Low-altitude unmanned aerial vehicle detection system false alarm judgment method

Abstract

The invention provides a low-altitude unmanned aerial vehicle detection system false alarm judgment method, and belongs to the technical field of unmanned aerial vehicle judgment, and the method comprises the steps: obtaining an original unmanned aerial vehicle data set and a to-be-judged unmanned aerial vehicle data set from a low-altitude unmanned aerial vehicle detection system through an unmanned aerial vehicle data collection device; preprocessing the original unmanned aerial vehicle data set and the to-be-determined unmanned aerial vehicle data set to obtain a first preprocessed unmanned aerial vehicle data set and a second preprocessed unmanned aerial vehicle data set; and constructing a training model, and performing model analysis on the training model through the first preprocessed unmanned aerial vehicle data set to obtain a false alarm judgment model. According to the method, the unmanned aerial vehicle target and the interference source can be effectively distinguished, the accuracy of false alarm judgment is improved, the problem of model bias caused by low false alarm sample proportion is solved, and complex and changeable low-altitude environment requirements and strict engineering index requirements are met.

Description

Technical Field

[0001] The present invention mainly relates to the technical field of unmanned aerial vehicle (UAV) determination, and particularly to a method for false alarm determination of a low-altitude UAV detection system. Background Art

[0002] With the rapid development of the low-altitude economy, UAVs are increasingly widely used in various fields, and the surveillance and detection of low-altitude UAVs have become a key link in ensuring airspace safety. In a low-altitude UAV detection system, the detection of non-cooperative targets (UAVs without the RID / ADS-B broadcast function enabled) relies on multi-modal devices such as radar, optoelectronics, and electromagnetic detection. However, due to interference from complex environments (such as birds, electromagnetic noise, building reflections, etc.), the system is prone to false alarm phenomena. False alarms not only occupy system resources, interfere with the judgment of management personnel, but may also lead to misactivation of emergency responses, affecting the efficiency and safety of low-altitude airspace management.

[0003] Existing false alarm determination methods mostly adopt single-device linkage verification (such as linking an optoelectronic camera to confirm after the detection system discovers a target), do not fully utilize the complementarity of multi-modal detection data, and the determination model lacks the ability of dynamic optimization, making it difficult to adapt to the complex and changeable low-altitude environment. At the same time, the network structure of traditional models has insufficient ability to extract interference source features, and the loss function design does not fully consider the class imbalance problem of false alarm determination (the proportion of false alarm samples is usually low), resulting in low accuracy of false alarm recognition and poor generalization ability of the model, and unable to meet the strict requirement that the false alarm probability of the low-altitude UAV detection system ≤ 5%. Therefore, there is an urgent need for a false alarm determination method that can accurately extract multi-modal features, dynamically optimize the determination model, and adapt to the class imbalance scenario. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for false alarm determination of a low-altitude UAV detection system in view of the deficiencies of the prior art.

[0005] The technical solution of the present invention to solve the above technical problem is as follows: A method for false alarm determination of a low-altitude UAV detection system includes the following steps: Obtain the original UAV data set and the to-be-determined UAV data set from the low-altitude UAV detection system through a UAV data acquisition device; Preprocess the original UAV data set and the to-be-determined UAV data set respectively to obtain the first preprocessed UAV data set corresponding to the original UAV data set and the second preprocessed UAV data set corresponding to the to-be-determined UAV data set; Construct a training model, and perform model analysis on the training model through the first preprocessed UAV data set to obtain a false alarm determination model; Predict the second preprocessed UAV data set through the false alarm determination model to obtain a false alarm probability; The false alarm determination result of the low-altitude drone detection system is obtained by analyzing the false alarm probability based on the data acquisition device of the drone.

[0006] Another technical solution of the present invention to solve the above-mentioned technical problems is as follows: A false alarm determination device for a low-altitude unmanned aerial vehicle detection system, comprising: The dataset acquisition module is used to obtain the raw UAV dataset and the UAV dataset to be judged from the low-altitude UAV detection system through the UAV data acquisition device. The preprocessing module is used to preprocess the original drone dataset and the drone dataset to be judged respectively to obtain a first preprocessed drone dataset corresponding to the original drone dataset and a second preprocessed drone dataset corresponding to the drone dataset to be judged. The model analysis module is used to construct a training model and perform model analysis on the training model using the first preprocessed UAV dataset to obtain a false alarm determination model. The prediction module is used to predict the false alarm probability of the second preprocessed UAV dataset using the false alarm determination model. The judgment result acquisition module is used to analyze the false alarm probability based on the UAV data acquisition device to obtain the false alarm judgment result of the low-altitude UAV detection system.

[0007] Based on the above-mentioned false alarm determination method for low-altitude unmanned aerial vehicle (UAV) detection systems, the present invention also provides a false alarm determination system for low-altitude UAV detection systems.

[0008] Another technical solution of the present invention to solve the above-mentioned technical problems is as follows: a false alarm determination system for a low-altitude unmanned aerial vehicle (UAV) detection system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the false alarm determination method for a low-altitude UAV detection system as described above.

[0009] Based on the above-mentioned false alarm determination method for low-altitude unmanned aerial vehicle (UAV) detection systems, the present invention also provides a computer-readable storage medium.

[0010] Another technical solution of the present invention to solve the above-mentioned technical problems is as follows: a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the false alarm determination method of the low-altitude unmanned aerial vehicle detection system as described above.

[0011] The beneficial effects of this invention are as follows: The original UAV dataset and the dataset to be judged are obtained from the low-altitude UAV detection system through a UAV data acquisition device. Preprocessing of the original UAV dataset and the dataset to be judged yields a first preprocessed UAV dataset and a second preprocessed UAV dataset. Model analysis of the training model using the first preprocessed UAV dataset yields a false alarm judgment model. The false alarm probability is obtained by predicting the second preprocessed UAV dataset using the false alarm judgment model. Based on the judgment analysis of the false alarm probability by the UAV data acquisition device, the false alarm judgment result of the low-altitude UAV detection system is obtained. This effectively distinguishes UAV targets from interference sources, improves the accuracy of false alarm judgment, solves the model bias problem caused by the low proportion of false alarm samples, and meets the needs of complex and ever-changing low-altitude environments and stringent engineering performance requirements. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating the false alarm determination method for a low-altitude unmanned aerial vehicle (UAV) detection system provided in an embodiment of the present invention. Figure 2 This is a block diagram of a false alarm determination device for a low-altitude unmanned aerial vehicle (UAV) detection system provided in an embodiment of the present invention. Detailed Implementation

[0013] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0014] Figure 1 This is a flowchart illustrating the false alarm determination method for a low-altitude unmanned aerial vehicle (UAV) detection system provided in an embodiment of the present invention.

