Unmanned aerial vehicle pilot signal identification method and device and computer equipment

Abstract

The application relates to a UAV pilot signal identification method, device and computer equipment. The method comprises the following steps: obtaining a training sample, constructing a pilot signal identification model, performing multi-scale feature extraction on an input time-frequency spectrum through a feature extraction unit, dividing an initial feature map into non-overlapping local frequency blocks according to continuous frequency segments, independently completing channel dimension adaptive weight distribution in the blocks, and outputting an enhanced multi-scale feature map; performing bidirectional cross-scale feature interaction and local dynamic weighted fusion through a feature fusion unit, outputting a multi-scale fusion feature map, completing boundary box regression and classification prediction through a detection output unit, and outputting an identification result; using a training sample, training the pilot signal identification model by using a loss function, inputting a to-be-identified signal time-frequency spectrum into the trained model, and obtaining a UAV pilot signal identification result. The method can balance model lightweight and identification accuracy and robustness in a complex electromagnetic environment.

Description

Technical Field

[0001] This application relates to the field of signal recognition technology, and in particular to a method, apparatus and computer equipment for identifying pilot signals of unmanned aerial vehicles. Background Technology

[0002] With the continuous development of UAV monitoring and low-altitude security technologies, UAV signal recognition methods based on time-frequency maps and deep learning have gradually become a research hotspot. The image transmission signals widely used in civilian UAVs are mostly based on OFDM multi-carrier modulation, and during transmission, pilot signals of a fixed format are embedded for channel estimation and synchronization. These pilot signals appear as narrow-band, short-duration, low-energy small targets in the time-frequency map, and the overall time-frequency maps of different UAV models under the same protocol are highly similar, distinguishable only by subtle frequency domain structures of the pilot signals. Currently, traditional UAV pilot signal recognition schemes often feed the time-frequency maps constructed by short-time Fourier transform or wavelet transform into convolutional neural networks for feature extraction and classification, or, based on this, introduce time-frequency map enhancement preprocessing, improved feature pyramids, lightweight channel attention mechanisms, adjusting anchor frame scale and adding high-resolution detection heads, model quantization and structural simplification, and hybrid loss function optimization, in order to improve the detection and recognition of weak pilot targets.

[0003] However, existing technologies are difficult to adapt to the inherent characteristics of narrowband, short-term, low-energy operation and complex electromagnetic environment. High-frequency subtle frequency domain features are easily lost in multi-scale sampling. Small differences between different models of the same protocol cannot be effectively extracted. Local point-like electromagnetic noise is difficult to filter accurately. Effective signal channels are often confused with noise channels. There is a lack of information exchange and coordination between modules. The problem of extreme sample imbalance has never been properly solved. Ultimately, it is impossible to meet the dual requirements of real-time and high-precision detection under the computing power constraints of edge devices. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, and computer equipment for identifying pilot signals of unmanned aerial vehicles (UAVs) to address the aforementioned technical problems.

[0005] A method for identifying pilot signals of unmanned aerial vehicles (UAVs), the method comprising: Obtain training samples; the training samples include time-frequency spectra labeled with UAV pilot signals; A pilot signal recognition model is constructed; the pilot signal recognition model includes a feature extraction unit, a feature fusion unit, and a detection output unit. The feature extraction unit performs multi-scale feature extraction on the input time-frequency spectrum to obtain multiple initial feature maps with different resolutions. Each initial feature map is divided into multiple non-overlapping local frequency blocks according to continuous frequency segments. Adaptive weight allocation of the channel dimension is performed independently within each local frequency block. During model training, the channel weights are optimized according to the inverse gradient corresponding to the high-level semantic features output by the feature fusion unit, and the enhanced multi-scale feature map is output. The feature fusion unit performs frequency offset alignment and feature position registration on the enhanced multi-scale feature map. Then, it performs bidirectional cross-scale feature interaction from top to bottom and bottom to top on the aligned feature map. During the interaction, learnable non-negative weights are assigned to the local feature range for dynamic weighted fusion, and a multi-scale fused feature map is output. The detection output unit performs bounding box regression and classification prediction on the multi-scale fused feature map and outputs the recognition result of the pilot signal. Using the training samples, the pilot signal recognition model is trained using a pre-set loss function to obtain the trained pilot signal recognition model; The time-frequency spectrum corresponding to the UAV communication signal to be identified is input into the trained pilot signal recognition model, and the recognition result of the UAV pilot signal in the time-frequency spectrum is output.

