Unmanned aerial vehicle detection technology based on deep learning and multi-source information feature fusion

Abstract

The present application relates to a deep learning-based multi-source information feature fusion unmanned aerial vehicle detection technology, comprising the following steps: step S1, data acquisition; step S2, feature extraction; step S3, feature fusion; step S4, feature training; step S5, model optimization; step S6, generating an adaptive adjustment detection model. A deep learning and modal alignment method is used to realize multi-level feature extraction and fusion. On the basis of deep learning, deep and shallow features of information are extracted, and a modal alignment method is used to realize the fusion of different sensor features in a time stamp alignment manner. The multi-layer fusion features are input, the correlation between the multi-level features and the target detection is learned through the weight of the neurons in the neural network, a multi-sensor feature method based on neuron weight learning is formed, and thus an adaptive adjustment detection model is constructed, the environmental adaptability of the model is improved, the robustness of the model is enhanced, and the demand of complex unmanned aerial vehicle detection tasks is met.

Description

UAV detection technology based on deep learning and multi-source information feature fusion Technical Field

[0001] This invention relates to the field of unmanned aerial vehicles (UAVs), and in particular to a UAV detection technology based on deep learning and multi-source information feature fusion. Background Technology

[0002] Drone detection technology is becoming increasingly intelligent, and environmental perception is the primary step for drones to achieve autonomous detection and the foundation for scene recognition, providing essential information such as the position, speed, and trajectory of multiple targets in the drone's vicinity. Drone detection has evolved from a single detection method to a fusion approach using multiple detectors, including radar, photoelectric tracking devices, and radio equipment. This approach enables comprehensive perception of the position, speed, and trajectory of surrounding drones, achieving complementarity in perception range and information completeness. This provides an effective solution for reliable perception in complex environments, effectively reducing the impact of multipath interference and background noise in complex urban environments. However, the differences in the design principles and physical characteristics of various detectors result in significant variations in their descriptions and representations of the environment. Effectively fusing the information features from various sensors remains a key challenge in current drone detection technology.

[0003] Traditional information feature extraction methods are mostly manually set, such as short-time Fourier transform, and typically only extract shallow features, making them unsuitable for the complex detection tasks of UAVs. Furthermore, because the sampling periods of data output from various sensors are inconsistent, it is impossible to achieve the fusion of features from different sensors by aligning timestamps.

[0004] Therefore, existing technology cannot meet our needs. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides an adaptively updated UAV detection technology based on deep learning and multi-source information feature fusion.

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

[0007] The UAV detection technology based on deep learning and multi-source information feature fusion includes the following steps:

[0008] Step S1, Data Acquisition: Collect at least two types of data;

[0009] Step S2, Feature Extraction: Use a convolutional neural network to extract multi-scale features for each type of data;

[0010] Step S3, Feature Fusion: Align all extracted features modally;

[0011] Step S4, Feature Training: The fused features are fed into the BP neural network for training to generate the initial model;

[0012] Step S5, model optimization, using ensemble learning to optimize the parameters in the BP neural network.

[0013] Step S6: Generate an adaptively adjusted detection model.

[0014] As a preferred embodiment, the data collected in step S1 includes radar detection data, photoelectric detection data, and wireless communication detection data.

[0015] As a preferred embodiment, it is further defined that: between step S1 and step S2, data preprocessing is performed on the collected data. The data preprocessing extracts the required parts for each type of data. The required parts include the size, shape, three-dimensional orientation, and communication fingerprint of the UAV under test.

[0016] As a preferred embodiment, the extraction process in step S2 is further defined as follows:

[0017] S21, Data Extraction: Multi-scale convolutional layers are used to extract the spatial features of the data.

[0018] S22, Dimensionality Reduction: Spatial features are reduced in dimensionality through pooling layers to obtain shallow and deep feature maps;

[0019] S23, Fusion: The shallow feature map and the deep feature map are fused through the fusion layer to obtain the multimodal sample feature map.

[0020] As a preferred embodiment, the fusion process in step S3 is further defined as follows:

[0021] S31, Target Optimization:

[0022] L(O,P,Q,R|X,Y,Z)=α1L1(O,R|X)+α2L2(P,R|Y)++α3L3(Q,R|Z) (Formula 1)

[0023] in, and These are feature maps of different multimodal samples.

[0024] N is the number of samples.