[0015] like Figure 1 As shown, a method for determining false alarms in a low-altitude unmanned aerial vehicle (UAV) detection system includes the following steps: S1: Obtain the raw drone dataset and the drone dataset to be judged from the low-altitude drone detection system through the drone data acquisition device; S2: Preprocess the original drone dataset and the drone dataset to be judged respectively to obtain a first preprocessed drone dataset corresponding to the original drone dataset and a second preprocessed drone dataset corresponding to the drone dataset to be judged. S3: Construct a training model, and perform model analysis on the training model using the first preprocessed UAV dataset to obtain a false alarm determination model; S4: The false alarm probability is obtained by predicting the second preprocessed UAV dataset using the false alarm determination model; S5: Based on the analysis of the false alarm probability by the UAV data acquisition device, the false alarm determination result of the low-altitude UAV detection system is obtained.

[0016] In the above embodiments, the original UAV dataset and the UAV dataset to be judged are obtained from the low-altitude UAV detection system through the UAV data acquisition device. The original UAV dataset and the UAV dataset to be judged are preprocessed to obtain a first preprocessed UAV dataset and a second preprocessed UAV dataset. The false alarm judgment model is obtained by model analysis of the training model through the first preprocessed UAV dataset. The false alarm probability is obtained by predicting the second preprocessed UAV dataset through the false alarm judgment model. The false alarm judgment result of the low-altitude UAV detection system is obtained based on the judgment analysis of the false alarm probability by the UAV data acquisition device. This can effectively distinguish UAV targets from interference sources, improve the accuracy of false alarm judgment, solve the model bias problem caused by the low proportion of false alarm samples, and meet the needs of complex and ever-changing low-altitude environment and stringent engineering performance requirements.

[0017] Optionally, as an embodiment of the present invention, the UAV data acquisition device includes a 5G-A device, a TDOA device, an AOA device, an optoelectronic camera, an electromagnetic detection device, a protocol parsing device, and a radar device; The process of obtaining the raw UAV dataset and the UAV dataset to be judged from the low-altitude UAV detection system through the UAV data acquisition device includes: The 5G-A device obtains the location data to be determined and multiple raw location data from the low-altitude drone detection system. The TDOA device obtains the time-domain data to be determined and multiple raw time-domain data from the low-altitude unmanned aerial vehicle detection system. The AOA device obtains the angle data to be determined and multiple raw angle data from the low-altitude unmanned aerial vehicle detection system. The photoelectric camera obtains the photoelectric image to be determined and multiple raw photoelectric images from the low-altitude UAV detection system. The electromagnetic detection device obtains the electromagnetic data to be determined and multiple raw electromagnetic data from the low-altitude UAV detection system. The protocol parsing device obtains the communication protocol of the drone to be determined and multiple original drone communication protocols from the low-altitude drone detection system. The radar device obtains radar data to be determined and multiple raw radar data from the low-altitude UAV detection system. The raw UAV dataset includes multiple raw positioning data, multiple raw time-domain data, multiple raw angle data, multiple raw photoelectric images, multiple raw electromagnetic data, multiple raw UAV communication protocols, and multiple raw radar data. The UAV dataset to be determined includes the positioning data to be determined, the time-domain data to be determined, the angle data to be determined, the photoelectric images to be determined, the electromagnetic data to be determined, the UAV communication protocols to be determined, and the radar data to be determined.

[0018] Specifically, 5G-A equipment collects high-precision latitude and longitude, altitude, signal propagation delay, signal strength (RSRP), channel state information (CSI), and Doppler frequency offset of the target (i.e., the positioning data to be determined and multiple raw positioning data). TDOA equipment: collects timestamp data of the target signal arriving at multiple TDOA base stations and time synchronization deviation correction data between base stations (i.e., time domain data to be determined and multiple raw time domain data). AOA equipment: collects the horizontal angle of arrival, vertical angle of arrival, and angular resolution data of the target signal (i.e., the angle data to be determined and multiple raw angle data).

[0019] It should be understood that the photoelectric camera: acquires high-definition video images of the target area (frame rate ≥ 30 frames / second) (i.e., the photoelectric image to be judged and multiple original photoelectric images), and records the target shape, motion trajectory, color features, etc. in the image.

[0020] It should be understood that the raw data (i.e., the raw UAV dataset and the UAV dataset to be judged) of the radar, photoelectric, electromagnetic detection, protocol parsing, ADS-B, Beidou and other equipment in the low-altitude UAV detection system are collected within the specified test period, including: target number, target type, timestamp, detection result, coordinate position, speed, heading angle, flight status, video images, raw signal data, etc.

[0021] Specifically, electromagnetic detection equipment: collects electromagnetic signal spectrum data (i.e., electromagnetic data to be determined and multiple raw electromagnetic data) of the target area, and extracts features such as signal frequency, bandwidth, and modulation method; Protocol parsing module (i.e., protocol parsing device): detects whether there is a drone-specific communication protocol (such as DJI-specific protocol, general image transmission protocol, etc.) (i.e., the drone communication protocol to be determined and multiple original drone communication protocols); Radar equipment: Supplement the collection of target data such as speed, heading angle, and radar cross-section (RCS) (i.e., radar data to be determined and multiple raw radar data).

[0022] In the above embodiments, the raw drone dataset and the drone dataset to be judged are obtained from the low-altitude drone detection system by the drone data acquisition device, which can effectively distinguish between drone targets and interference sources, improve the accuracy of false alarm judgment, and solve the model bias problem caused by the low proportion of false alarm samples.

[0023] Optionally, as an embodiment of the present invention, the process of preprocessing the original drone dataset and the drone dataset to be judged respectively to obtain a first preprocessed drone dataset corresponding to the original drone dataset and a second preprocessed drone dataset corresponding to the drone dataset to be judged includes: Alignment processing is performed on the data in the original drone dataset and the data in the drone dataset to be judged, respectively, to obtain a first aligned drone dataset corresponding to the original drone dataset and a second aligned drone dataset corresponding to the drone dataset to be judged; Coordinate transformations are performed on the first aligned UAV dataset and the second aligned UAV dataset respectively to obtain a first transformed UAV dataset corresponding to the original UAV dataset and a second transformed UAV dataset corresponding to the UAV dataset to be judged. Outlier processing is performed on the first transformed drone dataset and the second transformed drone dataset respectively to obtain a first preprocessed drone dataset corresponding to the original drone dataset and a second preprocessed drone dataset corresponding to the drone dataset to be judged.

[0024] It should be understood that timestamp alignment of data from different devices is achieved based on GNSS-PPS or IEEE1588 PTP time synchronization mechanisms.

[0025] Specifically, coordinate transformation tools are used to convert latitude and longitude data (i.e., the first aligned UAV dataset and the second aligned UAV dataset) into UTM coordinates to complete coordinate system one.

[0026] It should be understood that invalid data with a missing rate exceeding 10% or a signal strength below a threshold should be removed.

