Low, slow and small unmanned aerial vehicle radar identification method

Abstract

The application relates to the technical field of radar target recognition, and discloses a low, slow and small unmanned aerial vehicle radar recognition method, steps of which are as follows: S1, a radar echo preprocessing stage; S2, a micro-Doppler feature extraction stage; S3, a classification reasoning stage; and S4, a model training and deployment stage; the application completes a whole-process special design according to the hardware characteristics of a low-cost quasi-coherent magnetron radar, can adapt to the signal characteristics of the radar system and the computing power constraints of an embedded platform, can complete end-to-end intelligent recognition without relying on high-computing-power processing equipment and high-performance full-phase radar platforms, greatly reduces the deployment threshold and hardware cost of low, slow and small target intelligent recognition technology, can be directly integrated into a terminal processing system of an existing low-cost low-altitude surveillance radar, realizes localized real-time operation of the recognition algorithm, and does not need to rely on cloud computing power support.

Description

Technical Field

[0001] This invention relates to the field of radar target recognition technology, and in particular to a radar recognition method for low-speed, small unmanned aerial vehicles (UAVs). Background Technology

[0002] Currently, low-altitude, slow-moving, and small targets, represented by consumer-grade and industrial-grade drones, pose a prominent low-altitude security threat in key protection scenarios such as military strongholds, civilian airports, and major public events, due to their ease of operation, strong concealment, and low acquisition threshold.

[0003] To build a wide-area low-altitude surveillance network, low-cost quasi-coherent magnetron radar has become the core choice for large-scale deployment of low-altitude surveillance sensors due to its performance and cost advantages that are suitable for low-altitude detection scenarios. The accurate identification technology for low, slow and small targets for this radar platform is the core technical link to ensure the effective operation of the low-altitude surveillance system and to achieve accurate handling of threat targets.

[0004] Existing radar identification technologies for low, slow, and small targets are mainly divided into two categories, neither of which can adapt to the dual core constraints of low-cost quasi-coherent magnetron radar platforms. The first category is traditional feature engineering-based identification methods. These methods rely on manually designed feature indicators to complete target classification. They can achieve basic identification results in simple background environments, but in complex scenarios with strong clutter interference, the robustness of the features is insufficient, and they are easily affected by background clutter, resulting in a high false alarm rate. They cannot adapt to the signal quality constraints of quasi-coherent radar. Summary of the Invention

[0005] The purpose of this invention is to provide a radar identification method for low-speed, small unmanned aerial vehicles (UAVs) to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A radar identification method for low-speed, small unmanned aerial vehicles (UAVs) includes the following steps:

[0008] S1. In the radar echo preprocessing stage, pulse compression processing is performed on the original radar I / Q sampling data to construct the coherent processing interval, motion compensation operation is performed, and a two-dimensional range Doppler image is generated.

[0009] S2. Micro-Doppler feature extraction stage: Candidate target detection is completed in the distance Doppler map, the complex baseband time series corresponding to the target is extracted, short-time Fourier transform is performed, and standardized micro-Doppler spectrum is generated.

[0010] S3, Classification Reasoning Stage: The micro-Doppler spectrum is normalized and input into the lightweight CA-DSNet network to complete feature extraction and calculation, and outputs the probability value corresponding to the target category;

[0011] S4. In the model training and deployment phase, complete the standardized construction of the dataset, perform teacher network pre-training, optimize the training of the lightweight network through the joint loss function, and complete the model format conversion and embedded platform deployment.

[0012] As a further improvement to this technical solution: in the radar echo preprocessing stage, pulse compression processing is performed on the original I / Q sampling data of the radar to generate a one-dimensional range image; 8 to 16 pulses are selected to form a single coherent processing interval; phase alignment processing is performed on the target echo within the coherent processing interval to complete motion compensation; and fast Fourier transform is performed on the echo sequence of each range unit to generate a two-dimensional range Doppler image.

