A multi-unmanned aerial vehicle aliasing spectrum signal recognition method based on convolutional neural network and box dimension feature fusion
By fusing convolutional neural networks with box-dimensional features, the problem of low signal recognition accuracy of UAVs in multi-signal aliasing scenarios is solved, and high-precision recognition and strong generalization ability for multiple types of UAVs are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ANHUI UNIV
- Filing Date
- 2026-03-07
- Publication Date
- 2026-06-09
AI Technical Summary
Existing drone signal recognition methods suffer from low accuracy and poor generalization ability in scenarios with multiple signals overlapping, making it difficult to accurately distinguish between different types of drones.
A method based on convolutional neural network and box dimension feature fusion is adopted to identify the aliased spectrum signals of multiple UAVs through size normalization, multi-scale sliding window segmentation, convolutional neural network feature extraction, box dimension feature extraction and feature fusion.
It improves recognition accuracy and robustness, can flexibly handle input maps of different sizes, and can accurately identify multiple types of UAVs, with high accuracy and strong generalization ability.
Smart Images
Figure CN122174045A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of signal processing and pattern recognition technology, and in particular to a method for recognizing multi-UAV aliased spectrum signals based on the fusion of convolutional neural networks and box-dimensional features. Background Technology
[0002] With the rapid development of the civilian drone industry, drones are increasingly used in aerial photography, logistics, agriculture, and other fields, but this has also brought about problems such as airspace security and privacy violations. The detection and identification of drone communication signals is of great significance for low-altitude safety monitoring. Drones typically communicate using frequency bands such as 2.4GHz, 5.8GHz, and 915MHz, and their time-frequency spectra contain rich signal characteristics that can be used for drone type identification.
[0003] Currently, drone signal recognition methods mainly include traditional methods based on feature engineering and methods based on deep learning. However, these methods have the following drawbacks: Traditional methods based on feature engineering, such as time-frequency feature extraction and statistical feature analysis, require manual feature design, have limited feature representation capabilities, and are difficult to capture complex signal patterns; while methods based on single deep learning can automatically learn features, they often ignore the geometric distribution characteristics of the signal, resulting in low recognition accuracy in multi-signal aliasing scenarios; and in multi-signal aliasing scenarios, when multiple drone signals are superimposed in the same time-frequency spectrum, existing methods struggle to accurately distinguish between different types of drones, leading to a significant decrease in recognition accuracy. Summary of the Invention
[0004] To address the issues of low recognition accuracy and poor generalization ability in existing technologies under multi-signal aliasing scenarios, the present invention aims to provide a multi-UAV aliasing spectrum signal recognition method based on the fusion of convolutional neural networks and box-dimensional features, which is highly accurate, robust, and anti-interference.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for identifying multi-UAV aliased spectral signals based on the fusion of convolutional neural networks and box-dimensional features, comprising:
[0006] (1) The size of the multi-UAV aliased time-frequency map to be identified is normalized, and the normalized map is segmented by a multi-scale sliding window to obtain a sub-map of uniform size.
[0007] (2) Construct a convolutional neural network feature extraction module to extract features from the segmented sub-maps. Adaptively weight the features through a dual mechanism of channel attention and spatial attention to obtain the enhanced CNN feature vector.
[0008] (3) Construct a box dimension feature extraction module, use the multi-scale box counting method to quantify the geometric distribution characteristics of the sub-map, and obtain the box dimension feature vector by statistically analyzing the distribution sparsity of signal features at different scale grids;
[0009] (4) Construct a feature fusion and classification module, splice and fuse CNN feature vectors and box dimension feature vectors, perform multi-label classification through a fully connected network, output the probability of UAV types in the spectrum, and realize the identification of multiple types of UAVs in the aliased spectrum signal.
[0010] Step (1) includes the following steps: normalizing the size of the multi-UAV aliased time-frequency map to be identified means adjusting the input map to a standard size of 512×512 pixels; for mixed maps with a width greater than 512 pixels, a sliding window strategy is used to segment them according to a width of 512 pixels, and the segmentation formula is as follows:
[0011] (1)
[0012] in, Indicates the number of sub-maps after segmentation; Indicates the width of the input graph; This indicates the rounding up operation;
[0013] The segmented sub-maps are normalized to map pixel values to the [0,1] interval. The normalization formula is as follows:
[0014] (2)
[0015] in, Represents the original pixel value; and These represent the minimum and maximum pixel values in the graph, respectively. To prevent division by zero of small constants; This represents the normalized pixel value.
