A multi-dimensional feature data multi-level fusion unmanned aerial vehicle target type identification method

Abstract

The application discloses a multi-dimensional feature data multi-stage fusion unmanned aerial vehicle target type identification method, relates to the unmanned aerial vehicle identification technical field, and comprises the following steps: acquiring multi-modal sensor data, constructing a double-branch multi-stage fusion identification architecture, multi-modal signal feature fusion identification branch processing, photoelectric video identification branch processing, decision-level fusion and result output; the application constructs a double-branch multi-stage fusion identification architecture, respectively processes photoelectric video data and multi-modal signal data, the multi-modal signal feature fusion identification branch adopts an independent modal encoder to map each modal feature to a unified space, and performs feature-level weighted fusion based on a cross-modal attention mechanism, adaptively captures complementary information between different modes, the photoelectric video identification branch efficiently detects small targets based on an improved YOLOv11 network, the identification result is dynamically weighted and fused, and the identification precision and system robustness of the unmanned aerial vehicle target type in a complex environment are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of drone identification technology, and in particular to a drone target type identification method based on multi-dimensional feature data multi-level fusion. Background Technology

[0002] With the rapid development of drone technology and the continuous reduction in cost, its application in various fields is becoming more and more widespread. However, the low-altitude safety issues that have arisen are also becoming more prominent. Rapid and accurate detection and identification of drones has become crucial in fields such as low-altitude security.

[0003] Most existing drone identification methods rely on single sensor data, such as visible light video or radar signals. However, single-modal methods have significant limitations: optical video is susceptible to lighting and weather conditions; radar data lacks fine visual features; and radio signals have poor stability in complex electromagnetic environments.

[0004] Although existing research has attempted to fuse multimodal data to compensate for the limitations of a single sensor, current fusion methods typically suffer from the following problems: 1. Shallow fusion level: Most methods only perform simple fusion at the data layer (original data splicing) or the decision layer (simple voting or weighted averaging), failing to fully explore the complementarity of different modal data at the feature level, resulting in limited fusion effect.

[0005] 2. Poor performance in detecting small targets: Long-distance, small-sized drone targets account for a very small percentage of features in optoelectronic video data. Current mainstream target detectors (such as the YOLO series) typically use multiple stride convolution and pooling operations to extract high-level semantic features, which may cause drone targets to be compressed to 1-2 pixels or even completely lost in deep feature maps, easily leading to missed detections and false detections.

[0006] 3. Poor environmental adaptability: In actual deployments, interference from complex environments frequently leads to the loss of some sensor data (such as video signal loss or radar signal interruption). Existing methods lack robust processing mechanisms for modal data loss; when data for a particular modality is missing, system performance deteriorates sharply, even causing detection and identification results to fail.

[0007] Therefore, this invention proposes a multi-dimensional feature data multi-level fusion method for UAV target type identification to solve the problems existing in the prior art. Summary of the Invention

[0008] To address the aforementioned problems, the present invention aims to propose a multi-dimensional feature data multi-level fusion method for UAV target type identification. By constructing a dual-branch identification architecture, photoelectric video data and multimodal signal data are processed separately, and deep fusion is performed at the feature level and decision level respectively, ultimately achieving high-precision and highly robust UAV type identification.

[0009] To achieve the objectives of this invention, the present invention is implemented through the following technical solution: a method for identifying UAV target types through multi-level fusion of multi-dimensional feature data, comprising the following steps: S1. Acquire multimodal sensor data Acquire multimodal sensor data, including photoelectric video data, radar track data, radio signal data, and photoelectric track data; S2. Construct a dual-branch, multi-level fusion recognition architecture A dual-branch, multi-level fusion recognition architecture is constructed, which includes a multimodal signal feature fusion recognition branch and an optoelectronic video recognition branch. S3, Multimodal signal feature fusion and recognition branch processing In the multimodal signal feature fusion and recognition branch, radar track data, radio signal data, and photoelectric track data are processed by available modality discrimination, feature extraction, feature encoding, and cross-modal attention fusion, and the first recognition result and its first confidence level are output. S4, Photoelectric Video Recognition Branch Processing In the optoelectronic video recognition branch, the improved YOLOv11 network is used to perform UAV target detection and recognition processing on video frames in optoelectronic video data, and outputs a second recognition result and its second confidence level. S5, Decision-level Fusion and Result Output The first and second identification results are fused at the decision level. The weighting coefficients are dynamically adjusted according to the first and second confidence levels. The identification results output by the multimodal signal feature fusion identification branch and the photoelectric video identification branch are weighted and fused to output the final UAV target type identification result.