[0013] As a further improvement to this technical solution: In the micro-Doppler feature extraction stage, constant false alarm rate (CFAR) detection is performed in the range-Doppler image to screen candidate target points with a signal-to-noise ratio greater than 8dB, and the range cell corresponding to the target is locked; the complex baseband time series corresponding to the coherent processing interval within the range cell corresponding to the target is extracted; a short-time Fourier transform is performed on the complex baseband time series to generate a single-channel micro-Doppler spectrum with a size of 64×64; the calculation expression for the short-time Fourier transform is as follows: ,in, Given a complex baseband time series as input. This is a sliding window function with a window size of 256 points. For the Doppler frequency dimension, In terms of time dimension, For the output microDoppler spectrum, the overlap rate of the window function is set to 75%.

[0014] As a further improvement to this technical solution: In the classification reasoning stage, the input micro-Doppler spectrum is linearly normalized to map the values ​​to the interval between 0 and 1; the normalized micro-Doppler spectrum is input into a lightweight CA-DSNet network, which sequentially passes through a standard convolutional initial layer, three sets of depthwise separable convolutional modules, three sets of SE channel attention modules, a global average pooling unit, a Dropout regularization unit, and a fully connected layer; the probability values ​​corresponding to four types of targets—quadcopter drones, fixed-wing drones, birds, and kites—are output through the fully connected layer.

[0015] As a further improvement to this technical solution: the standard convolutional initial layer of the lightweight CA-DSNet network has 32 output channels, a kernel size of 3×3, and a stride of 2; the three sets of depthwise separable convolutional modules are connected in sequence, with the number of output channels of the three sets of modules being 64, 128, and 256, respectively. Each set of depthwise separable convolutional modules consists of depthwise convolution and pointwise convolution. The depthwise convolution performs convolution operation independently on each input channel, and the pointwise convolution completes the feature fusion and transformation of the channel dimension through 1×1 convolution.

[0016] As a further improvement to this technical solution: the three SE channel attention modules correspond one-to-one with the three depthwise separable convolutional modules, and are embedded in the output of each depthwise separable convolutional module. The channel weight calculation expression for the SE channel attention module is as follows: ,in, For the first The weight values ​​corresponding to each feature channel It is the Sigmoid activation function. It is the ReLU activation function. This is the weight matrix of the first fully connected layer, with a dimensionality reduction ratio set to 8. This is the weight matrix of the second fully connected layer. The channel descriptor is obtained by global average pooling for the c-th feature channel. The weight of each channel is multiplied by the corresponding feature channel one by one to complete the adaptive weighting of the feature map.

[0017] As a further improvement to this technical solution: during the model training and deployment phase, the collected radar echo data is labeled and divided to generate standardized training, validation, and test sets; a ResNet-18 network is selected as the teacher network to complete pre-training and generate high-precision soft labels; a lightweight CA-DSNet network is used as the student network, and the network parameters are iteratively optimized through a joint loss function; the trained model is converted to ONNX format to complete operator adaptation and deployment on the embedded platform.

[0018] As a further improvement to this technical solution: the calculation expression of the joint loss function is as follows: ,in, This represents the total loss value during the training process. These are the weighting coefficients for the cross-entropy loss. These are the weighting coefficients for the KL divergence loss. The cross-entropy loss between the student network output and the real label is used. The KL divergence loss is used to compare the student network output with the teacher network soft labels. The training process iteratively updates the student network parameters by minimizing the total loss value.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] 1. This invention provides a dedicated design for the entire process of low-cost quasi-coherent magnetron radar, which is adapted to the signal characteristics of this type of radar and the computing power constraints of the embedded platform. It can complete end-to-end intelligent identification without relying on high-performance processing equipment and high-performance fully coherent radar platforms, which greatly reduces the deployment threshold and hardware cost of intelligent identification technology for low, slow and small targets. It can be directly integrated into the terminal processing system of existing low-cost low-altitude surveillance radar to realize the local real-time operation of the identification algorithm without relying on cloud computing power support.