[0016] Step (2) specifically includes the following steps:
[0017] (2a) The convolutional neural network feature extraction module adopts a five-layer convolutional network structure. The kernel size of each layer is 3×3, the stride is 2, the padding is 1, and the number of channels is 32, 64, 128, 256, and 512 respectively. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. The batch normalization formula is as follows:
[0018] (3)
[0019] in, Indicates the convolution output; and These represent the batch mean and variance, respectively. and These are learnable parameters; This is the normalized output;
[0020] (2b) The channel attention module learns the importance weights of each feature channel through global average pooling and global max pooling. The formula for calculating the channel attention weights is as follows:
[0021] (4)
[0022] in, Indicates the input feature map; Indicates a fully connected layer; This represents the Sigmoid activation function; Indicates channel attention weights; and These represent average pooling and max pooling, respectively. The channel attention module first compresses the spatial dimension to 1×1 through global average pooling and global max pooling, then learns the dependencies between channels through a shared fully connected network, and finally outputs the weight coefficients of each channel through the Sigmoid activation function to achieve adaptive feature enhancement of the channel dimension.
[0023] (2c) The spatial attention module learns the importance weights of spatial locations by aggregating across channels. The formula for calculating the spatial attention weights is as follows:
[0024] (5)
[0025] in, and This represents average pooling and max pooling along the channel dimension. This indicates a channel splicing operation. This represents the convolution operation. The spatial attention weights are represented by the spatial attention module, which aggregates channel information into two two-dimensional feature maps by performing average pooling and max pooling operations along the channel dimension. After concatenation, the spatial attention map is generated by 7×7 convolution, thus achieving adaptive feature enhancement in the spatial dimension.
[0026] (2d) Compress the attention-weighted feature map into a 512-dimensional CNN feature vector through global average pooling.
[0027] Step (3) specifically includes the following steps:
[0028] (3a) The box dimension feature extraction module adopts a multi-scale box counting method, setting six different scale grids with grid sizes of 2×2, 4×4, 8×8, 16×16, 32×32, and 64×64, respectively. The box counting method is based on the fractal geometry principle. By covering the spectrum of grids at different scales, the number of boxes containing signal features is counted, thereby quantifying the geometric distribution characteristics of the signal.
[0029] (3b) Calculate the feature threshold for the normalized map. The threshold is the 75th percentile of the map pixel value. The calculation formula is as follows:
[0030] (6)
[0031] in, Represents the set of normalized spectral pixel values. This represents the 75th percentile operation. The feature threshold is represented by a quantile threshold instead of a fixed threshold, which can adapt to the spectrum of different signal intensities and improve the robustness of feature extraction.
[0032] (3c) For each scale The grid is used to count the number of boxes containing significant signal features, and the following criteria are used for judgment:
[0033] (7)
[0034] in, Indicates position The grid cells at that location, and These represent the maximum and mean values within the grid, respectively. This indicates whether the grid contains features; by simultaneously judging the maximum and mean values, it is possible to effectively distinguish between real signal features and noise interference.
[0035] (3d) Count the number of effective boxes at each scale to form a 6-dimensional box-dimensional feature vector. The calculation formula is as follows:
[0036] (8)
[0037] in, Representing scale The number of valid boxes below; This indicates whether the grid contains features; the number of boxes at different scales reflects the distribution characteristics of the signal at different resolutions. Small-scale grids can capture detailed features, while large-scale grids can reflect the overall distribution.
[0038] (3e) The 6-dimensional box-dimensional feature vector is processed through a two-layer fully connected network. The first layer expands the dimension from 6 to 64, and the second layer expands it from 64 to 128, resulting in a 128-dimensional geometric feature vector.
