192] 2.1 Ablation experiment:

[0193] (1) Structure level: In order to comprehensively evaluate the role of each module in the modulation recognition performance in the framework of the present application, the present application carries out systematic ablation experiment analysis from the perspective of structural composition. As shown in Table 2, first, a benchmark model (Supervised) is constructed which does not contain any self-supervised mechanism or structural enhancement module, and the model is completely based on the traditional supervised training method, and the test accuracy is 83.18%, which serves as the basis for subsequent comparison. On this basis, the cross-time-frequency pooling module (CTFP) designed in the present application is integrated into each stage of ResNet-50, and the accuracy is improved to 85.07%. The module captures long-range context information on the time axis and the frequency axis respectively, and models the fusion of the two kinds of information, making up for the problem of neglecting the axial structure in the process of processing the two-dimensional time-frequency graph by the conventional two-dimensional pooling, effectively improving the sensitivity and capture ability of the network to the modulation mode. Further, under the premise of introducing the structural enhancement module, the influence of different pre-training targets on the model performance is explored. Based on the structure containing the CTFP module, the contrastive learning loss and the mask reconstruction loss are introduced for self-supervised training. The results show that: when the contrastive loss is used alone, the model recognition accuracy reaches 85.62%, indicating that by constructing the contrast relationship between the time-frequency views, the model can learn more discriminative feature representations; when the reconstruction loss is used for pre-training alone, the accuracy also reaches 85.47%, which is also better than the model without pre-training, indicating that the local recovery task based on the mask can guide the model to learn the structural details of the time-frequency graph and strengthen the local feature modeling. The above results show that even if only a single self-supervised target is used, significant performance improvement can be achieved. Finally, the complete CMIR training framework of the present application is constructed, and the contrastive learning and reconstruction mechanism are integrated into a unified self-supervised strategy, and the model trained by the training framework has an accuracy of 86.35%. Compared with the schemes using contrastive loss and reconstruction loss respectively, the joint scheme brings performance gains of 0.73% and 0.88% respectively, verifying the complementarity of the two in the feature learning process: the contrastive target helps to learn the global discriminative structure, while the reconstruction target improves the modeling ability of the model for fine-grained structure. The two mechanisms work together to significantly enhance the representation generalization ability of the encoder under the condition of no labeled data.

[0194] Table 2 Ablation experiment at the structure level

[0195]

[0196] (2) Masking level: On the basis of the foregoing, to further verify the role and stability of the designed masking mechanism in self-supervised training, the influence of different masking ratios on model performance was analyzed in depth. 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 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%, which shows that even without masking, the image reconstruction task itself can regularize the encoder to some extent and promote it to learn potential features. However, as the masking ratio increases, the model performance shows slight fluctuations, and the highest accuracy of 86.35% is achieved when the masking ratio is 0.75. This experimental phenomenon shows that in this task scenario, a higher masking ratio enhances the feature learning ability of the model. This result can be attributed to the fact that a high masking ratio significantly increases the complexity of the reconstruction process, forcing the model to rely more on the global structure information of the input data to complete the reconstruction, thereby learning more abstract and generalizable latent representations.

[0197] Table 3 Masking level ablation experiment

[0198]

[0199] 2.2 Signal-to-noise ratio and class accuracy analysis:

[0200] To comprehensively examine the recognition robustness and generalization performance of the method in complex noise environments, the classification performance of the model under different sample conditions was analyzed in depth from the dimensions of signal-to-noise ratio and modulation signal type. Figure 6The classification accuracy of the model for 10 types of modulation signals under 11 SNR settings (ranging from -16 dB to 4 dB) is shown. Overall, the model performs well in the medium and high SNR region (i.e., SNR ≥ -6 dB), with a classification accuracy of more than 98% for most modulation types, reflecting the method's strong anti-interference ability and adaptability to complex channel conditions in real environments. In low SNR environments (such as -16 dB and -14 dB), although noise significantly affects the clarity of the modulation structure, the model still maintains some distinguishing ability, indicating that the adopted self-supervised pre-training strategy effectively enhances the model's robustness in noisy backgrounds. From the performance differences of various modulation types, some modulation methods maintain high recognition rates throughout the entire SNR range. For example, eqfm, lfm, ns, and frank exhibit stability under all SNR conditions, with eqfm achieving nearly 100% accuracy in almost all test scenarios, indicating that its time-frequency features are more prominent and easier for the model to accurately model and identify. In contrast, bpsk, 4fsk, fsk, and lfm-bpsk modulation methods have significantly lower accuracy in low SNR conditions (especially -16 dB to -12 dB), with 4fsk and bpsk only reaching about 30.00% and 18.75% at -14 dB. This result shows that these types have more blurred time-frequency graphs under severe noise interference, posing greater challenges to feature extraction. However, with the improvement of SNR, the recognition performance of related types quickly recovers, and at -6 dB and above, it basically restores to nearly 100% accuracy. In addition, for combined modulation types (such as lfm-bpsk and sfm-bpsk), although their recognition effect is slightly lower than that of single modulation signals under extremely low SNR conditions, when the SNR rises to around -6 dB, the accuracy quickly improves to about 90%, showing that the model has strong modeling and discrimination ability in analyzing mixed modulation structures.

[0201] 2.3 Effect of different encoders on model performance:

[0202] To further evaluate the universality and scalability of the proposed self-supervised framework, different mainstream encoder structures were selected for comparative experiments under the same training configuration and loss mechanism to investigate their modulation recognition performance under different SNR conditions, such as Figure 7The selected encoders include the classic convolutional structure ResNet18, the standard Vision Transformer (ViT), the Swin Transformer based on hierarchical modeling, and the ResNet18 with the integrated CTFP module (ResNet18+CTFP) proposed in the present application. The above four architectures represent the mainstream structure system of current image modeling, i.e., traditional CNN, global attention Transformer, local sliding window Transformer, and structure-enhanced convolutional network combined with frequency-time perception mechanism. The experimental results are shown in the table. Under the complete self-supervised training mechanism, the four different types of encoders all exhibit excellent recognition ability under high signal-to-noise ratio conditions (SNR≥0 dB). Among them, the classification accuracy of ResNet18, ViT and SwinT is stable at about 99%, indicating that the present method has good downstream adaptability and can realize effective modulation signal classification combined with various backbone networks. However, in the environment with low signal-to-noise ratio (SNR≤−10 dB), the performance gap between different network architectures gradually appears, reflecting the structural differences in noise resistance and feature robustness. Specifically, under the two extreme scenarios of −14 dB and −16 dB, the accuracy of ViT is 46.75% and 32.75% respectively, slightly better than SwinT (49.60%, 35.20%), but both are inferior to the traditional convolutional architecture ResNet18 (54.50%, 39.25%). This indicates that although the Transformer architecture has strong global modeling capability, its sensitivity to local details will limit its extraction effect of key information under strong noise interference. It is worth emphasizing that when the cross-time-frequency pooling module (CTFP) is integrated into ResNet18, the performance of the model is further improved in the full SNR range. Especially in the three extreme channel conditions of −16 dB, −14 dB and −12 dB, the accuracy is 43.00%, 57.50% and 76.00% respectively, which is about 3% to 4% higher than the basic 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 the structural information in the time and frequency dimensions, thereby improving the overall discrimination effect.

[0203] In summary, although the Transformer structure has superior performance in natural image tasks, its performance in radar time-frequency graph modeling is slightly unstable. This phenomenon shows that its original architecture has not fully adapted to the high non-stationarity and low redundancy structure specific to radar data, and there is still a certain generalization bottleneck. The ResNet structure combined with the CTFP module exhibits stronger adaptability in high noise and complex modulation background, further verifying the practicality and effectiveness of the proposed structure design.