[0021] 2. This invention combines the physical characteristics of micro-Doppler features with the feature extraction advantages of deep learning. It can stably extract effective identification features of low, slow, and small targets in complex clutter environments, effectively improving the accuracy and stability of target identification, reducing the false alarm rate in complex scenarios, and accurately classifying and distinguishing multiple typical targets in low-altitude scenarios. It provides stable and reliable core identification technology support for the large-scale deployment of wide-area low-altitude surveillance networks and can effectively meet the low-altitude security protection needs of various key scenarios.

[0022] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it according to the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Specific embodiments of the present invention are given in detail below with reference to the accompanying drawings. Attached Figure Description

[0023] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0024] Figure 1 This is a schematic diagram of the structure of a radar identification method for low-speed, small unmanned aerial vehicles (UAVs). Detailed Implementation

[0025] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are for illustrative purposes only and are not intended to limit the scope of the invention. The invention is described more specifically in the following paragraphs by way of example with reference to the accompanying drawings. It should be noted that the drawings are in a very simplified form and use non-precise proportions, and are only used to facilitate and clarify the illustration of the embodiments of the present invention.

[0026] Please see Figure 1 In this embodiment of the invention, a radar identification method for low-speed, small unmanned aerial vehicles (UAVs) includes the following steps:

[0027] S1. In the radar echo preprocessing stage, pulse compression processing is performed on the original radar I / Q sampling data to construct the coherent processing interval, motion compensation operation is performed, and a two-dimensional range Doppler image is generated.

[0028] S2. Micro-Doppler feature extraction stage: Candidate target detection is completed in the distance Doppler map, the complex baseband time series corresponding to the target is extracted, short-time Fourier transform is performed, and standardized micro-Doppler spectrum is generated.

[0029] S3, Classification Reasoning Stage: The micro-Doppler spectrum is normalized and input into the lightweight CA-DSNet network to complete feature extraction and calculation, and outputs the probability value corresponding to the target category;

[0030] S4. Model training and deployment phase: Complete the standardized construction of the dataset, perform teacher network pre-training, optimize the training of the lightweight network through the joint loss function, and complete the model format conversion and embedded platform deployment.

[0031] Specifically, the radar echo preprocessing stage is a fundamental step in radar signal processing. Its function is to convert the raw baseband data acquired by the radar into a range-Doppler two-dimensional feature map that can be used for target detection, thereby completing the energy aggregation and mapping of the target in the range and velocity dimensions.

[0032] The micro-Doppler feature extraction stage is the core step in feature generation of this method. Its function is to locate the target position from the two-dimensional distance Doppler image, extract the corresponding time-domain echo sequence of the target, and generate a micro-Doppler spectrum that can characterize the micro-motion characteristics of the target through time-frequency transformation, providing core feature input for subsequent classification.

[0033] The classification and reasoning stage is the core of the target identification process in this method. Its function is to standardize the generated micro-Doppler spectrum, and then use a dedicated lightweight neural network to complete feature extraction and classification calculation, outputting the probability result of the target's category, thus realizing the type determination of low, slow and small targets.

[0034] The model training and deployment phase is the implementation and deployment stage of this method. Its role is to build a standardized dataset to complete the training and optimization of the dedicated network, and at the same time complete the model format conversion and hardware adaptation to ensure that the method can run stably on the embedded platform.

[0035] In the radar echo preprocessing stage, pulse compression processing is performed on the original I / Q sampling data of the radar to generate a one-dimensional range image; 8 to 16 pulses are selected to form a single coherent processing interval; phase alignment processing is performed on the target echo within the coherent processing interval to complete motion compensation; fast Fourier transform is performed on the echo sequence of each range cell to generate a two-dimensional range Doppler image.

[0036] Specifically, pulse compression processing aims to resolve the contradiction between radar peak power and detection range while ensuring radar range resolution. By matching and filtering the linear frequency modulated signal, wide pulses are compressed into narrow pulses, improving the signal-to-noise ratio of the echo. The generated one-dimensional range profile can characterize the energy distribution of target echoes at different distances.