[0010] A further improvement lies in the following: In step S3, the specific steps for processing radar track data, radio signal data, and electro-optical track data are as follows: Available mode discrimination: Detect the existence of radar track data, radio signal data and electro-optical track data, eliminate unusable modes, and only perform subsequent processing on available mode data; Feature vector initialization: Assign a fixed-length feature vector to each modality. If data for one modality is missing, set all of its feature vectors to zero. At the same time, generate a modality mask vector to mark the availability of each modality, which is used to mask the missing modality during subsequent attention fusion. Feature extraction: Extract radar track features from radar track data, extract electro-optical track features from electro-optical track data, extract radio signal features from radio signal data, and combine the extracted features into an initial feature vector of fixed length; Feature encoding: An independent modality encoder is set up for each available modality. Each modality encoder maps the features of each modality to a unified 128-dimensional feature space through a linear transformation layer, a batch normalization layer, a ReLU activation function layer, a Dropout layer and a residual block sequence. Cross-modal attention fusion: The encoded feature vectors of each modality are stacked into a three-dimensional tensor, which is then input into a multi-head self-attention layer to calculate the cross-modal attention score. The attention scores of the effective modalities are normalized by softmax to obtain the fusion weight. The features of each modality are weighted and fused, and then passed through residual connection, layer normalization and global average pooling to output the fused feature vector. Classification output: The fused feature vector is input into a multi-layer fully connected classification network, which outputs the first recognition result and its first confidence score.

[0011] Further improvements are made in the following aspects: the radar track features include statistics on slant range, radial velocity, azimuth angle, elevation angle and their rate of change; the photoelectric track features include position mean, range, standard deviation, rate of change and their trajectory fitting features; and the radio signal features include azimuth angle, mean, standard deviation, mode and their fluctuation features of uplink and downlink frequencies.

[0012] Further improvements are made in the following ways: In the cross-modal attention fusion, the input features are weighted by the importance of learnable feature selection, the feature vectors are mapped to a unified 128-dimensional space using linear and normalization layers, BatchNorm and Dropout are used to prevent overfitting, residual block sequences are used for feature encoding, all modal features are stacked, Transformer multi-head self-attention is introduced to calculate attention scores, attention weights are generated, the features are weighted and fused, and the fused feature vector is output through residual linking, layer normalization and global average pooling.

[0013] Further improvements are made in the following aspects: The photoelectric video recognition branch adopts an improved YOLOv11 network, including a Backbone feature extraction unit, a Neck feature fusion unit, and a Head detection head. The Backbone feature extraction unit replaces the strided convolution and pooling layers in YOLOv11 with the SPDConv module to preserve the detailed information of small targets. The Neck feature fusion unit adopts the SABS-FPN structure and introduces the SBA module to achieve bidirectional fusion of shallow boundary information and deep semantic information. The Head detection head uses the Focaler-PIoUv2 loss function to replace the original CIoU loss, thereby improving the localization accuracy of small targets and the learning ability of difficult samples.

[0014] A further improvement is made in that the SPDConv module includes a spatial-to-depth (SPD) layer and a non-stretch convolutional (Conv) layer. The spatial-to-depth layer reduces each spatial dimension of the input feature map to the channel dimension while preserving the information within each channel. The non-stretch convolutional layer performs convolution processing on the transformed features to compress the channel dimension, as expressed in the following expression: in, f 0,0 Indicates input X Width and height dimensions start from 0, and... S End, based on step size scale This step involves truncating the input feature map. X Segment along the width and height dimensions.

