A deep learning-based low signal-to-noise ratio multi-rotor unmanned aerial vehicle radar signal classification method and system

Abstract

A low signal-to-noise ratio multi-rotor unmanned aerial vehicle radar signal classification method and system based on deep learning belong to the field of complex domain signal classification. The problem that current methods lack effective classification performance in low signal-to-noise ratio or complex environment is solved. The method comprises the following steps: arranging radar echo data into a two-channel one-dimensional time sequence signal, the first channel being real part data and the second channel being imaginary part data, forming an input tensor, inputting into a C-ELTSNet network, containing four residual shrinkage modules and four fully connected layers, each module performing: twice one-dimensional convolution operation, followed by batch normalization and LeakyReLU activation after convolution to obtain a feature map; noise filtering is performed through an MCAT module; the filtered features are added to the input features through residual connection, and are output to four fully connected layers for feature compression and class mapping, and the type of unmanned aerial vehicle is output according to the C-ELTSNet network. It is used in the field of unmanned aerial vehicle classification.

Description

Technical Field

[0001] This invention belongs to the field of complex domain signal classification, and in particular relates to a method for classifying radar signals of low signal-to-noise ratio multi-rotor UAVs based on deep learning. Background Technology

[0002] Current drone detection technologies are mostly based on image and radio frequency (RF) signals. Image signals remain relatively efficient for short-range drone detection, but their accuracy drops sharply under special weather conditions such as fog, clouds, and rain. With advancements in communication technology, RF signals are increasingly susceptible to external interference and deception by spoofing signals. Radar signals, compared to other signals, have significant advantages in drone detection and identification. They are capable of operating day and night, are less affected by weather and other environmental factors, have long detection ranges, and are rich in target information such as range, micro-Doppler, and spatial position. However, due to the small size of drones, their highly random flight paths, and their ability to hover, drone classification based on radar cross-section (RCS), range-Doppler information (RD information), and flight path information has limitations and unreliability. Nevertheless, different types of drones have different and relatively fixed physical structures (propeller length, propeller rotation frequency, propeller spacing, etc.) and self-motion characteristics. These different physical structures generate unique micro-Doppler characteristics during motion; therefore, the micro-Doppler characteristics of radar signals can serve as a reliable basis for distinguishing different types of drones.

[0003] Currently, various deep learning techniques have been successfully applied to the field of radar detection for unmanned aerial vehicles (UAVs). Han et al. improved the feature extraction capability of the network by using a dual-stream convolutional transformer (DCT) model combined with self-attention and convolution. However, the signal-to-noise ratio (SNR) of the signal in this paper was only 30 dB, and the detection performance in complex environments or under low SNR conditions was not considered. To improve the image quality of micro-Doppler features of radar signals in the preprocessing stage, Park et al. proposed a new micro-Doppler (MDS) image generation technique, and used a lightweight CNN network for accurate classification of small UAVs at long distances. However, this study also improved the clarity of micro-Doppler features by increasing the radar's operating frequency, and the SNR was also at a high level. Sanjoy Basak et al. first used Anderson-Darling (AD) GoF to check for the presence of signals in the signal, and then used an improved YoLo framework to classify the signals in the STFT image based on the detection results. However, the above methods also require a high SNR and depend on high image clarity.

[0004] Regarding feature extraction and training strategy optimization, Priti Mandal et al. effectively improved the classification performance of flying targets by combining the feature extraction capabilities of convolutional neural networks with the powerful training strategy optimization capabilities of the Shuffled Frog Leap (CNN-SFL) algorithm. LV YE et al. designed a two-dimensional deep fusion network based on raw radar cross section (RCS) data and 2D Grammy angle field (GAF) data of UAVs. They used two ResNets to extract different forms of data features and fused them to classify UAVs. To enable the network to focus on effective information, Zheng Si et al. utilized the unique spatiotemporal dimension of SNNs and constructed a multi-dimensional attention module, proposing a spiking neural network algorithm based on multi-attention fusion. Harish Chandra Kumawat et al. designed and implemented a novel lightweight DCNN model using their proposed deep convolutional module, which uses fewer parameters for deep feature extraction and classification. To reduce false positive rates, Wang et al. designed a dual-head CNN architecture, with one CNN head determining target presence and the other regressing target offset. They investigated a non-maximum suppression (NMS) mechanism comprised of probability-based initial recognition, density-based recognition, and voting-based regression. B. Saha et al. proposed an LSTM-256 network architecture based on LSTM and a two-layer dense structure, achieving an optimal balance between classification performance and computational cost. AHMED N. SAYED et al. proposed a machine learning algorithm based on mechanical control information (MCML), incorporating mechanical control information into UAV classification, significantly improving the classification performance of various machine learning algorithms. Neda Rojhani et al. proposed a random signal augmentation algorithm based on a scene deterministic augmentation model to reduce the risk of overfitting in CNN networks and preserve important physical features of the data, applying the augmented data to the training of CNN networks.