[0037] The purpose of constructing the coherent processing interval is to select a fixed number of continuous pulses as the smallest data unit for a single Doppler analysis. The selection range of 8 to 16 pulses is adapted to the phase stabilization time of the quasi-coherent magnetron radar, so as to ensure the accuracy of Doppler analysis while avoiding the interference of phase noise accumulation on the analysis results.

[0038] Motion compensation is used to compensate for the Doppler frequency shift caused by the uniform motion of the target. Phase alignment processing is used to eliminate the influence of the target's motion on the echo phase, while preserving the micro-doppler features of the target's rotor, wings and other components, thus providing clean echo data for subsequent micro-doppler feature extraction.

[0039] Fast Fourier Transform (FFT) processing converts the time-domain echo sequence of each range unit to the frequency domain, completes the Doppler dimension analysis, and generates a two-dimensional range-Doppler map that can simultaneously represent the target's range and velocity information, enabling preliminary localization and detection of aerial targets.

[0040] In the micro-Doppler feature extraction stage, constant false alarm rate (CFAR) detection is performed on the range-Doppler image to screen candidate target points with a signal-to-noise ratio (SNR) greater than 8 dB, thus locking the range cell corresponding to the target. The complex baseband time series corresponding to the coherent processing interval within the range cell corresponding to the target is extracted. A short-time Fourier transform (SFT) is performed on the complex baseband time series to generate a single-channel micro-Doppler spectrum with a size of 64×64. The calculation expression for the SFT is as follows: ,in, Given a complex baseband time series as input. This is a sliding window function with a window size of 256 points. For the Doppler frequency dimension, In terms of time dimension, For the output microDoppler spectrum, the overlap ratio of the window function is set to 75%;

[0041] Specifically, constant false alarm rate (CFAR) detection adaptively sets a detection threshold in the range Doppler image to filter out target echoes that meet the signal-to-noise ratio (SNR) requirements from background clutter and noise. A screening threshold with an SNR greater than 8 dB can filter out most invalid clutter points while ensuring the target detection rate, reducing the unnecessary overhead of subsequent calculations.

[0042] Complex baseband time series extraction is used to locate the range cell where the detected candidate target is located and extract the complete time-domain complex baseband echo data within that range cell corresponding to the coherent processing interval. This sequence contains information on phase and frequency changes caused by the target's micro-motion and is the basic input data for generating the micro-Doppler spectrum.

[0043] The short-time Fourier transform converts the complex baseband echo sequence in the time domain into a two-dimensional spectrum in the time and frequency domain. It also characterizes the frequency variation of the echo signal over time, accurately capturing the micro-Doppler frequency changes caused by micro-motion characteristics such as rotor rotation and wing vibration. It is the core transformation method for generating micro-Doppler features.

[0044] In the formula, the window length of the sliding window function is set to 256 points to match the number of pulses in the coherent processing interval and ensure the temporal resolution of a single transformation; the window function overlap rate is set to 75% to improve the temporal continuity of the time-frequency transformation and avoid feature breaks in the spectrum; the final generated 64×64 single-channel micro-Doppler spectrum, while ensuring feature integrity, adapts to the input size requirements of the subsequent lightweight network and controls the computational load of subsequent inference.

[0045] In the classification reasoning stage, the input micro-Doppler spectrum is linearly normalized to map the values ​​to the interval between 0 and 1. The normalized micro-Doppler spectrum is then input into a lightweight CA-DSNet network, passing sequentially through a standard convolutional initial layer, three sets of depthwise separable convolutional modules, three sets of SE channel attention modules, a global average pooling unit, a Dropout regularization unit, and a fully connected layer. The fully connected layer outputs the probability values ​​corresponding to four target categories: quadcopter drones, fixed-wing drones, birds, and kites.

[0046] Specifically, linear normalization maps the pixel values ​​of the micro-Doppler spectrum to a uniform range of 0 to 1, eliminating the interference of differences in spectrum value amplitude on network feature extraction, ensuring the consistency of the distribution of network input data, and improving the stability of network inference.

[0047] The lightweight CA-DSNet network is a dedicated classification and inference network for this method. It is adapted to the computing power constraints of embedded platforms and is specially designed for the feature characteristics of micro-Doppler spectra. Each component of the network sequentially completes the initial feature extraction, depth extraction, adaptive weighting, dimensionality compression, regularization, and classification output.