[0015] A further improvement is that the SABS-FPN structure includes an SBA module, which selectively aggregates shallow and deep features through two RAU units. The RAU process is represented as follows: in, T 1 and T 2 is the input feature, obtained through two linear mappings and sigmoid function and The input features are processed to reduce the channel dimension to 32, resulting in a feature map. and , This represents point-by-point multiplication. This means subtracting features The reverse operation refines the imprecise and coarse estimates into accurate and complete prediction maps. RAU uses a 1×1 kernel convolution operation as a linear mapping process. The overall process of SBA can be represented as follows: in, F S These are shallow features. F b For deep features, C 3×3 This represents a 3×3 convolutional layer.

[0016] A further improvement is that the Focaler-PIoUv2 loss function is the Focaler-IoU loss applied to the PIOUv2 bounding box regression loss function, and its definition is: U=1- + in, The reconstructed Focaler-IoU is defined as: in, By adjusting d and u The value of makes Focus on different regression samples, The penalty factor for adaptive target size is defined as follows: in, , , , These are the absolute values ​​of the distances between the corresponding edges of the predicted bounding box and the target bounding box, respectively. and This represents the width and height of the target bounding box. Since the denominator of this coefficient depends only on the size of the target bounding box and is independent of the size of the anchor box and the smallest outer bounding box within the target box, it is used in the loss function. p As a penalty factor, it does not lead to an increase in the anchor frame size; PIoUv2 adds a non-monotonic attention mechanism to the PIOU architecture. Its attention function and loss are defined as follows: in, Represents the attention function. q As a penalty factor, used to measure anchor frame quality. It is a hyperparameter that controls the behavior of the attention function, and its value is 1.3.

[0017] A further improvement is made in S5, where the specific steps of the decision-level fusion are as follows: the recognition results output by the multimodal signal feature fusion recognition branch and the photoelectric video recognition branch are weighted and averaged, the weights are dynamically adjusted according to the confidence of each branch, and the category with the highest probability is output as the final recognition result, along with a comprehensive confidence score.

[0018] A further improvement is that the method supports target recognition based on available modal data even when some modal data is missing, demonstrating robustness and environmental adaptability.

[0019] The beneficial effects of this invention are as follows: Deep fusion of multi-source features improves recognition accuracy: This invention constructs a dual-branch + multi-level fusion architecture, fusing different modal data at the feature level and decision level respectively. For multimodal signal features, independent modal encoders map various features to a unified space, making similar features more clustered and dissimilar features more separated. A cross-modal attention mechanism is introduced to establish global context awareness and dynamically learn the importance weights of each modal feature. Refined and adaptive weighting replaces traditional fixed-weight fusion, fully exploring the complementary information of different modal data. At the same time, the fusion of photoelectric and video recognition results at the decision level further improves the accuracy and reliability of the final recognition.

[0020] Superior small target detection capability: In the optoelectronic video recognition branch, this invention uses SPDConv to replace the strided convolution and pooling layers of the YOLOv11 feature extraction network, constructs the SABS-FPN network structure, and uses the SBA module to selectively aggregate boundary information and semantic information. Combined with PIoUv2 and Focaler-IoU loss, it effectively reduces information loss, enhances the extraction of important details, and significantly improves the model's feature preservation ability, feature fusion efficiency and positioning accuracy for distant and small-sized UAV targets.

[0021] High flexibility and system robustness: This invention designs a dual-branch independent operating architecture and uses a modal masking mechanism to process multimodal data. When some modal data (such as radar signals, photoelectric video or radio signals) are missing due to environmental interference, the system can automatically detect available modalities, and use zero-filling and masking to shield the missing modalities. It can complete feature extraction, encoding and fusion based only on the effective modalities and still output reliable recognition results. This mechanism enables the system to have strong adaptability in complex dynamic environments and greatly improves its engineering practical value. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the UAV target type identification method based on multi-dimensional feature data multi-level fusion of the present invention. Figure 2 This is an overall architecture diagram of the UAV target type identification method based on multi-dimensional feature data multi-level fusion of the present invention; Figure 3 This is a schematic diagram of the photoelectric video recognition branch SBA structure of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] It should be noted that the technical means not described in detail in the following embodiments are all conventional means in the art, are not the key points of the invention, and will not be elaborated upon.