[0006] A drone pilot signal identification device, the device comprising: A sample acquisition module is used to acquire training samples; the training samples include time-frequency spectra labeled with UAV pilot signals; The model building module is used to build a pilot signal recognition model; the pilot signal recognition model includes a feature extraction unit, a feature fusion unit, and a detection output unit. The feature enhancement processing module is used to perform multi-scale feature extraction on the input time-frequency spectrum through the feature extraction unit to obtain multiple initial feature maps with different resolutions. Each initial feature map is divided into multiple non-overlapping local frequency blocks according to continuous frequency segments. Adaptive weight allocation of the channel dimension is performed independently within each local frequency block. During model training, the channel weights are optimized according to the inverse gradient corresponding to the high-level semantic features output by the feature fusion unit, and the enhanced multi-scale feature map is output. The multi-scale feature fusion module is used to perform frequency offset alignment and feature position registration on the enhanced multi-scale feature map through the feature fusion unit, and then perform bidirectional cross-scale feature interaction from top to bottom and from bottom to top on the aligned feature map. During the interaction process, learnable non-negative weights are assigned to the local feature range for dynamic weighted fusion, and the multi-scale fused feature map is output. The signal recognition module is used to perform bounding box regression and classification prediction on the multi-scale fused feature map through the detection output unit, and output the recognition result of the pilot signal; The model training module is used to train the pilot signal recognition model using the training samples and a pre-set loss function to obtain the trained pilot signal recognition model. The result output module is used to input the time-frequency spectrum corresponding to the UAV communication signal to be identified into the trained pilot signal recognition model, and output the recognition result of the UAV pilot signal in the time-frequency spectrum.

[0007] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps: Obtain training samples; the training samples include time-frequency spectra labeled with UAV pilot signals; A pilot signal recognition model is constructed; the pilot signal recognition model includes a feature extraction unit, a feature fusion unit, and a detection output unit. The feature extraction unit performs multi-scale feature extraction on the input time-frequency spectrum to obtain multiple initial feature maps with different resolutions. Each initial feature map is divided into multiple non-overlapping local frequency blocks according to continuous frequency segments. Adaptive weight allocation of the channel dimension is performed independently within each local frequency block. During model training, the channel weights are optimized according to the inverse gradient corresponding to the high-level semantic features output by the feature fusion unit, and the enhanced multi-scale feature map is output. The feature fusion unit performs frequency offset alignment and feature position registration on the enhanced multi-scale feature map. Then, it performs bidirectional cross-scale feature interaction from top to bottom and bottom to top on the aligned feature map. During the interaction, learnable non-negative weights are assigned to the local feature range for dynamic weighted fusion, and a multi-scale fused feature map is output. The detection output unit performs bounding box regression and classification prediction on the multi-scale fused feature map and outputs the recognition result of the pilot signal. Using the training samples, the pilot signal recognition model is trained using a pre-set loss function to obtain the trained pilot signal recognition model; The time-frequency spectrum corresponding to the UAV communication signal to be identified is input into the trained pilot signal recognition model, and the recognition result of the UAV pilot signal in the time-frequency spectrum is output.

[0008] The aforementioned UAV pilot signal recognition method, device, and computer equipment extract multi-scale features from the time-frequency spectrum and perform local channel adaptive weight allocation using frequency domain segmentation. Simultaneously, relying on the back gradient of high-level semantic features to dynamically optimize channel weights, they can accurately enhance the effective features corresponding to the pilot signal and suppress redundant noise features. While ensuring the model's lightweight characteristics and without increasing computational overhead, they significantly improve the feature representation capability of weak pilot signals, effectively addressing signal detection errors in low signal-to-noise ratio scenarios. By performing frequency offset alignment and feature position registration on the enhanced feature map, and combining dynamic weighted fusion of local feature ranges to achieve bidirectional cross-scale feature interaction, they can preserve subtle frequency domain differences in pilot signals, significantly improving the ability to finely distinguish pilot signals from different UAV models using the same protocol. Furthermore, this invention uses a preset loss function to complete iterative model training, ensuring overall model convergence and detection stability. The embodiments of this invention can significantly improve the recognition accuracy and robustness of UAV pilot signals in complex electromagnetic environments while maintaining model lightweightness, meeting the real-time detection needs of embedded edge devices. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating a method for identifying pilot signals of a drone in one embodiment; Figure 2 This is a schematic diagram of the BEF-YOLOv8 structure in one embodiment; Figure 3 This is a schematic diagram of the workflow of a frequency domain block-based local ECA module in one embodiment; Figure 4 This is a comparative diagram of the workflows of FPN, PANet, and improved BiFPN in one embodiment; Figure 5 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0011] In one embodiment, such as Figure 1 As shown, a method for identifying pilot signals of a UAV is provided, including the following steps: Step 102: Obtain training samples. The training samples include time-frequency spectra labeled with UAV pilot signals.