[0048] The standard convolutional initial layer is used to perform initial feature extraction and downsampling on the input normalized spectral map, reduce the size of the feature map, increase the number of feature channels, and lay the foundation for subsequent deep feature extraction.

[0049] The depthwise separable convolution module is used to extract deep features from the feature map. While retaining the feature extraction capability, it significantly reduces the number of network parameters and computational cost, adapting to the computing power limitations of embedded platforms.

[0050] The SE channel attention module adaptively weights different channels of the feature map, strengthens channels containing effective micro-motion features, suppresses channels containing noise and invalid information, improves the network's ability to extract core features, and adapts to the scene characteristics of quasi-coherent radar micro-Doppler feature blurring.

[0051] The global average pooling unit compresses a two-dimensional feature map into a one-dimensional feature vector, significantly reducing the dimensionality of the features, reducing the number of parameters in subsequent fully connected layers, and avoiding the over-sensitivity of fully connected layers to the spatial location of the feature map, thereby improving the generalization ability of the network.

[0052] Dropout regularization units randomly deactivate some neurons during network training to prevent the network from overfitting the training data and improve the network's generalization ability in different scenarios.

[0053] The fully connected layer maps a one-dimensional feature vector to a probability value corresponding to the target category. The four output categories correspond to four typical aerial targets in low-speed, small-scale scenarios, thus achieving accurate target classification.

[0054] The standard convolutional initial layer of the lightweight CA-DSNet network has 32 output channels, a kernel size of 3×3, and a stride of 2. Three sets of depthwise separable convolutional modules are connected in sequence, with 64, 128, and 256 output channels respectively. Each set of depthwise separable convolutional modules consists of depthwise convolution and pointwise convolution. The depthwise convolution performs convolution operation independently on each input channel, and the pointwise convolution completes the feature fusion and transformation of the channel dimension through 1×1 convolution.

[0055] Specifically, the standard convolutional initial layer's 32 output channels can complete the dimensional expansion of basic features in the initial stage of the network. The 3×3 convolutional kernel size is adapted to the local feature extraction requirements of micro-Doppler spectra. The stride of 2 can achieve feature map downsampling, reducing the input 64×64 feature map size to 32×32, thus reducing the computational load of subsequent modules.

[0056] Three sets of depthwise separable convolutional modules are connected in sequence, and the number of output channels gradually increases from 64 to 256, which can realize the layer-by-layer extraction and dimensional expansion of features, from shallow edge and texture features to deep micro-motion features and category features, ensuring the feature extraction capability of the network.

[0057] Each group of depthwise separable convolutional modules consists of two parts: depthwise convolution and pointwise convolution. Depthwise convolution performs convolution operations independently on each input channel, and each convolutional kernel is only responsible for feature extraction of one feature channel, which greatly reduces the computational cost of convolution operations. Pointwise convolution transforms and fuses the feature map output by depthwise convolution through a 1×1 convolutional kernel, integrating feature information from different channels. While reducing the computational cost and parameter count, it retains the feature fusion capability of standard convolution, adapting to the computing power constraints of embedded platforms.

[0058] The three SE channel attention modules correspond one-to-one with the three depthwise separable convolutional modules, and are embedded in the output of each depthwise separable convolutional module. The channel weight calculation expression for the SE channel attention module is as follows: ,in, For the first The weight values ​​corresponding to each feature channel It is the Sigmoid activation function. It is the ReLU activation function. This is the weight matrix of the first fully connected layer, with a dimensionality reduction ratio set to 8. This is the weight matrix of the second fully connected layer. The channel descriptor is obtained by global average pooling for the c-th feature channel. The weight of each channel is multiplied by the corresponding feature channel one by one to complete the adaptive weighting of the feature map.