[0027] In the above embodiments, the original drone dataset and the drone dataset to be judged are preprocessed to obtain the first preprocessed drone dataset and the second preprocessed drone dataset, which realizes the timestamp alignment of data from different devices and completes the removal of abnormal data.

[0028] Optionally, as an embodiment of the present invention, the process of performing model analysis on the trained model using the first preprocessed UAV dataset to obtain a false alarm determination model includes: Feature extraction analysis is performed on the first preprocessed UAV dataset to obtain multiple original positioning feature vectors, multiple original time domain feature vectors, multiple original angle feature vectors, multiple original image feature vectors, multiple original electromagnetic signal feature vectors, multiple original UAV communication feature vectors, and multiple original radar signal feature vectors. The training model is trained using all the original positioning feature vectors, all the original time-domain feature vectors, all the original angle feature vectors, all the original image feature vectors, all the original electromagnetic signal feature vectors, all the original UAV communication feature vectors, and all the original radar signal feature vectors to obtain a false alarm determination model.

[0029] In the above embodiments, the training model is analyzed using the first preprocessed UAV dataset to obtain a false alarm determination model, which can effectively distinguish between UAV targets and interference sources, improve the accuracy of false alarm determination, solve the model bias problem caused by the low proportion of false alarm samples, and meet the requirements of complex and ever-changing low-altitude environment and stringent engineering indicators.

[0030] Optionally, as an embodiment of the present invention, the first preprocessed UAV dataset includes multiple preprocessed positioning data, multiple preprocessed time-domain data, multiple preprocessed angle data, multiple preprocessed photoelectric images, multiple preprocessed electromagnetic data, multiple preprocessed UAV communication protocols, and multiple preprocessed radar data. The process of performing feature extraction analysis on the first preprocessed UAV dataset to obtain multiple original positioning feature vectors, multiple original time-domain feature vectors, multiple original angle feature vectors, multiple original image feature vectors, multiple original electromagnetic signal feature vectors, multiple original UAV communication feature vectors, and multiple original radar signal feature vectors includes: The standard deviation of each of the preprocessed positioning data is calculated using the sliding window standard deviation algorithm to obtain the calculated positioning data corresponding to each of the preprocessed positioning data. The Kalman filter algorithm is used to filter each of the calculated positioning data to obtain the filtered positioning data corresponding to each of the preprocessed positioning data. Each of the filtered positioning data is processed by Fast Fourier Transform to obtain the original positioning feature vector corresponding to each of the preprocessed positioning data. Outlier processing is performed on each of the preprocessed time-domain data to obtain the time-domain data to be processed corresponding to each of the preprocessed time-domain data. The linear interpolation calibration algorithm is used to calibrate each of the time-domain data to be processed, so as to obtain calibrated time-domain data corresponding to each of the preprocessed time-domain data. Wavelet transform is performed on each of the calibrated time-domain data to extract the original time-domain feature vectors corresponding to each of the preprocessed time-domain data. The median filtering algorithm is used to filter each of the preprocessed angle data to obtain the filtered angle data corresponding to each of the preprocessed angle data. Each of the filtered angle data is subjected to first-order difference processing to obtain the difference-processed angle data corresponding to each of the preprocessed angle data. Pearson correlation coefficients are calculated for each of the differentially processed angle data, and the calculation results are used as the original angle feature vectors to obtain the original angle feature vectors corresponding to each of the preprocessed angle data. Feature extraction is performed on each of the preprocessed photoelectric images using a pre-built ResNet50 model to obtain the original image feature vector corresponding to each of the preprocessed photoelectric images. Wavelet transform is performed on each of the preprocessed electromagnetic data to extract the original electromagnetic signal feature vector corresponding to each of the preprocessed electromagnetic data. The pre-constructed convolutional neural network is used to extract features from each of the pre-processed UAV communication protocols to obtain the original UAV communication feature vectors corresponding to each of the pre-processed UAV communication protocols. Wavelet transform is performed on each of the preprocessed radar data to extract the original radar signal feature vector corresponding to each of the preprocessed radar data.

[0031] In the above embodiments, feature extraction and analysis are performed on the first preprocessed UAV dataset to obtain multiple original positioning feature vectors, multiple original time domain feature vectors, multiple original angle feature vectors, multiple original image feature vectors, multiple original electromagnetic signal feature vectors, multiple original UAV communication feature vectors, and multiple original radar signal feature vectors. By making full use of the complementarity of the data from each device, the UAV target and the interference source can be effectively distinguished, thereby improving the accuracy of false alarm identification.

[0032] Optionally, as an embodiment of the present invention, the training model includes a first fully connected layer, an attention network, a feature enhancement layer, a second fully connected layer, and a Softmax function layer; The process of training the training model using all the original positioning feature vectors, all the original time-domain feature vectors, all the original angle feature vectors, all the original image feature vectors, all the original electromagnetic signal feature vectors, all the original UAV communication feature vectors, and all the original radar signal feature vectors to obtain the false alarm determination model includes: The first fully connected layer performs mapping processing on each of the original positioning feature vectors, original temporal feature vectors, original angle feature vectors, original image feature vectors, original electromagnetic signal feature vectors, original UAV communication feature vectors, and original radar signal feature vectors to obtain a first mapped positioning feature vector corresponding to each of the original positioning feature vectors, a first mapped temporal feature vector corresponding to each of the original temporal feature vectors, a first mapped angle feature vector corresponding to each of the original angle feature vectors, a first mapped image feature vector corresponding to each of the original image feature vectors, a first mapped electromagnetic signal feature vector corresponding to each of the original electromagnetic signal feature vectors, a first mapped UAV communication feature vector corresponding to each of the original UAV communication feature vectors, and a first mapped radar signal feature vector corresponding to each of the original radar signal feature vectors. The attention network is used to perform fusion analysis on each of the first mapped localization feature vectors, the first mapped time-domain feature vectors corresponding to each of the original time-domain feature vectors, the first mapped angle feature vectors corresponding to each of the original angle feature vectors, the first mapped image feature vectors corresponding to each of the original image feature vectors, the first mapped electromagnetic signal feature vectors corresponding to each of the original electromagnetic signal feature vectors, the first mapped UAV communication feature vectors corresponding to each of the original UAV communication feature vectors, and the first mapped radar signal feature vectors corresponding to each of the original radar signal feature vectors, to obtain multiple original fused feature vectors. The feature enhancement layer performs feature enhancement analysis on each of the original fused feature vectors to obtain the enhanced fused feature vectors corresponding to each of the original fused feature vectors. The second fully connected layer is used to map each of the enhanced fused feature vectors to obtain the mapped fused feature vectors corresponding to each of the original fused feature vectors. The Softmax function layer is used to predict each of the mapped and fused feature vectors to obtain the prediction probability corresponding to each of the original fused feature vectors. Import multiple real labels, perform loss function analysis on all real labels and all predicted probabilities, and obtain the target loss function; The training model is updated with parameters according to the target loss function. The training model is then rebuilt after the parameter update until a preset number of iterations is reached. The training model with updated parameters is then used as the false alarm detection model.

