Unmanned aerial vehicle identification method based on ensemble learning and application

Through the integrated learning framework and the drone recognition method of multi-scale convolution and global local attention mechanism, the problem of poor drone recognition effect in the existing technology is solved, and the drone with high accuracy and stability is achieved Identification, suitable for complex environments.

CN119068899APending Publication Date: 2024-12-03SOUTHWEAT UNIV OF SCI & TECH
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202411138054.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-19
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The existing drone identification methods have poor recognition effects in low-altitude flight, occlusion, poor light and environmental interference, and there are problems of information loss and manual feature selection based on audio signal processing.

Method used

Using an integrated learning-based drone recognition method, the original audio signal is directly processed through two first-level expert models (IMGLA-ModelA and IMGLA-ModelB) and a metaclassifier Stacking integrated learning framework, and uses multi-scale convolution and global local attention mechanism extract features, reduce manual intervention, and improve adaptability and accuracy.

Benefits of technology

It significantly improves the accuracy and stability of drone identification, can effectively identify drones in complex environments, reduce the misjudgment rate, meet real-time requirements, and reduce the impact on background noise.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119068899A_ABST
    Figure CN119068899A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of unmanned aerial vehicle identification, in particular to an unmanned aerial vehicle identification method based on ensemble learning and application, and a model provided by the invention is composed of two first-level expert models and a meta classifier. Firstly, two first-level expert models perform feature extraction on data, then an obtained classification result is input into a meta classifier, then the meta classifier performs integration fusion on results of the first-level models, and finally, unmanned aerial vehicle monitoring and recognition results are output. A multi-scale global local attention module is added to each of the two primary expert models on the basis of a residual structure and a depth separable convolution structure. The method is compared with other methods for processing one-dimensional signals on the basis of a self-made unmanned aerial vehicle data set. Experiments verify the effectiveness and superiority of the model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of drone identification, and in particular to a drone identification method based on ensemble learning and its application. Background Art

[0002] As a product of the deep integration of aviation technology and information technology, drones have developed rapidly in recent years and are widely used in logistics, agriculture and other fields. Their high flexibility and autonomous operation capabilities enable drones to play an important role in promoting new production models and service industries. However, the popularity of drones has also brought challenges in terms of privacy, security and regulations. There is an urgent need to find a balance between technological progress and social needs.

[0003] At present, the main methods for drone identification include radar-based methods, radio frequency-based methods, video-based methods, infrared-based methods, and acoustic signal-based methods. Martins et al. proposed a method based on radar cross section (RCS), which uses radar to provide reflectivity of the target to identify drones. When the signal-to-noise ratio is 10dB, the drone classification accuracy reaches 97.43%. Daiki et al. proposed a deep learning drone identification method based on radar images, and the five-category accuracy of drones reached 90%. However, the radar recognition effect on low-altitude drones is general, and drones may not be effectively detected. The radio frequency-based method can effectively detect the low-altitude flight signals of drones. Basak et al. proposed a method for detection and classification through frequency domain fingerprints, which can effectively handle possible radio frequency interference and has certain versatility and robustness. Phuc et al. proposed a method based on active tracking and passive monitoring of radio frequency signals to achieve drone classification and identification. When the radio frequency power emitted by the drone is low or the electromagnetic interference is strong, the recognition accuracy of the radio frequency method will be greatly affected. Michael et al. proposed a moving target detection and sensor motion compensation recognition algorithm, with a recognition accuracy of 93%. However, in the case of obstruction and poor lighting (such as at night), the optical video system may not be able to effectively capture the image features of the drone. Infrared video is easily interfered by other thermal radiation sources, resulting in an increased probability of false detection, especially at long distances. The above methods are easily affected by the environment or have their own limitations. The acoustic-based method can serve as a supplement to the drone identification method.

[0004] Maclean et al. trained CNN by converting the focus to spectrum and used it directly for soundscape prediction. The network reduces the size of the spectral feature vector through convolution, uses residual blocks to accelerate convergence and prevent gradient vanishing / explosion, and uses a deep structure to extract details in the spectrogram. The model accurately classifies bird calls.

[0005] Shi et al. proposed an acoustic signal feature detection method that combines multi-microphone detection with Bayesian filtering and Gaussian mixture models. It can effectively detect and locate drones. Brendan et al. proposed a two-dimensional analysis model that uses the harmonic properties of the acoustic signal generated by the propeller-driven aircraft to increase the detection distance to identify drones in advance and avoid them in time. Nie et al. proposed a drone detection and positioning system based on multi-dimensional signal features. The average detection rate is 95.58% when combining multiple features. Andrea et al. proposed a drone audio signal recognition method based on support vector machines, which can effectively identify drones and issue warnings.

[0006] At present, the methods for processing audio signals mainly include spectrum analysis, feature extraction, pattern recognition, deep learning, etc. The audio signal processing methods based on spectrum analysis, feature extraction and pattern recognition require manual feature selection, are highly dependent on prior knowledge, and have problems such as large data dimensions and insufficient generalization ability. The advantage of deep learning methods over the above methods is that they can automatically learn and extract effective features from data without too much manual intervention, and can better adapt to the complexity and diversity of data. Deep learning methods are usually able to process large-scale data and have strong generalization ability, so they can often achieve better performance when processing audio signals. The main methods based on deep learning include: Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Transformer, etc. When processing audio signals, networks such as RNN and Transformers may face problems such as sequence length limitations, insufficient information capture, too many parameters, the need for a large amount of labeled data, and overfitting. Ross et al. combined speech perception features with speech generation features (MFCC and LPC features) and learned a joint feature space through CNN to recognize the entire speech chain. This type of method combines signal processing (such as MFCC and STFT, wavelet transform, GFCC) and deep learning. However, the signal processing process also requires the problem of manual feature selection, which may cause information loss, including information compression, spatial information loss, and ignoring details. Although it is possible to extract features that are useful for the task, it is also necessary to find a balance between information retention and model performance.

