Intelligent identification method and system based on unmanned aerial vehicle countermeasure

Abstract

The application discloses an intelligent identification method and system based on unmanned aerial vehicle countermeasures, relates to the technical field of unmanned aerial vehicles, and comprises the following steps: calculating modal initial weights based on an MLP supervised learning method, calculating attention weights through a Transformer attention mechanism, and fusing feature vectors based on the modal initial weights and the attention weights by adopting a collaborative attention mechanism. By introducing the collaborative attention mechanism, combining the MLP supervised learning and the Transformer attention mechanism, and realizing adaptive modal weight adjustment based on an environment vector and threat intelligence, the problem of decreased recognition accuracy of traditional fixed weight fusion in a complex environment is effectively solved, the identification capability for disguised unmanned aerial vehicles is enhanced, the false positive rate and the false negative rate are significantly reduced, the robustness of feature extraction in diversified scenarios is ensured, and the environmental adaptability, the recognition accuracy and the anti-deception capability of the unmanned aerial vehicle countermeasure system are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicles, in particular to an intelligent identification method and system based on unmanned aerial vehicle countermeasures. BACKGROUND

[0002] Unmanned aerial vehicle countermeasure technology has developed rapidly in the fields of aviation safety, military defense and public safety in recent years. With the popularization of unmanned aerial vehicle technology and the widespread application of low-cost unmanned aerial vehicles, the demand for intelligent identification and countermeasures against unmanned aerial vehicles is increasingly urgent. Existing technologies mainly rely on multi-modal sensor fusion and deep learning models to achieve unmanned aerial vehicle detection and classification. However, existing technologies still have limitations in the dynamic nature of multi-modal fusion, environmental adaptability and detection ability of disguised targets. Traditional multi-modal fusion methods lack sufficient consideration of environmental changes and dynamic relationships between modalities, resulting in decreased recognition accuracy in complex environments (such as strong electromagnetic interference or low visibility). Existing methods lack effective anti-deception mechanisms when dealing with disguised unmanned aerial vehicles or abnormal behavior, and are easily affected by interference signals or fake features, resulting in high false positive and false negative rates. Existing unmanned aerial vehicle countermeasure technologies have problems of insufficient environmental adaptability and limited identification ability of disguised targets in multi-modal feature fusion and anti-deception detection. The present application introduces a collaborative attention mechanism, combines MLP supervised learning and Transformer attention mechanism to dynamically fuse radar, radio frequency, visible light and infrared modal features, and adaptively adjusts the weights based on the environment vector and threat intelligence, solving the dynamic nature of feature fusion in complex environments and improving the anti-deception ability through consistency detection and anomaly analysis, significantly improving the recognition accuracy and countermeasure efficiency. SUMMARY

[0003] In view of the above existing problems, the present application is proposed.

[0004] Therefore, the present application provides an intelligent identification method and system based on unmanned aerial vehicle countermeasures, solving the problem of insufficient environmental adaptability and limited identification ability of disguised targets in multi-modal feature fusion and anti-deception detection of existing unmanned aerial vehicle countermeasure technologies.

[0005] To solve the above technical problems, the present application provides the following technical solutions:

[0006] In a first aspect, the present application provides an intelligent identification method based on unmanned aerial vehicle countermeasures, which includes collecting real-time environmental parameters, generating an environment vector, calculating an initial threat level based on the environment vector and threat intelligence, and generating a multi-modal data set based on the initial threat level.

[0007] The time-frequency spectrum of the radar echo is extracted using a short-time Fourier transform, the time-frequency spectrum is input into a MobileNetV3 model to extract radar signal features, a 1D-CNN model is used to extract features of the radio frequency signal, an EfficientNet-B0 model is used to extract visual features of the visible light image, an IRNet is used to extract thermal imaging features of the infrared image, and a feature vector set is generated;

[0008] The initial weights of the modes are calculated based on an MLP supervised learning method, the attention weights are calculated through a Transformer attention mechanism, and the feature vectors are fused based on the initial weights of the modes and the attention weights through a collaborative attention mechanism;

[0009] The real-time classification and threat level are generated based on the fused feature vectors using a Transformer model, the countermeasure priority is calculated through the threat level and the real-time classification result, and the anti-deception detection result and the countermeasure instruction are generated based on the countermeasure priority;

[0010] Based on the anti-deception detection result and the countermeasure instruction, dynamic countermeasure instruction execution and effect feedback optimization are performed.

[0011] As a preferred scheme of the intelligent identification method based on unmanned aerial vehicle countermeasures, the real-time environment parameters are collected to generate an environment vector, the environment vector is combined with threat intelligence to calculate an initial threat level, and a multi-modal data set is generated based on the initial threat level. The environment vector is normalized to obtain an environment vector The environment vector is combined with threat intelligence to calculate an initial threat level According to The sampling frequency is adjusted, the data is aligned using a high-precision GPS clock based on the multi-modal data after the sampling frequency is adjusted, and a data set with a unified timestamp is generated .

