Radar and communication signal integrated identification method and device based on YOLOv13

Abstract

The invention provides a radar and communication signal integrated identification method and device based on YOLOv13, and relates to the technical field of signal modulation identification. The method comprises the following steps: converting radar and original communication signals into a unified three-channel time-frequency characteristic pattern; constructing an improved YOLOv13 network, wherein the network integrates a bimodal feature fusion module, an improved BiFPN + + multi-scale fusion module and an anchor-frame-free detection module based on a task alignment structure; training and lightening the network by using a data set to obtain an identification model; and finally, outputting the category, confidence and position of the signal through the model. According to the method, the problems of difficulty in multi-system signal adaptation, difficulty in dense signal instance level positioning and difficulty in model real-time deployment are solved, and the method has the advantages of high recognition precision, high adaptability and good real-time performance.

Description

Technical Field

[0001] This application relates to the field of signal modulation recognition technology, and more specifically, to a method and apparatus for integrated recognition of radar and communication signals based on YOLOv13. Background Technology

[0002] With the continuous development of modern electronics, intelligent communications, and integrated sensing systems, the demand for collaborative radar and communication functions within the same platform has significantly increased. As spectrum resources become increasingly scarce and scenarios involving multiple services operating in parallel become more prevalent, the boundaries between radar and communication systems are blurring, exhibiting a high degree of convergence in waveform structure, modulation methods, and signal processing frameworks. Against this backdrop, Integrated Radar and Communication (IRC) technology has emerged. By fusing detection and communication functions on the same hardware platform or carrier, it effectively improves spectrum utilization efficiency, system integration, and multi-task collaborative capabilities, and has become an important direction for the development of integrated electronic systems. Signal integration identification, as a key component, aims to automatically identify and classify different types of radar and communication signals in a shared electromagnetic environment, providing fundamental support for spectrum management, interference detection, and collaborative sensing.

[0003] In existing research, deep learning models such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) have been used to automatically extract salient features from time-frequency graphs, enabling end-to-end intelligent recognition. These methods can be trained with large-scale data to learn the time-frequency structure of complex signals, playing a crucial role in improving the accuracy and stability of integrated recognition. Hybrid waveforms from multiple systems, such as OFDM, LFM, and NOMA, are used in integrated systems. By optimizing spectrum planning, joint modulation methods, and time-frequency resource allocation strategies, different signal types can behave collaboratively within shared frequency bands. Supporting spectrum sensing and signal recognition technologies further enhance the system's adaptability to complex environments, providing a theoretical basis and feasible practical path for integrated recognition.

[0004] While existing integrated signal recognition technologies have made significant progress in radar and communication signal classification using traditional signal processing methods, time-frequency analysis-based deep learning methods, multi-task learning frameworks, and spectrum sharing mechanisms, they still have obvious limitations. Traditional recognition methods based on sparse reconstruction and compressed sensing mainly rely on manually constructed features and specific signal models, making them difficult to adapt to dynamic electromagnetic environments with multiple systems and multiple sources of signals coexisting. Deep learning classification models based on time-frequency maps such as STFT and wavelet transform can usually only output a single category label, failing to perform spatial localization and instance-level recognition of multiple coexisting signals in the time-frequency map, thus limiting their application in dense signal environments. Furthermore, current deep learning methods lack a unified framework for cross-system signal representation, often training models separately for single systems such as OFDM, LFM, and MIMO, making it difficult to be compatible with multiple waveform structures. Moreover, existing classification models are large in scale and have limited inference speed, making it difficult to meet the rapid recognition requirements in scenarios emphasizing real-time performance, such as electronic warfare or high-speed vehicle communication. At the same time, since integrated signal datasets generally rely on simulation generation and lack sufficient coverage of real-world scene samples, the generalization ability of the model is limited when faced with strong interference, low signal-to-noise ratio, or complex changes in signal morphology.

[0005] Therefore, there is an urgent need for a method and apparatus for integrated identification of radar and communication signals based on YOLOv13 to solve one of the aforementioned technical problems. Summary of the Invention

[0006] The purpose of this application is to provide a method and apparatus for integrated identification of radar and communication signals based on YOLOv13, which can solve at least one of the aforementioned technical problems. The specific solution is as follows:

[0007] According to a specific embodiment of this application, this application provides a method for integrated identification of radar and communication signals based on YOLOv13, including:

[0008] Construct a dataset based on three-channel time-frequency feature maps;

[0009] Build a network based on the improved YOLOv13;

[0010] Specifically, a dual-modal feature fusion module is introduced into the backbone of the network, an improved BiFPN++ multi-scale feature fusion module is introduced into the neck of the network, and an anchorless detection module based on task alignment structure is introduced into the detection head of the network.

[0011] The network is trained using the training dataset to obtain an integrated signal recognition model;

[0012] The original signal to be identified is converted into a corresponding three-channel time-frequency feature map, which is then input into the integrated signal recognition model. The model outputs the signal category, confidence level, and position in the time-frequency map.