[0025] See Figure 1 , Figure 2 , Figure 3 This embodiment provides a method for UAV target type identification by multi-dimensional feature data fusion at multiple levels. This method can read photoelectric, radar, and radio data, and perform feature extraction and encoding using different branches. It fuses different modal data at the feature level and decision level, and outputs UAV type identification results. The specific implementation steps include: S1. Acquire multimodal sensor data Read raw data from multiple sensors, including: Optoelectronic video data: a continuous video stream from a visible light or infrared camera; Radar track data: includes target slant range, azimuth, elevation angle, slant range change rate, azimuth change rate, and elevation angle change rate; Radio signal data: including the azimuth angle of the detected UAV communication signals, the target uplink frequency, and the downlink frequency; Electro-optical trajectory data: azimuth and elevation angles of the target obtained through preliminary electro-optical detection.

[0026] The system detects whether any modal data is missing. If so, processing for that modality is abandoned, and subsequent recognition is performed using only the remaining modal data. For example, if photoelectric video data cannot be acquired during the current recognition process, the photoelectric video recognition branch is abandoned, and the remaining available modal data is processed to output the recognition result. This operation ensures that even when some modal data is missing due to complex environmental factors, the system can still effectively recognize the data using the remaining modal data, improving the system's environmental adaptability.

[0027] Data preprocessing is performed on radar, radio, and electro-optical track data. Outlier data is removed through an outlier detection mechanism to avoid the influence of outliers. The data is then standardized using the Z-Score method to avoid the influence of different distribution intervals on the identification results, eliminate the influence of dimensions, and improve data quality.

[0028] S2. Construct a dual-branch, multi-level fusion recognition architecture Build as Figure 2 The dual-branch, multi-level fusion recognition architecture shown includes: Multimodal signal feature fusion and recognition branch: used to process radar track data, radio signal data and electro-optical track data, fuse multi-source signals at the feature level, and output the first recognition result and its confidence level; Optoelectronic video recognition branch: Used to process optoelectronic video data, perform target detection and recognition based on the improved YOLOv11 network, and output a second recognition result and its confidence level.

[0029] The two branches run in parallel and are eventually merged at the decision level.

[0030] S3, Multimodal signal feature fusion and recognition branch processing This step corresponds to Figure 2 The left branch specifically includes the following sub-steps: S3.1 Available Mode Identification: Detect the existence of radar track data, radio signal data and electro-optical track data, eliminate unusable modes, and only process the available mode data.

[0031] S3.2 Feature Extraction: Extract radar track features from radar track data, extract electro-optical track features from electro-optical track data, and extract radio signal features from radio signal data. Specific features are as follows:

[0032] Radar track features: For radar track time series data, the core of feature extraction is to discover the differences in behavioral characteristics, motion change patterns, and maneuverability of different UAV targets. Therefore, this invention extracts target features from multiple dimensions, including basic kinematic features, temporal dynamic trend features, and motion patterns. The specific features and logic are shown in Table 1 below: Table 1 Radar Data Characteristics Electro-optical track characteristics: Electro-optical data characteristics: For electro-optical track data, the core is to extract the target's dynamic change trend, differences in motion patterns, etc. The extracted features are shown in Table 2 below: Table 2 Characteristics of Electro-optical Track Data Radio signal characteristics: Radio signal data includes "uplink frequency", "downlink frequency", and "azimuth". The analysis mainly focuses on the static statistical characteristics and dynamic change patterns of the target. The extracted features and their significance are shown in Table 3 below: Table 3 Radio Data Characteristics The extracted features are combined into an initial feature vector of fixed length. If data for a certain modality is missing, its feature vector is set to zero, and a modality mask vector is generated to indicate the availability of each modality.