[0005] In summary, although existing research has made some progress in the field of UAV classification, many current methods only consider the signal classification problem of UAVs under simple backgrounds and high signal-to-noise ratios. However, in practical applications, there are problems such as complex electromagnetic backgrounds, low signal discernibility, and various interferences. Therefore, most current research results lack effective classification performance in low signal-to-noise ratio or complex environments. Summary of the Invention

[0006] In view of this, the present invention aims to propose a method and system for classifying radar signals of low signal-to-noise ratio multi-rotor UAVs based on deep learning, so as to solve the problem that current methods lack effective classification performance in low signal-to-noise ratio or complex environments.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A deep learning-based method for classifying radar signals of low signal-to-noise ratio multi-rotor UAVs, the method comprising:

[0009] Acquire radar echo data and organize it into a dual-channel one-dimensional time-series signal, where the first channel is the real part data and the second channel is the imaginary part data, forming an input tensor;

[0010] The input tensor is fed into the C-ELTSNet network, which sequentially comprises:

[0011] Four residual shrinking modules and four fully connected layers, each module is executed sequentially:

[0012] The input features are subjected to two one-dimensional convolution operations, followed by batch normalization and LeakyReLU activation after each convolution to obtain the feature map. x 2;

[0013] Noise filtering is performed on the feature map using a multi-channel learnable adaptive thresholding module;

[0014] The filtered features are added to the input features through residual connections, activated by LeakyReLU, and then output to four fully connected layers. These four fully connected layers are used for feature compression and class mapping.

[0015] The drone type is output based on the C-ELTSNet network.

[0016] Furthermore, a preferred method is proposed, wherein the noise filtering of the feature map through a multi-channel learnable adaptive threshold module includes:

[0017] Generation and Feature Map x 2. Adaptive threshold map of the same dimension V After being compressed by two layers of convolution, the gated vector is output through the Sigmoid function. y ;

[0018] Based on the gate vector y Perform element-wise noise filtering:

[0019]

[0020] in, The output features after soft thresholding. For the input feature map, the first The first channel, the first Feature elements at each time step This is the gating threshold.

[0021] Furthermore, a preferred method is proposed, wherein the gating vector generation method is as follows:

[0022]

[0023] in, This represents the LeakyReLU activation function. This represents the Sigmoid function. and These are the weight parameters for adaptive learnable neurons.

[0024] Furthermore, a preferred embodiment is proposed, wherein the residual connection of the residual shrinkage module adopts the following method:

[0025] When the number of input and output channels is inconsistent, a 1×1 convolutional layer is used to transform the input features through channels, and batch normalization is used to adjust the feature distribution.

[0026] When the number of channels is the same, perform identity mapping directly.

[0027] Furthermore, a preferred method is proposed, in which the input features are subjected to two one-dimensional convolution operations, followed by batch normalization and LeakyReLU activation after each convolution to obtain the feature map. x 2, including:

[0028]

[0029] in, x 1 represents the feature vector group.

[0030] Furthermore, a preferred method is proposed, wherein the radar echo data is generated based on the Martin-Mulgrew physical model, with a signal-to-noise ratio ranging from -15dB to 15dB, and includes micro-Doppler characteristic signals of at least 5 types of UAVs.

[0031] Based on the same inventive concept, this invention also proposes a low signal-to-noise ratio multi-rotor UAV radar signal classification system based on deep learning, the system comprising:

[0032] The data processing module is used to acquire radar echo data and organize the radar echo data into a dual-channel one-dimensional time-series signal, where the first channel is the real part data and the second channel is the imaginary part data, forming an input tensor.

[0033] The C-ELTSNet classification module is used to input the input tensor into the C-ELTSNet network. The network sequentially comprises four residual shrinking modules and four fully connected layers. Each module sequentially performs the following operations: two one-dimensional convolution operations on the input features, followed by batch normalization and Leaky ReLU activation after each convolution to obtain the feature map.x 2; Noise filtering is performed on the feature map through a multi-channel learnable adaptive threshold module; The filtered features are added to the input features through residual connections, activated by LeakyReLU, and output to four fully connected layers for feature compression and class mapping.

[0034] The output module is used to output the drone type based on the C-ELTSNet network.

[0035] Based on the same inventive concept, the present invention also proposes a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes a deep learning-based low signal-to-noise ratio multi-rotor UAV radar signal classification method according to any one of the above claims.

[0036] Based on the same inventive concept, the present invention also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of a deep learning-based low signal-to-noise ratio multi-rotor UAV radar signal classification method as described in any one of the above descriptions.

[0037] Compared with the prior art, the beneficial effects of the present invention are:

[0038] The method proposed in this invention, through the synergistic effect of a complex domain dual-channel input structure (real / imaginary part processing) and a multi-channel learnable adaptive threshold module (MCAT), improves the UAV classification accuracy from 45.8% (DenseNet201) to 49.7% under extreme conditions with a signal-to-noise ratio as low as -15dB. Especially in the low signal-to-noise ratio range of -15dB to -5dB, the average accuracy improvement is ≥3.9%, solving the classification failure problem caused by signal distortion in existing technologies.

[0039] The proposed method utilizes the MCAT module to generate a dual-dimensional adaptive threshold map (instead of a global or channel-level threshold) through end-to-end learning, achieving for the first time a refined modeling of noise distribution. Ablation experiments demonstrate that this module delivers a 5.3% improvement in absolute accuracy at -15dB, significantly outperforming traditional fixed threshold or attention mechanisms, thus overcoming the robustness bottleneck of deep learning models in complex noisy environments.

[0040] The method proposed in this invention maintains high accuracy (average OA 82.5%) while compressing the number of model parameters to 8.48M, with a single-sample inference time of ≤1.183ms, and an overall computational cost of only 40% of ResNet34. This characteristic of "higher accuracy and lower complexity" violates the traditional deep learning model's accuracy-complexity proportionality law, providing a technical foundation for deployment in embedded devices.