[0013] Furthermore, the method for constructing the three-channel time-frequency feature map includes:

[0014] Preprocessing of raw signals to be processed in radar and communication systems;

[0015] The preprocessed signal is subjected to time-frequency transformation to obtain a complex time-frequency matrix;

[0016] Amplitude, instantaneous frequency, and phase information are extracted from the complex time-frequency matrix to form three feature channels: amplitude spectrum channel, instantaneous frequency channel, and phase spectrum channel, respectively.

[0017] The three feature channels are normalized and their sizes are adjusted, and then spliced ​​together to form the three-channel time-frequency feature map.

[0018] Furthermore, the dual-modal feature fusion module includes:

[0019] Structural modality branches are used to extract global structural features from input features;

[0020] Texture modality branch, used to extract local texture features of input features;

[0021] The fusion unit is used to perform weighted fusion of the global structural features and the local texture features using learnable weights.

[0022] Furthermore, the backbone of the network also includes:

[0023] Replace some standard convolutions with lightweight modules that include depthwise separable convolutions;

[0024] Introducing dynamic convolution into deep networks;

[0025] Channel attention and spatial attention mechanisms are introduced at the output layer of the backbone network.

[0026] Furthermore, the improved BiFPN++ multi-scale feature fusion module includes:

[0027] The input terminal is used to receive feature maps of multiple scales from different levels of the backbone network;

[0028] A path used for bidirectional information transfer and fusion among the multiple scale feature maps;

[0029] At the fusion node, a weighted fusion method with learnable weights is adopted.

[0030] Furthermore, the anchorless frame detection module based on the task alignment structure includes:

[0031] The classification branch is used to predict the probability of the target category at each spatial location;

[0032] The regression branch is used to predict the target bounding box information for each spatial location;

[0033] The task alignment structure is used to dynamically adjust the optimization weights of classification and regression tasks during training based on the quality of the prediction task.

[0034] Furthermore, the method also includes a model lightweighting step:

[0035] Depthwise separable convolutions are used in the network structure;

[0036] After training, the network is pruned in a structured manner.

[0037] Furthermore, the construction of the dataset based on the three-channel time-frequency feature map includes:

[0038] Obtain labeled raw radar and communication signal samples;

[0039] Each of the original signal samples is constructed into a corresponding three-channel time-frequency feature map;

[0040] All constructed and labeled three-channel time-frequency feature maps constitute the training dataset used for network training.

[0041] Further, training the network using the training dataset includes:

[0042] The training dataset is input into the network for forward propagation to obtain the prediction results;

[0043] Calculate the loss between the predicted result and the true label;

[0044] Network parameters are updated via backpropagation.

[0045] This application also provides a device for integrated radar and communication signal identification based on YOLOv13, comprising:

[0046] The data preprocessing unit is used to convert the raw signal to be identified into a three-channel time-frequency feature map;

[0047] Network building unit, used to build the network based on the improved YOLOv13;

[0048] The model training unit is used to train the network using the training dataset to obtain the integrated signal recognition model;

[0049] The signal recognition unit is used to process the input three-channel time-frequency feature map using the integrated signal recognition model and output the recognition result.

[0050] Compared with the prior art, the above-described solutions of this application have at least the following beneficial effects:

[0051] 1. This application discloses a method and apparatus for integrated recognition of radar and communication signals based on YOLOv13. First, a standardized preprocessing and feature construction process is designed to uniformly convert heterogeneous radar and communication raw signals into three-channel time-frequency feature maps containing amplitude, instantaneous frequency, and phase information, providing consistent input across different systems for subsequent processing. Second, a dual-modal feature fusion module is introduced into the network backbone. Through parallel structural and texture branches, the global continuous structure and local modulation details of the signal are extracted respectively, and adaptive fusion is performed using learnable weights. This achieves efficient representation and feature extraction of signals from multiple systems such as LFM, OFDM, and QAM within a unified framework, significantly improving the model's cross-system adaptability and feature representation capabilities in complex dynamic electromagnetic environments.

[0052] 2. This application discloses a method and apparatus for integrated recognition of radar and communication signals based on YOLOv13. It reconstructs the integrated recognition task into a target detection problem in a time-frequency image and employs an anchor-free detection module based on a task-aligned structure. This module directly predicts the center point and boundary of signal instances, eliminating the limitation of preset anchor frames. At its core, through a Task-Aligned structure, it dynamically adjusts the weights of classification and localization losses during training based on prediction quality, forcing the two sub-tasks to optimize collaboratively. This enables the recognition method to simultaneously detect, locate, and classify multiple signal instances in the same time-frequency image, significantly improving the robustness and localization accuracy in complex scenarios such as signal aliasing, low signal-to-noise ratio, and weak signals, achieving a leap from "overall classification" to "instance-level perception."