[0039] Step (4) specifically includes the following steps:
[0040] (4a) The feature fusion module concatenates the 512-dimensional CNN feature vector with the 128-dimensional box-shaped feature vector to form a 640-dimensional fused feature vector. The concatenation formula is as follows:
[0041] (9)
[0042] in, Represents the CNN feature vector; Represents the eigenvectors of the box dimension; This represents a vector concatenation operation; feature fusion achieves a complementary combination of deep features and geometric features, CNN features capture complex texture and pattern information, and box-dimensional features characterize the geometric distribution characteristics of the signal;
[0043] (4b) The fused feature vectors are classified through a two-layer fully connected network. The first layer compresses the dimension from 640 to 256. ReLU activation function and Dropout regularization are used between layers. The formula for ReLU activation function is as follows:
[0044] (10)
[0045] in, This represents the input to the neuron; This represents the neuron's output; the Dropout ratio is set to 0.3 to effectively prevent overfitting; the second layer maps the dimension from 256 to the number of classes. ;
[0046] (4c) The output layer uses the Sigmoid activation function for multi-label classification, outputting the probability of existence of each drone type. The calculation formula is as follows:
[0047] (11)
[0048] in, Indicates the first Linear output of drone-like systems; Indicates the first The probability of the existence of drone-like objects; using the Sigmoid activation function instead of Softmax; supporting multi-label classification; that is, multiple drone types may exist in the same map;
[0049] (4d) For multiple sub-maps after sliding window segmentation, predictions are performed separately, and the average probability is taken as the final prediction result. The fusion formula is as follows:
[0050] (12)
[0051] in, Indicates the number of sub-maps; Indicates the first Predicted probability of individual sub-maps; This represents the final probability after fusion; the sliding window fusion strategy can effectively handle mixed maps of different widths while maintaining sensitivity to local features;
[0052] (4e) Select the one with the highest probability. Using the identified types as the recognition results, the system can accurately identify multiple types of drones in mixed spectrum signals.
[0053] As can be seen from the above technical solution, the beneficial effects of the present invention are as follows: First, the present invention effectively integrates deep texture features and geometric fractal features through the dual-branch extraction of convolutional neural networks and box-dimensional features, thereby improving the expressive power and recognition of features; Second, the dual attention mechanism adopted in the present invention achieves adaptive enhancement of key features through the synergistic effect of channel attention and spatial attention, thereby improving the model's attention to important information; Third, the sliding window inference mechanism adopted in the present invention can flexibly handle input maps of different sizes, and achieves effective recognition of mixed maps of two, three, or even more types, with extremely high accuracy and strong generalization ability. Attached Figure Description
[0054] Figure 1 This is a flowchart of the method of the present invention;
[0055] Figure 2 This is a structural diagram of the dual attention mechanism in this invention;
[0056] Figure 3 This is a schematic diagram of box-dimensional feature extraction in this invention; Detailed Implementation
[0057] like Figure 1 As shown, a method for identifying aliased spectral signals of multiple unmanned aerial vehicles (UAVs) based on the fusion of convolutional neural networks and box-dimensional features includes:
[0058] (1) The size of the multi-UAV aliased time-frequency map to be identified is normalized, and the normalized map is segmented by a multi-scale sliding window to obtain a sub-map of uniform size.
[0059] (2) Construct a convolutional neural network feature extraction module to extract features from the segmented sub-maps. Adaptively weight the features through a dual mechanism of channel attention and spatial attention to obtain the enhanced CNN feature vector.
[0060] (3) Construct a box dimension feature extraction module, use the multi-scale box counting method to quantify the geometric distribution characteristics of the sub-map, and obtain the box dimension feature vector by statistically analyzing the distribution sparsity of signal features at different scale grids;
[0061] (4) Construct a feature fusion and classification module, splice and fuse CNN feature vectors and box dimension feature vectors, perform multi-label classification through a fully connected network, output the probability of UAV types in the spectrum, and realize the recognition of multiple types of UAVs in the mixed spectrum signal.
[0062] The step (1) includes the following steps: the size normalization processing of the multi-UAV aliased time-frequency map to be identified refers to adjusting the input map to a standard size of 512×512 pixels; for the mixed map with a width greater than 512 pixels, a sliding window strategy is used to divide it into segments with a width of 512 pixels.
[0063] For a mixed spectrum of two different frequency bands, the width is 1024 pixels and it is divided into 2 sub-spectrums; for a mixed spectrum of three different frequency bands, the width is 1536 pixels and it is divided into 3 sub-spectrums.