[0041] The method proposed in this invention abandons the preprocessing step of converting complex signals to time-frequency images in existing technologies, and directly performs dual-channel modeling on one-dimensional complex sequences, avoiding phase information loss caused by transformations such as STFT. Experiments have shown that this design enables the DJI Matrice 300 RTK to achieve a classification accuracy of 99.8% at a signal-to-noise ratio of 15dB, significantly improving the sensitivity to micro-motion features.

[0042] The method proposed in this invention reveals for the first time through systematic experiments that a four-layer residual contraction structure is the optimal configuration: fewer than four layers result in insufficient feature extraction, while more than four layers lead to a decrease in high signal-to-noise ratio performance due to excessive filtering (five layers reduce performance by 2.8% compared to four layers at 15dB). This discovery overturns the traditional understanding that "deeper networks perform better" and establishes a new lightweight architecture. Attached Figure Description

[0043] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0044] Figure 1 This is a flowchart illustrating a low signal-to-noise ratio (SNR) multi-rotor UAV radar signal classification method based on deep learning, as described in this invention.

[0045] Figure 2 This is a framework diagram of a low signal-to-noise ratio multi-rotor UAV radar signal classification technology based on deep learning, as described in this invention.

[0046] Figure 3 This is a schematic diagram of the signal simulation results described in this invention;

[0047] Figure 4 This diagram illustrates the classification accuracy of C-ELTSNet for different types of drones under different signal-to-noise ratios as described in this invention.

[0048] Figure 5 This is a schematic diagram showing the comparison results of the model described in this invention with other existing models in terms of Mean OA, Macro-F1, and Kappa.

[0049] Figure 6 This is a schematic diagram of the C-ELTSNet confusion matrix described in this invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other, and the described embodiments are only some embodiments of the present invention, not all embodiments.

[0051] Implementation Method 1: A low signal-to-noise ratio (SNR) multi-rotor UAV radar signal classification method based on deep learning, the method comprising:

[0052] Acquire radar echo data and organize it into a dual-channel one-dimensional time-series signal, where the first channel is the real part data and the second channel is the imaginary part data, forming an input tensor;

[0053] The input tensor is fed into the C-ELTSNet network, which sequentially comprises:

[0054] Four residual shrinking modules and four fully connected layers, each module is executed sequentially:

[0055] The input features are subjected to two one-dimensional convolution operations, followed by batch normalization and LeakyReLU activation after each convolution to obtain the feature map. x 2;

[0056] Noise filtering is performed on the feature map using a multi-channel learnable adaptive thresholding module;

[0057] The filtered features are added to the input features through residual connections, activated by LeakyReLU, and then output to four fully connected layers. These four fully connected layers are used for feature compression and class mapping.

[0058] The drone type is output based on the C-ELTSNet network.

[0059] This implementation proposes the C-ELTSNet model for UAV classification tasks under low signal-to-noise ratio (SNR) conditions. This model presents new challenges to the accuracy and complexity of current deep learning-based UAV classification systems. The model employs a multi-channel learnable adaptive threshold module proposed in this invention. By learning noise patterns at each position of each channel, an adaptive threshold is formed, thereby achieving refined noise reduction and feature extraction. Simultaneously, the model introduces an adaptive residual contraction technique, using internal convolution and threshold filtering to perform residual connections, extracting clearer and deeper features. Therefore, this network can effectively extract and aggregate clear signal features, significantly improving classification accuracy with lower computational cost.

[0060] Implementation Method Two: This implementation method further defines the low signal-to-noise ratio (SNR) multi-rotor UAV radar signal classification method based on deep learning described in Implementation Method One. The step of filtering noise from the feature map using a multi-channel learnable adaptive threshold module includes:

[0061] Generation and Feature Map x 2. Adaptive threshold map of the same dimension V After being compressed by two layers of convolution, the gated vector is output through the Sigmoid function. y ;

[0062] Element-wise noise filtering is performed based on the gate vector y:

[0063]

[0064] in, The output features after soft thresholding. For the input feature map, the first The first channel, the first Feature elements at each time step This is the gating threshold.

[0065] Implementation Method 3: This implementation method further defines the deep learning-based low signal-to-noise ratio multi-rotor UAV radar signal classification method described in Implementation Method 2. The gating vector generation method is as follows:

[0066]

[0067] in, This represents the LeakyReLU activation function. This represents the Sigmoid function. and These are the weight parameters for adaptive learnable neurons.

[0068] Implementation Method Four: This implementation method further defines the low signal-to-noise ratio (SNR) multi-rotor UAV radar signal classification method based on deep learning described in Implementation Method One. The residual connection of the residual shrinking module adopts the following method:

[0069] When the number of input and output channels is inconsistent, a 1×1 convolutional layer is used to transform the input features through channels, and batch normalization is used to adjust the feature distribution.

[0070] When the number of channels is the same, perform identity mapping directly.

[0071] Implementation Method 5: This implementation method further defines the low signal-to-noise ratio (SNR) multi-rotor UAV radar signal classification method based on deep learning described in Implementation Method 1. The input features are subjected to two one-dimensional convolution operations, followed by batch normalization and LeakyReLU activation after each convolution to obtain the feature map. x 2, including:

[0072]

[0073] in, x 1 represents the feature vector group.