[0053] 3. This application discloses a method and apparatus for integrated radar and communication signal recognition based on YOLOv13. It features comprehensive lightweight improvements at the network structure level, including the use of depthwise separable convolutions in the backbone and neck region, and post-training structured pruning to remove redundant channels. Simultaneously, at key locations, such as after the backbone output and after multi-scale fusion, channel and spatial attention modules, such as CBAM and lightweight channel attention, are embedded to focus the network on key time-frequency regions and feature channels. These lightweight improvements effectively reduce the number of model parameters and computational complexity, while the attention mechanism ensures that the model's ability to distinguish key signal features such as "frequency band stripes" and "frequency hopping points" is enhanced rather than weakened during the lightweighting process. Attached Figure Description

[0054] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. In the drawings:

[0055] Figure 1 This is a flowchart illustrating a method for integrated identification of radar and communication signals based on YOLOv13, as described in an embodiment of this application. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0057] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to limit the application. The singular forms “a,” “said,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms, and “multiple” generally includes at least two unless the context clearly indicates otherwise.

[0058] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0059] It should be understood that although the terms first, second, third, etc., may be used in the embodiments of this application, these descriptions should not be limited to these terms. These terms are only used to distinguish the descriptions. For example, first may also be referred to as second without departing from the scope of the embodiments of this application, and similarly, second may also be referred to as first.

[0060] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device that includes said element.

[0061] The optional embodiments of this application are described in detail below with reference to the accompanying drawings.

[0062] Example 1:

[0063] like Figure 1 As shown, a method for integrated identification of radar and communication signals based on YOLOv13 includes:

[0064] S1: Construct a dataset based on three-channel time-frequency feature maps.

[0065] This step aims to construct a dataset for model training, with its core samples being three-channel time-frequency feature maps. The specific construction process is as follows:

[0066] First, a set of original signal samples to be processed is collected or generated by simulation software, including various types of radar and communication signals such as LFM, NLFM, pulse train, QAM, PSK, OFDM, FHSS, etc., and each sample is labeled with its signal category and the precise bounding box position in the time-frequency diagram.

[0067] a. For each labeled raw signal sample to be processed, a corresponding three-channel time-frequency feature map is constructed through a unified preprocessing and feature extraction process. This construction process includes:

[0068] The original signal undergoes standardization preprocessing including DC bias removal, band-limited filtering, and energy normalization. The energy normalization is expressed as follows:

[0069]

[0070] Where x[n] represents the original discrete signal sequence; N represents the total length of the signal; μ x Represents the DC component of the signal; i represents the loop index for the summation operation; This represents the normalized signal sequence.

[0071] b. To obtain a two-dimensional representation reflecting the non-stationary structure, a Short-Time Fourier Transform (STFT) is performed on the normalized signal using a Hamming window with a window length of 512, a frame shift of 128, and 1024 FFT points to obtain the complex time-frequency matrix. The expression for the Short-Time Fourier Transform is:

[0072]

[0073] Where X(t,k) represents the complex time-frequency matrix; t represents the time frame index; k represents the frequency index; w[] represents the window function; R represents the frame shift, which determines the degree of overlap between adjacent time frames; n-tR represents the translation of the window function; N represents the discrete Fourier transform kernel; f represents the number of FFT points; j represents the imaginary unit.

[0074] In the technical solution of this application embodiment, a Hamming window is selected to obtain lower sidelobe leakage and make the time-frequency graph clearer. The window length L = 512, which is compatible with the large bandwidth structure of LFM and the fine-grained subcarrier information of OFDM. The frame shift R = L / 4 makes the frequency graph smooth and continuous in the time direction. The FFT point count is 1024, providing sufficient frequency domain resolution.

[0075] c. After taking the modulus |X(t,k)|, logarithmically compressing, and normalizing the complex time-frequency matrix, the amplitude spectrum channel is obtained. In this embodiment, the amplitude spectrum channel characterizes the energy density distribution of the signal, which is the most basic and intuitive feature, showing when and in what frequency band the signal appears. In this embodiment, the expression for generating the amplitude spectrum channel is:

[0076] S(t,k)=log(|X(t,k)|+∈)

[0077] Where S(t,k) represents the logarithmic amplitude time-frequency plot; ∈ represents a very small positive constant, used to avoid logarithmic singularity; |X(t,k)| represents taking the modulus of X(t,k).

[0078] The phase spectrum is calculated from the complex time-frequency matrix, and the partial derivative of the phase with respect to time t is calculated to obtain the instantaneous frequency channel. In this embodiment, the complex time-frequency matrix characterizes the instantaneous rate of change of the signal frequency. For radar signals such as linear frequency modulation (LFM), the instantaneous frequency changes linearly with time, and this channel will show clear slanted stripes; for communication signals, the frequency jumps or modulation information are also reflected in this channel, providing a key dimension of frequency dynamic characteristics.

[0079] The phase angle is extracted from the complex time-frequency matrix to obtain the phase spectrum channel. In this embodiment, by characterizing the phase structure of the signal, the phase information is crucial for identifying certain modulation methods, such as PSK phase modulation, and for analyzing the coherence of the signal. It is relatively independent of the amplitude channel and provides the "texture" and "structure" information of the signal.