[0012] As a preferred scheme of the intelligent identification method based on unmanned aerial vehicle countermeasures, the time-frequency spectrum of the radar echo is extracted using a short-time Fourier transform, the time-frequency spectrum is input into a MobileNetV3 model to extract radar signal features, a 1D-CNN model is used to extract features of the radio frequency signal, and a feature vector set is generated. The features are extracted from the radar echo, a time-frequency spectrum is generated using a short-time Fourier transform (STFT), the spectrum is input into a pre-trained MobileNetV3 model, the network extracts spatial and frequency features of the radar signal through convolution operation, and a fixed-dimension feature vector is output ;

[0013] RF signal Feature extraction, signal energy and spectral entropy , when the electromagnetic noise level N ) , the RF signal is severely disturbed, the signal energy reliability is reduced, and the spectral entropy weight, enhance the representation of signal frequency distribution, use 1D-CNN model, output the feature vector of RF signal ;

[0014] Visible light image Feature extraction, if the light intensity u, skip processing to avoid invalid calculation, otherwise use the pre-trained EfficientNet-B0 model to extract the high-order visual of visible light image , use IRNet to extract the thermal imaging features of infrared image , generate a feature vector set through adaptive feature extraction .

[0015] As a preferred scheme of the intelligent identification method based on unmanned aerial vehicle countermeasure of the application, wherein: the feature vector based on the modal initial weight and the attention weight adopts the collaborative attention mechanism, which is based on the MLP supervised learning method to calculate the modal initial weight , the data-driven attention weight is calculated through the Transformer attention mechanism , the initial weight of each mode is adopted by the collaborative attention mechanism , according to the environment vector and the threat level Adaptive adjustment, using data-driven attention weight , based on the dynamic importance of modal features , multiply and as the comprehensive weight, and the corresponding modal feature Weighted sum to generate a fusion feature vector .

[0016] As a preferred scheme of the intelligent identification method based on unmanned aerial vehicle countermeasure provided by the application, the use of the Transformer model for real-time classification and threat level generation based on the fusion feature vector is a full connection neural network trained by an unmanned aerial vehicle data set, a cross-entropy loss is used, a time sequence fusion feature is input into a lightweight Transformer model, an intention vector is generated by analyzing the unmanned aerial vehicle behavior mode, and a dynamic threat level is calculated through a full connection layer to reflect the real-time threat degree, a contrast loss is used in combination with a cross-entropy loss to optimize the model, the ability to distinguish disguised targets is enhanced, and the result output includes a category probability, a dynamic threat level, and an intention vector.

[0017] As a preferred scheme of the intelligent identification method based on unmanned aerial vehicle countermeasure provided by the application, the use of the Transformer model for real-time classification and threat level generation based on the fusion feature vector is a full connection neural network trained by an unmanned aerial vehicle data set, a cross-entropy loss is used, a time sequence fusion feature is input into a lightweight Transformer model, an intention vector is generated by analyzing the unmanned aerial vehicle behavior mode, and a dynamic threat level is calculated through a full connection layer to reflect the real-time threat degree, a contrast loss is used in combination with a cross-entropy loss to optimize the model, the ability to distinguish disguised targets is enhanced, and the result output includes a category probability, a dynamic threat level, and an intention vector.

[0018] As a preferred scheme of the intelligent identification method based on unmanned aerial vehicle countermeasure provided by the application, the use of the Transformer model for real-time classification and threat level generation based on the fusion feature vector is a full connection neural network trained by an unmanned aerial vehicle data set, a cross-entropy loss is used, a time sequence fusion feature is input into a lightweight Transformer model, an intention vector is generated by analyzing the unmanned aerial vehicle behavior mode, and a dynamic threat level is calculated through a full connection layer to reflect the real-time threat degree, a contrast loss is used in combination with a cross-entropy loss to optimize the model, the ability to distinguish disguised targets is enhanced, and the result output includes a category probability, a dynamic threat level, and an intention vector.

[0019] In a second aspect, the application provides an intelligent identification system based on unmanned aerial vehicle countermeasures, which comprises,

[0020] An environment perception and threat initialization module is used to collect real-time environment parameters, generate an environment vector, and calculate an initial threat level in combination with threat intelligence to generate a multi-modal data set.

[0021] A multi-modal feature extraction module is used to extract each modal feature and generate a feature vector set.

[0022] a feature fusion and weight calculation module, configured to combine the normalized environment vector and the initial threat level of the environment perception module, and generate a fused feature vector using a collaborative attention mechanism;

[0023] a classification and threat assessment module, configured to perform real-time classification and threat level assessment using a Transformer model and a fully connected neural network;

[0024] an anti-spoofing detection and countermeasure decision module, configured to perform anti-spoofing detection and countermeasure instruction generation on the fused feature vector of the feature fusion module and the modal features.

[0025] In a third aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and wherein the computer program, when executed by the processor, implements any step of the intelligent identification method based on UAV countermeasures according to the first aspect of the present application.

[0026] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any step of the intelligent identification method based on UAV countermeasures according to the first aspect of the present application.