[0059] Specifically, the three SE channel attention modules correspond one-to-one with the three depthwise separable convolutional modules and are embedded at the output of each depthwise separable convolutional module. This enables adaptive weighting of the feature channels immediately after feature extraction at each layer, continuously strengthening effective features and suppressing ineffective features during feature propagation, thereby improving the feature extraction efficiency of each layer of the network.

[0060] The core computation of the SE channel attention module consists of two steps. The first step is global feature aggregation, which uses global average pooling to compress the two-dimensional feature map of each feature channel into a single value, generating the channel descriptor for the corresponding channel. The first step is to characterize the overall feature information of the channel; the second step is to adaptively generate channel weights, which generates the weight value corresponding to each channel through two fully connected layers and activation functions, so as to evaluate the importance of different channels.

[0061] In the formula, the dimensionality reduction ratio is set to 8, which can reduce the number of parameters and computation of the fully connected layer while ensuring the accuracy of channel weight evaluation, thus adapting to the lightweight design requirements of the network; the Sigmoid activation function can map the generated weight values ​​to the 0 to 1 interval, realizing weighted control of feature channels; the ReLU activation function can improve the nonlinear expression ability of the two fully connected layers and ensure the accuracy of channel importance evaluation.

[0062] The generated channel weights are multiplied channel by channel with the original feature map to complete the adaptive weighting of the feature map. This allows the network to focus on the effective channels containing the target's micro-Doppler features, reduce the interference of invalid features caused by quasi-coherent radar phase noise, and improve the network's ability to recognize fuzzy features.

[0063] During the model training and deployment phase, the collected radar echo data is labeled and divided to generate standardized training, validation, and test sets. The ResNet-18 network is selected as the teacher network to complete pre-training and generate high-precision soft labels. The lightweight CA-DSNet network is used as the student network, and the network parameters are iteratively optimized through a joint loss function. The trained model is converted into ONNX format to complete operator adaptation and deployment on the embedded platform.

[0064] Specifically, the purpose of dataset labeling and partitioning is to label the radar echo data collected in the field according to the target type, and to divide it into training set, validation set and test set. The training set is used for iterative updates of network parameters, the validation set is used for model accuracy verification and hyperparameter adjustment during the training process, and the test set is used for final performance testing of the model after training to ensure the model's generalization ability and recognition accuracy.

[0065] The purpose of ResNet-18 teacher network pre-training is to learn the mapping relationship between micro-Doppler features and categories in the dataset through a high-precision deep network, and generate high-precision soft labels. Compared with hard labels, soft labels contain more similarity information between categories, which can provide richer guidance for the training of lightweight student networks.

[0066] The purpose of joint loss optimization training for student networks is to allow student networks to learn the category information of real labels and the feature distribution information of soft labels from teacher networks simultaneously through the joint loss function. This significantly reduces the number of network parameters while preserving the high-precision recognition ability of teacher networks, thus solving the problem of insufficient recognition accuracy in lightweight networks.

[0067] ONNX format conversion and embedded platform adaptation aims to convert trained models from frameworks such as PyTorch into the ONNX universal open-source format, eliminating model compatibility issues between different frameworks. At the same time, it adapts and optimizes operators for the NPU of embedded platforms, ensuring that the model can run stably and efficiently on embedded platforms and meet the requirements of real-time recognition.

[0068] The formula for calculating the joint loss function is as follows: ,in, This represents the total loss value during the training process. These are the weighting coefficients for the cross-entropy loss. is the weight coefficient of the KL divergence loss, is the cross-entropy loss between the output of the student network and the true label, is the KL divergence loss between the output of the student network and the soft label of the teacher network. The training process completes the iterative update of the student network parameters by minimizing the total loss value;

[0069] Specifically, the joint loss function is composed of two parts, the cross-entropy loss and the KL divergence loss, which can simultaneously constrain the classification accuracy and feature distribution of the student network, and achieve the training effect of knowledge distillation;

[0070] Cross-entropy loss is used to constrain the difference between the output of the student network and the true labeled label, ensuring that the student network can learn the correct class classification information. It is the basic loss function for the training of the classification network;