[0074] Implementation Method Six: This implementation method further defines the radar signal classification method for low signal-to-noise ratio multi-rotor UAVs based on deep learning described in Implementation Method One. The radar echo data is generated based on the Martin-Mulgrew physical model, with a signal-to-noise ratio range of -15dB to 15dB, and includes micro-Doppler characteristic signals of at least 5 types of UAVs.

[0075] Implementation Method Seven: A low signal-to-noise ratio multi-rotor UAV radar signal classification system based on deep learning, the system comprising:

[0076] The data processing module is used to acquire radar echo data and organize the radar echo data into a dual-channel one-dimensional time-series signal, where the first channel is the real part data and the second channel is the imaginary part data, forming an input tensor.

[0077] The C-ELTSNet classification module is used to input the input tensor into the C-ELTSNet network. The network sequentially includes four residual shrinking modules. Each module sequentially performs two one-dimensional convolution operations on the input features, followed by batch normalization and LeakyReLU activation after each convolution to obtain the feature map. x 2; Noise filtering is performed on the feature map through a multi-channel learnable adaptive thresholding module; the filtered features are added to the input features through residual connections, and the output is activated by LeakyReLU;

[0078] The output module is used to output the drone type based on the C-ELTSNet network.

[0079] Implementation Method 8: A computer device according to this implementation method includes a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes a deep learning-based low signal-to-noise ratio multi-rotor UAV radar signal classification method according to any one of Implementation Methods 1 to 6.

[0080] Implementation Method Nine: A computer-readable storage medium according to this implementation method stores a computer program, which, when executed by a processor, performs the steps of a deep learning-based low signal-to-noise ratio multi-rotor UAV radar signal classification method as described in any one of Implementation Methods One to Six.

[0081] Implementation Method 10, see below Figures 1 to 6 This embodiment describes a specific example of the deep learning-based low signal-to-noise ratio multi-rotor UAV radar signal classification method described in Embodiment 1. It also serves to explain Embodiments 2 through 6. Specifically:

[0082] Step 1: Organize the radar echo data into a dual-channel format, with one channel containing real data and the other channel containing imaginary data.

[0083] Step 2: Input the organized radar data into the established network.

[0084] Specifically, this implementation addresses the problem of one-dimensional radar echo signal recognition by designing a deep network structure, C-ELTSNet, based on residual shrinkage and multi-channel learnable adaptive thresholds. This network effectively suppresses noise and captures key features while maintaining a low number of model parameters, thereby improving UAV classification performance under low signal-to-noise ratio conditions.

[0085] To preserve the original complex signal completely and avoid information loss during data conversion, this implementation uses dual-channel one-dimensional radar echo data as input, with one channel representing the real part and the other the imaginary part. Research indicates that this approach allows the network to directly perceive amplitude information and capture phase relationships crucial for classification. Furthermore, the separation of real and imaginary inputs helps the network automatically discover complementary patterns, improving robustness to weak targets and complex interference, thereby achieving higher classification accuracy.

[0086] The main body of the C-ELTSNet network consists of four residual shrinking modules, multiple downsampling layers, and a series of fully connected classification layers. The input radar one-dimensional signal batch X∈R^(B×C×L), where B is the batch size, C is the number of channels, and L is the data length. First, the signal passes through the first residual shrinking module, which internally consists of a residual layer, a convolutional normalization module, and a multi-channel adaptive learnable thresholding module. The signal first passes through the main branch, where temporal features are extracted through two "convolution-BatchNorm-LeakyReLU" combinations. Then, the multi-channel adaptive thresholding module learns the gating coefficients at each position in each channel and shrinks low-amplitude noise. Simultaneously, the residual branch uses 1×1 convolution and BatchNorm to adjust the number of channels. The two branches are summed element-wise at the end, and then activated by LeakyReLU once, outputting a feature tensor of shape R^(B×C×L). Next, the feature tensor is downsampled by half in the time dimension using a max-pooling layer, reducing computation and parameter count, expanding the receptive field of the convolutional layers, and enhancing the model's focus on important parts of the signal. Simultaneously, a Dropout layer inserted after each pooling step suppresses overfitting by randomly discarding some features. After four such processing steps with differentiated parameter designs for each step, the network successfully extracts the feature information of the radar signal. The extracted feature information is then passed through a three-layer fully connected network and further refined into a higher-order representation using LeakyReLU activation. Finally, a linear layer completes the classification and outputs the result.

[0087] The signal processing in this embodiment combines deep feature extraction, adaptive noise suppression, and identity residual path to enhance useful signals while maintaining smooth gradients, thereby improving the network's ability to discriminate radar echoes and its training stability. The construction ideas for each key module will be described in detail below.

[0088] The residual shrinkage module is the main component of this invention. Each module consists of two one-dimensional convolutional layers (Conv 1D), a corresponding batch normalization layer (BatchNorm 1D), and an activation function (LeakyReLU). An additional multi-channel learnable adaptive thresholding module is introduced at the output for feature sparsity and noise suppression. The data processing flow in this module specifically includes the following steps:

[0089] To effectively extract locally representative features from the input multi-channel time-series signal, a one-dimensional convolution (Conv1d) operation is first performed on the input features. Specifically, this module uses a convolution kernel with a kernel size of kernel_size to perform sliding convolution on the input features to capture local feature information and short-term dependencies. Simultaneously, to improve the training stability and convergence speed of the network, batch normalization (BN) is introduced immediately after the convolutional layer. Batch normalization standardizes the feature distribution of each channel in each batch to a similar scale, effectively alleviating the gradient vanishing and gradient exploding problems in deep network training, making the model easier to optimize and generalize. Subsequently, the normalized output is activated by the LeakyReLU function to further enhance the model's non-linear expressive power and avoid the neuron dying problem. The LeakyReLU function is defined as follows:

[0090]

[0091] in, It is a small positive coefficient (usually 0.01 or 0.1) to ensure that the neuron retains a certain level of activity when the input is negative, thus avoiding the information loss problem caused by the traditional ReLU activation function outputting zero when the input is negative.