[0027] The present application has the following beneficial effects: by introducing a collaborative attention mechanism, combining MLP supervised learning and Transformer attention mechanism, realizing adaptive modal weight adjustment based on environment vector and threat intelligence, effectively solving the problem of recognition accuracy decline of traditional fixed weight fusion in complex environment (such as strong electromagnetic interference or low visibility), enhancing the recognition ability of disguised UAV through calculation of inter-modal cosine similarity and One-Class SVM anomaly detection, significantly reducing the false positive rate and the false negative rate, and improving the generalization ability of single modal feature extraction, the present application optimizes the radar, radio frequency, visible light and infrared feature extraction process, combines STFT, 1D-CNN, MobileNetV3, EfficientNet-B0 and IRNet model, ensures the robustness of feature extraction in diversified scenarios, significantly improves the environmental adaptability, recognition accuracy and anti-spoofing ability of the UAV countermeasure system, and provides a more efficient and reliable solution for aviation safety and public safety. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0029] Fig. 1 Flow chart of the intelligent identification method based on unmanned aerial vehicle countermeasures in embodiment 1.

[0030] Fig. 2 Structure diagram of the intelligent identification system based on unmanned aerial vehicle countermeasures in embodiment 1.

[0031] Fig. 3 Multimodal feature fusion data structure diagram of the intelligent identification method based on unmanned aerial vehicle countermeasures in embodiment 1. DETAILED DESCRIPTION

[0032] In order to make the above objectives, characteristics and advantages of the present application more apparent, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0033] In the following description, a large number of specific details are set forth in order to facilitate a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the connotation of the present application, therefore the present application is not limited by the specific embodiments disclosed below.

[0034] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an independent or alternative embodiment that excludes other embodiments.

[0035] Embodiment 1, reference Figs. 1-3 , for the first embodiment of the present application, the embodiment provides an intelligent identification method based on unmanned aerial vehicle countermeasures, including the following steps:

[0036] S1, collect real-time environmental parameters to generate an environmental vector, and calculate an initial threat level based on the environmental vector combined with threat intelligence, and generate a multimodal data set based on the initial threat level;

[0037] Use short-time Fourier transform to extract radar echo time-frequency spectrum, input the time-frequency spectrum into MobileNetV3 model to extract radar signal features, use 1D-CNN model to extract features of radio frequency signals, use EfficientNet-B0 model to extract visual features of visible light images, use IRNet to extract infrared image thermal imaging features, and generate a feature vector set;

[0038] Specifically, real-time environmental parameters are collected to generate an environmental vector, which is combined with threat intelligence to calculate an initial threat level. Based on the initial threat level, a multi-modal data set is generated for the system (a comprehensive software and hardware platform for detecting, identifying, and countering UAVs). The UAV countermeasure ground station is equipped with a millimeter wave radar (to detect distance, speed, and angle), a radio frequency receiver (to capture communication and control signals), a high-resolution visible light camera (RGB image), a long-wave infrared camera (thermal imaging), a light sensor (TSL2561, accuracy ±5 lux), an electromagnetic interference detector (spectrum analyzer, accuracy ±0.1 dBm), and a weather station (visibility meter, accuracy ±0.1 km) to collect real-time environmental parameters and generate an environmental vector , :

[0039] ,

[0040] where is the discrete time point at the current time, t is the time variable, and i is the time series index, is the light intensity, representing the ambient light level at the current time , reflecting day and night or weather conditions (such as sunny, foggy), measured directly by the light sensor, is the electromagnetic noise level, representing the electromagnetic environment interference intensity at the current time , reflecting the reliability of radio frequency signals, measured by the electromagnetic interference detector, is the visibility, representing the atmospheric visibility at the current time , affecting the performance of optical and infrared sensors, measured by the visibility meter of the weather station, is the wind speed, representing the wind intensity at the current time , affecting the UAV trajectory and radar signal stability, measured by the wind speed sensor of the weather station;

[0041] The environmental vector is combined with threat intelligence (external information related to UAV activities, including regional UAV activity frequency (such as the number of UAV appearances per hour) and known enemy UAV feature library (such as model, signal characteristics, behavior pattern)) to calculate the initial threat level :

[0042] ,

[0043] where σ is the Sigmoid function, ensuring the output is in [0,1], r is the transpose operation of the vector, is the weight vector of the normalized environmental vector , obtained through MLP supervised learning method, Threat intelligence characteristics (activity frequency and match with a feature database) are generated through queries to intelligence databases (such as military intelligence databases) to reflect the level of potential threats within a region. The threat intelligence feature vector contains the frequency of UAV activity in the region and the matching degree of the feature library. It is generated by querying the intelligence database. B is the bias term with an initial value of 0. It is obtained by training on historical data (including known UAV activities and environmental conditions) through MLP supervised learning.

[0044] according to The sampling frequency is adjusted (e.g., 25Hz for high threat > threshold; otherwise 15Hz). Based on the multimodal data after adjusting the sampling frequency, the data is aligned using a high-precision GPS clock (accuracy ±10ns) to generate a dataset with a unified timestamp. :

[0045] ,

[0046] in, Radar echo (intensity, range, velocity). It is a radio frequency signal (spectral characteristics). Infrared image, This is a visible light image.