[0025] (N1+P1), (N2+P2), and (N3+P3) correspond to the dimensions of the multimodal sample feature maps, respectively.

[0026] O∈m×f P∈ m×f and Q∈ m×f It is a matrix of different feature maps.

[0027] R∈ f×m It is about learning a shared space of latent features.

[0028] L1, L2, and L3 are the decomposition loss coefficients, which are: ∑ i L i =1;

[0029] S32, Generate hash code: Use threshold transformation to generate the corresponding hash code B from the feature map of the multimodal sample;

[0030] S33: Modal alignment:

[0031]

[0032] Where X, Y, and Z represent multimodal sample feature maps.

[0033] O, P, and Q are matrices representing different feature maps.

[0034] B represents the hash code;

[0035] S34: Multimodal Sample Feature Map Reconstruction:

[0036] b O =B T O(Formula 3)

[0037] b P =B T P(Formula 4)

[0038] b Q =B T Q(Formula 5)

[0039] Among them, b O b P b Q These are the binary vector representations of the feature maps of different multimodal samples.

[0040] As a preferred embodiment, it is further defined that O, P, Q are matrices obtained by linear projection.

[0041] As a preferred embodiment, it is further defined as follows: S35: generating a semantic representation of feature-level fusion: A = [b O b P b Q ].

[0042] As a preferred embodiment, the training process in step S4 is further defined as follows:

[0043] S41: Weight Learning

[0044]

[0045] y = f(w1v1 + w2v2 + ... + w i v i )(Formula 7)

[0046] In this system, the weight between each neuron is w, the value of each neuron is v, and the output value y of the activation function f(x) is either 0 or 1. When y = 1, it means that a drone has been detected; otherwise, it is 0.

[0047] S42: Generate the initial model for offline data.

[0048] As a preferred embodiment, the optimization process in step S5 is further defined as follows:

[0049] S51, Model Building: Create an online detection model;

[0050] S52, Model Testing: Test the online detection model and the initial offline data model, and calculate the accuracy respectively;

[0051] S53, Determine the weight W:

[0052] W t =C×W t-1 (Formula 8)

[0053]

[0054] Wherein, the initial weight W is 0.5, t is time, K is a constant, and K is 1%-5%;

[0055] When the adaptive weight adjustment coefficient C at time t is greater than 1.5, the weight W at time t+1 is equal to the weight W at time t-1.

[0056] As a preferred embodiment, the method is further defined as follows: the model establishment in step S51 includes;

[0057] Step S511, Online Data Acquisition: Acquire at least two types of online data;

[0058] Step S512, Feature Extraction: Use a convolutional neural network to extract multi-scale features from each type of online data;

[0059] Step S513, Feature Fusion: Align all extracted features modally;

[0060] Step S514, Feature Training: The fused features are fed into the BP neural network for training to generate an online detection model.

[0061] The beneficial effects of this invention are as follows: This invention uses deep learning and modality alignment to achieve the extraction and fusion of multi-level features. Based on deep learning, it extracts deep and shallow features of information and uses modality alignment to constrain the information of multiple sensors to a unified spatiotemporal space by ensuring consistency between modalities, thereby achieving the fusion of features from different sensors and ultimately forming multi-layer fused features. This is achieved by fusion of features from different sensors through timestamp alignment. The multi-layer fused features are then input into a BP neural network. The correlation between multi-level features and UAV target detection is learned through the weights of neurons in the neural network, forming a multi-sensor feature method based on neuron weight learning, thereby constructing an initial model for UAV detection. After training, the weights of each neuron in the neural network represent the semantic importance of each data source. This weight fusion criterion is not the traditional method of setting weights for different sensors, but rather a self-learning process based on the weights learned by the neurons, enabling rapid adaptation to changes in the local environment. Building upon this, an ensemble learning method is used to optimize the parameters of the BP neural network based on online UAV data, resulting in decision-level fusion of offline and online data, thus enabling adaptive updates to the UAV detection model. This unmanned detection method, employing feature-level and decision-level fusion, achieves the fusion of multi-sensor data at different granularities. Feature-level fusion focuses on features as a dimension, while decision-level fusion focuses on data type dimensions, forming different-dimensional detection model constructions. This approach has high application value, improves the model's environmental adaptability, enhances its robustness, and meets the needs of complex UAV detection tasks. Figure Description

[0062] Figure 1 is a framework diagram of the present invention;