[0064] The segmented sub-maps are normalized to map pixel values to the [0,1] interval. The normalization formula is as follows:
[0065] (2)
[0066] in, Represents the original pixel value; and These represent the minimum and maximum pixel values in the graph, respectively. Values ; This represents the normalized pixel value.
[0067] Step (2) specifically includes the following steps:
[0068] (2a) The convolutional neural network feature extraction module adopts a five-layer convolutional network structure. The kernel size of each layer is 3×3, the stride is 2, the padding is 1, and the number of channels is 32, 64, 128, 256, and 512 respectively. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. The batch normalization formula is as follows:
[0069] (3)
[0070] in, Indicates the convolution output; and These represent the batch mean and variance, respectively. and These are learnable parameters; This is the normalized output; Values Batch normalization can accelerate network training and improve the model's generalization ability.
[0071] (2b) The channel attention module learns the importance weights of each feature channel through global average pooling and global max pooling. The formula for calculating the channel attention weights is as follows:
[0072] (4)
[0073] in, This represents the input feature map; its dimension is... ; Indicates batch size; and Indicates the height and width of the feature map; This represents a two-layer fully connected network; and These represent average pooling and max pooling, respectively; the hidden layer dimension is... ; This represents the Sigmoid activation function. Represents the channel attention weights, with dimensions of The channel attention module learns the importance weights of each channel and adaptively enhances useful channels while suppressing redundant channels.
[0074] (2c) The spatial attention module learns the importance weights of spatial locations by aggregating across channels. The formula for calculating the spatial attention weights is as follows:
[0075] (5)
[0076] in, and This represents average pooling and max pooling along the channel dimension, generating two dimensions. Feature map; This indicates a channel concatenation operation, with the concatenated dimension being... ; This represents a 7×7 convolution operation. Represents spatial attention weights, with dimension 1. The spatial attention module learns the importance weights of each spatial location, enabling the model to focus more on regions with significant signal features.
[0077] (2d) The attention-weighted feature map is compressed into a 512-dimensional CNN feature vector using global average pooling. The global average pooling formula is as follows:
[0078] (6)
[0079] in, This represents the attention-weighted feature map. This represents the feature vector after pooling.
[0080] Step (3) specifically includes the following steps:
[0081] (3a) The box dimension feature extraction module adopts a multi-scale box counting method, setting six different scale grids with grid sizes of 2×2, 4×4, 8×8, 16×16, 32×32, and 64×64, respectively. The box counting method is based on the fractal geometry principle. By covering the spectrum of grids at different scales, the number of boxes containing signal features is counted, thereby quantifying the geometric distribution characteristics of the signal.
[0082] (3b) Calculate the feature threshold for the normalized map. The threshold is the 75th percentile of the map pixel value. The calculation formula is as follows:
[0083] (7)
[0084] in, Represents the set of normalized spectral pixel values. This represents the 75th percentile operation. The feature threshold is represented by a quantile threshold instead of a fixed threshold, which can adapt to the spectrum of different signal intensities and improve the robustness of feature extraction.
[0085] (3c) For each scale The grid is used to count the number of boxes containing significant signal features, and the following criteria are used for judgment:
[0086] (8)
[0087] in, Indicates position The grid cells at that location, and These represent the maximum and mean values within the grid, respectively. This indicates whether the grid contains features; by simultaneously judging the maximum value and the mean value, it is possible to effectively distinguish between real signal features and noise interference. The maximum value condition ensures the existence of a significant signal, while the mean value condition ensures that the signal has a certain distribution range.
[0088] (3d) Count the number of effective boxes at each scale to form a 6-dimensional box-dimensional feature vector. The calculation formula is as follows:
[0089] (9)
[0090] in, Representing scale The number of effective boxes at different scales reflects the distribution characteristics of the signal at different resolutions. Small-scale grids can capture detailed features, while large-scale grids can reflect the overall distribution.
[0091] (3e) The 6-dimensional box-dimensional feature vector is processed through a two-layer fully connected network. The first layer expands the dimension from 6 to 64 and uses the ReLU activation function. The second layer expands the dimension from 64 to 128 and uses the ReLU activation function to obtain a 128-dimensional geometric feature vector.