[0204] 2.4 Comparison with other methods:

[0205] To verify the performance advantage of the proposed model in the radar signal modulation recognition task, several representative neural network architectures are selected for comparative evaluation. These comparative methods include SVM, ResNet, MobileNetV2, ConvNeXt, MobileViT, VAN, LPINe and MAPNet. These methods are widely used in computer vision and radar signal processing fields, each with its own characteristics, and can represent different model design ideas and characteristics.

[0206] Under the full supervision condition (100% labeled data), the modulation recognition ability of various representative methods under different signal-to-noise ratio levels is systematically evaluated. As shown in Table 4, the results show that traditional methods such as support vector machines are almost unable to effectively complete the classification task in low SNR environments, with an accuracy of only 25.12% and less than 35% at −16 dB and −10 dB. The main reason is that it highly depends on artificial static features and is difficult to model dynamic time-frequency patterns under noise. In contrast, deep learning models significantly outperform SVM in overall performance. For example, ResNet and ConvNeXt, which are based on residual structures, maintain stable performance in medium to high signal-to-noise ratio conditions, but still show significant performance degradation at low SNR, with an accuracy of 54.5% and 48% at −14 dB, respectively. MobileNetV2 performs well in high SNR intervals, but its accuracy drops rapidly in low signal-to-noise ratio scenarios, indicating that lightweight structures lack robustness under noise interference. Models with attention mechanisms such as VAN and MobileViT have certain recognition ability in the SNR range of −6 dB to +4 dB, but their performance fluctuates significantly in extreme noise scenarios. For example, MobileViT achieves an accuracy of only 42.13% at −16 dB, indicating that its cross-channel modeling capability is insufficient to effectively resist the performance loss caused by feature degradation. Specialized structures for the field, such as MAPNet and LPINet, achieve good results in signal-to-noise ratio conditions of −4 dB and above, but also face performance decline problems in low SNR, indicating that their embedded prior mechanisms are difficult to maintain stability in extreme environments. LSYNet, as an existing self-supervised method, has relatively good robustness at lower SNR, but its overall performance is still inferior to the proposed CMIR. In summary, CMIR achieves the best or near-optimal classification performance in the full signal-to-noise ratio range, especially in scenarios of −10 dB and above, with an accuracy of over 90%, and 100% at 0 dB. In extremely low SNR (−16 dB, −14 dB), the accuracy reaches 43.00% and 57.50%, respectively, significantly outperforming other comparative models.

[0207] Table 4 Recognition accuracy of different methods under different labeled training data

[0208]

[0209] Under the weakly supervised condition of using only 10% labeled samples, the performance of each model generally declines, but the robustness difference between different structures is more obvious, which helps to reflect their generalization ability under limited supervision. Full supervision networks such as ResNet, ConvNeXt and MobileNetV2 have a significant decline in recognition accuracy when the training data is scarce, for example, the accuracy drops to 71.13%, 33.63% and 66.88% at -10dB, indicating that such architectures have a high degree of dependence on labeled data. The performance of attention-driven models such as MobileViT is more unstable and prone to overfitting after the training set is reduced, with an accuracy of 47.88% at -12dB. Although MAPNet and LPINet incorporate domain prior, their structural advantages are difficult to fully exploit under limited training data, with accuracies of 26.13% and 31.38% at -16dB, respectively. LSYNet has a contrastive learning mechanism, but it still has shortcomings under the dual interference of extreme noise and insufficient data, as it does not combine with the reconstruction target. Its performance is relatively stable at -10dB and -8dB (accuracies of 75.23% and 84.14%, respectively), but it still has shortcomings under the dual interference of extreme noise and insufficient data. In contrast, CMIR effectively alleviates the model degradation problem caused by weak supervision with its self-supervised training strategy that combines joint contrast and reconstruction. The accuracies are 76.38%, 86.38% and 92.75% at -10dB, -8dB and -6dB, respectively; even at -16dB, the accuracy remains at 34.75%. In the medium to high signal-to-noise ratio range (0dB and above), the accuracy is always stable at more than 97%, fully surpassing all comparison methods.