[0033] Preferably, the preset number of iterations can be 50.

[0034] It should be understood that multiple parallel input branches are set up to receive preprocessed structured data features (i.e., the original UAV communication feature vector), image features (i.e., the original image feature vector), and signal timing features (i.e., the original electromagnetic signal feature vector and the original radar signal feature vector), respectively. Each input branch maps the features to a unified dimension (512-dimensional) through a fully connected layer (i.e., the first fully connected layer).

[0035] Specifically, a fully connected layer (i.e., the second fully connected layer) is used to map the fused and enhanced features (i.e., the enhanced fused feature vector) to a 2D output (non-false alarm / false alarm), and the class probability (i.e., the prediction probability) is output through the Softmax function (i.e., the Softmax function layer).

[0036] In the above embodiments, a false alarm determination model is obtained by training the training model with the original positioning feature vector, original time domain feature vector, original angle feature vector, original image feature vector, original electromagnetic signal feature vector, original UAV communication feature vector, and original radar signal feature vector. This model can effectively distinguish between UAV targets and interference sources, improve the accuracy of false alarm determination, solve the model bias problem caused by the low proportion of false alarm samples, and meet the requirements of complex and ever-changing low-altitude environments and stringent engineering performance requirements.

[0037] Optionally, as an embodiment of the present invention, the attention network includes multiple third fully connected layers and a Sigmoid activation function layer; The process of fusing and analyzing each of the first mapped localization feature vectors, the first mapped time-domain feature vectors corresponding to each of the original time-domain feature vectors, the first mapped angle feature vectors corresponding to each of the original angle feature vectors, the first mapped image feature vectors corresponding to each of the original image feature vectors, the first mapped electromagnetic signal feature vectors corresponding to each of the original electromagnetic signal feature vectors, the first mapped UAV communication feature vectors corresponding to each of the original UAV communication feature vectors, and the first mapped radar signal feature vectors corresponding to each of the original radar signal feature vectors through the attention network to obtain multiple original fused feature vectors includes: The first mapped positioning feature vector, the first mapped time-domain feature vector corresponding to each of the original time-domain feature vectors, the first mapped angle feature vector corresponding to each of the original angle feature vectors, the first mapped image feature vector corresponding to each of the original image feature vectors, the first mapped electromagnetic signal feature vector corresponding to each of the original electromagnetic signal feature vectors, the first mapped UAV communication feature vector corresponding to each of the original UAV communication feature vectors, and the first mapped radar signal feature vector corresponding to each of the original radar signal feature vectors are mapped through multiple third fully connected layers. This results in the second mapped positioning feature vector corresponding to each of the original positioning feature vectors, the second mapped time-domain feature vector corresponding to each of the original time-domain feature vectors, the second mapped angle feature vector corresponding to each of the original angle feature vectors, the second mapped image feature vector corresponding to each of the original image feature vectors, the second mapped electromagnetic signal feature vector corresponding to each of the original electromagnetic signal feature vectors, the second mapped UAV communication feature vector corresponding to each of the original UAV communication feature vectors, and the second mapped radar signal feature vector corresponding to each of the original radar signal feature vectors. The Sigmoid activation function layer performs weighting on each of the second-mapped positioning feature vectors, the second-mapped time-domain feature vectors corresponding to each of the original time-domain feature vectors, the second-mapped angle feature vectors corresponding to each of the original angle feature vectors, the second-mapped image feature vectors corresponding to each of the original image feature vectors, the second-mapped electromagnetic signal feature vectors corresponding to each of the original electromagnetic signal feature vectors, the second-mapped UAV communication feature vectors corresponding to each of the original UAV communication feature vectors, and the second-mapped radar signal feature vectors corresponding to each of the original radar signal feature vectors. This yields positioning attention weights corresponding to each of the original positioning feature vectors, time-domain attention weights corresponding to each of the original time-domain feature vectors, angle attention weights corresponding to each of the original angle feature vectors, image attention weights corresponding to each of the original image feature vectors, electromagnetic attention weights corresponding to each of the original electromagnetic signal feature vectors, UAV communication attention weights corresponding to each of the original UAV communication feature vectors, and radar attention weights corresponding to each of the original radar signal feature vectors. The first formula is used to calculate the following: for each of the first mapped positioning feature vectors, the positioning attention weights corresponding to each of the original positioning feature vectors, the first mapped time-domain feature vectors corresponding to each of the original time-domain feature vectors, the time-domain attention weights corresponding to each of the original time-domain feature vectors, the first mapped angle feature vectors corresponding to each of the original angle feature vectors, the angle attention weights corresponding to each of the original angle feature vectors, the first mapped image feature vectors corresponding to each of the original electromagnetic signal feature vectors, the first mapped electromagnetic signal feature vectors corresponding to each of the original electromagnetic signal feature vectors, the first mapped UAV communication feature vectors corresponding to each of the original UAV communication feature vectors, the first mapped radar signal feature vectors corresponding to each of the original radar signal feature vectors, the image attention weights corresponding to each of the original image feature vectors, the electromagnetic attention weights corresponding to each of the original electromagnetic signal feature vectors, the UAV communication attention weights corresponding to each of the original UAV communication feature vectors, and the radar attention weights corresponding to each of the original radar signal feature vectors, multiple original fused feature vectors are obtained. The first formula is: , in, For the first One original fused feature vector, For the first Each positioning attention weight, For the first Each temporal attention weight, For the first Attention weights from various angles For the first Image attention weights, For the first Each electromagnetic attention weight, For the first Individual drone communication attention weights, For the first Each radar attention weight, For the first After the first mapping, the feature vector is located. For the first The first mapped temporal feature vector For the first The first mapped angular feature vector For the first The first mapped image feature vector For the first The first mapped electromagnetic signal feature vector For the first The first mapped UAV communication feature vector For the first The first mapped radar signal feature vector.

[0038] Specifically, an attention mechanism is used to weight and fuse multimodal features, and the calculation formula is as follows: , in, For the first The feature vector of a modality For the first Attention weights for each modality For the number of modes, It is obtained through adaptive learning via an attention subnetwork (composed of two fully connected layers (i.e., the third fully connected layer) + a Sigmoid activation function (i.e., a Sigmoid activation function layer)) to highlight effective modal features and suppress interfering modal features.

[0039] In the above embodiments, multiple original fused feature vectors are obtained by fusing and analyzing the first mapped localization feature vector, the first mapped temporal feature vector, the first mapped angle feature vector, the first mapped image feature vector, the first mapped electromagnetic signal feature vector, the first mapped UAV communication feature vector, and the first mapped radar signal feature vector through an attention network. This highlights the effective modal features and suppresses the interfering modal features, and fuses multimodal features. It solves the model bias problem caused by the low proportion of false alarm samples, and meets the requirements of complex and ever-changing low-altitude environments and stringent engineering performance requirements.