[0012] Pilot signals for unmanned aerial vehicles (UAVs) are fixed-format signals embedded in UAV OFDM communication systems. They are used for channel estimation, carrier synchronization, and symbol synchronization, and appear as narrow-band, short-duration stripes in time-frequency spectra. Time-frequency spectra are generated by using time-frequency analysis methods such as short-time Fourier transform to convert one-dimensional time-domain communication signals into a two-dimensional time-frequency distribution image, which can intuitively show the energy distribution of the signal at different times and frequencies.

[0013] It is understandable that obtaining time-frequency spectra labeled with UAV pilot signals as training samples can provide the model with clear learning objectives and benchmarks, enabling the model to learn the typical time-frequency characteristics of pilot signals of different types of UAVs, and laying a data foundation for subsequent high-precision and high-robust pilot signal recognition.

[0014] Step 104: Construct the pilot signal recognition model. The pilot signal recognition model includes a feature extraction unit, a feature fusion unit, and a detection output unit.

[0015] The feature extraction unit is the backbone of the model, responsible for extracting feature representations at different levels from the input time-frequency spectrum and generating multiple initial feature maps with decreasing resolution. The feature fusion unit is the neck of the model, responsible for integrating complementary information from features at different resolutions to generate fused features that combine semantic and detailed information. The detection output unit is the head of the model, responsible for predicting bounding boxes and classifying targets based on the fused features.

[0016] This invention is based on the YOLOv8 lightweight network architecture. By simplifying the backbone structure, optimizing the feature transfer path, and simplifying the detection head design, this network significantly reduces the number of model parameters and computational overhead while maintaining the basic perception capability of weak targets in the time-frequency spectrum. This effectively ensures the overall inference speed of the model and can be directly adapted to the deployment requirements of embedded hardware platforms such as FPGA and ARM. It is understood that this invention is not limited to the YOLOv8 architecture; other single-stage target detection models with multi-scale feature extraction capabilities, such as YOLOv5, are also applicable.

[0017] Step 106: The input time-frequency spectrum is subjected to multi-scale feature extraction by the feature extraction unit to obtain multiple initial feature maps with different resolutions. Each initial feature map is divided into multiple non-overlapping local frequency blocks according to continuous frequency segments. Adaptive weight allocation of channel dimension is performed independently in each local frequency block. During model training, the channel weights are optimized according to the inverse gradient corresponding to the high-level semantic features output by the feature fusion unit, and the enhanced multi-scale feature map is output.

[0018] Multi-scale feature extraction extracts pilot signal features of varying durations and energy intensities from the input image through convolution and downsampling operations at different levels, generating multiple feature maps with progressively decreasing resolution. Local frequency block partitioning divides the feature map into continuous sub-regions along the frequency axis, ensuring each sub-region corresponds to a continuous frequency range. Adaptive weight allocation along the channel dimension automatically learns the importance of each feature channel for pilot signal recognition through an attention mechanism, assigning different weight coefficients to different channels to enhance effective feature channels and suppress redundant channels.

[0019] It is understandable that localized attention computation based on continuous frequency bands can accurately distinguish between the continuous frequency channel containing the pilot signal and the discretely distributed noise channel, avoiding the problem of global attention simultaneously amplifying the noise channel. During the training phase, the weights are dynamically adjusted based on the high-level semantic features output by the feature fusion unit, achieving bidirectional interaction of multi-scale information. This allows the front-end attention module to utilize global semantic information to optimize weight allocation, further improving the signal-to-noise ratio of the effective signal. Thus, without significantly increasing computational load, the feature representation capability for weak pilot signals is significantly enhanced.