[0210] 2.5 Complexity Analysis:

[0211] In addition to ensuring the model has excellent recognition performance, its computational overhead and parameter size are also key factors that must be considered in the actual deployment process. To evaluate the feasibility of the proposed method in terms of resource usage, the computational complexity (measured in FLOPs) and parameter quantity of various comparative models under the standard input dimension (1x64x64) were quantitatively compared, as shown in Table 5. The evaluation objects cover typical convolutional architectures, lightweight networks, Transformer series models, modulation recognition dedicated structures, and self-supervised learning methods, among other categories. The relevant statistical results are shown in the table. The FLOPs of the proposed CMIR method is about 274.85M, and the total number of parameters is 18.31M, which is in the middle to high position among the overall comparative models. In contrast, MobileNetV2 and MobileViT have significantly lower computational costs, with FLOPs of 26.06M and 21.43M, respectively, and parameter quantities controlled within 3M, reflecting their deployment advantages on edge devices and mobile platforms. However, at the same time, these lightweight models are significantly inferior to performance-oriented architectures in terms of accuracy. Among high-complexity models, ConvNeXt-tiny has the highest computational demand, reaching 362.88M FLOPs, while LSYNet and LPINet have FLOPs of 183.54M and 232.59M, respectively, which are also at a relatively high level. In terms of parameter size, ConvNeXt-tiny and LSYNet have parameter quantities of 27.8M and 14.75M, respectively, which are significantly higher than the structure proposed in this invention. It is worth noting that although LPINet and MAPNet are extremely compact in terms of parameter size (less than 0.3M), they have obvious performance disadvantages in previous recognition experiments, indicating that their lightweight design has an accuracy bottleneck. Comprehensive evaluation shows that the proposed method ensures leading recognition accuracy without using highly complex or bloated network structures, achieving a good balance between performance and resource consumption. Compared to ultra-lightweight models, the resource occupation is slightly increased, but the structure still maintains a modular, clear, and easy-to-understand design style, which is beneficial to engineering landing and maintenance.

[0212] Table 5 Comparison of FLOPs and Params of different models

[0213]

[0214] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the specific embodiments of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application. Any modification or equivalent replacement that does not depart from the spirit and scope of the present application should be covered within the protection scope of the claims of the present application.

Claims

1. A radar signal modulation identification method of self-supervised contrastive mask reconstruction, characterized in that, The method comprises the following steps: Step 1, obtaining radar signals and constructing a radar modulation signal dataset containing labeled data and unlabeled data, and performing improved multiple synchronous compression transform time-frequency analysis processing on the radar modulation signals to convert them into time-frequency images; Step 2, building a self-supervised contrastive mask image reconstruction model, the self-supervised contrastive mask image reconstruction model comprising an online branch and a target branch, the online branch comprising an improved ResNet encoder integrated with a cross-time-frequency pooling module, a reconstruction decoder and a projection head, and the target branch comprising a target encoder and a projection head whose parameters are updated through an exponential moving average mechanism; Step 3, pre-training the self-supervised contrastive mask image reconstruction model using the unlabeled data, applying data enhancement and random mask operation to the time-frequency images to generate double views, inputting the double views into the online branch and the target branch respectively, updating the model parameters through joint optimization of reconstruction loss and contrast loss, and obtaining pre-training weights; Step 4, migrating the pre-training weights to a downstream recognition network, fine-tuning the network using a small amount of labeled data after freezing part of the encoder parameters, and obtaining a radar signal modulation recognition model.

2. The method of claim 1, wherein, The step 1 comprises: Step 11, generating multiple radar modulation signals according to radar signals; Step 12, performing time-frequency transform on the radar modulation signals by using an IMSST method, the IMSST decomposing signals through multi-scale analysis and adjusting time-frequency energy distribution through synchronous extrusion to make energy gather on actual time-frequency tracks; Step 13, obtaining time-frequency images with a size of 1×64×64 after time-frequency transform processing; Step 14, integrating all obtained time-frequency images to form a radar modulation signal time-frequency image dataset containing labeled data and unlabeled data.