[0071] KL divergence loss is used to constrain the difference between the output distribution of the student network and the soft label distribution of the teacher network output, enabling the student network to learn the association information between the feature extraction logic of the teacher network and the classes, and replicating the high-precision recognition ability of the teacher network under a lightweight structure;

[0072] Weight coefficient and are used to adjust the proportion of the two parts of the loss in the total loss, and can balance the classification accuracy and the effect of knowledge distillation according to the training requirements, achieving the optimal training effect of the student network;

[0073] In the training process, through the gradient descent algorithm, the parameters of the student network are continuously iteratively updated to minimize the total loss value, making the output of the student network continuously approach the true label and the soft label of the teacher network, and finally obtaining a network model with both lightweight and high recognition accuracy.

[0074] The usage method and working principle of the present invention are as follows:

[0075] Usage method: First, deploy the lightweight recognition model optimized by training on the ARM and NPU architecture embedded processing platform supporting the low-cost quasi-coherent magnetron radar. The radar system collects the original echo baseband data of the monitored airspace in real time and transmits it to the processing platform. The processing platform first completes the preprocessing of the radar echo and generates the range-Doppler map according to the preset process, then locks the candidate targets from the range-Doppler map and generates the corresponding micro-Doppler spectrogram. Subsequently, the standardized micro-Doppler spectrogram is input into the pre-deployed lightweight neural network for inference calculation, and finally the classification and recognition result of the target is output. After associating the recognition result with the target track information, it is output to the back-end disposal system. The entire recognition process is executed cyclically with the sampling refresh rate of the radar to achieve continuous real-time recognition of low, slow, and small targets in the monitored airspace.

[0076] Working Principle: Classification and recognition are achieved based on the unique micro-motion characteristics corresponding to the micro-Doppler features generated by the structure and motion patterns of small, slow, and low-speed targets. First, radar echo preprocessing is used to compensate for the target's motion, eliminating interference from the target's motion on the micro-motion characteristics. Then, a short-time Fourier transform is used to convert the target's time-domain echo sequence into a time-frequency domain micro-Doppler spectrum, fully preserving the target's micro-motion characteristic information. Subsequently, a lightweight, deeply separable convolutional network specifically designed for micro-Doppler features is used to extract features. A channel attention mechanism adaptively strengthens the weights of effective feature channels, suppressing invalid feature interference from noise and clutter. Simultaneously, a knowledge distillation training strategy significantly compresses the number of network parameters and computational load while retaining the high-precision network's feature extraction capabilities. Finally, accurate classification of different types of small, slow, and low-speed targets is achieved on resource-constrained embedded platforms.

[0077] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Those skilled in the art can readily implement the present invention based on the description and drawings above. However, any modifications, alterations, and variations made by those skilled in the art without departing from the scope of the present invention using the disclosed technical content are equivalent embodiments of the present invention. Furthermore, any modifications, alterations, and variations made to the above embodiments based on the essential technology of the present invention are still within the protection scope of the present invention.

Claims

1. A radar identification method for low-speed, small unmanned aerial vehicles (UAVs), characterized in that, Includes the following steps: S1. In the radar echo preprocessing stage, pulse compression processing is performed on the original radar I / Q sampling data to construct the coherent processing interval, motion compensation operation is performed, and a two-dimensional range Doppler image is generated. S2. Micro-Doppler feature extraction stage: Candidate target detection is completed in the distance Doppler map, the complex baseband time series corresponding to the target is extracted, short-time Fourier transform is performed, and standardized micro-Doppler spectrum is generated. S3, Classification Reasoning Stage: The micro-Doppler spectrum is normalized and input into the lightweight CA-DSNet network to complete feature extraction and calculation, and outputs the probability value corresponding to the target category; S4. In the model training and deployment phase, complete the standardized construction of the dataset, perform teacher network pre-training, optimize the training of the lightweight network through the joint loss function, and complete the model format conversion and embedded platform deployment.