[0007] In summary, the detection and identification of drones has become particularly important. Radar-based methods are not easy to monitor drones flying at low altitudes, and video-based methods have high requirements for imaging quality. However, drone detection methods based on acoustic signals can make up for the shortcomings of these traditional drone detection methods. Therefore, the present invention proposes a drone identification method and application based on ensemble learning, which accurately identifies drones through drone audio signals, aiming to supplement the shortcomings of other methods. Summary of the invention

[0008] The purpose of the present invention is to provide a method and application of drone identification based on ensemble learning. By adopting an ensemble learning framework, the original audio signal of the drone can be directly processed without manual feature extraction. Specifically, the proposed model consists of two first-level expert models and a meta-classifier. First, the two first-level expert models extract features from the data respectively, and then input the obtained classification results into the meta-classifier. Next, the meta-classifier integrates and fuses the results of the first-level models, and finally outputs the results of drone monitoring and identification. The two first-level expert models add multi-scale global-local attention modules based on the residual structure and the deep separable convolution structure respectively. The method of the present invention is compared with other methods for processing one-dimensional signals on a self-made drone dataset. The experiment verifies the effectiveness and superiority of the model.

[0009] In order to achieve the above technical objectives and the above technical effects, the present invention is implemented through the following technical solutions:

[0010] A drone identification method based on ensemble learning includes the following steps:

[0011] S1: Different types of drones are selected to construct a dataset, and the flight sounds of drones at different distances are recorded at a sampling rate of 44100 Hz. Background noises such as thunder and rain, wind, and bird calls are introduced to increase the complexity of the dataset, forming a "no drone" class sample, and the recorded data are randomly divided into a training set (64%), a validation set (16%), and a test set (20%).

[0012] S2: Extract features from the collected audio signals and directly extract features from the original audio signals. A multi-scale convolution structure is introduced, and convolution kernels of different sizes (such as 3×3, 5×5, 7×7) are used to perform convolution operations on the same input signal to capture diverse frequency components. Combining local and global attention mechanisms, the features are weighted after multi-scale convolution. The local attention mechanism enhances the model's sensitivity to key information, while the global attention mechanism learns the dependencies between channels through global average pooling, improving the overall feature expression capability.

[0013] S3: Construct an integrated learning model (IMGLA-Net) based on multi-scale convolution and global-local attention. IMGLA-NET adopts the Stacking integrated learning framework and includes two first-level expert models IMGLA-ModelA and IMGLA-ModelB and a meta-classifier. IMGLA-ModelA uses an improved residual network combined with multi-scale convolution and local-global attention. IMGLA-ModelB is based on a deep separable convolutional network and improves computational efficiency by reducing the number of parameters.

[0014] S4: In the feature fusion stage, the output features of the two expert models are connected and passed into the meta-classifier to achieve the final recognition decision; when evaluating the performance of the model, the precision, recall and F1 score indicators are used, and the classification effect and robustness of the model are analyzed through the confusion matrix.

[0015] Furthermore, the step S1 specifically includes:

[0016] S1.1: Drone selection: To ensure the diversity and representativeness of the data, four different models of drones were selected: DIVA, SG970, SG106 and DJI Mavic3, marked as drones 1-4 respectively;

[0017] S1.2: Recording equipment selection: Use a dual-channel microphone for audio recording, and set the sampling rate to 44100 Hz to ensure that the captured audio signal has sufficient frequency resolution and dynamic range;

[0018] S1.3: Recording environment selection: Select a quiet environment for audio recording, and record the sound of the drone flying at different distances (0 to 50 meters, 10 meters interval). Specific settings include: Vertical distance: 1.7 meters from the microphone to the ground. Horizontal distance: When the drone is flying, keep it within two meters to avoid fluctuations in sound quality caused by distance changes.

[0019] S1.4: Introducing noise samples: During the recording process, record quiet environmental sounds as samples of the no drone category; in order to enhance the robustness of the model, add background noises such as thunderstorm, wind, and bird calls to form a "no drone" category; ensure the diversity of "no drone" samples to reduce the probability of misjudgment; samples in this category include: thunderstorm (representing environmental noise), wind (light wind, strong wind contrast), and bird calls (the calls of various birds);

[0020] S1.5: Dataset division: The collected audio dataset is randomly divided into training set (64%), validation set (16%), and test set (20%) for subsequent training and evaluation.

[0021] Furthermore, the step S2 specifically includes:

[0022] S2.1: Perform fast Fourier transform (FFT) on the collected audio signal to extract spectral features. FFT can not only reveal the energy distribution of different frequency components, but also provide a combined perspective of time domain and frequency domain for subsequent feature analysis;

[0023] S2.2: Use a multi-scale convolution structure, set convolution kernels of different sizes (such as 3x3, 5x5, 7x7), perform convolution operations on the same input signal, and globally learn the audio signal; by comparing the feature maps under different receptive fields, capture the subtle differences and dynamic characteristics of frequency components;

[0024] S2.3: Introduce local attention in the multi-scale convolutional feature map, and enhance the model's sensitivity to key information by calculating the weighted sum of local features; apply the Softmax function to calculate the attention weight, combine the convolutional features for recalibration, adopt the global attention mechanism in SE-Net, obtain the global information of the features through global average pooling, and use the fully connected layer to learn the dependencies between channels.

[0025] Furthermore, the step S3 specifically includes:

[0026] IMGLA-NET adopts the Stacking ensemble learning framework, which includes two first-level expert models IMGLA-ModelA and IMGLA-ModelB and a meta-classifier;

[0027] The overall structure includes:

[0028] Input layer: receives the pre-processed audio from the original audio signal;

[0029] First-level expert model A (IMGLA-ModelA): Improved based on residual network, convolution adopts padding strategy, including one-dimensional multi-scale residual local attention mechanism module and global attention mechanism module; adds global attention mechanism based on SE-Net, uses 1D convolution to extract local features, and each convolution branch is configured with convolution kernels of different sizes; the output layer is Softmax classification layer, which converts the output into probability distribution; the specific steps of SENet attention mechanism include: Squeeze: compress the feature map to a one-dimensional vector through global average pooling; Excitation: learn the relationship between channels through one or two layers of fully connected networks, generate channel weights, and apply them to the input feature map;

[0030] IMGLA-ModelB: Improved based on Xception, the convolution adopts padding strategy; achieves lighter feature extraction; reduces the number of parameters and improves computational efficiency through the depth-separable convolution structure; combines deep convolution and point-by-point convolution, first operates independently on each channel, and then fuses features through 1x1 convolution. Introduces local attention mechanism, and improves the model's ability to capture key information through feature screening;

[0031] Meta-classifier: combines the outputs of two expert models to make the final decision.