3. The method of claim 1, wherein, The step 2 comprises: Step 21, designing the overall framework of the self-supervised contrastive mask image reconstruction model, and specifying the composition and connection relationship of the online branch and the target branch; Step 22, constructing an improved ResNet encoder integrated with a cross-time-frequency pooling module, the improved ResNet encoder comprising five consecutive stages; Step 23, designing a reconstruction decoder of the online branch, the reconstruction decoder being composed of five layers of deconvolution modules, each layer comprising a transpose convolution, a batch normalization and a ReLU activation function, and recovering feature space size layer by layer to finally output a reconstructed image with a size consistent with that of the input time-frequency image; Step 24, designing a projection head of the online branch and the target branch, the projection head being a two-layer fully connected network, the first layer mapping 512-dimensional high-dimensional features output by the encoder to 256-dimensional features, and the second layer mapping to 128-dimensional features for projecting the features to a contrast space; Step 25, setting a parameter updating mechanism of the target encoder: the parameters of the target encoder are updated synchronously from the improved ResNet encoder of the online branch through an EMA mechanism; Step 26, constructing a joint loss function: ， wherein with are weight coefficients of reconstruction loss and contrast loss, respectively, with are reconstruction loss and contrast loss, respectively.

4. The method of claim 3, wherein, In step 21, the online branch includes an improved ResNet encoder integrated with a cross-time-frequency pooling module, a reconstruction decoder, and a projection head; the target branch includes a target encoder and a projection head, and the target encoder has the same structure as the improved ResNet encoder, and the parameters of the target encoder are updated by the exponential moving average mechanism to synchronize the parameters of the improved ResNet encoder of the online branch; The 1×64×64 time-frequency graph obtained by transforming the radar signal through IMSST is processed through two types of enhancement operations to generate two views, wherein the view with a random mask is input into the online branch, and the unmasked view after data enhancement is input into the target branch.

5. The method of claim 3, wherein, The specific process of step 22 is as follows: Based on the ResNet-18 architecture, five consecutive stages layer1-layer5 are set, the layer1 stage performs initial feature extraction on the input time-frequency graph through convolutional layers and maximum pooling layers, and the layer2 to layer5 stages perform deep feature modeling through residual blocks, accompanied by an increase in the number of channels in the manner of 64→128→256→512 and a decrease in the resolution of the feature map; After the 3×3 convolution of the last residual unit in each stage of layer2 to layer5, and after all residual blocks in the layer5 stage, a CTFP module is inserted; the CTFP module performs one-dimensional strip pooling in the time and frequency dimensions respectively, and enhances the time-frequency correlation of the features by fusing the pooling results, which specifically includes: performing horizontal strip pooling on the input feature map, output is: where x is the input feature map, and W is the feature map width. performing vertical strip pooling on the input feature map, output is: where H is the feature map height; pooling results of horizontal strips pooling results of vertical strips fuse the results by 1x1 convolution, activate by sigmoid, and then multiply element-wise with the original feature map to obtain enhanced features: ， ， where, denotes element-wise multiplication, is a sigmoid function, is a 1 x 1 convolution, is the output of cross-time-frequency pooling, C is the channel dimension of the feature map.

6. The method of claim 3, wherein, The construction of the joint loss function includes: The reconstruction loss is defined only for the pixel error in the mask area: ， wherein denotes the position of the occlusion, and denotes the pixel value of the original image and the model reconstructed image at the i-th position, respectively; The InfoNCE loss is used as the contrast loss: ， wherein , is the same sample after different data augmentation, and is the cosine similarity, is the temperature coefficient, is the total number of contrast samples; Two learnable weight parameters are introduced for contrastive loss and reconstruction loss respectively and to build the final joint loss function.