[0063] Figure 2 is a schematic diagram of feature extraction from multi-source data based on deep learning;

[0064] Figure 3 is a schematic diagram of the process of extracting latent semantic representation hash codes from multi-source data based on deep learning;

[0065] Figure 4 is a schematic diagram of a BP neural network;

[0066] Figure 5 is a schematic diagram of model optimization. Detailed Implementation Methods

[0067] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0068] As shown in Figures 1 to 5, the UAV detection technology based on deep learning and multi-source information feature fusion includes the following steps:

[0069] Step S1, Data Acquisition: Collect at least two types of data;

[0070] Step S2, Feature Extraction: Use a convolutional neural network to extract multi-scale features for each type of data;

[0071] Step S3, Feature Fusion: Align all extracted features modally;

[0072] Step S4, Feature Training: The fused features are fed into the BP neural network for training to generate the initial model;

[0073] Step S5, model optimization: Optimize the parameters in the BP neural network using ensemble learning;

[0074] Step S6: Generate an adaptively adjusted detection model.

[0075] This invention employs deep learning and modality alignment to extract and fuse multi-level features. Based on deep learning, it extracts deep and shallow features of information and uses modality alignment to constrain information from multiple sensors to a unified spatiotemporal space by ensuring consistency between modalities. This achieves the fusion of features from different sensors, ultimately forming multi-layer fused features. The fusion of features from different sensors is achieved through timestamp alignment. These multi-layer fused features are then input into a BP neural network. The weights of neurons in the neural network learn the correlation between multi-level features and UAV target detection, forming a multi-sensor feature method based on neuron weight learning, thereby constructing an initial model for UAV detection. After training, the weights of each neuron in the neural network represent the semantic importance of each data source. This weight fusion criterion is not the traditional method of setting weights for different sensors, but rather a self-learning process based on the weights learned by the neurons, enabling rapid adaptation to changes in the local environment. Building upon this, an ensemble learning method is used to optimize the parameters of the BP neural network based on online UAV data, resulting in decision-level fusion of offline and online data, thus enabling adaptive updates to the UAV detection model. This unmanned detection method, employing feature-level and decision-level fusion, achieves the fusion of multi-sensor data at different granularities. Feature-level fusion focuses on features as a dimension, while decision-level fusion focuses on data type dimensions, forming different-dimensional detection model constructions. This approach has high application value, improves the model's environmental adaptability, enhances its robustness, and meets the needs of complex UAV detection tasks.

[0076] In this embodiment, the data collected in step S1 includes radar detection data, photoelectric detection data, and wireless communication detection data. Furthermore, radar detection is highly reliable for detecting moving targets, such as aperture radar (SAR) and lidar. It has a very long detection range and highly accurate azimuth, especially for large, high-speed moving objects, and is therefore frequently used for detecting flying objects. However, real-world drones are relatively small, and when their flight speed is slow, the accuracy of radar detection is significantly reduced. Therefore, other data sources are needed to ensure detection accuracy. To compensate for radar detection in scenarios where drones fly at low speeds, this invention adds photoelectric detection and wireless communication detection. Photoelectric detection includes infrared detection. Infrared detection, as a passive detection method that receives infrared light signals radiated or reflected by the target, utilizes the difference between the target and background images to achieve target detection. Therefore, noise, target, and background image grayscale values ​​in the infrared imaging are identified, and then the target grayscale value is extracted. Wireless communication detection utilizes the fingerprint information generated during the drone's communication process to identify individual drones. Because each drone has different hardware, its transmitted signals differ from other aircraft, resulting in a unique communication fingerprint for each drone. Collecting these three types of detection data is necessary to meet the needs of drone detection.

[0077] In this embodiment, between steps S1 and S2, data preprocessing is also performed on the collected data. Data preprocessing extracts the necessary portions for each type of data, including the size, shape, three-dimensional orientation, and communication fingerprint of the UAV under test. Multi-source data preprocessing first involves extracting the region of interest from multiple sensors; that is, detecting a suspicious flying object triggers an alarm in the information system and extracts information about the region of interest. Then, radar and thermal infrared indexers estimate the object's size, shape, and three-dimensional orientation information, while wireless detection acquires the UAV's communication fingerprint.

[0078] As shown in Figure 2, in this embodiment, the extraction process in step S2 includes:

[0079] S21, Data Extraction: Multi-scale convolutional layers are used to extract the spatial features of the data.