[0032] Furthermore, the step S4 specifically includes:

[0033] S4.1: IMGLA-ModelA and IMGLA-ModelB are trained independently, and the hyperparameters are tuned using the cross-validation method to ensure the generalization ability of the model; the learning rate is set to 0.001, the batch size is 32, the optimizer is Adam, and the loss function is cross entropy loss;

[0034] The cross entropy loss is defined as:

[0035]

[0036] Where C is the total number of categories y i is the actual distribution (one-hot encoding) value on the i-th category, is the probability of the i-th category predicted by the model.

[0037] S4.2: Feature fusion: The output features of IMGLA-ModelA and IMGLA-ModelB are concatenated in the feature dimension and input into the meta-classifier (logistic regression or random forest, etc. can be selected) for classification decision.

[0038] Furthermore, the model performance evaluation indicators in step S4 use accuracy, precision, recall and F1 score. Samples are divided into true positive (TP), false positive (FP), true negative (TN) and false negative (FN). Accuracy: refers to the ratio of the number of samples correctly classified by the classifier to the total number of samples, that is, the proportion of all correctly classified samples to the total number of samples. The formula is as follows:

[0039]

[0040] Precision: refers to the ratio of the number of samples correctly classified as positive examples by the classifier to the number of samples classified as positive examples among all samples classified as positive examples. The formula is as follows:

[0041]

[0042] Recall rate: refers to the ratio of the number of samples correctly classified as positive examples by the classifier to the total number of true positive examples, also known as recall rate. The formula is as follows:

[0043]

[0044] F1 score: It takes into account both precision and recall, and is the harmonic mean of precision and recall. The formula is as follows:

[0045]

[0046] On the other hand, the present invention proposes the application of the above method in identifying drones based on acoustic signals.

[0047] Beneficial effects of the present invention:

[0048] Existing technologies are mainly aimed at traditional radar, radio frequency, infrared and image recognition solutions in deep learning. The present invention proposes to use raw sound signal recognition. Beneficial effects: Compared with other deep learning methods, it has higher accuracy with fewer or similar parameters. Compared with other traditional methods, the equipment layout is simpler and the cost is lower. The inherent mechanism of this method: The one-dimensional audio signal drone recognition method based on deep learning automatically extracts the time and frequency characteristics of the audio signal through a neural network model to distinguish drones from other sound sources. The method involves the steps of preprocessing, feature extraction, model training and classification of audio data, and uses the convolutional neural network IMGLA-Net model to capture the complex patterns of signal features, thereby achieving efficient drone detection and recognition.

[0049] The present invention adopts an integrated learning framework (IMGLA-Net) so that the model does not need to rely on manual feature extraction and directly processes the original audio signal. This design greatly improves the adaptability and accuracy of the model. In traditional audio signal processing methods, it is often necessary to rely on expert knowledge for feature selection, which is not only time-consuming but also prone to information loss. Through integrated learning, the model can automatically identify and extract data features related to drone features, reducing the deviations that may be caused by human intervention. IMGLA-Net consists of two first-level expert models (IMGLA-ModelA and IMGLA-ModelB) and a meta-classifier. The former independently processes the input data and extracts features, and the latter fuses the prediction results of the two models. Through the Stacking mode, this process effectively integrates the advantages of different models in feature extraction and classification, thereby significantly improving the accuracy and stability of recognition.

[0050] The multi-scale convolution and global-local attention mechanisms introduced in the model of the present invention significantly enhance the feature capture capability of audio signals. The multi-scale convolution structure captures the diverse frequency components in the audio signal by setting convolution kernels of different sizes, and can effectively extract the subtle differences contained in the signal. This feature enables the model to remain efficient when processing complex audio signals. At the same time, the global-local attention mechanism further improves the model's sensitivity to key information. Local attention helps the model focus on important frequencies in the audio by weighting different features, while global attention learns the dependencies between channels on a global scale, thereby improving the overall feature expression capability. Through this mechanism, the model can better suppress the impact of background noise on the extraction of drone audio features, ensuring its effectiveness in complex environments.

[0051] In the present invention, IMGLA-ModelA and IMGLA-ModelB respectively adopt improved residual networks and deep separable convolutions, so that the models not only have good accuracy, but also have excellent computational efficiency. IMGLA-ModelA solves the common gradient vanishing problem of deep networks in training through the residual structure, so that the network can still learn effectively at a deeper level. This design ensures that the model will not be restricted by depth when capturing complex audio features. IMGLA-ModelB greatly reduces the number of parameters through deep separable convolutions, reduces the computational burden, and makes the model more suitable for real-time operation in resource-constrained environments. This efficient feature extraction mechanism enables the final classifier to complete the recognition task of the drone in a shorter time while ensuring the accuracy, meeting the real-time requirements in practical applications.

[0052] The present invention significantly improves the robustness of the model in various environments by designing a training data set covering diverse background noises. The introduced "no drone" class samples include a variety of background sounds such as thunder and rain sounds and bird calls, ensuring that the model can learn the impact of different noise conditions on drone audio signals during the training phase. Through such a diverse data set, the model not only improves the recognition ability of the target signal, but also reduces the misjudgment rate of background noise. Specifically, through diversified training, the model can effectively distinguish between environmental noise and characteristic sounds emitted by drones, so as to more accurately monitor and classify drones in practical applications. This design ensures that in complex environments, the model can still maintain a low misjudgment rate and a high recognition rate, greatly enhancing the practicality and reliability of drone monitoring technology.