[0047] By collecting real-time environmental parameters to generate normalized environmental vectors and combining them with threat intelligence to calculate the initial threat level using MLP supervised learning, adaptive sampling frequency adjustment and multimodal data alignment are achieved, enhancing the environmental adaptability of data collection. Compared with traditional fixed-weight fusion methods, by dynamically adjusting the sampling frequency (25Hz for high threats and 15Hz for low threats) and aligning the data with a high-precision GPS clock, the reliability of multimodal data in complex environments is effectively improved, the false alarm and false negative rates caused by environmental interference are reduced, and the robustness and recognition accuracy of the UAV countermeasure system are significantly improved.

[0048] Furthermore, short-time Fourier transform is used to extract the radar echo time spectrum, which is then input into the MobileNetV3 model to extract radar signal features. A 1D-CNN model is used to extract radio frequency signal features, an EfficientNet-B0 model is used to extract visual features of the visible light image, and IRNet is used to extract thermal imaging features of the infrared image. A feature vector set is then generated to correspond to the radar echo. Feature extraction is performed, and a time-frequency spectrum is generated using the Short Time Fourier Transform (STFT):

[0049] ,

[0050] in, For Hanning windows, the length is based on Adjustment (high threat 128 ms, low threat 64 ms), is the frequency variable, obtained by the Cooley-Tukey FFT method, is the complex exponential function, used to convert the time domain signal into the frequency domain, where ω is the angular frequency, τ is the time variable, and j is the imaginary unit, calculated by the short-time Fourier transform (STFT), is the infinitesimal time increment in the integral, approximated by numerical integration methods in actual calculations;

[0051] The input spectrum is input into the pre-trained MobileNetV3 model, and the network extracts the spatial and frequency features of the radar signal through convolution operations, outputting a fixed-dimensional feature vector:

[0052] ,

[0053] where, ∈ is the feature vector of the radar signal at time , representing the 96-dimensional features extracted after processing the STFT spectrum with MobileNetV3;

[0054] Feature extraction is performed on the radio frequency signal , and the signal energy and spectral entropy are calculated:

[0055] ,

[0056] where, is the signal energy, is the amplitude of the nth frequency component of the radio frequency signal in the frequency domain, calculated from the radio frequency signal by the fast Fourier transform (FFT), is the total number of frequency components, is the spectral entropy, calculated by the Shannon information entropy method, is the normalized probability distribution of the spectrum, representing the probability of frequency component q, obtained by normalizing the spectrum generated by the fast Fourier transform (FFT);

[0057] When the electromagnetic noise level N( ) is greater than the preset maximum value of electromagnetic noise (determined by experimental analysis on the radio frequency dataset), the radio frequency signal is severely disturbed, the signal energy reliability decreases, and the spectral entropy increases Weight, enhance the characterization of signal frequency distribution, using 1D-CNN model (customized model trained from scratch on radio frequency dataset, collect radio frequency dataset, including unmanned aerial vehicle communication and control signal samples (about 100,000, containing signal energy, spectral entropy features), and label signal type (such as enemy, commercial), model architecture design as three layers of convolution (64, 128, 96 neurons), using ReLU activation and maximum pooling, output 96-dimensional feature vector, with cross-entropy loss as the target, using Adam optimizer, training on the dataset for 50 epochs, 32 data per batch. Adjust the hyperparameters through 5-fold cross-validation to ensure the generalization ability of the model, finally achieve more than 95% classification accuracy on the validation set, generate a customized model suitable for radio frequency feature extraction), output the feature vector of the radio frequency signal :

[0058] ,

[0059] Feature extraction is performed on the visible light image If the light intensity u, indicates that the environment is too dark (such as night or extremely low light), the visible light camera image quality is poor, and the feature extraction is unreliable, the threshold value u is determined through experimental verification, and the processing is skipped to avoid invalid calculation, otherwise the pre-trained EfficientNet-B0 model is used to extract high-order visual features of the visible light image ∈ :

[0060] ,

[0061] Extract the thermal imaging features of the infrared image using IRNet (based on ResNet18, using a deep convolutional neural network, directly trained on the infrared unmanned aerial vehicle dataset, optimizing thermal texture feature extraction) :

[0062] ,

[0063] Generate a feature vector set through adaptive feature extraction .

[0064] ​The radar echo is subjected to short-time Fourier transform (STFT) combined with an adaptive Hanning window to generate a time-frequency spectrum and input MobileNetV3 to extract features, the radio frequency signal is subjected to calculation of signal energy and spectral entropy, and the spectral entropy weight is increased in a high electromagnetic noise environment, a customized 1D-CNN model (three layers of convolution, 95% classification accuracy) is combined to enhance the feature robustness, the visible light image is subjected to light threshold screening to avoid invalid calculation, and EfficientNet-B0 is used to extract high-order features, and the infrared image is subjected to extraction of thermal imaging features by using the optimized IRNet, the adaptive extraction and optimization of multi-modal features are realized, a unified 96-dimensional feature vector set is generated, compared with the traditional fixed feature extraction method, the feature extraction precision and reliability in a complex environment are significantly improved, and the false alarm rate and the missing alarm rate are reduced, so that more efficient and robust feature representation is provided for the anti-drone.