[0040] Optionally, as an embodiment of the present invention, the feature enhancement layer includes a plurality of sequentially arranged residual blocks, and the process of performing feature enhancement analysis on each of the original fused feature vectors through the feature enhancement layer to obtain the enhanced fused feature vectors corresponding to each of the original fused feature vectors includes: The first residual block performs feature enhancement processing on each of the original fused feature vectors, and the processing result is used as the input of the next residual block until the last residual block is passed, thereby obtaining the enhanced fused feature vectors corresponding to each of the original fused feature vectors. The process of performing feature enhancement processing on each of the original fused feature vectors through the first residual block is specifically as follows: The residual block includes multiple convolutional layers, batch normalization layers, and ReLU activation function layers. The multiple convolutional layers extract features from each of the original fused feature vectors to obtain extracted fused feature vectors corresponding to each of the original fused feature vectors. The batch normalization layer normalizes each of the extracted fused feature vectors to obtain normalized fused feature vectors corresponding to each of the original fused feature vectors. Each of the original fused feature vectors is element-wise added to the normalized fused feature vectors corresponding to each original fused feature vector to obtain the summed fused feature vectors corresponding to each original fused feature vector. The ReLU activation function layer is used to map each of the summed and fused feature vectors to obtain the processed fused feature vectors corresponding to each of the original fused feature vectors.

[0041] Preferably, the number of residual blocks can be 3.

[0042] Specifically, the feature enhancement layer consists of three residual blocks. Each residual block contains two convolutional layers (3×3 kernel size, stride 1), a batch normalization layer (BN layer), and a ReLU activation function (i.e., ReLU activation function layer). The residual connections alleviate the gradient vanishing problem in deep networks and enhance the model's ability to extract complex interference features.

[0043] In the above embodiments, feature enhancement analysis is performed on each original fused feature vector through the feature enhancement layer to obtain enhanced fused feature vectors, which alleviates the gradient vanishing problem in deep networks, strengthens the model's ability to extract complex interference features, and improves the model's generalization ability.

[0044] Optionally, as an embodiment of the present invention, the process of performing loss function analysis on all the true labels and all the predicted probabilities to obtain the target loss function includes: The target loss function is obtained by calculating all the true labels and all the predicted probabilities using the second equation, which is: , in, , , in, Let be the target loss function. For balance coefficient, For the first loss function, For the second loss function, For the first A real label, To predict the number of probabilities, For the first Each predicted probability, For category weights, For focusing parameters.

[0045] Specifically, to address the class imbalance problem (few false alarm samples and many non-false alarm samples) in false alarm determination, a weighted hybrid loss function is designed, combining the advantages of cross-entropy loss and focal loss, as shown in the following formula: , in, The cross-entropy loss (i.e., the first loss function) is calculated using the following formula: , The label represents the true alarm (1 for non-false alarm, 0 for false alarm). Predict the probability for the model (i.e., predict the probability). The improved focal loss (i.e., the second loss function) is calculated using the following formula: , This is the focus parameter (value 2). By assigning class weights (setting the weight of false alarm samples to 5 and the weight of non-false alarm samples to 1), and adjusting the weights and focusing parameters, the loss contribution of majority class samples is reduced, thereby increasing the model's attention to minority class false alarm samples. This is a balancing factor (value 0.4), used to adjust the ratio of cross-entropy loss to focal loss.

[0046] In the above embodiments, loss function analysis is performed on all real labels and all predicted probabilities to obtain the target loss function, which can adapt to the class imbalance scenario, solve the model bias problem caused by the low proportion of false alarm samples, and improve the recognition accuracy of minority class false alarm samples.

[0047] Optionally, as an embodiment of the present invention, the UAV data acquisition device includes a photoelectric camera, and the process of determining and analyzing the false alarm probability based on the UAV data acquisition device to obtain the false alarm determination result of the low-altitude UAV detection system includes: Determine whether the false alarm probability is greater than or equal to a preset probability. If so, use the first preset determination result as the false alarm determination result of the low-altitude UAV detection system. If not, obtain the photoelectric image to be processed from the low-altitude UAV detection system through the photoelectric camera. Determine whether the photoelectric image to be processed meets a preset condition. If yes, the first preset determination result is used as the false alarm determination result of the low-altitude UAV detection system. If no, the second preset determination result is used as the false alarm determination result of the low-altitude UAV detection system. The preset condition is that there is no UAV target in the photoelectric image to be processed or there is a preset non-UAV target in the photoelectric image to be processed.

[0048] Preferably, the preset probability can be 0.8.

[0049] It should be understood that the non-drone target can be a bird.

[0050] Specifically, the model outputs the probability p of a target being a false alarm (i.e., false alarm probability). If p ≥ 0.8, it is judged as a false alarm (i.e., the first preset judgment result); if p < 0.8, the photoelectric camera is linked to perform a second confirmation on the grid where the target is located. If the photoelectric camera does not detect the target or detects a non-UAV target (such as a bird), it is judged as a false alarm (i.e., the first preset judgment result); otherwise, it is judged as a non-false alarm (i.e., the second preset judgment result).

[0051] In the above embodiments, the false alarm probability is determined and analyzed by the UAV data acquisition device to obtain the false alarm determination result of the low-altitude UAV detection system. This solves the model bias problem caused by the low proportion of false alarm samples and meets the requirements of complex and ever-changing low-altitude environment and stringent engineering indicators.

[0052] Optionally, as another embodiment of the present invention, the present invention relates to the field of low-altitude surveillance and detection technology, specifically to a method for determining false alarms in a low-altitude unmanned aerial vehicle (UAV) detection system. This method is applicable to the identification and optimization of false alarms in the detection process of non-cooperative targets in an urban low-altitude integrated management and service platform, and can effectively reduce the false alarm probability of the low-altitude UAV detection system and improve the detection accuracy.

[0053] Optionally, as another embodiment of the present invention, the purpose of the present invention is to provide a false alarm determination method for a low-altitude unmanned aerial vehicle (UAV) detection system. By optimizing the network structure and loss function of the determination model and combining the fusion characteristics of multimodal detection data, the accuracy of false alarm identification and the generalization ability of the model are improved, and the probability of false alarms is effectively controlled, providing reliable technical support for low-altitude airspace management.

[0054] Optionally, as another embodiment of the present invention, the present invention further includes: automatically storing the original signal data of the detection device that triggers the false alarm event, environmental snapshot data, and photoelectric camera images to form a structured false alarm negative sample library; every 24 hours, using an incremental training method, adding newly added false alarm samples to the training set to dynamically update the model; after the update, verifying the model performance through a test set to ensure that the model's accuracy in identifying similar interference sources is improved without reducing the overall detection performance of the system.