[0080] d. Perform zero-mean unit variance normalization on each of the three channels, and use bilinear interpolation to uniformly adjust the size of each channel to 640×640 pixels. Finally, stitch them together along the channel dimensions to form the final three-channel time-frequency feature map 640×640×3. The normalization formula for each channel is:

[0081]

[0082] in, This represents the normalized single-channel time-frequency plot; This represents the mean of the data in that channel. This represents the standard deviation of the data in that channel.

[0083] After performing zero-mean unit variance normalization on each of the three channels, the expressions for the amplitude spectrum channel, instantaneous frequency channel, and phase spectrum channel are as follows:

[0084]

[0085] P(t,k)=argX(t,k)

[0086] Where C1(t,k) represents the amplitude spectrum channel; F(t,k) represents the instantaneous frequency channel; and P(t,k) represents the phase spectrum channel.

[0087] The final expression for the three-channel time-frequency characteristic map is:

[0088]

[0089] Where I(t,k) represents the three-channel time-frequency characteristic map.

[0090] All the three-channel time-frequency feature maps and their annotation information generated above are organized and expanded in size through data augmentation techniques such as random flipping and rotation, ultimately forming a training dataset for improving the training of YOLOv13 networks.

[0091] In this embodiment, by forming three independent feature channels—amplitude spectrum channel, instantaneous frequency channel, and phase spectrum channel—the principle of the RGB three-channel representation of a color image is mimicked, providing the YOLOv13 network with a multi-dimensional signal representation with low information redundancy. Compared to using only a single amplitude map, this joint input of energy, frequency variation, and phase structure greatly enhances the model's ability to distinguish similar waveforms, identify modulation types, and maintain robustness at low signal-to-noise ratios. Furthermore, it maps heterogeneous signals to a consistent two-dimensional time-frequency feature representation, addressing the differences in waveform structure, energy distribution, and modulation methods among different signals.

[0092] S2: Build a network based on the improved YOLOv13.

[0093] This step constructs an improved YOLOv13 network specifically for time-frequency plot signal detection. This network builds upon the original YOLOv13 framework, with targeted enhancements to the backbone, neck, and head.

[0094] The technical solution of this application embodiment introduces a dual-modal feature fusion module in the backbone of the network. This module feeds the input feature maps into two branches in parallel: the structural modality branch uses convolutional kernels with a large receptive field to focus on extracting global structural features of radar signals, such as continuous bands and slope changes, as presented in the time-frequency map; the texture modality branch uses smaller convolutional kernels with increased depth to focus on extracting local detail features of communication signals, such as modulation texture, subcarrier groups, and energy level transitions. Subsequently, the feature maps output from the two branches are weighted and fused channel by channel using learnable weight vectors.

[0095] The three-channel time-frequency feature map I(t,k,c) is decomposed into two types of information sources: global structural features and local texture features. The initial convolutional feature expressions for both are as follows:

[0096] F s =f s (I),F m =f m (I)

[0097] Among them, f s (I) represents the convolutional network operation of the structural modal branch; f m (I) represents the convolutional network operation of the texture modality branch; F s F m These represent the extracted structural feature map and texture feature map, respectively.

[0098] Furthermore, both branches will output corresponding feature tensors at different levels l of the network, expressed as:

[0099]

[0100] Among them, H l W l C l These represent the height, width, and number of channels of the feature map at layer l, respectively. These represent the structural modality branch and the texture modality branch, respectively, which output corresponding feature tensors at different layers l of the network. This indicates that each element in the tensor is a real number.

[0101] In this embodiment, the convolution kernel design of the structural mode branch is more biased towards a larger receptive field in order to capture the overall shape, slope changes and energy envelope structure of the broadband radar signal; while the texture mode branch uses a smaller convolution kernel and appropriately increases the depth in order to capture the microstructure texture of the communication modulation signal, such as energy level transitions and carrier spacing.

[0102] Subsequently, using learnable weight vectors, the output feature maps of the two branches at the same level are fused channel-by-channel with weights. Through a learnable linear fusion method, deep unification is achieved while maintaining modal integrity. The fusion formula is as follows:

[0103]

[0104] Where, α (l) β (l) represents the weight parameters automatically learned by the network for the structural mode branch and the texture mode branch, respectively, used to adaptively balance the contributions of structural information and texture information; ⊙ represents channel-wise multiplication, i.e., Hadamard product; This represents the multi-level features after fusion.

[0105] In the technical solution of this application embodiment, the dual-modal feature fusion module decomposes the input unified time-frequency feature map into two modalities with different focuses. Features are extracted by two independent but synchronously optimizable convolutional branches. Then, the two are unified in the deep semantic space through a learnable fusion mechanism. This solves the technical problem that in the integrated recognition task of radar and communication signals, it is often difficult to simultaneously take into account the structural differences and texture details of the two types of signals by relying solely on the time-frequency features of a single modality.

[0106] In this embodiment, some standard CSP (Cross Stage Partial) modules are replaced in the backbone network with lightweight bottleneck modules containing depthwise separable convolutions to reduce computational cost; dynamic convolutions are introduced in the deep layers to adaptively fuse the responses of multiple convolution kernels based on the input features; and channel attention modules and spatial attention modules (CBAM) are introduced after the backbone output layer to enhance key channels and important spatial regions of the feature map.