[0092] In summary, the convolutional feature extraction module proposed in this embodiment can be specifically described as follows:

[0093]

[0094] Then, repeat the same operation process, that is, for The same convolution, batch normalization, and LeakyReLU activation operations are applied again to further extract higher-level local feature information. The final output is denoted as... :

[0095]

[0096] Through the two consecutive convolution and nonlinear transformation operations described above, the rich local patterns and hidden features in time series data are deeply mined and expressed, thereby improving the discriminative ability of feature representation and helping the network achieve more accurate signal classification and recognition results.

[0097] After performing the convolution and normalization operations described above, the feature map typically contains both valid signal and noise components. To further improve the robustness and purity of feature representation, this implementation also proposes and employs a multi-channel learnable adaptive thresholding module (MCAT). The core idea of ​​MCAT is to achieve more refined and adaptive denoising and feature selection on the input features by learning data-driven channel and position-dependent adaptive thresholds.

[0098] Specifically, this module operates on the output feature map of the second convolutional layer. Using this as input, a refined adaptive threshold map is independently learned in both channel and time dimensions through an adaptive neuron population. Subsequently, the threshold map After extracting global information, feature mapping is first performed using two consecutive convolutional layers (neither containing bias terms), ultimately yielding a gating vector consistent with the number of input feature channels. The specific mapping operation formula is defined as follows:

[0099]

[0100] in, This represents the LeakyReLU activation function. This represents the Sigmoid function. and These are the weight parameters of the adaptive learnable neuron (the size of which is determined by the scaling factor reduction), which are automatically optimized through a data-driven approach.

[0101] The gating vector obtained above This can be viewed as an adaptive control threshold for the feature elements at each time step in each channel, thus playing a role in dynamic adjustment and effective noise reduction on the feature map. To achieve precise noise removal, this implementation proposes a "precise soft-thresholding" operation, which applies a specific threshold to the input feature map. x In step 2, each element is executed sequentially, and the specific operation is defined as follows:

[0102]

[0103] in, The output features after soft thresholding. For the input feature map, the first The first channel, the first Feature elements at each time step The threshold value mentioned above is used for gating.

[0104] When input features The amplitude is lower than the gating threshold at the corresponding position. When the input feature amplitude is above the gating threshold, the output feature will be set to zero; when the input feature amplitude is above the gating threshold, it will be moderately preserved and enhanced. This precise adaptive thresholding operation can effectively reduce the interference of low-amplitude noise components on the model, while preserving the practically meaningful high-amplitude effective information, thereby improving the signal-to-noise ratio of the feature map.

[0105] By introducing and optimizing the adaptive threshold module described above, the network structure proposed in this embodiment significantly improves the feature representation ability and generalization performance in complex noisy environments.

[0106] In deep networks, the vanishing gradient problem often adversely affects the training process as the number of layers increases, thus limiting the model's representational power and convergence speed. To address this issue, this invention introduces a residual connection (Shortcut Connection) in each residual shrinking module. This aims to effectively alleviate the vanishing gradient phenomenon and enhance the expressive power of features by directly passing input features to subsequent layers.

[0107] Specifically, in the design of residual connections, if the input features within the module and the output features after adaptive thresholding (denoted as...) If there is a mismatch in the number of channels, a 1×1 convolutional layer is used to transform the channels of the input features, and batch normalization is used to adjust the feature distribution to be consistent with the output. Conversely, when the number of channels in the input and output are the same, identity mapping is performed directly. This flexible design ensures that skip connections can be successfully built regardless of whether the number of input and output channels are matched, ensuring lossless transfer of feature information.

[0108] Finally, the input features after channel adjustment are added element-wise to the output features after adaptive thresholding, as shown in the following formula:

[0109]

[0110] in, This represents the input features after necessary channel transformation and normalization operations.

[0111] Next, the features after addition will be... The residual module's final output is obtained by passing it through the LeakyReLU activation function again. This residual connection design not only better facilitates gradient propagation during backpropagation but also makes the network optimization process smoother, easier to converge, and preserves key information from the original input.

[0112] The MultiChannel Adaptive Threshold (MCAT) module proposed in this embodiment is one of the core components of the C-ELTSNet network structure. It aims to finely and adaptively adjust the threshold in the feature map through a data-driven approach. Compared to traditional fixed or global adaptive thresholding methods, this module can more efficiently learn independent and adaptive thresholds at various locations of multi-channel signal features, thereby enhancing the model's robustness and generalization ability in complex noisy environments.

[0113] The core component of the MCAT module is a learnable population of neurons, consisting of convolutional layers, activation functions, and normalization layers. Specifically, the module receives a three-dimensional tensor of shape (B, C, L) as input. The neuron population is designed to have the same number of output channels as the input channels, thus directly outputting a three-dimensional adaptive threshold map with the same shape (B, C, L).