[0055] Optionally, as another embodiment of the present invention, the present invention further includes: recording each judgment result and the corresponding multimodal data, storing them in a false alarm log library, for subsequent interference source analysis and model optimization.

[0056] Alternatively, as another embodiment of the present invention, compared with the prior art, the present invention has the following beneficial effects: 1. Multimodal feature fusion improves recognition accuracy: By using an attention mechanism to weightedly fuse multimodal data features from radar, photoelectric, and electromagnetic sources, the complementary nature of data from various devices is fully utilized to effectively distinguish UAV targets from interference sources (birds, electromagnetic noise, etc.), thereby improving the accuracy of false alarm identification. 2. Network structure optimization enhances feature extraction capability: The feature enhancement layer composed of residual blocks is used to alleviate the gradient vanishing problem in deep networks, enhance the model's ability to extract complex interference features, and improve the model's generalization ability. 3. Loss function adapted to class imbalance scenarios: The designed weighted hybrid loss function solves the model bias problem caused by the low proportion of false alarm samples by adjusting the class weights and setting the focus parameters, thereby improving the recognition accuracy of minority false alarm samples; 4. Online update mechanism enables dynamic optimization: Based on online incremental training of the false alarm negative sample library, the model can adaptively learn new interference source features, continuously reduce the false alarm rate, and meet the needs of complex and ever-changing low-altitude environments. 5. Meets stringent engineering requirements: Through the above technical solutions, the false alarm probability of the low-altitude UAV detection system can be ≤5%, and the alarm delay can be ≤5 seconds, which fully meets the technical requirements of low-altitude airspace management.

[0057] Figure 2 This is a block diagram of a false alarm determination device for a low-altitude unmanned aerial vehicle (UAV) detection system, provided in an embodiment of the present invention.

[0058] Alternatively, as another embodiment of the present invention, such as Figure 2 As shown, a false alarm determination device for a low-altitude unmanned aerial vehicle (UAV) detection system includes: The dataset acquisition module is used to obtain the raw UAV dataset and the UAV dataset to be judged from the low-altitude UAV detection system through the UAV data acquisition device. The preprocessing module is used to preprocess the original drone dataset and the drone dataset to be judged respectively to obtain a first preprocessed drone dataset corresponding to the original drone dataset and a second preprocessed drone dataset corresponding to the drone dataset to be judged. The model analysis module is used to construct a training model and perform model analysis on the training model using the first preprocessed UAV dataset to obtain a false alarm determination model. The prediction module is used to predict the false alarm probability of the second preprocessed UAV dataset using the false alarm determination model. The judgment result acquisition module is used to analyze the false alarm probability based on the UAV data acquisition device to obtain the false alarm judgment result of the low-altitude UAV detection system.

[0059] Optionally, another embodiment of the present invention provides a false alarm determination system for a low-altitude unmanned aerial vehicle (UAV) detection system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the false alarm determination method for a low-altitude UAV detection system as described above. This system can be a computer or similar system.

[0060] Optionally, another embodiment of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the false alarm determination method for a low-altitude unmanned aerial vehicle (UAV) detection system as described above.

[0061] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0062] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described apparatus and unit can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0063] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0064] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of the present invention, depending on actual needs.

[0065] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0066] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0067] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for determining false alarms in a low-altitude unmanned aerial vehicle (UAV) detection system, characterized in that, Includes the following steps: The raw drone dataset and the drone dataset to be judged are obtained from the low-altitude drone detection system using a drone data acquisition device. The original drone dataset and the drone dataset to be judged are preprocessed respectively to obtain a first preprocessed drone dataset corresponding to the original drone dataset and a second preprocessed drone dataset corresponding to the drone dataset to be judged. A training model is constructed, and the training model is analyzed using the first preprocessed UAV dataset to obtain a false alarm determination model. The false alarm probability is obtained by predicting the second preprocessed UAV dataset using the false alarm determination model. The false alarm determination result of the low-altitude drone detection system is obtained by analyzing the false alarm probability based on the data acquisition device of the drone.

2. The false alarm determination method for a low-altitude unmanned aerial vehicle (UAV) detection system according to claim 1, characterized in that, The UAV data acquisition device includes 5G-A equipment, TDOA equipment, AOA equipment, photoelectric camera, electromagnetic detection equipment, protocol parsing equipment, and radar equipment; The process of obtaining the raw UAV dataset and the UAV dataset to be judged from the low-altitude UAV detection system through the UAV data acquisition device includes: The 5G-A device obtains the location data to be determined and multiple raw location data from the low-altitude drone detection system. The TDOA device obtains the time-domain data to be determined and multiple raw time-domain data from the low-altitude unmanned aerial vehicle detection system. The AOA device obtains the angle data to be determined and multiple raw angle data from the low-altitude unmanned aerial vehicle detection system. The photoelectric camera obtains the photoelectric image to be determined and multiple raw photoelectric images from the low-altitude UAV detection system. The electromagnetic detection device obtains the electromagnetic data to be determined and multiple raw electromagnetic data from the low-altitude UAV detection system. The protocol parsing device obtains the communication protocol of the drone to be determined and multiple original drone communication protocols from the low-altitude drone detection system. The radar device obtains radar data to be determined and multiple raw radar data from the low-altitude UAV detection system. The raw UAV dataset includes multiple raw positioning data, multiple raw time-domain data, multiple raw angle data, multiple raw photoelectric images, multiple raw electromagnetic data, multiple raw UAV communication protocols, and multiple raw radar data. The UAV dataset to be determined includes the positioning data to be determined, the time-domain data to be determined, the angle data to be determined, the photoelectric images to be determined, the electromagnetic data to be determined, the UAV communication protocols to be determined, and the radar data to be determined.

3. The false alarm determination method for a low-altitude unmanned aerial vehicle (UAV) detection system according to claim 1, characterized in that, The process of preprocessing the original drone dataset and the drone dataset to be judged to obtain a first preprocessed drone dataset corresponding to the original drone dataset and a second preprocessed drone dataset corresponding to the drone dataset to be judged includes: Alignment processing is performed on the data in the original drone dataset and the data in the drone dataset to be judged, respectively, to obtain a first aligned drone dataset corresponding to the original drone dataset and a second aligned drone dataset corresponding to the drone dataset to be judged; Coordinate transformations are performed on the first aligned UAV dataset and the second aligned UAV dataset respectively to obtain a first transformed UAV dataset corresponding to the original UAV dataset and a second transformed UAV dataset corresponding to the UAV dataset to be judged. Outlier processing is performed on the first transformed drone dataset and the second transformed drone dataset respectively to obtain a first preprocessed drone dataset corresponding to the original drone dataset and a second preprocessed drone dataset corresponding to the drone dataset to be judged.