[0053] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0055] Figure 1 Schematic diagram of drone and noise audio waveform;

[0056] Figure 2 Schematic diagram of drone and noise spectrum;

[0057] Figure 3 This is a schematic diagram of the overall structure of the model;

[0058] Figure 4 It is the schematic diagram of IMGLA-ModelA structure;

[0059] Figure 5 It is a schematic diagram of the IMGLA-ModelA submodule structure;

[0060] Figure 6 This is a schematic diagram of the SENet structure;

[0061] Figure 7 It is the schematic diagram of IMGLA-ModelB structure;

[0062] Figure 8 Schematic diagram of the depth-separable convolution structure;

[0063] Fig. 9 It is a schematic diagram of the accuracy error line of different lengths;

[0064] Fig.10 It is a schematic diagram of the accuracy error curve of S1-S11;

[0065] Fig.11 It is a schematic diagram of the S1-S11Loss value error curve;

[0066] Fig.12 A schematic diagram showing the confusion matrix before and after integration;

[0067] Fig.13 This is a schematic diagram of the module's own ablation experiment accuracy error curve;

[0068] Fig.14 This is a schematic diagram of the error curve of the module's own ablation experiment Loss value;

[0069] Fig.15 Schematic diagram of the accuracy error curve of different attention mechanisms;

[0070] Fig.16Schematic diagram of the error curve of Loss value of different attention mechanisms;

[0071] Fig.17 Schematic diagram of the accuracy error curve for different numbers of residual blocks;

[0072] Fig.18 Schematic diagram of the error curve of Loss value with different numbers of residual blocks;

[0073] Fig.19 Schematic diagram of the accuracy error curve for different numbers of fully connected layers;

[0074] Fig. 20 Schematic diagram of the error curve of Loss value for different numbers of fully connected layers;

[0075] Fig.21 Schematic diagram of the accuracy error curve of different integrated learning frameworks;

[0076] Fig. 22 Schematic diagram of the Loss value error curve of different integrated learning frameworks;

[0077] Fig.23 Schematic diagram of the accuracy error curve of different methods;

[0078] Fig.24 Schematic diagram of the Loss value error curve of different methods;

[0079] Fig.25 Schematic diagram of interference test under different signal-to-noise ratios. DETAILED DESCRIPTION

[0080] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0081] Example 1

[0082] The drone identification method based on ensemble learning described in this embodiment includes the following steps:

[0083] S1: Different types of drones are selected to construct a dataset, and the flight sounds of drones at different distances are recorded at a sampling rate of 44100 Hz. Background noises such as thunder and rain, wind, and bird calls are introduced to increase the complexity of the dataset, forming a "no drone" class sample, and the recorded data are randomly divided into a training set (64%), a validation set (16%), and a test set (20%).

[0084] S2: Extract features from the collected original audio signals; introduce a multi-scale convolution structure, and use convolution kernels of different sizes (such as 3×3, 5×5, 7×7) to perform convolution operations on the same input signal to capture diverse frequency components. Combining local and global attention mechanisms, weight the features after multi-scale convolution. The local attention mechanism enhances the model's sensitivity to key information, while the global attention mechanism learns the dependencies between channels through global average pooling, improving the overall feature expression capability.

[0085] S3: Construct an integrated learning model (IMGLA-Net) based on multi-scale convolution and global-local attention. IMGLA-NET adopts the Stacking integrated learning framework and includes two first-level expert models IMGLA-ModelA and IMGLA-ModelB and a meta-classifier. IMGLA-ModelA uses an improved residual network combined with multi-scale convolution and local-global attention. IMGLA-ModelB is based on a deep separable convolutional network and improves computational efficiency by reducing the number of parameters.

[0086] S4: In the feature fusion stage, the output features of the two expert models are connected and passed into the meta-classifier to achieve the final recognition decision; when evaluating the performance of the model, the precision, recall rate, and F1 score indicators are used, and the classification effect and robustness of the model are analyzed through the confusion matrix and ROC curve.

[0087] Example 2

[0088] This study uses four types of drones, namely DIVA, SG970, SG106, and DJI Mavic 3 (corresponding to drone 1, drone 2, drone 3, and drone 4, respectively) for drone identification. The size and weight of the drones are shown in Table 1.

[0089] Table 1 UAV parameters

[0090]

[0091]

[0092] The recording device is a dual-channel microphone with a sampling rate of 44100Hz. During the recording process, the vertical distance between the microphone and the ground is 1.7m, and the distance of the drone flying up and down, left and right is controlled to about two meters. The drone is 0 meters away from the microphone (almost close to the microphone) and 50 meters away at most. The audio is recorded at an interval of ten meters. The duration of the recorded audio is shown in Table 2.

[0093] Table 2 Audio recording duration at different distances from drones

[0094] UAV 0m 10m 20m 30m 40m 50m UAV1 46S 42S 50S 52S 47S 43S UAV2 45S 43S 47S 42S 48S 49S UAV3 47S 48S 46S 52S 51S 50S UAV4 46S 52S 56S 51S 46S 48S

[0095] UAVs generate specific audio signals during flight, and these audio signals have unique spectrum and time domain characteristics. By analyzing these audio signals, we can extract the audio signal features related to the characteristics of the UAV. The waveforms of the four UAV audio files are shown in the figure below. Figure 1 (a) shows. During the recording process, the sound of a quiet environment is recorded as a sample of the no drone category. In order to improve the robustness of the recognition model and prevent the model from misidentifying other sounds as the presence of a drone, samples such as thunderstorm mixed sounds, bird calls, and wind sounds are added to the no drone category. The waveform of the sample without a drone category is shown in Figure 1 (b) The frequencies of Noise-1 and Noise-3 are lower than those of UAV and their amplitudes are also lower in some locations, while the waveform of Noise-2 is close to the more complex UAV audio.

[0096] The audio signal samples of the four drones are transformed by FFT to obtain a spectrum diagram. The spectrum diagram of the drones shows that the frequency distribution range of the drones is wide and the frequency components are complex. The frequencies of drones 1, 2, and 4 have higher energy in the range of 4kHz to 12kHz, and most of the signal energy is concentrated in this frequency band. Drone 3 has higher energy in the range of 3KHz-5KHz and 6KHz-9KHz, which indirectly reflects the differences between the drones. In view of the complex frequency components of drones, this embodiment uses a multi-scale convolution structure to extract the key features contained in different frequency components. The spectrum diagram of the drone is as follows: Figure 2 (a) shown.