[0114] After the threshold map is generated, the module constrains it using the Sigmoid activation function to ensure that the adaptively generated threshold always remains within the range of [0,1]. This design not only prevents computational instability caused by threshold values ​​that are too large or too small, but also ensures stable performance of the module during training and inference. The Sigmoid function formula is shown below:

[0115]

[0116] in, Indicates the input signal. It is the natural constant (Euler constant, approximately 2.71828).

[0117] After generating the adaptive threshold, the module then applies soft thresholding to the input signal. In this embodiment, the soft thresholding function formula is as follows:

[0118]

[0119] in, It is the input characteristic signal. It is an adaptive threshold dynamically generated by the aforementioned learnable neuron population.

[0120] Specifically, the above formula gradually suppresses features with signal strength below a threshold to zero, thereby achieving effective noise removal; at the same time, it retains important information in the signal that is above the threshold. This soft-threshold-based design can remove noise while preserving the continuity and richness of the signal, which is more in line with the needs of signal processing in real-world scenarios.

[0121] Compared to traditional methods that use global thresholding or global average pooling to generate a uniform single threshold for each channel, the MCAT module has significant advantages in its mechanism. Traditional global methods ignore the differences in features at different locations within the same channel in noisy environments, easily leading to over-suppression or information loss at certain locations. The MCAT module, however, allows each feature location in each channel to adaptively obtain its unique threshold, thus enabling more precise adjustment of signal processing based on data features in both spatial location and channel dimension. This more flexible and nuanced feature selection method gives the model stronger adaptability and better feature generalization ability.

[0122] Furthermore, the threshold learning process is fully embedded in the end-to-end training framework of the model, enabling the network to automatically find the optimal threshold configuration through supervised learning optimization objectives, without the need for manual parameter tuning. This not only improves the automation level of signal processing but also significantly reduces the difficulty of model tuning, making this module more practical in real-world applications.

[0123] Based on the above design, the Multi-Channel Learnable Adaptive Threshold Module (MCAT) can achieve efficient adaptive adjustment of signal features in complex noisy environments through a refined, position-sensitive, and channel-sensitive threshold learning method. It can play a role in improving model robustness, enhancing feature selection ability, and improving generalization performance.

[0124] To verify the effectiveness of this method, a simulation experiment was also conducted in this implementation. The machine configuration used in this experiment was as follows: AMD Ryzen 9 7950X 16-core CPU, 64GB RAM, 1TB + 512GB hard drive, equipped with NVIDIA RTX 4090 graphics card (24GB VRAM); the operating system was Ubuntu 20.04, Python version was 3.11, the deep learning framework was PyTorch 2.3.0, and CUDA version was 12.2.

[0125] In this embodiment, a UAV micro-Doppler feature dataset was first generated based on the Martin-Mulgrew physical model. The dataset contains radar echo data at seven different signal-to-noise ratios (SNRs): -15dB, -10dB, -5dB, 0dB, 5dB, 10dB, and 15dB. For each SNR condition, the dataset includes radar echo data from five different types of UAVs: DJI Mavic Air 2, DJI Mavic Mini, DJI Matrice 300 RTK, DJI Phantom 4, and ParrotDisco. 1000 data points were simulated for each UAV, with each data point containing 2000 complex-valued sampling points, totaling 5000 data points across the five UAVs and 35000 data points across the seven SNRs. The composition of the dataset is shown in Table 1, and the signal simulation results are as follows: Figure 3 As shown.

[0126] For carrier-related parameters, this dataset selects X-band radar, which combines high penetration and high Doppler resolution in everyday scenarios, with wavelengths... It is 2.998cm, and the frequency is... The GHz frequency is 10 GHz, and the sampling frequency is set to 10 kHz.

[0127] For the structural parameters of the UAV, the simulation process in this dataset is based on real data from five UAVs, with the number of propellers (N) and the distance between the propeller root and the center of rotation fixed. ), the distance between the propeller tip and the center of rotation ( ), propeller rotation frequency ( The four parameters are shown in Table 2.

[0128] Table 1 Dataset Composition

[0129]

[0130] The main spatial position parameters of radar include echo amplitude ( ), distance between drones and radar ( ), radial velocity relative to the radar rotation center ( ), the angle between the radar and the drone propeller plane ( ) and propeller blade spacing ( In this dataset, the parameters in this part are random values ​​to simulate the radar echo of the UAV under different motion states. The value range and random rules of the relevant parameters are shown in Table 3.

[0131] Table 2 Relevant parameters for various types of UAVs

[0132]

[0133] Table 3. Relative position parameters of UAV radar

[0134]

[0135] The main evaluation metrics used in this experiment are: accuracy, overall recognition accuracy, Macro-F1, and Kappa, and their calculation methods are as follows:

[0136]

[0137] in, It is a category The number of samples correctly predicted in the middle. It is a category The number of samples that were incorrectly predicted.

[0138]

[0139] in It is the total number of categories. That is the total number of samples.

[0140]

[0141] in, , , It is a category The middle was incorrectly predicted as the category. Number of samples in other categories It is a category The actual category is... The number of samples that were incorrectly classified as other categories.

[0142]

[0143] in, .

[0144] This experiment compares C-ELTSNet with existing popular methods, including CNN-SFL, DenseNet201, MobileNetV2, SENet50, VGG16, HQNN, and ResNet34.