[0107] In this embodiment, an improved BiFPN++ multi-scale feature fusion module is introduced at the neck of the network to replace the feature pyramid structure in the original network. This module receives feature map inputs from three different scales of the backbone network, corresponding to feature maps downsampled by 8x, 16x, and 32x, denoted as P3, P4, and P5.

[0108] The fusion process involves two key paths:

[0109] The top-down approach involves first upsampling the semantically rich P5 feature map and fusing it with P4 features of the same scale from the backbone; then, upsampling the fused P4 features and fusing them with P3 features from the backbone. This approach transmits high-level semantic information to low-level features.

[0110] Bottom-up path: The P3 features fused from the top-down path are downsampled and then fused again with the fused P4 features; similarly, the fused P4 features are downsampled and then fused again with the P5 features. This path feeds back the precise location information from the lower layers to the higher layers.

[0111] The fusion process has two critical paths, and each path's fusion node employs a learnable weighted fusion method. The fusion formula for each path is expressed as follows:

[0112]

[0113] Where A and B represent the two input feature maps to be fused, and their specific references depend on the position of the node in the bidirectional path; w1 and w2 represent trainable positive weights used to dynamically adjust the relative importance of the two input features; Conv represents a standard convolution operation used to further process the weighted fused features; φ(A,B) represents the new fused feature map output after the learnable weighted fusion.

[0114] The technical solution in this application embodiment constructs a network based on an improved YOLOv13, which can dynamically and adaptively balance feature contributions from different levels and paths, thereby more effectively fusing multi-scale information from wide-bandwidth radar signals to fine communication modulation textures in the time-frequency map.

[0115] In the detection head of the network, an anchor-free detection module based on a task-aligned structure is introduced to completely replace the original anchor-based multi-scale prediction structure of YOLOv13. This module uses an anchor-free mechanism to directly predict the center point of the target and the distance from the center point to the four sides of the bounding box.

[0116] For the BiFPN++ multi-scale feature fusion module improved from the neck network, each scale feature map output is...

[0117] Wherein, Ps represents the scale, any one of P3, P4, and P5 in the embodiments of this application; H s represents the height of the feature; W s represents the width of the feature; C represents the number of channels in the feature;

[0118] The anchorless detection module based on task alignment structure first refines features through shared convolutional layers, and then splits into two branches in parallel: a classification branch and a regression branch.

[0119] Classification branch prediction: Outputs the probability that the target exists at each spatial location (i, j). and specific multi-class probability distributions

[0120] Regression branch: Outputs the distances (l, t, r, b) in four directions of the target bounding box at each spatial location (i, j), assuming that location is the target's center point. (i，j) And a center-ness score C (i，j) Where l, t, r, and b represent the offsets from the center point to the left, top, right, and bottom boundaries, respectively.

[0121] Based on the four directional distances of the target bounding box predicted by the regression branch, the spatial location (i, j) is mapped back to the center coordinates (x, j) of the original time-frequency map. ij ,y ij This allows direct decoding of the absolute coordinates of the prediction box in the original time-frequency graph:

[0122] x min =x ij -l (i,j) ,

[0123] y min =y ij -t (i,j) ,

[0124] x max =x ij +r (i,j) ,

[0125] y max =y ij +b (i,j)

[0126] Where, x min y min These represent the x and y coordinates of the top-left corner of the bounding box, respectively; x max y max These represent the x and y coordinates of the bottom right corner of the bounding box, respectively; (i,j) t represents the distance from (i, j) to the left boundary; (i,j) Represents the distance from (i, j) to the upper boundary; r (i,j) b represents the distance from (i, j) to the right boundary; (i,j) Represents the distance from (i, j) to the lower boundary; (x ij ,y ij) represent the horizontal and vertical coordinates of the center point on the original input time-frequency map (640×640) where the feature map position (i,j) is mapped back to the original input time-frequency map.

[0127] Centerness score C (i，j) The expression is:

[0128]

[0129] The technical solution of this application embodiment clearly demonstrates the decoding logic of anchorless detection through formulas: using the fixed point (x) mapped from the feature map position... ij ,y ij Using a hypothetical center, the coordinates of the top left and bottom right corners of the target box are directly obtained by adding or subtracting the four directional offsets l, t, r, and b predicted at that position. This completes the conversion from "relative offset" to "absolute positioning," eliminating the need for predefined anchor frames of different sizes and proportions in traditional methods. This improves the adaptability to irregular, multi-scale time-frequency signals. The anchor-free mechanism avoids the limitations of predefined anchor frame size and shape, making it particularly suitable for radar signals (such as broadband LFM) and communication signals (such as dense subcarriers) with irregular shapes and large scale spans in time-frequency maps.