[0080] S22, Dimensionality Reduction: Spatial features are reduced in dimensionality through pooling layers to obtain shallow and deep feature maps;

[0081] S23, Fusion: The shallow feature map and the deep feature map are fused through the fusion layer to obtain the multimodal sample feature map.

[0082] Since different sensor data have different modalities, before performing modal alignment on all extracted features in step S3 (feature fusion), semantic mapping of the data is required. This means mapping the original data to a certain space to provide a foundation for feature-level fusion of multiple data sources. Therefore, to fully learn the features of multi-source data, as shown in Figure 2, this invention employs a multi-scale feature extraction method during the feature extraction process using CNN (Convolutional Neural Network). This involves deep and shallow data extraction for the three types of data mentioned above: radar detection data, infrared detection data, and wireless communication detection data.

[0083] This invention employs multi-scale convolutional layers to extract spatial features from alarm time-series radar data, thermal infrared data, and wireless data. Since the convolutional layers have different sizes, different scales of convolutional layers are used to extract spatial features during the CNN extraction process. Then, pooling layers are used to reduce the dimensionality of the feature maps. Finally, a fusion layer is used to fuse the semantic information of shallow and deep data, essentially superimposing the shallow and deep feature maps. This involves superimposing and fusing feature maps extracted from different convolutional layers for the same input data to obtain a multimodal sample feature map. By fusing and analyzing shallow feature information and deep semantic information, the richness of feature information extraction is enhanced.

[0084] As shown in Figure 3, in this embodiment, the fusion process in step S3 includes:

[0085] S31, Target Optimization:

[0086] L(O,P,Q,R|X,Y,Z)=α1L1(O,R|X)+α2L2(P,R|Y)++α3L3(Q,R|Z) (Formula 1)

[0087] in, and These are feature maps of different multimodal samples.

[0088] N is the number of samples.

[0089] (N1+P1), (N2+P2), and (N3+P3) correspond to the dimensions of the multimodal sample feature maps, respectively.

[0090] O∈ m×f P∈ m×f and Q∈ m×f It is a matrix of different feature maps.

[0091] R∈ f×m It is about learning a shared space of latent features.

[0092] L1, L2, and L3 are the decomposition loss coefficients, which are: ∑ i L i =1;

[0093] S32, Generate hash code: Use threshold transformation to generate the corresponding hash code B from the feature map of the multimodal sample;

[0094] S33: Modal alignment:

[0095]

[0096] Where X, Y, and Z represent multimodal sample feature maps.

[0097] O, P, and Q are matrices representing different feature maps.

[0098] B represents the hash code;

[0099] S34: Multimodal Sample Feature Map Reconstruction:

[0100] b O =B T O(Formula 3)

[0101] b P =B T P(Formula 4)

[0102] b Q =B T Q(Formula 5)

[0103] Among them, b O b P b Q These are the binary vector representations of the feature maps of different multimodal samples.

[0104] Where X, Y, and Z represent the multimodal sample feature maps of radar detection data, infrared detection data, and wireless communication detection data, respectively. To simplify the solution process of Equation 1, modal alignment, or semantic consistency between modalities, is used to learn the latent associations between multimodal samples. Since the temporal granularity of radar detection data, infrared detection data, and wireless communication detection data is different, feature fusion is impossible without feature alignment. This method not only solves the problem of temporal asynchrony caused by the difference in acquisition frequencies of different sensors to a certain extent, but also simplifies the latent semantic representation through the association between modalities, thereby achieving semantically consistent representation between modalities.

[0105] In this embodiment, O, P, and Q are matrices obtained through linear projection. The feature maps of the three multimodal samples are projected onto a multidimensional space using vectors. R in Formula 1 and B in Formula 2 are matrices. The relationship between R and B is that after threshold transformation, many values ​​in R that do not meet the threshold value are changed to 0, thus obtaining B. As shown in Figure 3, blank cells represent 0, and dark-colored cells represent 1.

[0106] In this embodiment, it also includes S35: generating a semantic representation of feature-level fusion: A = [b O b P b Q Feature-level fusion focuses on features as dimensions. The binary representations of the three data features mentioned above can be concatenated to achieve feature-level fusion, forming a semantic representation A based on feature-level fusion of the three data sources.