[0097] The spectrum of the noise category is as follows Figure 2 (b) shows that Blank means the background sound is recorded in a quiet environment, and this part of the sound is more reflected as a slight wind sound. The frequency distribution of this slight wind sound also appears in the four drone spectra shown in (a). In order to reduce the impact of this part of wind sound on the drone recognition results, this embodiment adds an attention mechanism in the feature extraction process so that the proposed model pays more attention to drone-related features. The remaining samples without drone categories are shown as Wind, Bird, and Thunderstorm in (b).

[0098] Example 3

[0099] This embodiment proposes an integrated learning model for processing one-dimensional audio signals: an integrated learning model (IMGLA-Net) based on multi-scale convolution and global local attention integration, which is used for drone recognition. The IMGLA-NET model adopts an integrated learning framework of the Stacking mode, which mainly includes two first-level expert models and a meta-classifier. The two first-level expert models in IMGLA-NET are IMGLA-ModelA and IMGLA-ModelB. The two first-level expert models add local global attention mechanisms and multi-scale convolutions, which can not only extract richer and more diverse features from audio signals, but also accurately capture subtle differences and key information in different audio signals. The preliminary classification results obtained by IMGLA-ModelA and IMGLA-ModelB are further input into the meta-classifier. By learning and fusing these meta-features, the meta-classifier combines the recognition advantages of different first-level expert models for different drones, thereby achieving more accurate recognition of drone audio signals. This IMGLA-NET based on the integrated learning framework shows excellent performance and provides new ideas and directions for the research and application of drone recognition technology. The overall structure of IMGLA-NET is as follows Figure 3 shown.

[0100] Stacking

[0101] Ensemble learning significantly improves the accuracy and stability of predictions by combining multiple independently trained weak learners. The core idea is that the advantages of different learners can be integrated to enhance the performance of the model. Ensemble learning can be mainly divided into three methods: bagging, boosting, and stacking. Stacking first trains multiple weak learners, and then uses the prediction results of these base learners as new features and inputs them into the meta-learner for the final prediction. Stacking makes full use of the advantages of each weak learner and further improves the performance of the entire ensemble learning framework.

[0102] In the Stacking mode, the meta-classifier plays a key role in integrating the prediction results of weak classifiers. After the output results of IMGLA-ModelA and IMGLA-ModelB are connected in the feature dimension, they are sent to the fully connected layer to map the connected results to the required output space, such as outputting the classification results in the classification task. The Stacking model, with its ability to combine the prediction results of multiple models, can capture complex patterns and relationships, thereby significantly improving the overall performance of classification or regression tasks. It reduces the error of a single model and enhances the accuracy and robustness of the overall prediction.

[0103] IMGLA-ModelA

[0104] This embodiment designs two first-level expert models: IMGLA-ModelA and IMGLA-ModelB. IMGLA-ModelA is improved based on the residual network, while IMGLA-ModelB is improved based on Xception. All convolutions in IMGLA-ModelA and IMGLA-ModelA use the padding strategy. Both integrate multi-scale convolution and local global attention mechanism to improve the prediction accuracy of the model.

[0105] IMGLA-ModelA is mainly composed of a one-dimensional multi-scale residual local attention mechanism module and a global attention mechanism module. Its specific structure is as follows: Figure 4 The one-dimensional multi-scale residual local attention mechanism module is improved on the basis of the residual module. In view of the complex frequency components of drone audio signals, a multi-scale convolution structure is added to the module, and a local attention scheme is used to screen the features of convolution branches of different scales. Its specific structure is shown in Figure 5 (a) As shown. A global attention mechanism based on SE-Net is added to the last part of IMGLA-ModelA, and the residual block structure is as follows Figure 5 (b) The parameters of IMGLA-ModelA are shown in Table 3.

[0106] Table 3 IMGLA-ModelA parameters

[0107] Layer Input Output Kernel Stride Padding Convld 1*600 64*600 3 1 1 MaxPool - - 3 2 1 MSC 64*300 64*300 [3,5,7] 1 Kernel / / 2 Layer1 66*300 66*300 3 1 1 Layer2 66*300 128*150 3 2 1 Layer3 128*150 256*75 3 2 1 Layer4 256*75 512*37 3 2 1 AvgPool1d 512*37 512*1 - - - Linear 512*1 5*1 - - -

[0108] The attention mechanism aims to improve the model's ability to understand input data through multi-level attention. The main purpose of local attention is to further screen different channels after multi-scale convolution filtering features, thereby enhancing the model's sensitivity to local information. The global attention mechanism integrates the global information of the entire feature map to improve the overall representation ability and performance of the model. This embodiment uses the SENet attention mechanism. SENet learns the weights of features based on the loss, so that effective feature channels obtain larger weights, and feature maps with small or invalid effects obtain smaller weights. It mainly consists of two parts: Squeeze and Excitation. The "Squeeze" step uses global average pooling to compress spatial information, and the "Excitation" step uses a fully connected network to learn the dependencies between channels, thereby adaptively readjusting channel weights and improving model performance. The SENet structure is as follows: Figure 6 shown.

[0109] IMGLA-ModelB

[0110] IMGLA-ModelB utilizes depthwise separable convolutions by decomposing standard convolutions into depthwise convolutions and pointwise convolutions. Depthwise convolutions operate independently on each channel, while pointwise convolutions use 1x1 convolutions to combine channels. Depthwise separable convolutions not only improve computational efficiency, but also reduce the size of the model while maintaining model performance, making it more suitable for resource-constrained devices. See the IMGLA-ModelB structure for details. Figure 7 , where the depthwise separable convolution structure is shown in Figure 8 , see the multi-scale local attention structure Figure 5 (a), the parameters of IMGLA-ModelB are shown in Table 4.

[0111] Table 4 IMGLA-ModelB parameters

[0112]

[0113] Example 4

[0114] This part verifies the effectiveness and superiority of the proposed IMGLA-NET model through three experiments: sample length experiment, ablation experiment, and comparative experiment. The sample length experiment explores the impact of different sample lengths on accuracy. The ablation experiment conducts a comprehensive analysis of the model's own structure, including comparative experiments of different integrated structures, performance comparison experiments of different first-level expert models, performance comparison experiments of convolution modules, and performance comparison experiments of attention mechanisms. The comparative experiment conducts a comprehensive comparative analysis with other excellent methods, demonstrating the significant advantages and application potential of the model in processing drone audio signals.