[0065] S2, the initial weight of the mode is calculated based on the MLP supervised learning method, the attention weight is calculated through the Transformer attention mechanism, and the feature vector is fused based on the initial weight of the mode and the attention weight through the collaborative attention mechanism;

[0066] The fusion feature vector is used to perform real-time classification and generate a threat level by using a Transformer model;

[0067] Specifically, the initial weight of the mode is calculated based on the MLP supervised learning method, the attention weight is calculated through the Transformer attention mechanism, and the feature vector is fused based on the initial weight of the mode and the attention weight through the collaborative attention mechanism. :

[0068] ,

[0069] wherein, is an adaptive function of the mode j, realized by MLP (2 layers, 128 neurons), is a splicing vector, is an exponential function, that is, , (such as e≈2.718), is used to convert the MLP output to a positive value, so as to ensure the weight normalization, is an adaptive function of the mode k∈{R,F,V,I}, outputting a scalar value, indicating the initial weight contribution of the mode k, the MLP is trained through supervised learning (based on an environment and threat data set, cross-entropy loss, Adam optimizer), and the weight is optimized to reflect the importance of the mode in different environments and threats;

[0070] Data-driven attention weights are calculated using the Transformer attention mechanism (single-layer, 4-head encoder). :

[0071] ,

[0072] Where W is the weight matrix in the Transformer attention mechanism, b is the bias vector, W and b are obtained by training the Transformer encoder (single layer, 4 heads), and tanh is the hyperbolic tangent activation function. Mapping the input to [−1,1] enhances feature representation. For a moment A single modal feature vector, derived from the feature vector set F( ), The weight vectors in the Transformer attention mechanism are obtained by supervised learning training on a multimodal dataset using a Transformer encoder.

[0073] A collaborative attention mechanism is used to determine the initial weights of each modality. According to the environment vector and threat level Adaptive adjustment, leveraging data-driven attention weights Based on modal features The dynamic importance of it will and Multiplication is used as a comprehensive weight, and then combined with the corresponding modal features. Weighted summation to generate a fused feature vector .

[0074] By combining environmental vectors and threat levels with MLP-based supervised learning to calculate initial modality weights, the importance of modalities under different environments (such as illumination and noise) is reflected. The Transformer attention mechanism is used to calculate data-driven attention weights to capture the dynamic importance of modal features. The initial weights are multiplied with the attention weights through a collaborative attention mechanism to generate comprehensive weights. The radar, radio frequency, visible light, and infrared features are weighted and summed to generate a 96-dimensional fused feature vector. This overcomes the limitation of traditional fixed-weight fusion methods in responding insufficiently to environmental changes and significantly improves the robustness of feature fusion and recognition accuracy in complex environments.

[0075] Further, based on the fusion feature vector, a Transformer model is used for real-time classification and threat level generation. A fully connected neural network (three layers, 256-64-class number) is trained on the drone dataset, using cross-entropy loss. The time series fusion features are input into a lightweight Transformer model (two layers, four heads), which analyzes the drone behavior patterns, generates an intent vector, and calculates the dynamic threat level through a fully connected layer (the first five frames of time series fusion feature vectors are input into a lightweight Transformer model (two layers, four attention heads). The model is trained on the drone behavior dataset. The Transformer processes sequences through multi-head self-attention mechanisms, capturing temporal dependencies and behavior patterns such as trajectory changes or communication frequencies, generating an intent vector representing the context of the drone's behavior. The intent vector is input into a fully connected neural network layer, which applies trained weight matrices and biases to transform the vector into a high-dimensional representation. The Softmax activation function is then used to calculate the probability distribution of different threat levels (such as low, medium, and high). The network is optimized through supervised learning on labeled threat scenarios to ensure accurate threat assessment, outputting a dynamic threat level, which is a probability distribution representing the severity of the drone's current threat. The real-time threat level is reflected, and the contrastive loss (InfoNCE) is used in combination with the cross-entropy loss to optimize the model (design the model structure, including a feature extractor (such as a convolutional neural network) to generate high-dimensional feature representations, and a fully connected layer to output classification results. The feature extractor converts input data (such as images or text) into feature vectors, while the fully connected layer maps the feature vectors to class probabilities, ensuring that the model can output both feature representations and classification results. Preprocess the dataset, including data augmentation (such as flipping, cropping) and normalization to improve model robustness. During training, calculate the cross-entropy loss for each batch of data to optimize classification performance, while using the contrastive loss to enhance feature discrimination by making positive sample pairs more similar and negative sample pairs more different. Combine these two losses through weighted combination, with weights adjustable according to experiments to balance classification accuracy and feature learning effect. Use an optimizer (such as Adam) to update model parameters, and adjust weights and hyperparameters on the validation set to optimize performance. Regularly evaluate the model's classification accuracy and feature representation quality on the test set to ensure model convergence and good generalization ability. The model is trained on the drone dataset, and the contrastive loss (InfoNCE) is used in combination with the cross-entropy loss to optimize the model (design the model structure, including a feature extractor (such as a convolutional neural network) to generate high-dimensional feature representations, and a fully connected layer to output classification results. The feature extractor converts input data (such as images or text) into feature vectors, while the fully connected layer maps the feature vectors to class probabilities, ensuring that the model can output both feature representations and classification results. Preprocess the dataset, including data augmentation (such as flipping, cropping) and normalization to improve model robustness. During training, calculate the cross-entropy loss for each batch of data to optimize classification performance, while using the contrastive loss to enhance feature discrimination by making positive sample pairs more similar and negative sample pairs more different. Combine these two losses through weighted combination, with weights adjustable according to experiments to balance classification accuracy and feature learning effect. Use an optimizer (such as Adam) to update model parameters, and adjust weights and hyperparameters on the validation set to optimize performance. Regularly evaluate the model's classification accuracy and feature representation quality on the test set to ensure model convergence and good generalization ability), enhancing the ability to distinguish between disguised targets. The output includes class probability, dynamic threat level, and intent vector.