[0145] During training, all model weights are randomly initialized and updated end-to-end using backpropagation based on mini-batch processing. The batch size is set to 8, the learning rate to 0.0001, the loss function is cross-entropy loss, and the optimizer is AdamW. For the training strategy, early stopping is used in this implementation, with a minimum of 20 epochs, a maximum of 500 epochs, and a loss function threshold of 0.0001. Training stops when the number of training epochs exceeds the minimum and the loss value falls below the threshold. Furthermore, this implementation mixes data at seven signal-to-noise ratios (SNR): -15dB, -10dB, -5dB, 0dB, 5dB, 10dB, and 15dB for training, and evaluates the model separately at each SNR during testing.

[0146] The adaptive threshold module included in this invention works with the backbone network to optimize the threshold control vector, enabling it to gradually learn to adaptively sparsify features of different channels and amplitudes during training, thereby improving classification performance and enhancing the robustness of the network.

[0147] This experiment uses a four-layer backbone network built on the ResidualShrinkingBlock1D module, supplemented by max pooling for downsampling. The input is a two-dimensional tensor. (Batch size 8, number of channels 2, length 2000), corresponding to one-dimensional sequence data with two channels. Following the fourth layer, the output feature map is flattened in both channel and spatial dimensions, then processed through several fully connected layers (FC) for feature extraction and dimensionality reduction, finally mapping it to the classification space. Furthermore, several Dropout layers (probability 0.5) are inserted between the convolutional and fully connected layers to mitigate overfitting and improve the model's generalization ability. The activation function is uniformly LeakyReLU, whose negative half-axis has a certain slope, which can alleviate the dead neuron problem while maintaining non-linear expressive power.

[0148] C-ELTSNet's classification performance for different types of drones at different signal-to-noise ratios is as follows: Figure 4 As shown. By Figure 4 It can be seen that when the signal-to-noise ratio is 15dB, the classification accuracy of various types of drones is relatively high. Among them, the model's classification accuracy for DJI Matrice 300RTK is significantly higher than that for other drone types, reaching approximately 99.8%.

[0149] The comparative experimental results with existing popular methods are shown in Table 4. It can be observed that when the signal-to-noise ratio drops to -15 dB, C-ELTSNet's classification accuracy is at least 3.9% higher than other methods. Furthermore, across the entire signal-to-noise ratio range, C-ELTSNet's overall average recognition accuracy is at least 2% higher than other existing popular methods.

[0150] Table 4 Overall classification accuracy results

[0151]

[0152] Furthermore, this implementation also compares the proposed model with several existing popular models in terms of Mean OA, Macro-F1, and Kappa metrics. Figure 5 As can be seen, the model proposed in this invention outperforms existing state-of-the-art methods in terms of Mean OA, Macro-F1, and Kappa metrics. This indicates that it not only has high noise robustness (Mean OA) and classification accuracy (Macro-F1), but also high performance in classification consistency (Kappa).

[0153] This embodiment also plots the confusion matrix of C-ELTSNet for UAV classification at various signal-to-noise ratios, such as... Figure 6 As shown in the figure, the network has learned the essential characteristics of each drone quite well, within -15dB. Within a signal-to-noise ratio (SNR) range of 15 dB, the classification accuracy for the correct drone type was the highest. However, when the SNR dropped to -10 dB and -15 dB, the confusion matrix revealed that the model tended to misclassify the signal as a DJI Matrice 300 RTK.

[0154] The network structure proposed in this invention is built based on multiple slaughter and shrinkage modules. Therefore, in order to determine the optimal number of these modules, this experiment conducted key ablation studies, and the experimental results are shown in Table 5.

[0155] Table 5 Overall recognition accuracy of different numbers of residual shrinkage modules

[0156]

[0157] Table 5 shows that when the number of layers is less than or equal to 4, the classification accuracy of the UAV shows an increasing trend under various signal-to-noise ratio (SNR) conditions, with the increase being particularly significant at extremely low SNRs (-15dB and -10dB). Specifically, the accuracy of 4 layers is 0.108 higher than that of 1 layer at -15dB. As the SNR increases, the increase gradually decreases, reaching only 0.023 at 15dB. When the number of layers reaches 5, the recognition accuracy actually decreases compared to 4 layers. Observation reveals that when the number of layers continues to increase, the maximum fluctuation of the recognition accuracy under low SNR conditions (-15dB-0dB) is only 0.016, and the overall trend is a slight decrease. This is because the noise filtering effect reaches its optimal level with the increase of the number of layers. On the other hand, under high SNR conditions, the recognition effect decreases significantly. This is because as the number of layers increases, much useful information is filtered out by the threshold. In summary, Table 5 shows that the average overall recognition rate is highest when the number of layers is 4, regardless of the signal-to-noise ratio. Generally speaking, when the number of layers is 4 or less, the model's feature extraction capability continuously improves, and the recognition accuracy gradually increases; however, when the number of layers is greater than 4, the feature extraction capability reaches its limit, and further increasing the number of layers may lead to over-filtering of many feature information, resulting in varying degrees of decline in recognition performance.

[0158] Based on the above analysis, the network achieves the best recognition performance with 4 layers. Therefore, a 4-layer residual shrinkage module is a reasonable choice.

[0159] To verify the effectiveness of the multi-channel learnable adaptive threshold module proposed in this invention, an ablation experiment was also conducted on this part. The classification accuracy after removing the multi-channel learnable adaptive threshold module was compared with that before removal, and the experimental results are shown in Table 6.

[0160] Table 6. Experimental results of ablation using a multi-channel learnable adaptive threshold module.