[0115] Experimental preparation

[0116] The distribution of the data set used in this embodiment is shown in Table 5, and the size of the data is 1*600. The data set contains 5 categories, namely, samples of 4 types of drones and samples without drones. Among them, the samples without drones not only include the situation where there is no drone flying during recording (that is, the background sound when collecting data), but also include another 3 sounds (bird calls, thunderstorms, and wind sounds) as interference samples. The experiments in this embodiment were repeated ten times. The final result takes the average of the ten repeated experiments.

[0117] Table 5 Dataset distribution

[0118]

[0119]

[0120] In this study, the PyTorch library in the Python programming language was used, and the processor of the experimental equipment was Intel(R) Xeon(R) Gold 5220CPU@2.20GHz. The graphics cards were NVIDIA GeForce RTX 3090 and NVIDIA RTXA4500. During the training process, the batch size of the experimental setting was 128, and each training iteration was 200 rounds. Adam was selected as the optimizer of the classifier.

[0121] The loss function used in the experiment of this embodiment is cross entropy loss (CrossEntropyLoss), which is a loss function widely used in classification tasks, especially for multi-classification problems in neural networks. The core idea is to guide the training of the model by measuring the difference between the predicted distribution and the actual distribution. Mathematically, the cross entropy loss calculates the entropy difference between the true distribution of the actual category (usually represented by one-hot encoding) and the probability distribution predicted by the model. The cross entropy loss is defined as:

[0122]

[0123] Where C is the total number of categories y i is the actual distribution (one-hot encoding) value on the i-th category, is the probability of the i-th category predicted by the model.

[0124] A notable feature of cross entropy loss is that it not only considers whether the prediction is correct or not, but also the confidence of the prediction. When the model's predicted probability is closer to the actual distribution, the loss is smaller; conversely, when the prediction deviates from the actual distribution, the loss increases. Therefore, cross entropy loss can effectively guide the model's parameter update and gradually improve the model's classification performance.

[0125] The model performance evaluation indicators include accuracy, precision, recall and F1 score. Samples are divided into true positive (TP), false positive (FP), true negative (TN) and false negative (FN). Accuracy: refers to the ratio of the number of samples correctly classified by the classifier to the total number of samples, that is, the proportion of all correctly classified samples to the total number of samples. The formula is as follows:

[0126]

[0127] Precision: refers to the ratio of the number of samples correctly classified as positive examples by the classifier to the number of samples classified as positive examples among all samples classified as positive examples. The formula is as follows:

[0128]

[0129] Recall rate: refers to the ratio of the number of samples correctly classified as positive examples by the classifier to the total number of true positive examples, also known as recall rate. The formula is as follows:

[0130]

[0131] F1 score: It takes into account both precision and recall, and is the harmonic mean of precision and recall. The formula is as follows:

[0132]

[0133] Data length experiment

[0134] This experiment selects four classic CNN networks, namely AlexNet, GoogLeNet, Xception, Vgg, and two improved models IMGLA-ModelA and IMGLA-ModelB. The experiment divides the audio length into 7 comparison groups from 300 to 900 according to the sampling points. The test results are shown in Table 6, and the accuracy error line graph of different lengths under different networks is shown in Fig. 9 shown.

[0135] Table 6 Accuracy of different networks with different lengths

[0136] 300 400 500 600 700 800 900 AlexNet 0.7789 0.8342 0.8345 0.8566 0.8552 0.8297 0.6682 Xception 0.6407 0.7612 0.7817 0.8003 0.7748 0.7387 0.6670 VGG 0.6316 0.6640 0.7248 0.8517 0.7455 0.7207 0.6967 GoogLeNet 0.6682 0.6808 0.7277 0.7453 0.7135 0.7080 0.6722 A-No-SEMSC 0.8472 0.8502 0.8807 0.8818 0.8658 0.8657 0.8495 B-No-SEMSC 0.8386 0.8552 0.8755 0.8892 0.8693 0.8618 0.8598 Model-A 0.8853 0.8917 0.8943 0.9128 0.8813 0.8653 0.8522 Model-B 0.9026 0.9162 0.9286 0.9349 0.9208 0.9195 0.8775

[0137] The experiments show that the accuracy of the sampling point length is 600, which is higher than that of other sampling points in the same network. Therefore, the subsequent experimental data use the sampling point length of 600.

[0138] Ablation experiment

[0139] In order to verify the rationality of the proposed model structure, this embodiment conducted the following ablation experiments: combination experiments of different first-level expert models, verification of the effectiveness and excellence of different sub-modules in the model, comparison of different attention mechanisms, comparison of the number of residual blocks, comparison of the number of fully connected layers in the meta-classifier, and comparison of different ensemble learning modes.

[0140] Combination Experiment

[0141] The choice of the first-level expert model has a great impact on the stacking mode. The stacking mode will be affected by the stability of the first-level expert model and the size of the difference in prediction results. When constructing a stacking model, multiple experiments are required to select the optimal first-level expert model. This embodiment selects the top four models with the highest average accuracy, namely AlexNet, VGG, IMGLA-ModelA, and IMGLA-ModelB (IMGL-ModelA and IMGLA-ModelB are the first-level expert models designed in this embodiment). The experiment combines the four models with the highest accuracy, with a total of three types and eleven different groups. The three combinations are a combination of two first-level expert models, a combination of three first-level expert models, and a combination of four first-level expert models. The model combinations are shown in Table 7. The test results are shown in Table 8, and the training process is shown in Table 8. Fig.10 , 11 shown.