[0076] The fusion feature vector is processed by a full connection neural network to output a UAV class probability, and a lightweight Transformer model is used to analyze the first five frame time sequence fusion features to capture behavior patterns such as UAV trajectory changes and communication frequencies to generate an intent vector, and then a full connection layer and a Softmax function are used to calculate a dynamic threat level, and a cross-entropy loss and a contrast loss are used to optimize the model to enhance the ability to distinguish disguised targets, overcome the limitations of traditional methods in identifying dynamic behaviors and disguised targets, improve the robustness of the model through data enhancement and normalization preprocessing, and significantly improve the precision and real-time performance of UAV classification and threat assessment in complex environments, providing more efficient and reliable decision-making basis for UAV countermeasures.

[0077] S3, calculating a countermeasure priority through the threat level and the real-time classification result, and generating an anti-deception detection result and a countermeasure instruction based on the countermeasure priority;

[0078] Based on the anti-deception detection result and the countermeasure instruction, dynamic countermeasure instruction execution and effect feedback optimization are performed.

[0079] Specifically, the countermeasure priority is calculated based on the threat level and classification result, the anti-deception detection result and countermeasure instruction are generated based on the countermeasure priority, the cosine similarity between each modal feature is calculated, the feature consistency is evaluated (for each pair of modal features (such as radar and radio frequency, visible light and infrared), the cosine similarity of the feature vectors is calculated, the dot product operation is performed on the two feature vectors, and the product of the norms is divided to obtain a similarity score in the range of [-1, 1], the similarity scores of all modal pairs are summarized to calculate the average consistency score), if the average consistency score is lower than a preset threshold (such as 0.75), it is marked as a potential camouflage target, a One-Class SVM model (based on a radial basis kernel) is used to analyze the fusion feature vector to detect abnormal behavior, and a negative value indicates abnormality, the countermeasure priority is calculated by combining the threat level and classification result (category probability, dynamic threat level and intention vector), the countermeasure priority is calculated based on the classification result, the system receives input data (such as text or file), performs in-depth processing through context analysis capability, generates category probability using classification algorithm to reflect the possibility of input data belonging to each predefined category, evaluates the dynamic threat level to quantify potential risks by combining real-time networking information provided by the proxy system, generates an intention vector through an intention recognition mechanism to capture the potential intention of the input content, integrates the category probability, threat level and intention vector to calculate the countermeasure priority to determine the urgency and priority order of the countermeasures, for example, high threat level and malicious intention vector input will generate high countermeasure priority, prompting the system to take immediate action such as isolating the threat source or strengthening defense measures), if the priority exceeds the threshold and there is no abnormality, a countermeasure instruction containing the target location (provided by radar), category and threat level is generated and sent to the radio frequency jammer, if deception is detected, the weight of the abnormal modal is reduced and the fusion feature vector step is returned to re-fuse, the output results are anti-deception detection result (whether it is a deception) and countermeasure instruction (target location, category, threat level, priority), and precise countermeasures are achieved.

[0080] The cosine similarity between each modal feature is calculated to evaluate the feature consistency, if the average consistency score is lower than a preset threshold, it is marked as a potential camouflage target, the limitations of traditional methods in identifying camouflage UAVs and static countermeasure decision are overcome, and the accuracy of anti-deception detection and dynamic adaptability of countermeasure instruction in complex environments are significantly improved, providing a more accurate and reliable solution for UAV countermeasures.

[0081] Further, based on the anti-fraud detection result and countermeasure instruction, dynamic countermeasure instruction execution and effect feedback optimization are performed to generate an optimized countermeasure strategy using an adaptive countermeasure execution and feedback optimization method. According to the target position and category in the countermeasure instruction (referring to the physical or virtual coordinates of the detected target (such as geographic location or network address), and the category refers to the type of the target (such as unmanned aerial vehicle, malware)), a radio frequency jammer is activated to perform a countermeasure operation (such as signal suppression or physical interception), real-time collection of countermeasure effect data (through radar and infrared sensors to monitor the target state change), and comparison with the category probability and threat level to evaluate the success rate of the countermeasure. If the countermeasure fails and the disguise is detected, adjust the modal weight, regenerate the fusion feature vector and reclassify, if successful, record the effect data to update the training data set of the Transformer model, the result outputs the optimized countermeasure strategy (updated instruction) and effect evaluation report (success rate and target state), realizes dynamic closed-loop countermeasure optimization.