[0092] Step (4) specifically includes the following steps:
[0093] (4a) The feature fusion module concatenates the 512-dimensional CNN feature vector with the 128-dimensional box-shaped feature vector to form a 640-dimensional fused feature vector. The concatenation formula is as follows:
[0094] (10)
[0095] in, Represents the CNN feature vector. Represents the eigenvectors of the box dimension. This represents a vector concatenation operation; feature fusion achieves a complementary combination of deep features and geometric features, CNN features capture complex texture and pattern information, and box-dimensional features characterize the geometric distribution characteristics of the signal;
[0096] (4b) The fused feature vectors are classified through a two-layer fully connected network. The first layer compresses the dimension from 640 to 256. ReLU activation function and Dropout regularization are used between layers. The formula for ReLU activation function is as follows:
[0097] (11)
[0098] in, Indicates the input to the neuron. This represents the neuron's output; the ReLU activation function has the advantages of simple computation and mitigating gradient vanishing; the Dropout ratio is set to 0.3, which effectively prevents overfitting; the second layer maps the dimension from 256 to the number of classes. , The value is 24, corresponding to 24 types of drones;
[0099] (4c) The output layer uses the Sigmoid activation function for multi-label classification, outputting the probability of existence of each drone type. The calculation formula is as follows:
[0100] (12)
[0101] in, Indicates the first Linear output of drone-like systems Indicates the first The probability of the existence of drones; using the Sigmoid activation function instead of Softmax, it can support multi-label classification, that is, multiple drone types may exist in the same map;
[0102] (4d) For multiple sub-maps after sliding window segmentation, predictions are performed separately, and the average probability is taken as the final prediction result. The fusion formula is as follows:
[0103] (13)
[0104] in, Indicates the number of subgraphs. Indicates the first Predicted probability of individual sub-maps This represents the final probability after fusion; the sliding window fusion strategy can effectively handle mixed maps of different widths while maintaining sensitivity to local features;
[0105] (4e) Select the one with the highest probability. Each type is used as the identification result. The number of mixed spectra is set according to the type of mixed spectra, for two types of mixed spectra For three types of mixed maps This enables accurate identification of multiple types of drones in mixed spectrum signals.
[0106] like Figure 2 As shown, the dual attention mechanism includes a channel attention module and a spatial attention module. The channel attention module first compresses the spatial dimension through global average pooling and global max pooling, and then learns the channel weights through a shared fully connected network. The spatial attention module aggregates channel information through channel-dimensional pooling operations, and then generates a spatial weight map through convolution.
[0107] like Figure 3 As shown, the box dimension feature extraction process includes five steps: map normalization, threshold calculation, multi-scale grid division, feature determination, and box counting. For a 512×512 input map, there are 65,536 grid cells at the 2×2 scale and 64 grid cells at the 64×64 scale.
Claims
1. A method for identifying aliased spectral signals from multiple unmanned aerial vehicles (UAVs) based on the fusion of convolutional neural networks and box-dimensional features, characterized by: (1) The size of the multi-UAV aliased time-frequency map to be identified is normalized, and the normalized map is segmented by a multi-scale sliding window to obtain a sub-map of uniform size. (2) Construct a convolutional neural network feature extraction module to extract features from the segmented sub-maps. Adaptively weight the features through a dual mechanism of channel attention and spatial attention to obtain the enhanced CNN feature vector. (3) Construct a box dimension feature extraction module, use the multi-scale box counting method to quantify the geometric distribution characteristics of the sub-map, and obtain the box dimension feature vector by statistically analyzing the distribution sparsity of signal features at different scale grids; (4) Construct a feature fusion and classification module, splice and fuse CNN feature vectors and box dimension feature vectors, perform multi-label classification through a fully connected network, output the probability of UAV types in the spectrum, and realize the identification of multi-UAV aliased spectrum signals.
2. The method for multi-type identification of UAV spectral signals based on the fusion of convolutional neural networks and box-dimensional features according to claim 1, characterized in that: Step (1) includes the following steps: The process of normalizing the size of the multi-UAV aliased time-frequency map to be identified refers to adjusting the input map to a standard size of 512×512 pixels; for mixed maps with a width greater than 512 pixels, a sliding window strategy is used to segment them into 512-pixel widths. The segmentation formula is as follows: (1) in, Indicates the number of sub-maps after segmentation; Indicates the width of the input graph; This indicates the rounding up operation; The segmented sub-maps are normalized to map pixel values to the [0,1] interval. The normalization formula is as follows: (2) in, Represents the original pixel value; and These represent the minimum and maximum pixel values in the graph, respectively. To prevent small constants from being divided by zero, they are generally taken as 1e-8; This represents the normalized pixel value.