[0130] The task-aligned, anchorless bounding box detection module dynamically adjusts the classification loss L at each scale during training based on the task alignment quality of each predicted sample, i.e., the intersection-over-union (IoU) between the predicted and ground truth bounding boxes, or a similarity metric for the modulated signal. cls,s With regression loss L reg,s weight w cls,s and w reg,s The expression for the total loss at this scale is:

[0131] L s =w cls,s ·L cls,s +w reg,s ·L reg,s

[0132] Among them, w cls,s The classification loss weights at scale s are dynamically changing coefficients used to scale the classification task loss L during training. cls,s Contribution to total loss; w reg,s The weight representing the regression localization loss at scale s is a dynamically changing coefficient used to scale the regression task loss L during training. reg,s Contribution to the total loss.

[0133] The technical solution of this application embodiment, through this dynamic weight adjustment, allows the model to automatically focus more attention on difficult samples that are accurately classified and located during optimization, forcing the two sub-tasks of classification and localization to work more closely together and align during the learning process, thereby significantly improving the sensitivity and robustness of recognition in complex scenarios such as weakly modulated signals and low signal-to-noise ratio signals.

[0134] Its core is the task alignment structure, which dynamically calculates and adjusts the weights between the classification loss and the bounding box regression loss during training based on the alignment quality (IoU) between each predicted bounding box and the ground truth bounding box. This allows the model to automatically focus on samples that are more accurately classified and localized during optimization, thereby achieving efficient collaborative optimization of classification and localization tasks.

[0135] S3: Train the network using the training dataset to obtain an integrated signal recognition model.

[0136] This step involves end-to-end training on the improved network constructed above, and lightweight optimizations are implemented.

[0137] First, the training dataset constructed by S1, containing three-channel time-frequency feature maps and their annotations, is input into the network for iterative training. After forward propagation, the loss between the network output and the ground truth annotations is calculated. The total loss function integrates the classification loss of the anchorless box detection module. The technical solution of this application uses Focal Loss to handle class imbalance, regression loss CIoU Loss or Smooth L1 Loss to optimize bounding box positions, and centrality loss to optimize centrality prediction. The training process uses backpropagation and gradient descent algorithms. In this application embodiment, AdamW is used to update all network parameters, and the task alignment structure automatically participates in and optimizes the weights of each loss term.

[0138] The technical solution of this application embodiment implements further lightweighting and attention enhancement after the main training is completed. Lightweighting includes: using depthwise separable convolutions in the network structure; and performing structured pruning on the network after training.

[0139] Attention Enhancement: In addition to introducing CBAM attention after the backbone output, a lightweight channel attention module is added after the output of each scale feature map (P3, P4, P5) in the neck network BiFPN++. This module first performs global average pooling (GAP) on the feature map Ps to obtain the channel description vector z, then generates the channel attention weight vector a through a two-layer fully connected network with non-linear activation. Finally, it performs channel weighting on the original features and applies lightweight channel attention again to the fused P3, P4, and P5. The specific expression is as follows:

[0140] z = GAP(P)s ),

[0141] a=σ(W2δ(W1z)),

[0142]

[0143] Wherein, GAP represents the global average pooling operation; z represents the channel description vector obtained after the GAP operation; a represents the calculated channel attention weight vector; W1 and W2 represent the weight matrices of the two fully connected layers for reducing the number of channels and restoring the original number of channels, respectively; δ represents the non-linear activation function, which is ReLU; σ represents the Sigmoid activation function. denoted as the output feature map after channel attention weighted modulation; a(c) represents all spatial locations of the same channel c multiplied by the same weight; h represents the height direction index of the feature map; w represents the width direction index of the feature map; c represents the channel direction index of the feature map.

[0144] The technical solution of this application embodiment performs lightweight channel attention on the fused P3, P4, and P5 again through attention enhancement, which enables the network to adaptively enhance channels sensitive to key signal features such as "band stripes", "subcarrier groups", and "frequency hopping points".

[0145] After model convergence, structured pruning techniques are employed. By evaluating the channel importance of each convolutional layer in the network, this embodiment removes channels with lower importance based on the L1 norm of the channel weights, thereby pruning the number of channels from C to C'. The pruning process aims to minimize performance loss, and is expressed as:

[0146]

[0147] Where k represents the number of pruned channels, and fine-tuning is usually required after pruning to restore model accuracy; Loss represents the loss function; Prune(Θ,k) represents the pruning operation function.

[0148] The technical solution of this application embodiment, through the above training and lightweighting process, ultimately obtains a high-precision, low-power, and small-size integrated signal recognition model. While maintaining excellent cross-system signal recognition and positioning capabilities, the recognition model possesses lightweight characteristics with the same structure but different weight distribution, enabling its efficient deployment on embedded platforms such as radar communication all-in-one machines, UAV edge computing nodes, or vehicle-mounted intelligent terminals, which have strict limitations on computing power and power consumption, thus meeting real-time processing requirements.

[0149] S4: Convert the original signal to be identified into a corresponding three-channel time-frequency feature map, input it into the integrated signal recognition model, and output the recognition result.