[0082] By dynamically adjusting the modal weight in multi-modal feature fusion, regenerating the fusion feature vector and triggering reclassification, the subsequent identification and countermeasure decision are optimized. If the countermeasure is successful, the effect data is recorded to update the training data set of the Transformer model, further improving the generalization ability and threat assessment accuracy of the model. Compared with the traditional static countermeasure process, the closed-loop feedback mechanism significantly enhances the environmental adaptability and anti-fraud ability of the system, ensuring the efficiency and reliability of the countermeasure strategy in complex and variable scenarios.

[0083] The embodiment also provides an intelligent identification system based on unmanned aerial vehicle countermeasure, which comprises an environment perception and threat initialization module for collecting real-time environment parameters, generating an environment vector and calculating an initial threat level combined with threat intelligence to generate a multi-modal data set;

[0084] A multi-modal feature extraction module for extracting each modal feature and generating a feature vector set;

[0085] A feature fusion and weight calculation module for combining the normalized environment vector of the environment perception module and the initial threat level to generate a fusion feature vector using a collaborative attention mechanism;

[0086] A classification and threat assessment module for real-time classification and threat level assessment using a Transformer model and a fully connected neural network;

[0087] An anti-fraud detection and countermeasure decision module for the fusion feature vector and each modal feature of the feature fusion module to perform anti-fraud detection and countermeasure instruction generation.

[0088] The embodiment also provides a computer device suitable for the intelligent identification method based on anti-unmanned aerial vehicle.

[0089] The computer device can be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is configured to perform wired or wireless communication with an external terminal. The wireless communication can be implemented through WIFI, an operator network, near field communication (NFC) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, a trackball or a touchpad arranged on the shell of the computer device, or an external keyboard, a touchpad or a mouse.

[0090] The embodiment also provides a storage medium having a computer program stored thereon. The program is executed by a processor to implement the intelligent identification method and system based on anti-unmanned aerial vehicle. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk or an optical disk.

[0091] To sum up, by introducing the synergistic attention mechanism, combining MLP supervised learning and Transformer attention mechanism, the adaptive modal weight adjustment based on environment vector and threat intelligence is realized, the problem of recognition accuracy decline of traditional fixed weight fusion under complex environment (such as strong electromagnetic interference or low visibility) is effectively solved, the recognition ability to disguised unmanned aerial vehicle is enhanced through the calculation of inter-modal cosine similarity and One-Class SVM anomaly detection, the false positive rate and the false negative rate are significantly reduced, and the generalization ability of single modal feature extraction is insufficient, the radar, radio frequency, visible light and infrared feature extraction process are optimized, combined with STFT, 1D-CNN, MobileNetV3, EfficientNet-B0 and IRNet model, the robustness of feature extraction in diversified scenes is ensured, the environmental adaptability, recognition accuracy and anti-deception ability of unmanned aerial vehicle countermeasure system are significantly improved, and a more efficient and reliable solution is provided for aviation safety and public safety.