3. The method for identifying multi-UAV aliased spectral signals based on the fusion of convolutional neural networks and box-dimensional features according to claim 1, characterized in that: Step (2) specifically includes the following steps: (2a) The convolutional neural network feature extraction module adopts a five-layer convolutional network structure. The kernel size of each layer is 3×3, the stride is 2, the padding is 1, and the number of channels is 32, 64, 128, 256, and 512 respectively. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. The batch normalization formula is as follows: (3) in, Indicates the convolution output; and These represent the batch mean and variance, respectively. and These are learnable parameters; This is the normalized output; (2b) The channel attention module learns the importance weights of each feature channel through global average pooling and global max pooling. The formula for calculating the channel attention weights is as follows: (4) in, Indicates the input feature map; Indicates a fully connected layer; This represents the Sigmoid activation function; Indicates channel attention weights; and These represent average pooling and max pooling, respectively. (2c) The spatial attention module learns the importance weights of spatial locations by aggregating across channels. The formula for calculating the spatial attention weights is as follows: (5) in, and This represents average pooling and max pooling along the channel dimension; This indicates a channel splicing operation; Indicates the convolution operation; Indicates the input feature map; Indicates spatial attention weights; (2d) Compress the attention-weighted feature map into a 512-dimensional CNN feature vector through global average pooling.
4. The method for identifying multi-UAV aliased spectral signals based on the fusion of convolutional neural networks and box-dimensional features according to claim 1, characterized in that: Step (3) specifically includes the following steps: (3a) The box dimension feature extraction module adopts the multi-scale box counting method and sets up 6 different scale grids with grid sizes of 2×2, 4×4, 8×8, 16×16, 32×32 and 64×64 respectively; (3b) Calculate the feature threshold for the normalized map. The threshold is the 75th percentile of the map pixel value. The calculation formula is as follows: (6) in, Represents the set of normalized spectral pixel values; Represents 75th percentile operations; Indicates the feature threshold; (3c) For each scale The grid is used to count the number of boxes containing significant signal features, and the following criteria are used for judgment: (7) in, Indicates position The grid cells at the location; and These represent the maximum and mean values within the grid, respectively. Indicates whether the grid contains features; (3d) Count the number of effective boxes at each scale to form a 6-dimensional box-dimensional feature vector. The calculation formula is as follows: (8) in, Representing scale The number of valid boxes below; (3e) The 6-dimensional box-dimensional feature vector is processed through a two-layer fully connected network. The first layer expands the dimension from 6 to 64, and the second layer expands it from 64 to 128, resulting in a 128-dimensional geometric feature vector.
5. The method for identifying multi-UAV aliased spectral signals based on the fusion of convolutional neural networks and box-dimensional features according to claim 1, characterized in that: Step (4) specifically includes the following steps: (4a) The feature fusion module concatenates the 512-dimensional CNN feature vector with the 128-dimensional box-shaped feature vector to form a 640-dimensional fused feature vector. The concatenation formula is as follows: (9) in, Represents the CNN feature vector; Represents the eigenvectors of the box dimension; This represents a vector concatenation operation; (4b) The fused feature vectors are classified through a two-layer fully connected network. The first layer compresses the dimension from 640 to 256, and the second layer maps the dimension from 256 to the number of categories. ReLU activation function and Dropout regularization are used between layers, and the Dropout ratio is set to 0.3; (4c) The output layer uses the Sigmoid activation function for multi-label classification, outputting the probability of existence of each drone type. The calculation formula is as follows: (10) in, Indicates the first Linear output of drone-like systems; Indicates the first The probability of the existence of drone-like systems; (4d) For multiple sub-maps after sliding window segmentation, predictions are performed separately, and the average probability is taken as the final prediction result. The fusion formula is as follows: (11) in, Indicates the number of sub-maps; Indicates the first Predicted probability of individual sub-maps; This represents the final probability after fusion; (4e) Select the one with the highest probability. Using the identified types as the recognition results, the system can accurately identify multiple types of drones in mixed spectrum signals.