[0142] Table 7 Basic model combinations

[0143] CombineName Models S1 VGG+AlexNet S2 VGG+Model-A S3 VGG+Model-B S4 AlexNet+Model-A S5 AlexNet+Model-B S6 Model-A+ModelB S7 VGG+AlexNet+Model-A S8 VGG+AlexNet+Model-B S9 VGG+Model-A+Model-B S10 AlexNet+Model-A+Model-B S11 VGG+AlexNet+Model-A+Model-B

[0144] Table 8 Accuracy of different combination models

[0145] Norm S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 ACC 86.37 91.64 96.52 91.28 97.24 0.9811 92.62 97.01 96.29 97.37 97.55 Pre 8655 91.75 96.57 91.71 97.24 0.9811 92.85 97.01 96.32 97.40 97.55 Recall 86.37 91.64 96.52 91.28 97.24 0.9811 92.62 97.01 96.29 97.37 97.55 F1 86.39 91.64 96.53 91.26 97.23 0.9811 92.62 97.00 96.28 97.37 97.55

[0146] Experimental results show that IMGLA-NET (S6) exhibits the best performance among the tested groups, demonstrating its effectiveness and superiority on specific applications or datasets.

[0147] By analyzing the confusion matrix, it is found that IMGLA-ModelA and IMGLA-ModelB have their own advantages in identifying different types of drones. IMGLA-ModelA performs better than IMGLA-ModelB in the recognition accuracy of drones 1 and 4, but IMGLA-ModelB has better recognition accuracy in identifying drones 2 and 4. This embodiment uses the Stacking model to combine the IMGLA-NET-ModeA and IMGLA-ModelB models through ensemble learning Stacking model, which significantly improves the overall recognition effect. The confusion matrix is ​​as follows Fig.12 shown.

[0148] Module ablation experiment

[0149] The ablation experiments conducted in this embodiment are with or without local attention mechanism, with or without global attention mechanism, with or without attention mechanism, and with or without multi-scale convolution. The experimental results show that multi-core convolution improves the accuracy by 3.71%. The attention mechanism improves the accuracy by 1.46%. The local attention mechanism improves the accuracy by 1.29%, while the global attention mechanism improves the accuracy by 0.79%. The model IMGLA-NET proposed in this embodiment improves the performance of the model that only integrates the basic network by 5.03%. The test results are shown in Table 9, and the training process is shown in Table 9. Fig.13 , 14 shown.

[0150] Table 9 Self-module ablation experiment

[0151]

[0152] Different attention mechanisms

[0153] In order to explore the performance of different attention mechanisms, this experiment systematically compared various attention mechanisms, including Self-Attention, ECA, EMA, and CA. After rigorous experimental verification, the results show that the SE attention mechanism has shown superior performance in these comparisons. Fig.15 , 16 shown.

[0154] Table 10 Tests of different attention mechanisms

[0155]

[0156] Number of residual blocks

[0157] In this experiment, we compared the effects of different numbers of residual blocks on the model accuracy. The experiment set up a model containing 1 to 4 residual blocks to evaluate the effect of the number of residual blocks on the model performance. The experimental results are shown in Table 11 and the test results are shown in Fig.17 , 18 shown.

[0158] Table 11 Number of different residual fast

[0159]

[0160] Number of fully connected layers

[0161] This experiment explores the effect of the number of fully connected layers on the model accuracy. The experiment designed a structure containing 1 to 5 fully connected layers, and evaluated its performance by comparing the indicators obtained when using different fully connected structures. The experiment proves that the model works best when there is one fully connected layer. The test results are shown in Table 12, and the training process is shown in Fig. 20 , 21shown.

[0162] Table 12 Test results of different numbers of fully connected layers

[0163]

[0164] Comparison of Ensemble Learning Models

[0165] Ensemble learning is a machine learning method that improves overall performance by combining the predictions of multiple models. Its main types include bagging, boosting, and stacking. The results of four evaluation indicators of three ensemble models are shown in Table 13. The experimental results show that the stacking ensemble mode has a significant improvement in accuracy. The training process is shown in Fig.21 , 22 shown.

[0166] Table 13 Tests of different ensemble learning frameworks

[0167]

[0168] Comparative experiment

[0169] In this part of the experiment, the convolutional neural network-related algorithms based on one-dimensional signals are used for comparative experiments. It is worth noting that the research on using one-dimensional signals to process drone audio signals for recognition in the field of deep learning is still in its infancy. Therefore, some relatively excellent algorithms in the field of one-dimensional signal processing and recognition are used for comparison. The comparative algorithms include the following: one-dimensional convolutional neural network and long short-term memory (LSTM) network structure (CNN-LSTM) that recognizes speech emotions by learning deep emotional features. One-dimensional cascade network structure (1D-MFCC-MCFC) for analyzing multiple emotions in speech. Serial one-dimensional CNN and BiLSTM structure (1D-CNN-BiLSTM) for detecting voice phishing attacks. End-to-end one-dimensional convolution structure based on environmental sound classification (End-to-End-1D-CNN). And a deep separable and interpretable feature extraction network (AT-SincNet) that comprehensively judges whether the speech is pathological based on the classification results of each speech frame. The above algorithms have conducted the same comparative experiments on the same drone dataset. The experimental results show that the performance of the IMGLA-NET model on this experimental dataset is significantly better than that of other models compared, which fully verifies its effectiveness and superiority in the field of drone audio recognition. The test results are shown in Table 14, and the training process is shown in Fig.23 , 24 .

[0170] Table 14 Comparison of different methods

[0171]

[0172] Noise comparison experiment

[0173] In actual environments, the presence of noise is unavoidable in the collected drone audio signals. The presence of noise often has a negative impact on the accuracy of the data and the performance of the system. This experiment aims to analyze the impact of noise on the performance of the data processing system by simulating different levels of noise (measured by the signal-to-noise ratio SNR, ranging from 0 to 30, with 5 as one level). By comparing the performance indicators of the system under different noise levels, the robustness of each model is analyzed. The experiment proves that this model has good robustness. The error curves under different signal-to-noise ratios are shown in Figure 2. Fig.25 shown.