4. The false alarm determination method for a low-altitude unmanned aerial vehicle (UAV) detection system according to claim 1, characterized in that, The process of performing model analysis on the trained model using the first preprocessed UAV dataset to obtain the false alarm determination model includes: Feature extraction analysis is performed on the first preprocessed UAV dataset to obtain multiple original positioning feature vectors, multiple original time domain feature vectors, multiple original angle feature vectors, multiple original image feature vectors, multiple original electromagnetic signal feature vectors, multiple original UAV communication feature vectors, and multiple original radar signal feature vectors. The training model is trained using all the original positioning feature vectors, all the original time-domain feature vectors, all the original angle feature vectors, all the original image feature vectors, all the original electromagnetic signal feature vectors, all the original UAV communication feature vectors, and all the original radar signal feature vectors to obtain a false alarm determination model.

5. The false alarm determination method for a low-altitude unmanned aerial vehicle (UAV) detection system according to claim 4, characterized in that, The first preprocessed UAV dataset includes multiple preprocessed positioning data, multiple preprocessed time-domain data, multiple preprocessed angle data, multiple preprocessed photoelectric images, multiple preprocessed electromagnetic data, multiple preprocessed UAV communication protocols, and multiple preprocessed radar data. The process of performing feature extraction analysis on the first preprocessed UAV dataset to obtain multiple original positioning feature vectors, multiple original temporal feature vectors, multiple original angle feature vectors, multiple original image feature vectors, multiple original electromagnetic signal feature vectors, multiple original UAV communication feature vectors, and multiple original radar signal feature vectors includes: The standard deviation of each of the preprocessed positioning data is calculated using the sliding window standard deviation algorithm to obtain the calculated positioning data corresponding to each of the preprocessed positioning data. The Kalman filter algorithm is used to filter each of the calculated positioning data to obtain the filtered positioning data corresponding to each of the preprocessed positioning data. Each of the filtered positioning data is processed by Fast Fourier Transform to obtain the original positioning feature vector corresponding to each of the preprocessed positioning data. Outlier processing is performed on each of the preprocessed time-domain data to obtain the time-domain data to be processed corresponding to each of the preprocessed time-domain data. The linear interpolation calibration algorithm is used to calibrate each of the time-domain data to be processed, so as to obtain calibrated time-domain data corresponding to each of the preprocessed time-domain data. Wavelet transform is performed on each of the calibrated time-domain data to extract the original time-domain feature vectors corresponding to each of the preprocessed time-domain data. The median filtering algorithm is used to filter each of the preprocessed angle data to obtain the filtered angle data corresponding to each of the preprocessed angle data. Each of the filtered angle data is subjected to first-order difference processing to obtain the difference-processed angle data corresponding to each of the preprocessed angle data. Pearson correlation coefficients are calculated for each of the differentially processed angle data, and the calculation results are used as the original angle feature vectors to obtain the original angle feature vectors corresponding to each of the preprocessed angle data. Feature extraction is performed on each of the preprocessed photoelectric images using a pre-built ResNet50 model to obtain the original image feature vector corresponding to each of the preprocessed photoelectric images. Wavelet transform is performed on each of the preprocessed electromagnetic data to extract the original electromagnetic signal feature vector corresponding to each of the preprocessed electromagnetic data. The pre-constructed convolutional neural network is used to extract features from each of the pre-processed UAV communication protocols to obtain the original UAV communication feature vectors corresponding to each of the pre-processed UAV communication protocols. Wavelet transform is performed on each of the preprocessed radar data to extract the original radar signal feature vector corresponding to each of the preprocessed radar data.

6. The false alarm determination method for a low-altitude unmanned aerial vehicle (UAV) detection system according to claim 4, characterized in that, The training model includes a first fully connected layer, an attention network, a feature enhancement layer, a second fully connected layer, and a Softmax function layer; The process of training the training model using all the original positioning feature vectors, all the original time-domain feature vectors, all the original angle feature vectors, all the original image feature vectors, all the original electromagnetic signal feature vectors, all the original UAV communication feature vectors, and all the original radar signal feature vectors to obtain the false alarm determination model includes: The first fully connected layer performs mapping processing on each of the original positioning feature vectors, original temporal feature vectors, original angle feature vectors, original image feature vectors, original electromagnetic signal feature vectors, original UAV communication feature vectors, and original radar signal feature vectors to obtain a first mapped positioning feature vector corresponding to each of the original positioning feature vectors, a first mapped temporal feature vector corresponding to each of the original temporal feature vectors, a first mapped angle feature vector corresponding to each of the original angle feature vectors, a first mapped image feature vector corresponding to each of the original image feature vectors, a first mapped electromagnetic signal feature vector corresponding to each of the original electromagnetic signal feature vectors, a first mapped UAV communication feature vector corresponding to each of the original UAV communication feature vectors, and a first mapped radar signal feature vector corresponding to each of the original radar signal feature vectors. The attention network is used to perform fusion analysis on each of the first mapped localization feature vectors, the first mapped time-domain feature vectors corresponding to each of the original time-domain feature vectors, the first mapped angle feature vectors corresponding to each of the original angle feature vectors, the first mapped image feature vectors corresponding to each of the original image feature vectors, the first mapped electromagnetic signal feature vectors corresponding to each of the original electromagnetic signal feature vectors, the first mapped UAV communication feature vectors corresponding to each of the original UAV communication feature vectors, and the first mapped radar signal feature vectors corresponding to each of the original radar signal feature vectors, to obtain multiple original fused feature vectors. The feature enhancement layer performs feature enhancement analysis on each of the original fused feature vectors to obtain the enhanced fused feature vectors corresponding to each of the original fused feature vectors. The second fully connected layer is used to map each of the enhanced fused feature vectors to obtain the mapped fused feature vectors corresponding to each of the original fused feature vectors. The Softmax function layer is used to predict each of the mapped and fused feature vectors to obtain the prediction probability corresponding to each of the original fused feature vectors. Import multiple real labels, perform loss function analysis on all real labels and all predicted probabilities, and obtain the target loss function; The parameters of the training model are updated according to the target loss function. After the parameters are updated, the training model is rebuilt until a preset number of iterations is reached. Then, the training model with updated parameters is used as the false alarm determination model.