[0150] In practical applications, for a segment of raw radar or communication signal to be identified, it is first converted into a 640×640×3 three-channel time-frequency feature map using the same STFT parameters, normalization method, and size adjustment as described in S1. This feature map is then input into the recognition model trained in S3. The model performs forward inference through the network and ultimately outputs the category of all detected signal instances in the time-frequency map, such as "LFM radar" or "QAM communication," the confidence score for each category, and the bounding box coordinates of the region where each instance is located, along with the start-end time and start-end frequency information.

[0151] Example 2:

[0152] This embodiment provides a device for integrated identification of radar and communication signals based on YOLOv13, the device comprising:

[0153] The data preprocessing unit is used to convert the original signal to be identified into a three-channel time-frequency feature map; the network construction unit is used to construct the network based on the improved YOLOv13; the model training unit is used to train the network using the training dataset to obtain the integrated signal recognition model; and the signal recognition unit is used to process the input three-channel time-frequency feature map using the integrated signal recognition model and output the recognition result.

[0154] The data preprocessing unit receives the raw signal to be identified from the RF front-end and converts it into a three-channel time-frequency feature map in a unified format. The data preprocessing unit performs the following operations sequentially: DC removal, filtering, and energy normalization preprocessing on the raw signal; short-time Fourier transform of the preprocessed signal to obtain a complex time-frequency matrix; extraction of the amplitude spectrum, instantaneous frequency, and phase spectrum from this matrix, forming three feature channels respectively; and finally, normalization and size adjustment of each channel, stitching them together to form a three-channel time-frequency feature map with a size of 640×640×3, which serves as the standardized input for subsequent network processing.

[0155] The network building block constructs and instantiates a network based on the improved YOLOv13. This network features targeted enhancements to the original YOLOv13 architecture, specifically comprising three core substructures:

[0156] Dual-modal feature fusion module: This module extracts global structural features and local texture features from the input feature map through two parallel convolutional branches, and then performs adaptive fusion using learnable weights.

[0157] Improved BiFPN++ multi-scale feature fusion module: This module replaces the original feature pyramid network and fuses feature maps of different scales multiple times through a bidirectional path from top to bottom and bottom to top, and adopts a learnable weighted summation method at the fusion node.

[0158] Anchorless detection module based on task alignment structure: This module adopts a prediction mechanism without predefined anchor boxes, directly outputting the target's class probability, centrality, and bounding box offset. Its internal task alignment structure can dynamically coordinate the learning of classification and localization tasks during training.

[0159] The model training unit uses the data preprocessing unit to process a large number of labeled raw signal samples, constructing a training dataset consisting of three-channel time-frequency feature maps and their annotation information. Subsequently, this dataset is used to train the improved YOLOv13-based network constructed by the network construction unit. The training process updates network parameters by optimizing the comprehensive loss function, and lightweight compression of the trained model can be achieved through techniques such as structured pruning, ultimately resulting in an integrated signal recognition model suitable for real-time deployment.

[0160] The signal recognition unit loads the integrated signal recognition model obtained by the model training unit. After the data preprocessing unit converts the real-time acquired raw signal to be recognized into a three-channel time-frequency feature map, this unit uses the integrated signal recognition model to process and infer the feature map, and finally outputs a structured result list, which contains the category label, confidence score and bounding box coordinates of each detected signal instance in the time-frequency map.

[0161] The technical solution provided in this application, a method for integrated radar and communication signal recognition based on YOLOv13, effectively solves the problems of poor adaptability and insufficient feature representation capability of traditional methods in signal representation and feature extraction by constructing a unified three-channel time-frequency feature map and designing a dual-modal feature fusion mechanism. First, the original radar and communication signals are standardized and preprocessed, and converted into a two-dimensional time-frequency feature map containing amplitude, instantaneous frequency, and phase information, providing a consistent input across different systems for subsequent processing. Then, a dual-modal feature fusion module is introduced into the network backbone, using structural and texture branches to extract the global contour and local details of the signal respectively, and adaptively fused using learnable weights. This series of operations enables the model to process signals with multiple modulation styles such as LFM, NLFM, OFDM, and QAM within a unified framework, overcoming the limitations of existing methods that rely on specific signal models and struggle to consider both structural and texture features, significantly improving the system's signal representation robustness and cross-system adaptability in complex electromagnetic environments.

[0162] In terms of recognition and localization mechanisms, this invention fundamentally solves the problem that traditional classification methods cannot perform instance-level identification and localization of dense and aliased signals in time-frequency images by introducing an anchor-free detection paradigm based on task-aligned structure. This invention reconstructs the integrated recognition task into a target detection problem in time-frequency images, employing an anchor-free detection architecture based on an improved YOLOv13. This architecture directly predicts the center point and bounding box offset of signal instances, eliminating the dependence on predefined anchor boxes. Its core task-aligned structure dynamically adjusts the optimization weights of classification and regression tasks during training based on the alignment quality between the predicted and ground truth boxes, thus forcing the two sub-tasks to work closely together during the learning process. This design enables the system not only to output the signal category but also to simultaneously output the precise time-frequency locations of multiple coexisting signal instances, achieving a paradigm shift from "image-level classification" to "instance-level detection," greatly enhancing the recognition accuracy and scene adaptability in scenarios with aliased signals, weak signals, and low signal-to-noise ratios.