[0033] S3.3 Feature Encoding: An independent modality encoder is set up for each available modality. Each modality encoder adopts a deep neural network architecture, including a linear transformation layer, a batch normalization layer, a ReLU activation function layer, a Dropout layer, and a residual block sequence, mapping the features of each modality to a unified 128-dimensional feature space. This operation makes similar features more clustered in the encoded space, and different types of features more separated, while combining basic low-level features into higher-level, more abstract semantic concepts.

[0034] S3.4 Cross-modal attention fusion: The encoded feature vectors of each modality are stacked into a 3D tensor and input into the Transformer multi-head self-attention layer to calculate the cross-modal attention score. The modal mask is applied to set the attention score of unusable modalities to negative infinity, ensuring that the weights are 0 in the subsequent softmax. The attention scores of the effective modalities are normalized by softmax to obtain the fusion weights of each modality. The features of each modality are weighted and fused to generate a unified feature representation after fusion. Then, after residual connection, layer normalization and global average pooling, the fused feature vector is output.

[0035] This step establishes global context awareness through an attention mechanism, capturing various complex dependencies between features and adaptively assigning different feature importance based on attention weights. For example, under adverse weather conditions, when the quality of photoelectric signals deteriorates, the attention mechanism automatically reduces the weight of optical modes and increases the weight of radar and radio modes, ensuring that the system can maintain a high recognition accuracy even in complex environments.

[0036] S3.5 Classification Output: The fused feature vector is input into a multi-layer fully connected classification network for target classification. The output class probability vector serves as the first recognition result. Simultaneously, a reliability score for this result is calculated through an independent uncertainty prediction branch, serving as the first confidence level. During training, the weighted sum of classification loss, confidence loss, and uncertainty loss is used as the total loss function, with a weight coefficient of 0.6 for classification loss, 0.2 for confidence loss, and 0.2 for uncertainty loss.

[0037] S4, Photoelectric Video Recognition Branch Processing This step corresponds to Figure 2 The right-hand branch performs UAV target detection and recognition on video frames in optoelectronic video data based on an improved YOLOv11 network. Specifically, it includes:

[0038] S4.1 Backbone Feature Extraction Unit: The strided convolutional and pooling layers in YOLOv11 are replaced with the SPDConv (Symmetric Positive Definite Convolution) module. The SPDConv module consists of a spatial-to-depth layer and a non-strided convolutional layer. Taking scale=2 as an example, the input feature map X is divided along its width and height dimensions by a factor of 2, resulting in four sub-maps: f 0,0 , f 0,1 , f 1,0 , f 1,1 .

[0039] Connect the four sub-feature maps along the channel dimension to obtain the feature map. X' The spatial dimension is reduced to half of its original size, while the channel dimension is increased to four times its original size. Finally, a convolutional layer with a stride of 1 compresses the channel dimension to the target dimension. This module effectively reduces information loss and enhances the extraction of details from small targets.

[0040] S4.2 Neck Feature Fusion Unit: Employing the SABS-FPN (selectively aggregate the boundary and semantic-FPN) structure, it introduces the SBASBA (Selective Boundary Aggregation) module to achieve bidirectional fusion of shallow boundary information and deep semantic information. For example... Figure 3 As shown, the SBA module selectively aggregates shallow and deep features through two RAU units. The RAU process is represented as follows:

[0041] in, T 1 and T 2 represents the input features. and This indicates that through 1×1 convolution and sigmoid A linear mapping of a function, through two linear mappings and sigmoid The function processes the input features, reducing the channel dimension to 32, and obtains the feature map. and , This represents point-by-point multiplication. This means subtracting features The reverse operation refines the imprecise and coarse estimates into accurate and complete prediction maps. RAU uses a 1×1 kernel convolution operation as a linear mapping process. The overall process of SBA can be represented as follows: in, F S These are shallow features. F b For deep features, C 3×3 Represents a 3×3 convolutional layer S4.3 Head Detector: The original CIoU loss is replaced with the Focaler-PIoUv2 loss function. The Focaler-PIoUv2 loss function is defined as follows:

[0042] U=1- + in, The reconstructed Focaler-IoU is defined as: in, By adjusting d and u The value of makes Focus on different regression samples, The penalty factor for adaptive target size is defined as follows: in, , , , These are the absolute values ​​of the distances between the corresponding edges of the predicted bounding box and the target bounding box, respectively. and This represents the width and height of the target bounding box. Since the denominator of this coefficient depends only on the size of the target bounding box and is independent of the size of the anchor box and the smallest outer bounding box within the target box, it is used in the loss function. p As a penalty factor, it does not lead to an increase in the anchor frame size; PIoUv2 adds a non-monotonic attention mechanism to the PIOU architecture. Its attention function and loss are defined as follows: in, Represents the attention function. q As a penalty factor, used to measure anchor frame quality. It is a hyperparameter that controls the behavior of the attention function, and its value is 1.3.

[0043] This loss function makes the algorithm pay more attention to difficult samples, improving the accuracy of small target localization.

[0044] Finally, the photoelectric video recognition branch outputs the second recognition result (category probability vector) and its second confidence level.

[0045] S5, Decision-level Fusion and Result Output Based on the recognition results of the two different branches S3 and S4, a weighted average is calculated on the recognition results of the two branches. The weights are dynamically adjusted according to the confidence of each branch. The category with the highest output probability is taken as the final recognition result, and a comprehensive confidence score is attached.

[0046] When some modal data is missing, the system automatically and flexibly uses the remaining available modal data to complete UAV target identification. For example, when there is no photoelectric or video data, only the multimodal signal feature fusion identification branch result is output; when both radar tracks and photoelectric tracks are missing, the model performs preliminary identification based on photoelectric, video, and radio signals respectively, and then performs fusion at the decision level to output the final result.

[0047] 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 identifying UAV target types through multi-dimensional feature data fusion at multiple levels, characterized in that, Includes the following steps: S1. Acquire multimodal sensor data, which includes photoelectric video data, radar track data, radio signal data, and photoelectric track data; S2. Construct a dual-branch multi-level fusion recognition architecture, which includes a multimodal signal feature fusion recognition branch and an optoelectronic video recognition branch; S3. In the multimodal signal feature fusion and recognition branch, radar track data, radio signal data, and photoelectric track data are processed by available modality discrimination, feature extraction, feature encoding, and cross-modal attention fusion, and the first recognition result and its first confidence level are output. S4. In the photoelectric video recognition branch, the video frames in the photoelectric video data are processed for UAV target detection and recognition based on the improved YOLOv11 network, and the second recognition result and its second confidence level are output. S5. Perform decision-level fusion on the first and second identification results, dynamically adjust the weighting coefficients according to the first and second confidence levels, and perform weighted fusion on the identification results output by the multimodal signal feature fusion identification branch and the photoelectric video identification branch to output the final UAV target type identification result.

2. The method for UAV target type identification based on multi-level fusion of multi-dimensional feature data according to claim 1, characterized in that: In step S3, the specific steps for processing radar track data, radio signal data, and electro-optical track data are as follows: Detecting the existence of radar track data, radio signal data, and electro-optical track data; eliminating unusable modes; performing subsequent processing only on usable mode data; assigning a fixed-length feature vector to each mode; if one mode's data is missing, setting all its feature vectors to zero; simultaneously generating a mode mask vector to mark the availability of each mode, used to mask missing modes during subsequent attention fusion; extracting radar track features from radar track data; extracting electro-optical track features from electro-optical track data; extracting radio signal features from radio signal data; and combining the extracted features into a fixed-length initial feature set. For each available modality, an independent modality encoder is set. Each modality encoder maps the modal features to a unified 128-dimensional feature space through a linear transformation layer, a batch normalization layer, a ReLU activation function layer, a Dropout layer, and a residual block sequence. The encoded modal feature vectors are stacked into a three-dimensional tensor and input into a multi-head self-attention layer to calculate cross-modal attention scores. The attention scores of the effective modalities are normalized by softmax to obtain fusion weights. The features of each modality are weighted and fused, and then passed through residual connections, layer normalization, and global average pooling to output the fused feature vector. The fused feature vector is input into a multi-layer fully connected classification network to output the first recognition result and its first confidence score.