[0092] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. An intelligent identification method based on unmanned aerial vehicle countermeasures, characterized in that: Comprising, Collecting real-time environmental parameters to generate an environmental vector, the environmental vector combining threat intelligence to calculate an initial threat level, and generating a multi-modal data set based on the initial threat level; Using short-time Fourier transform to extract a radar echo time-frequency spectrum, inputting the time-frequency spectrum into a MobileNetV3 model to extract radar signal features, using a 1D-CNN model to extract features of a radio frequency signal, using an EfficientNet-B0 model to extract visual features of a visible light image, using an IRNet to extract thermal imaging features of an infrared image, and generating a feature vector set; Based on the MLP supervised learning method, the initial weights of the modal are calculated, the attention weights are calculated through the Transformer attention mechanism, and the feature vectors are fused based on the initial weights of the modal and the attention weights using the collaborative attention mechanism; Based on the fusion feature vector, a Transformer model is used for real-time classification and threat level generation, countermeasure priority is calculated through threat level and real-time classification result, and anti-deception detection result and countermeasure instruction are generated based on countermeasure priority; Based on the anti-deception detection result and the countermeasure instruction, dynamic countermeasure instruction execution and effect feedback optimization are performed. 2.The method of claim 1, wherein: The collection real-time environmental parameters, generate environment vector, environment vector combination threat intelligence calculation initial threat level, based on initial threat level generation multimodal data collection system deployed in unmanned aerial vehicle countermeasure ground station, equipped with millimeter wave radar, radio frequency receiver, high resolution visible light camera, long wave infrared camera, using light sensor, electromagnetic interference detector, weather station, collection real-time environmental parameters, generate environment vector , and carry out normalization processing, obtain environment vector , combine environment vector and threat intelligence, calculate initial threat level , according to , adjust sampling frequency, based on the adjusted sampling frequency multimodal data, using high precision GPS clock align data, generate uniform timestamp data collection . 3.The method of claim 2, wherein: The radar echo time-frequency spectrum is extracted by using short-time Fourier transform, the time-frequency spectrum is input into a MobileNetV3 model to extract radar signal characteristics, a 1D-CNN model is used to extract the characteristics of the radio frequency signal, and a characteristic vector set is generated to point to the radar echo Feature extraction is performed, a short-time Fourier transform (STFT) is used to generate a time-frequency spectrum, the spectrum is input into a pre-trained MobileNetV3 model, the network extracts spatial and frequency characteristics of the radar signal through convolution operation, and outputs a fixed-dimension characteristic vector ; Feature extraction is performed on the radio frequency signal The signal energy is calculated And the spectral entropy When the electromagnetic noise level N( ) The radio frequency signal is severely disturbed, the signal energy The reliability is reduced, and the spectral entropy is increased The weight enhances the characterization of the signal frequency distribution, and a 1D-CNN model is used to output the feature vector of the radio frequency signal ; For visible light images perform feature extraction. If the light intensity <u, skip the processing to avoid invalid calculations. Otherwise, use the pre-trained EfficientNet-B0 model to extract the high-order vision of visible light images , use IRNet to extract the thermal imaging features of infrared images , generate a set of feature vectors through adaptive feature extraction .​ 4.The method of claim 3, wherein: The collaborative attention mechanism is adopted to fuse the feature vectors based on the modal initial weight and the attention weight The data-driven attention weight is calculated through the Transformer attention mechanism The initial weight of each modal is adopted to fuse the feature vectors through the collaborative attention mechanism The environmental vector and the threat level are adaptively adjusted The data-driven attention weight is utilized The dynamic importance of the modal feature is used to multiply and to obtain the comprehensive weight The corresponding modal feature is weighted and summed to generate the fused feature vector . 5.The method of claim 4, wherein: The Transformer model is used for real-time classification and threat level generation based on the fusion feature vector, a full connection neural network is trained through a UAV data set, a cross-entropy loss is used, a time series fusion feature is input into a lightweight Transformer model, a UAV behavior mode is analyzed to generate an intention vector, and a dynamic threat level is calculated through a full connection layer to reflect the real-time threat degree. A contrast loss is used in combination with a cross-entropy loss to optimize the model and enhance the ability to distinguish disguised targets. The result output category probability, dynamic threat level and intention vector. 6.The method of claim 5, wherein the method further comprises: The countermeasure priority is calculated through the threat level and the classification result, and the anti-deception detection result and the countermeasure instruction are generated based on the countermeasure priority. The cosine similarity between the features of each modal is calculated to evaluate the consistency of the features. If the average consistency score is lower than a preset threshold, it is marked as a potential disguised target. The One-Class SVM model is used to analyze the fusion feature vector to detect abnormal behavior, and the negative value indicates abnormality. The countermeasure priority is calculated based on the threat level and the classification result. If the priority exceeds the threshold and there is no abnormality, a countermeasure instruction containing the target position, category and threat level is generated and sent to the radio frequency jammer. If a disguise is detected, the weight of the abnormal modal is reduced and the fusion feature vector step is performed again. The output result is the anti-deception detection result and the countermeasure instruction. 7.The method of claim 6, wherein: The anti-deception detection result and the countermeasure instruction are used to perform dynamic countermeasure instruction execution and effect feedback optimization. The adaptive countermeasure execution and feedback optimization method is used to generate an optimized countermeasure strategy. According to the target position and category in the countermeasure instruction, the radio frequency jammer is activated to perform the countermeasure operation. Real-time countermeasure effect data is collected and compared with the category probability and threat level to evaluate the success rate of the countermeasure. If the countermeasure fails and a disguise is detected, the modal weight is adjusted, the fusion feature vector is regenerated and reclassified. If the countermeasure is successful, the effect data is recorded to update the training data set of the Transformer model. The optimized countermeasure strategy and effect evaluation report are outputted, and the dynamic closed-loop countermeasure optimization is realized.

8. The intelligent identification system based on anti-UAV measures according to any one of claims 1-7, characterized in that: The method comprises the following steps: An environment perception and threat initialization module is used to collect real-time environment parameters, generate an environment vector, and calculate an initial threat level in combination with threat intelligence to generate a multi-modal data set. A multi-modal feature extraction module is used to extract modal features and generate a feature vector set. A feature fusion and weight calculation module is used to combine the normalized environment vector of the environment perception module and the initial threat level to generate a fusion feature vector using a collaborative attention mechanism. A classification and threat evaluation module is used to perform real-time classification and threat level evaluation using a Transformer model and a fully connected neural network. An anti-deception detection and countermeasure decision module is used to perform anti-deception detection and countermeasure instruction generation on the fusion feature vector and the modal features of the feature fusion module. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: The processor executes the computer program to realize the steps of the intelligent identification method based on unmanned aerial vehicle countermeasures according to any one of claims 1-7.

10. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to realize the steps of the intelligent identification method based on unmanned aerial vehicle countermeasures according to any one of claims 1-7.

Images