7. The false alarm determination method for a low-altitude unmanned aerial vehicle (UAV) detection system according to claim 6, characterized in that, The attention network includes multiple third fully connected layers and a Sigmoid activation function layer; The process of fusing and analyzing each of the first mapped localization feature vectors, the first mapped time-domain feature vectors corresponding to each of the original time-domain feature vectors, the first mapped angle feature vectors corresponding to each of the original angle feature vectors, the first mapped image feature vectors corresponding to each of the original image feature vectors, the first mapped electromagnetic signal feature vectors corresponding to each of the original electromagnetic signal feature vectors, the first mapped UAV communication feature vectors corresponding to each of the original UAV communication feature vectors, and the first mapped radar signal feature vectors corresponding to each of the original radar signal feature vectors through the attention network to obtain multiple original fused feature vectors includes: The first mapped positioning feature vector, the first mapped time-domain feature vector corresponding to each of the original time-domain feature vectors, the first mapped angle feature vector corresponding to each of the original angle feature vectors, the first mapped image feature vector corresponding to each of the original image feature vectors, the first mapped electromagnetic signal feature vector corresponding to each of the original electromagnetic signal feature vectors, the first mapped UAV communication feature vector corresponding to each of the original UAV communication feature vectors, and the first mapped radar signal feature vector corresponding to each of the original radar signal feature vectors are mapped through multiple third fully connected layers. This results in the second mapped positioning feature vector corresponding to each of the original positioning feature vectors, the second mapped time-domain feature vector corresponding to each of the original time-domain feature vectors, the second mapped angle feature vector corresponding to each of the original angle feature vectors, the second mapped image feature vector corresponding to each of the original image feature vectors, the second mapped electromagnetic signal feature vector corresponding to each of the original electromagnetic signal feature vectors, the second mapped UAV communication feature vector corresponding to each of the original UAV communication feature vectors, and the second mapped radar signal feature vector corresponding to each of the original radar signal feature vectors. The Sigmoid activation function layer performs weighting on each of the second-mapped positioning feature vectors, the second-mapped time-domain feature vectors corresponding to each of the original time-domain feature vectors, the second-mapped angle feature vectors corresponding to each of the original angle feature vectors, the second-mapped image feature vectors corresponding to each of the original image feature vectors, the second-mapped electromagnetic signal feature vectors corresponding to each of the original electromagnetic signal feature vectors, the second-mapped UAV communication feature vectors corresponding to each of the original UAV communication feature vectors, and the second-mapped radar signal feature vectors corresponding to each of the original radar signal feature vectors. This yields positioning attention weights corresponding to each of the original positioning feature vectors, time-domain attention weights corresponding to each of the original time-domain feature vectors, angle attention weights corresponding to each of the original angle feature vectors, image attention weights corresponding to each of the original image feature vectors, electromagnetic attention weights corresponding to each of the original electromagnetic signal feature vectors, UAV communication attention weights corresponding to each of the original UAV communication feature vectors, and radar attention weights corresponding to each of the original radar signal feature vectors. The first formula is used to calculate the following: for each of the first mapped positioning feature vectors, the positioning attention weights corresponding to each of the original positioning feature vectors, the first mapped time-domain feature vectors corresponding to each of the original time-domain feature vectors, the time-domain attention weights corresponding to each of the original time-domain feature vectors, the first mapped angle feature vectors corresponding to each of the original angle feature vectors, the angle attention weights corresponding to each of the original angle feature vectors, the first mapped image feature vectors corresponding to each of the original electromagnetic signal feature vectors, the first mapped electromagnetic signal feature vectors corresponding to each of the original electromagnetic signal feature vectors, the first mapped UAV communication feature vectors corresponding to each of the original UAV communication feature vectors, the first mapped radar signal feature vectors corresponding to each of the original radar signal feature vectors, the image attention weights corresponding to each of the original image feature vectors, the electromagnetic attention weights corresponding to each of the original electromagnetic signal feature vectors, the UAV communication attention weights corresponding to each of the original UAV communication feature vectors, and the radar attention weights corresponding to each of the original radar signal feature vectors, multiple original fused feature vectors are obtained. The first formula is: ， in, For the first One original fused feature vector, For the first Each positioning attention weight, For the first Each temporal attention weight, For the first Attention weights from various angles For the first Image attention weights, For the first Each electromagnetic attention weight, For the first Individual drone communication attention weights, For the first Each radar attention weight, For the first After the first mapping, the feature vector is located. For the first The first mapped temporal feature vector For the first The first mapped angular feature vector For the first The first mapped image feature vector For the first The first mapped electromagnetic signal feature vector For the first The first mapped UAV communication feature vector For the first The first mapped radar signal feature vector.

8. The false alarm determination method for a low-altitude unmanned aerial vehicle (UAV) detection system according to claim 6, characterized in that, The feature enhancement layer includes multiple residual blocks arranged in sequence. The process of performing feature enhancement analysis on each of the original fused feature vectors through the feature enhancement layer to obtain the enhanced fused feature vectors corresponding to each of the original fused feature vectors includes: The first residual block performs feature enhancement processing on each of the original fused feature vectors, and the processing result is used as the input of the next residual block until the last residual block is passed, thereby obtaining the enhanced fused feature vectors corresponding to each of the original fused feature vectors. The process of performing feature enhancement processing on each of the original fused feature vectors through the first residual block is specifically as follows: The residual block includes multiple convolutional layers, batch normalization layers, and ReLU activation function layers. The multiple convolutional layers respectively extract features from each of the original fused feature vectors to obtain the extracted fused feature vectors corresponding to each of the original fused feature vectors. The batch normalization layer normalizes each of the extracted fused feature vectors to obtain normalized fused feature vectors corresponding to each of the original fused feature vectors. Each of the original fused feature vectors is element-wise added to the normalized fused feature vectors corresponding to each original fused feature vector to obtain the summed fused feature vectors corresponding to each original fused feature vector. The ReLU activation function layer is used to map each of the summed and fused feature vectors to obtain the processed fused feature vectors corresponding to each of the original fused feature vectors.

9. The false alarm determination method for a low-altitude unmanned aerial vehicle (UAV) detection system according to claim 6, characterized in that, The process of performing loss function analysis on all the true labels and all the predicted probabilities to obtain the target loss function includes: The target loss function is obtained by calculating all the true labels and all the predicted probabilities using the second equation, which is: ， in, , , in, Let be the target loss function. For balance coefficient, For the first loss function, For the second loss function, For the first A real label, To predict the number of probabilities, For the first Each predicted probability, For category weights, For focusing parameters.

10. The false alarm determination method for a low-altitude unmanned aerial vehicle (UAV) detection system according to claim 1, characterized in that, The UAV data acquisition device includes a photoelectric camera, and the process of determining and analyzing the false alarm probability based on the UAV data acquisition device to obtain the false alarm determination result of the low-altitude UAV detection system includes: Determine whether the false alarm probability is greater than or equal to a preset probability. If so, use the first preset determination result as the false alarm determination result of the low-altitude UAV detection system. If not, obtain the photoelectric image to be processed from the low-altitude UAV detection system through the photoelectric camera. Determine whether the photoelectric image to be processed meets a preset condition. If yes, the first preset determination result is used as the false alarm determination result of the low-altitude UAV detection system. If no, the second preset determination result is used as the false alarm determination result of the low-altitude UAV detection system. The preset condition is that there is no UAV target in the photoelectric image to be processed or there is a preset non-UAV target in the photoelectric image to be processed.

Images