2. The radar identification method for low-speed, small unmanned aerial vehicles according to claim 1, characterized in that, In the radar echo preprocessing stage, pulse compression processing is performed on the original I / Q sampling data of the radar to generate a one-dimensional range image. Select 8 to 16 pulses to form a single coherent processing interval; Phase alignment processing is performed on the target echoes within the coherent processing interval to complete motion compensation; Fast Fourier Transform is performed on the echo sequence of each range cell to generate a two-dimensional range Doppler map.

3. The radar identification method for low-speed, small unmanned aerial vehicles according to claim 1, characterized in that, In the micro-Doppler feature extraction stage, constant false alarm rate (CFAR) detection is performed on the range-Doppler image to screen candidate target points with a signal-to-noise ratio (SNR) greater than 8 dB, and the corresponding range cells of the targets are locked. The complex baseband time series corresponding to the coherent processing interval within the range cell corresponding to the target is extracted. A short-time Fourier transform (SFT) is performed on the complex baseband time series to generate a single-channel micro-Doppler spectrum with a size of 64×64. The calculation expression for the SFT is as follows: ,in, Given a complex baseband time series as input. This is a sliding window function with a window size of 256 points. For the Doppler frequency dimension, In terms of time dimension, For the output microDoppler spectrum, the overlap rate of the window function is set to 75%.

4. The radar identification method for low-speed, small unmanned aerial vehicles according to claim 1, characterized in that, In the classification reasoning stage, the input micro-Doppler spectrum is linearly normalized to map the values ​​to the interval between 0 and 1. The normalized micro-Doppler spectrum is then input into a lightweight CA-DSNet network, which sequentially passes through a standard convolutional initialization layer, three sets of depthwise separable convolutional modules, three sets of SE channel attention modules, a global average pooling unit, a Dropout regularization unit, and a fully connected layer. The fully connected layer outputs the probability values ​​corresponding to four target categories: quadcopter drones, fixed-wing drones, birds, and kites.

5. A radar identification method for low-speed, small unmanned aerial vehicles (UAVs) according to claim 4, characterized in that, The standard convolutional initial layer of the lightweight CA-DSNet network has 32 output channels, a kernel size of 3×3, and a stride of 2. The three sets of depthwise separable convolutional modules are connected in sequence, with 64, 128, and 256 output channels respectively. Each set of depthwise separable convolutional modules consists of depthwise convolution and pointwise convolution. The depthwise convolution performs convolution operations independently on each input channel, while the pointwise convolution completes feature fusion and transformation of the channel dimension through 1×1 convolution.

6. The radar identification method for low-speed, small unmanned aerial vehicles according to claim 4, characterized in that, The three SE channel attention modules correspond one-to-one with the three depthwise separable convolutional modules, and are embedded in the output of each depthwise separable convolutional module. The channel weight calculation expression for the SE channel attention module is as follows: ,in, For the first The weight values ​​corresponding to each feature channel It is the Sigmoid activation function. It is the ReLU activation function. This is the weight matrix of the first fully connected layer, with a dimensionality reduction ratio set to 8. This is the weight matrix of the second fully connected layer. The channel descriptor is obtained by global average pooling for the c-th feature channel. The weight of each channel is multiplied by the corresponding feature channel one by one to complete the adaptive weighting of the feature map.

7. The radar identification method for low-speed, small unmanned aerial vehicles according to claim 1, characterized in that, During the model training and deployment phase, the collected radar echo data is labeled and divided to generate standardized training, validation, and test sets. A ResNet-18 network is selected as the teacher network to complete pre-training and generate high-precision soft labels. A lightweight CA-DSNet network is used as the student network, and the network parameters are iteratively optimized through a joint loss function. The trained model is converted to ONNX format to complete operator adaptation and deployment on the embedded platform.

8. A radar identification method for low-speed, small unmanned aerial vehicles according to claim 7, characterized in that, The calculation expression for the joint loss function is as follows: ,in, This represents the total loss value during the training process. These are the weighting coefficients for the cross-entropy loss. These are the weighting coefficients for the KL divergence loss. The cross-entropy loss between the student's network output and the real label is used. The KL divergence loss is used to compare the student network output with the teacher network soft labels. The training process iteratively updates the student network parameters by minimizing the total loss value.

Images