[0174] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A drone identification method based on ensemble learning, characterized in that: The following steps are involved: S1: Different types of drones are selected to construct a dataset, and the flight sounds of drones at different distances are recorded at a sampling rate of 44100 Hz. Background noises such as thunder and rain, wind, and bird calls are introduced to increase the complexity of the dataset, forming "no drone" samples, and the recorded data are randomly divided into training set, validation set, and test set; S2: Extract features from the collected original audio signals; introduce a multi-scale convolution structure and use convolution kernels of different sizes to perform convolution operations on the same input signal to capture diverse frequency components; combine local and global attention mechanisms to weight features after multi-scale convolution. The local attention mechanism enhances the model's sensitivity to key information, while the global attention mechanism learns the dependencies between channels through global average pooling to improve the overall feature expression capability; S3: Construct an integrated learning model based on multi-scale convolution and global-local attention integration. IMGLA-NET adopts the Stacking integrated learning framework, which includes two first-level expert models IMGLA-ModelA and IMGLA-ModelB and a meta-classifier; IMGLA-ModelA uses an improved residual network, combined with multi-scale convolution and local-global attention; IMGLA-ModelB is based on a deep separable convolutional network, which improves computational efficiency by reducing the number of parameters; S4: In the feature fusion stage, the output features of the two expert models are connected and passed into the meta-classifier to achieve the final recognition decision; when evaluating the performance of the model, the precision, recall and F1 score indicators are used, and the classification effect and robustness of the model are analyzed through the confusion matrix.

2. The method for identifying drones based on ensemble learning according to claim 1, characterized in that: The step S1 specifically includes: S1.1: Drone selection: To ensure the diversity and representativeness of the data, four different models of drones were selected: DIVA, SG970, SG106 and DJI Mavic3, marked as drones 1-4 respectively; S1.2: Recording equipment selection: Use a dual-channel microphone for audio recording, and set the sampling rate to 44100 Hz to ensure that the captured audio signal has sufficient frequency resolution and dynamic range; S1.3: Recording environment selection: Choose a quiet environment for audio recording, and record the sound of the drone flying at different distances; specific settings include: vertical distance: 1.7 meters from the microphone to the ground; horizontal distance: when the drone is flying, keep it within two meters to avoid fluctuations in sound quality caused by distance changes; S1.4: Introducing noise samples: During the recording process, record quiet environmental sounds as samples of the no drone category; in order to enhance the robustness of the model, add background noises such as thunder and rain, wind, and bird calls to form a "no drone" category; ensure the diversity of "no drone" samples, thereby reducing the probability of misjudgment; S1.5: Dataset division: The collected audio dataset is randomly divided into training set, validation set, and test set for subsequent training and evaluation.

3. The method for identifying drones based on ensemble learning according to claim 1, characterized in that: The step S2 specifically includes: S2.1: Perform fast Fourier transform on the collected audio signal to extract spectrum features; FFT performs spectrum analysis on the signal; S2.2: Use a multi-scale convolution structure, set convolution kernels of different sizes, perform convolution operations on the same input signal, and globally learn the audio signal; by comparing feature maps under different receptive fields, capture the subtle differences and dynamic characteristics of frequency components; S2.3: Introduce local attention in the multi-scale convolutional feature map, and enhance the model's sensitivity to key information by calculating the weighted sum of local features; apply the Softmax function to calculate the attention weight, combine the convolutional features for recalibration, adopt the global attention mechanism in SE-Net, obtain the global information of the features through global average pooling, and use the fully connected layer to learn the dependencies between channels.

4. The method for identifying drones based on ensemble learning according to claim 1, characterized in that: The step S3 specifically includes: IMGLA-NET adopts the Stacking ensemble learning framework, which includes two first-level expert models IMGLA-ModelA and IMGLA-ModelB and a meta-classifier; The overall structure includes: Input layer: receives the pre-processed audio from the original audio signal; Level 1 expert model A: Improved based on residual network, convolution adopts padding strategy, including one-dimensional multi-scale residual local attention mechanism module and global attention mechanism module; adds global attention mechanism based on SE-Net, uses 1D convolution to extract local features, and configures convolution kernels of different sizes for each convolution branch; the output layer is Softmax classification layer, which converts the output into probability distribution; the specific steps of SENet attention mechanism include: Squeeze: compress the feature map to a one-dimensional vector through global average pooling; Excitation: learn the relationship between channels through one or two layers of fully connected networks, generate channel weights, and apply them to the input feature map; Level 1 expert model B: Improved based on Xception, the convolution adopts padding strategy; achieves lighter feature extraction; reduces the number of parameters and improves computational efficiency through the depthwise separable convolution structure; combines deep convolution and pointwise convolution, first operates independently on each channel, and then fuses features through 1x1 convolution; introduces local attention mechanism, and improves the model's ability to capture key information through feature screening; Meta-classifier: combines the outputs of two expert models to make the final decision.

5. The method for identifying drones based on ensemble learning according to claim 1, characterized in that: The step S4 specifically includes: S4.1: IMGLA-ModelA and IMGLA-ModelB are trained independently, and the hyperparameters are tuned using the cross-validation method to ensure the generalization ability of the model; the learning rate is set to 0.001, the batch size is 32, the optimizer is Adam, and the loss function is cross entropy loss; The cross entropy loss is defined as: Where C is the total number of categories y i is the value actually distributed in the i-th category, is the probability of the i-th category predicted by the model; S4.2: Feature fusion: The output features of IMGLA-ModelA and IMGLA-ModelB are concatenated in the feature dimension and input into the meta-classifier for classification decision.

6. The method for identifying drones based on ensemble learning according to claim 1, characterized in that: In step S4, the model performance evaluation index adopts accuracy, precision, recall and F1 score; the samples are divided into true positive TP, false positive FP, true negative TN and false negative FN; Accuracy: refers to the ratio of the number of samples correctly classified by the classifier to the total number of samples, that is, the proportion of all correctly classified samples to the total number of samples; the formula is as follows: Precision: refers to the ratio of the number of samples correctly classified as positive examples by the classifier to the number of samples classified as positive examples among all samples classified as positive examples; the formula is as follows: Recall rate: refers to the ratio of the number of samples correctly classified as positive examples by the classifier to the total number of true positive examples, also known as the recall rate; the formula is as follows: F1 score: It takes into account both precision and recall, and is the harmonic mean of precision and recall. The formula is as follows:

7. Application of the method according to any one of claims 1 to 6 in identifying drones based on acoustic signals.

Citation Information

Cited By

  • Data processing system based on voice recognition intelligent electronic scale

    CN119338079A

  • Unmanned aerial vehicle countering method

    CN120760544A

  • Deep learning abnormal sound detection method with generalization ability

    CN121506194A