[0107] As shown in Figure 4, in this embodiment, the training process in step S4 includes:

[0108] S41: Weight Learning

[0109]

[0110] y = f(w1v1 + w2v2 + ... + w i v i )(Formula 7)

[0111] In this system, the weight between each neuron is w, the value of each neuron is v, and the output value y of the activation function f(x) is either 0 or 1. When y = 1, it means that a drone has been detected; otherwise, it is 0.

[0112] S42: Generate the initial model for offline data.

[0113] As shown in Figure 4, the BP neural network consists of an input layer, hidden layers, and an output layer. The input layer has three neurons, representing the semantic representation of the feature-level fusion of the three data sources. The hidden layer has eight neurons, and the last layer is the output layer with one neuron. The weight between each neuron is w, and the value of each neuron is v. After training, the weight between each neuron represents the importance of the semantics of each data source. This weight fusion criterion is not the traditional method of setting weights for different sensors, but rather a self-learning process based on the weights of the neurons, enabling it to quickly adapt to changes in the same city scenario.

[0114] As shown in Figure 5, in this embodiment, the optimization process in step S5 includes:

[0115] S51, Model Building: Create an online model;

[0116] S52, Model Testing: Test the online model and the initial model on offline data, and calculate the accuracy rate for each.

[0117] S53, Determine the weight W:

[0118] W t =C×W t-1 (Formula 8)

[0119]

[0120] Wherein, the initial weight W is 0.5, t is time, K is a constant, and K is 1%-5%;

[0121] When the adaptive weight adjustment coefficient C at time t is greater than 1.5, there is a large difference in the data at this time, and the weight W at time t+1 is equal to the weight W at time t-1.

[0122] To avoid the problem of poor robustness in single offline training, this invention utilizes online data combined with ensemble learning to continuously update model parameters. After extracting deep and shallow semantic features from a small amount of online data using deep learning, feature alignment is performed, and a backpropagation (BP) neural network is used to train a small number of samples online, forming an online data detection model. Subsequently, the online model and the initial offline data model are tested using online and offline drone detection test data, and accuracy statistics are performed. The accuracy of the two types of models is calculated, and their weights are determined based on the accuracy. The initial weights are all set to 0.5. As time iterates, the weight at each time step is related to the weight at the previous time step. Furthermore, if the adaptive weight adjustment coefficient C at time t is greater than 1.5, it indicates that there is a significant difference in the data at that time, and the weight at time t+1 is equal to the weight at time t-1.

[0123] The weights of the online and offline models based on ensemble learning are continuously adjusted according to the data distribution, thereby achieving decision-level fusion in the process of UAV detection and identification, and enhancing the model's adaptability. The online data at time t is used as the offline test set at time t+1 to continuously test the changes in the data distribution of the online dataset, thus enabling the adaptive adjustment of the detection model's parameters.

[0124] As shown in Figure 5, in this embodiment, the model establishment in step S51 includes:

[0125] Step S511, Online Data Acquisition: Acquire at least two types of online data;

[0126] Step S512, Feature Extraction: Use a convolutional neural network to extract multi-scale features from each type of online data;

[0127] Step S513, Feature Fusion: Align all extracted features modally;

[0128] Step S514, Feature Training: The fused features are fed into the BP neural network for training to generate an online model.

[0129] Furthermore, the creation of the online model, like the initial offline data model, also requires feature extraction, feature fusion, and feature training to form the same model. Subsequently, online and offline drone detection test data are used to test both the online and offline initial models, and accuracy statistics are performed. The accuracy of the two types of models is calculated, and the weights of the two models are determined based on the accuracy.

[0130] This invention builds upon three existing data sources, focusing on feature extraction, feature fusion, recognition model training, and recognition model optimization. This leads to a deep learning-based multi-source information feature fusion UAV detection technology, which includes both feature-level fusion and decision-level fusion. Feature-level fusion aligns features from different data sources across modalities, enabling cross-modal data retrieval and confining information from multiple sensors to a unified spatiotemporal space. This avoids feature inconsistencies caused by inconsistent acquisition granularity across different sensors. Decision-level fusion employs a backpropagation (BP) neural network to identify the importance of various multi-source features. It iterates the weights between neurons during data training, avoiding the model's lack of robustness caused by manually setting feature importance. Finally, considering the incompleteness of the training data, ensemble learning is used to optimize the model on online training data, enabling real-time updates of model parameters, improving environmental adaptability, and enhancing robustness.