7. The method of claim 1, wherein, The step 3 includes: Step 31, preparing a no-label time-frequency image dataset: 1×64×64 size time-frequency images are obtained by improved multiple synchronous compression transformation, and the images in the no-label time-frequency image dataset are derived from ten radar modulation signals and cover samples with signal-to-noise ratios from-16dB to 4dB; Step 32, generating double-view input: two types of processing are performed on each no-label time-frequency image in the dataset to construct a first view and a second view; Step 33, double-branch feature extraction and processing: the first view is input into the improved ResNet encoder integrated with the cross-time-frequency pooling module for online branch processing, and after extracting high-dimensional features, the features are 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 to map to the contrast space; the second view is input into the target encoder for target branch processing, and after extracting the features, the features are input into the projection head after global average pooling to map to the contrast space; Step 34, joint loss function calculation; Step 35, model parameter updating and pre-training weight acquisition: the total loss is optimized through back propagation to update the parameters of the improved ResNet encoder, the reconstruction decoder, and the projection head; the parameters of the target encoder and the projection head are updated through the EMA mechanism; the reconstruction loss and the contrast loss are jointly optimized to complete pre-training, and finally the parameters of the improved ResNet encoder are saved as pre-training weights.

8. The method of claim 7, wherein, The first view is an online branch input: first apply conventional data enhancement, then perform random masking operation, divide the image into 8x8 patches, and randomly mask 75% of the area; The second view is a target branch input: only apply the same conventional data enhancement as the first view, and do not perform masking.

9. The self-supervised contrastive mask-reconstruction radar signal modulation identification method according to any one of claims 1-8, characterized in that, The step 4 comprises: Step 41, constructing a downstream recognition network structure: the downstream recognition network takes an improved ResNet encoder integrated with a cross-time-frequency pooling module as a feature extraction backbone, and connects a lightweight classification head at the output end of the encoder; 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 matching the number of modulation types, for outputting probability distribution of each category; Step 42, migrating pre-training weights: migrating the improved ResNet encoder parameters obtained by pre-training the contrastive mask image reconstruction self-supervised model to the corresponding encoder of the downstream recognition network as initial weights; Step 43, freezing part of the encoder parameters: freezing all parameters of the first three stages of the improved ResNet encoder, and only keeping the parameters of the last two stages and the classification head parameters trainable; Step 44, preparing a fine-tuning dataset: using 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, covering signal-to-noise ratios from -16dB to 4dB, and the number of labeled samples is 50-200 for each signal per signal-to-noise ratio; Step 45, fine-tuning training process: taking the fine-tuning data set as input, outputting the category prediction result through the downstream recognition network, calculating the error between the prediction result and the label by using the cross-entropy loss function; using the Adam optimizer to optimize the cross-entropy loss, updating the parameters of the last two stages of the trainable encoder and the parameters of the classification head, setting the initial learning rate to 0.0001, the batch size to 64, the training round to 200 rounds, and the learning rate to decay from the initial value to 1x10 6 ; Step 46, obtaining a radar signal modulation recognition model: after fine-tuning training is completed, saving all parameters of the downstream recognition network to form a final radar signal modulation recognition model.

10. An apparatus for radar signal modulation identification of self-supervised contrastive mask reconstruction, characterized in that, Comprise: A dataset construction module for obtaining radar signals and constructing a radar modulation signal dataset containing labeled data and unlabeled data, performing improved multiple simultaneous compression transform time-frequency analysis processing on the radar modulation signals, and converting them into time-frequency images; A model building module for building a self-supervised contrastive mask image reconstruction model, the self-supervised contrastive mask image reconstruction model comprising an online branch and a target branch, the online branch comprising an improved ResNet encoder integrated with a cross-time-frequency pooling module, a reconstruction decoder, and a projection head, and the target branch comprising a target encoder and a projection head whose parameters are updated through an exponential moving average mechanism; A pre-training module for pre-training the self-supervised contrastive mask image reconstruction model using the unlabeled data, applying data enhancement and random masking operation on the time-frequency images to generate double views, inputting the double views into the online branch and the target branch respectively, updating the model parameters through joint optimization of reconstruction loss and contrast loss, and obtaining pre-training weights; A fine-tuning module for migrating the pre-training weights to a downstream recognition network, freezing part of the encoder parameters, and fine-tuning using a small amount of labeled data to obtain a radar signal modulation recognition model.

Images