[0163] Regarding model efficiency and deployment feasibility, this invention effectively addresses the challenges of high computational complexity and real-time processing on resource-constrained embedded platforms by employing a lightweight design and attention enhancement strategy throughout the network structure. The invention extensively replaces standard convolutions with depthwise separable convolutions in the network backbone and neck region, and removes redundant channels after training through performance-loss-oriented structured pruning, significantly reducing the model's parameter count and computational overhead. Simultaneously, channel attention and spatial attention modules are embedded at key network locations (such as after the backbone output and after multi-scale fusion nodes), enabling the model to adaptively focus on key time-frequency regions and feature channels for identification. These measures, while ensuring and even enhancing the model's ability to distinguish key features such as "band stripes" and "subcarrier groups," achieve model lightweighting and inference acceleration. This allows the proposed integrated high-precision recognition model to be practically deployed in radar communication all-in-one machines, drones, or vehicle terminals with strict limitations on computing power and power consumption, meeting the stringent real-time requirements of engineering applications and enhancing the practical value of this invention.

[0164] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0165] The units described in the embodiments of this application can be implemented in software or hardware. The names of the units are not, in some cases, limiting the scope of the unit itself.

Claims

1. A method for integrated identification of radar and communication signals based on YOLOv13, characterized in that, include: Construct a dataset based on three-channel time-frequency feature maps; Build a network based on the improved YOLOv13; Specifically, a dual-modal feature fusion module is introduced into the backbone of the network, an improved BiFPN++ multi-scale feature fusion module is introduced into the neck of the network, and an anchorless detection module based on task alignment structure is introduced into the detection head of the network. The network is trained using the training dataset to obtain an integrated signal recognition model; The original signal to be identified is converted into a corresponding three-channel time-frequency feature map, which is then input into the integrated signal recognition model. The model outputs the signal's category, confidence level, and position in the time-frequency map.

2. The method according to claim 1, characterized in that, The method for constructing the three-channel time-frequency feature map includes: Preprocessing of raw signals to be processed in radar and communication systems; The preprocessed signal is subjected to time-frequency transformation to obtain a complex time-frequency matrix; Amplitude, instantaneous frequency, and phase information are extracted from the complex time-frequency matrix to form three feature channels: amplitude spectrum channel, instantaneous frequency channel, and phase spectrum channel, respectively. The three feature channels are normalized and their sizes are adjusted, and then spliced ​​together to form the three-channel time-frequency feature map.

3. The method according to claim 1, characterized in that, The dual-modal feature fusion module includes: Structural modality branches are used to extract global structural features from input features; Texture modality branch, used to extract local texture features of input features; The fusion unit is used to perform weighted fusion of the global structural features and the local texture features using learnable weights.

4. The method according to claim 1, characterized in that, The backbone of the network also includes: Replace some standard convolutions with lightweight modules that include depthwise separable convolutions; Introducing dynamic convolution into deep networks; Channel attention and spatial attention mechanisms are introduced at the output layer of the backbone network.

5. The method according to claim 1, characterized in that, The improved BiFPN++ multi-scale feature fusion module includes: The input terminal is used to receive feature maps of multiple scales from different levels of the backbone network; A path used for bidirectional information transfer and fusion among the multiple scale feature maps; At the fusion node, a weighted fusion method with learnable weights is adopted.

6. The method according to claim 1, characterized in that, The anchorless frame detection module based on task alignment structure includes: The classification branch is used to predict the probability of the target category at each spatial location; The regression branch is used to predict the target bounding box information for each spatial location; The task alignment structure is used to dynamically adjust the optimization weights of classification and regression tasks during training based on the quality of the prediction task.

7. The method according to claim 1, characterized in that, The method also includes a model lightweighting step: Depthwise separable convolutions are used in the network structure; After training, the network is pruned in a structured manner.

8. The method according to claim 1, characterized in that, The construction of the dataset based on the three-channel time-frequency feature map includes: Obtain labeled raw radar and communication signal samples; Each of the original signal samples is constructed into a corresponding three-channel time-frequency feature map; All constructed and labeled three-channel time-frequency feature maps constitute the training dataset used for network training.

9. The method according to claim 1, characterized in that, Training the network using the training dataset includes: The training dataset is input into the network for forward propagation to obtain the prediction results; Calculate the loss between the predicted result and the true label; Network parameters are updated via backpropagation.

10. An apparatus for implementing the method for integrated identification of radar and communication signals based on YOLOv13 as described in any one of claims 1-9, characterized in that, include: The data preprocessing unit is used to convert the raw signal to be identified into a three-channel time-frequency feature map; Network building unit, used to build the network based on the improved YOLOv13; The model training unit is used to train the network using the training dataset to obtain the integrated signal recognition model; The signal recognition unit is used to process the input three-channel time-frequency feature map using the integrated signal recognition model and output the recognition result.

Images