Claims

1. A UAV detection technology based on multi-source information feature fusion using deep learning, characterized in that, Includes the following steps: Step S1, Data Acquisition: Acquire at least two types of data, including radar detection data, photoelectric detection data, and wireless communication detection data; Step S2, Feature Extraction: Extract multi-scale features for each type of data using a convolutional neural network; Step S3, Feature Fusion: Align all extracted features modally. The fusion process in step S3 includes: S31, target optimization: (Formula 1) Where, 、 and These are multimodal sample feature maps representing radar detection data, infrared detection data, and wireless communication detection data, respectively, where N is the number of samples. 、 and These correspond to the dimensions of the feature maps of the multimodal samples. 、 and These are matrices representing different feature maps, where O, P, and Q are matrices obtained through linear projection. It learns a common latent feature space, where L1, L2, and L3 are decomposition loss coefficients, and the decomposition loss coefficients are: S32, Generate hash code: Use a threshold transformation method to generate the corresponding hash code B from the feature map of the multimodal samples; S33, Modality alignment: (Formula 2) Where O, P, Q are matrices obtained by linear projection, X, Y, Z represent multimodal sample feature maps of radar detection data, infrared detection data, and wireless communication detection data, respectively, and B represents the hash code; the three multimodal sample feature maps are projected onto a multidimensional space by vector projection, where R in Formula 1 and B in Formula 2 are matrices. The relationship between R and B is that after threshold transformation, many values ​​in R that do not meet the threshold value are changed to 0, thus obtaining B; S34, multimodal sample feature map reconstruction: (Official 3) (Official 4) (Formula 5) Wherein, 、 、 These are the binary vector representations of feature maps of different multimodal samples; Step S4, Feature Training: The fused features are fed into the BP neural network for training to generate an initial model; Step S5, Model Optimization: The parameters in the BP neural network are optimized using ensemble learning; Step S6, Generate an adaptively adjusted detection model.

2. The UAV detection technology based on deep learning and multi-source information feature fusion as described in claim 1, characterized in that: Between steps S1 and S2, data preprocessing is performed on the collected data. The data preprocessing extracts the required parts for each type of data, including the size, shape, three-dimensional orientation, and communication fingerprint of the UAV under test.

3. The UAV detection technology based on deep learning and multi-source information feature fusion according to claim 1, characterized in that: The extraction process in step S2 includes: S21, data extraction: extracting spatial features of the data using multi-scale convolutional layers; S22, dimensionality reduction: reducing the spatial features through pooling layers to obtain shallow feature maps and deep feature maps; S23, fusion: fusing the shallow feature maps and deep feature maps through a fusion layer to obtain multimodal sample feature maps.

4. The UAV detection technology based on deep learning and multi-source information feature fusion according to claim 1, characterized in that: It also includes S35: generating semantic representations for feature-level fusion. Feature-level fusion focuses on features as dimensions. The binary representations of the three data features mentioned above can be concatenated to achieve feature-level fusion, forming a semantic representation A based on feature-level fusion of the three data sources.

5. The UAV detection technology based on deep learning and multi-source information feature fusion according to claim 1, characterized in that: The training process in step S4 includes: S41: Weight learning: (Official 6) (Formula 7) Where the weight between each neuron is w, the value of the neuron is v, and the output value y of the activation function f(x) is 0 or 1. When y=1, it means that the drone has been detected, otherwise it is 0; S42: Generate the initial model for offline data.

6. The UAV detection technology based on deep learning and multi-source information feature fusion according to claim 5, characterized in that: The optimization process in step S5 includes: S51, Model building: Creating an online detection model; S52, Model testing: Testing the online detection model and the initial offline data model, and calculating the accuracy respectively; S53, Determining weights W: (Official 8) (Formula 9) Where the initial weight W is 0.5, t is time, K is a constant, and K is 1%-5%; when the adaptive weight adjustment coefficient C at time t is greater than 1.5, the weight W at time t+1 is equal to the weight W at time t-1.

7. The UAV detection technology based on deep learning and multi-source information feature fusion according to claim 6, characterized in that: The model building in step S51 includes: Step S511, online data acquisition: acquiring at least two types of online data; Step S512, feature extraction: using a convolutional neural network to extract multi-scale features from each type of online data; Step S513, feature fusion: performing modal alignment on all extracted features. Step S514, Feature Training: The fused features are fed into the BP neural network for training to generate an online detection model.

Images