3. The method for UAV target type identification based on multi-level fusion of multi-dimensional feature data according to claim 2, characterized in that: The radar track features include statistics on slant range, radial velocity, azimuth, elevation angle and their rate of change; the photoelectric track features include position mean, range, standard deviation, rate of change and their trajectory fitting features; and the radio signal features include azimuth, mean, standard deviation, mode and their fluctuation features of uplink and downlink frequencies.

4. The method for UAV target type identification based on multi-level fusion of multi-dimensional feature data according to claim 1, characterized in that: The photoelectric video recognition branch adopts an improved YOLOv11 network, including a Backbone feature extraction unit, a Neck feature fusion unit, and a Head detection head. The Backbone feature extraction unit replaces the strided convolution and pooling layers in YOLOv11 with the SPDConv module. The Neck feature fusion unit adopts the SABS-FPN structure and introduces the SBA module to achieve bidirectional fusion of shallow boundary information and deep semantic information. The Head detection head uses the Focaler-PIoUv2 loss function to replace the original CIoU loss.

5. The method for UAV target type identification based on multi-level fusion of multi-dimensional feature data according to claim 4, characterized in that: The SPDConv module includes a spatial-to-depth layer and a non-stretched convolutional layer. The spatial-to-depth layer reduces each spatial dimension of the input feature map to the channel dimension while preserving the information within each channel. The non-stretched convolutional layer performs convolution processing on the transformed features to compress the channel dimension, as expressed in the following expression: in, f 0,0 Indicates input X Width and height dimensions start from 0, and... S End, based on step size scale This step involves truncating the input feature map. X Segment along the width and height dimensions.

6. The method for UAV target type identification based on multi-level fusion of multi-dimensional feature data according to claim 4, characterized in that: The SABS-FPN structure includes an SBA module, which selectively aggregates shallow and deep features through two RAU units. The RAU process is represented as follows: in, T 1 and T 2 is the input feature, obtained through two linear mappings and sigmoid function and The input features are processed to reduce the channel dimension to 32, resulting in a feature map. and , This represents point-by-point multiplication. This means subtracting features The reverse operation refines the imprecise and coarse estimates into accurate and complete prediction maps. RAU uses a 1×1 kernel convolution operation as a linear mapping process. The overall process of SBA can be represented as follows: in, F S These are shallow features. F b For deep features, C 3×3 This represents a 3×3 convolutional layer.

7. The method for UAV target type identification based on multi-level fusion of multi-dimensional feature data according to claim 4, characterized in that: The Focaler-PIoUv2 loss function is the Focaler-IoU loss applied to the PIOUv2 bounding box regression loss function, and it is defined as follows: U=1- + in, The reconstructed Focaler-IoU is defined as: in, By adjusting d and u The value of makes Focus on different regression samples, The penalty factor for adaptive target size is defined as follows: in, , , , These are the absolute values ​​of the distances between the corresponding edges of the predicted bounding box and the target bounding box, respectively. and These represent the width and height of the target bounding box, used in the loss function. p As a penalty factor, it does not lead to an increase in the anchor frame size; PIoUv2 adds a non-monotonic attention mechanism to the PIOU architecture. Its attention function and loss are defined as follows: in, Represents the attention function. q As a penalty factor, It is a hyperparameter that controls the behavior of the attention function.

8. The method for UAV target type identification based on multi-level fusion of multi-dimensional feature data according to claim 1, characterized in that: In step S5, the specific steps of the decision-level fusion are as follows: the recognition results output by the multimodal signal feature fusion recognition branch and the photoelectric video recognition branch are weighted and averaged. The weights are dynamically adjusted according to the confidence of each branch. The category with the highest output probability is used as the final recognition result, and a comprehensive confidence score is attached.

Images