[0161]

[0162] The ablation experiments show that adding the MCAT module improves the classification accuracy of the model to varying degrees across all signal-to-noise ratio (SNR) levels (from -15dB to 15dB). Specifically, the introduction of the MCAT block significantly improves performance in the low SNR region (e.g., -15dB, -10dB, -5dB). For example, at -15dB, the classification accuracy increased from 0.444 to 0.497, a 5.3 percentage point improvement; at -10dB and -5dB, the overall accuracy improved by 4.8 and 3.9 percentage points, respectively. This indicates that the MCAT module significantly enhances the model's robustness under low SNR conditions. In the medium-to-high SNR region (0dB to 15dB), the MCAT module also improves model performance, but the increase is relatively small. For example, at 15dB, the accuracy only improves by 0.3 percentage points. This indicates that the main advantage of the MCAT module lies in improving the model's signal representation ability in harsh environments (low SNR), thereby significantly enhancing the model's generalization and robustness under complex conditions. Looking at the overall performance across all SNR conditions, the average overall recognition accuracy of the model after using MCAT was 0.825, an increase of 2.5 percentage points. This demonstrates that the MCAT module can significantly improve classification performance across the entire SNR range.

[0163] To further evaluate the performance of the proposed network model, this section also tested the number of model parameters, computational complexity, and inference time. The experimental results are shown in Table 7.

[0164] Table 7 Experimental Results of Model Complexity and Inference Performance

[0165]

[0166] Based on the above experimental results, the C-ELTSNet proposed in this invention achieves the highest accuracy under low signal-to-noise ratio conditions while using the fewest parameters and computational overhead, thus achieving an optimal balance in performance across all aspects.

[0167] The specific embodiments of the present invention disclosed above are merely illustrative of the invention. These embodiments do not exhaustively describe all details, nor do they limit the invention to the specific embodiments described. Many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A method for classifying radar signals of low signal-to-noise ratio multi-rotor UAVs based on deep learning, characterized in that, The method includes: Acquire radar echo data and organize it into a dual-channel one-dimensional time-series signal, where the first channel is the real part data and the second channel is the imaginary part data, forming an input tensor; The input tensor is fed into the C-ELTSNet network, which sequentially comprises: Four residual shrinking modules and four fully connected layers, each module is executed sequentially: The input features are subjected to two one-dimensional convolution operations, followed by batch normalization and LeakyReLU activation after each convolution to obtain the feature map. x 2; Noise filtering is performed on the feature map using a multi-channel learnable adaptive thresholding module; The filtered features are added to the input features through residual connections, activated by LeakyReLU, and then output to four fully connected layers. These four fully connected layers are used for feature compression and class mapping. The drone type is output based on the C-ELTSNet network; The noise filtering of the feature map through a multi-channel learnable adaptive threshold module includes: Generation and Feature Map x 2. Adaptive threshold map of the same dimension V After being compressed by two layers of convolution, the gated vector is output through the Sigmoid function. y ; Based on the gate vector y Perform element-wise noise filtering: in, The output features after soft thresholding. For the input feature map, the first The first channel, the first Feature elements at each time step This is the gate threshold; The gating vector is generated in the following way: in, This represents the LeakyReLU activation function. This represents the Sigmoid function. and These are the weight parameters for adaptive learnable neurons; The residual connection of the residual shrinkage module adopts the following method: When the number of input and output channels is inconsistent, a 1×1 convolutional layer is used to transform the input features through channels, and batch normalization is used to adjust the feature distribution. When the number of channels is the same, perform identity mapping directly.

2. The method for classifying radar signals of low signal-to-noise ratio multi-rotor UAVs based on deep learning according to claim 1, characterized in that, The input features are subjected to two one-dimensional convolution operations, followed by batch normalization and Leaky ReLU activation after each convolution to obtain the feature map. x 2, include: in, x 1 represents the feature vector group.

3. The method for classifying radar signals of low signal-to-noise ratio multi-rotor UAVs based on deep learning according to claim 1, characterized in that, The radar echo data is generated based on the Martin-Mulgrew physical model, with a signal-to-noise ratio ranging from -15dB to 15dB, and includes micro-Doppler characteristic signals of at least five types of UAVs.

4. A radar signal classification system for low signal-to-noise ratio multi-rotor UAVs based on deep learning, characterized in that, The system is implemented based on the method of claim 1, and the system includes: The data processing module is used to acquire radar echo data and organize the radar echo data into a dual-channel one-dimensional time-series signal, where the first channel is the real part data and the second channel is the imaginary part data, forming an input tensor. The C-ELTSNet classification module is used to input the input tensor into the C-ELTSNet network. The network sequentially comprises four residual shrinking modules and four fully connected layers. Each module sequentially performs the following operations: two one-dimensional convolution operations on the input features, followed by batch normalization and Leaky ReLU activation after each convolution to obtain the feature map. x 2; Noise filtering is performed on the feature map through a multi-channel learnable adaptive threshold module; The filtered features are added to the input features through residual connections, activated by LeakyReLU, and output to four fully connected layers for feature compression and class mapping. The output module is used to output the drone type based on the C-ELTSNet network.

5. A computer device, characterized in that: It includes a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes a deep learning-based low signal-to-noise ratio multi-rotor UAV radar signal classification method according to any one of claims 1-3.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of a deep learning-based low signal-to-noise ratio multi-rotor UAV radar signal classification method as described in any one of claims 1-3.

Images