[0020] Step 108: Through the feature fusion unit, frequency offset alignment and feature position registration are performed on the enhanced multi-scale feature map. Then, bidirectional cross-scale feature interaction is performed on the aligned feature map from top to bottom and from bottom to top. During the interaction, learnable non-negative weights are assigned to the local feature range for dynamic weighted fusion, and a multi-scale fused feature map is output.

[0021] Frequency offset alignment and feature position registration correct the frequency position deviation of the same pilot signal in feature maps of different scales by calculating the correlation of features at adjacent scales in the frequency dimension, ensuring that features can accurately correspond during cross-scale fusion. Bidirectional cross-scale feature interaction simultaneously constructs two feature transmission paths: a top-down path that transmits high-level strong semantic features to lower levels, and a bottom-up path that transmits low-level fine-grained features to higher levels, achieving information complementarity between features of different resolutions. Local-range dynamic weighted fusion uses local regions of the feature map as the basic computational unit, assigning independent learnable weights to each local region to achieve refined feature fusion.

[0022] Understandably, frequency offset alignment and feature position registration solve the problem of losing minute frequency offset information due to feature misalignment in traditional fusion structures, fully preserving subtle frequency domain differences between different models using the same protocol. Local dynamic weighted fusion can specifically suppress randomly distributed local point-like electromagnetic noise, avoiding overall feature degradation caused by local high-energy noise in global weighting. The combination of bidirectional interaction and local weighting enables each layer of fused features to simultaneously possess high-level semantic discrimination capabilities and low-level high-frequency detail preservation capabilities, significantly improving the accuracy of distinguishing similar pilot signals.

[0023] Step 110: By detecting the output unit, bounding box regression and classification prediction are performed on the multi-scale fused feature map, and the recognition result of the pilot signal is output.

[0024] Bounding box regression predicts the location coordinates of an object in an input image, generating a rectangular bounding box that accurately surrounds the object. Classification prediction determines whether the region within the bounding box contains the object and its specific category.

[0025] It is understandable that simultaneously performing bounding box regression and classification prediction on multi-scale fused feature maps can achieve end-to-end detection of pilot signals, outputting the location and category information of the pilot signal in one go. Detection based on multi-scale fused features can take into account pilot signals of different intensities and durations, improving the detection rate of short-duration, narrow-band, low-energy, and weak pilot signals, and reducing the false negative rate.

[0026] Step 112: Using training samples, train the pilot signal recognition model using a pre-set loss function to obtain the trained pilot signal recognition model.

[0027] The loss function consists of two parts: bounding box regression loss and improved classification focusing loss. The classification focusing loss incorporates a structure-aware factor and a multimodal modulation factor, comprehensively adapting to the spatial structural features of the pilot signal and the imbalanced distribution of samples. It relies on a differentiated loss weighting mechanism to drive the model to prioritize learning and distinguishing more difficult samples. During backpropagation, gradient enhancement amplifies the gradient magnitudes of difficult samples and key regions of the pilot boundary, ensuring that the loss information is completely transmitted back to the network front end for parameter iteration and updates.

[0028] Step 114: Input the time-frequency spectrum corresponding to the UAV communication signal to be identified into the trained pilot signal recognition model, and output the recognition result of the UAV pilot signal in the time-frequency spectrum.

[0029] The drone communication signal to be identified can be a communication signal emitted by an unknown drone and collected by a radio receiving device in a real low-altitude security scenario.

[0030] It is understandable that by inputting the time-frequency spectrum corresponding to the communication signal of the UAV to be identified into the trained pilot signal recognition model, the detection and recognition of the pilot signal can be completed automatically and quickly without manual intervention. The entire inference process has high computational efficiency and low hardware resource consumption, which can meet the real-time monitoring needs of the low-altitude security field and can be directly deployed on embedded edge devices such as FPGA and ARM.

[0031] In the aforementioned UAV pilot signal recognition method, multi-scale feature extraction from the time-frequency spectrum and frequency domain segmentation are combined to achieve local channel adaptive weight allocation. Simultaneously, channel weights are dynamically optimized using the inverse gradient of high-level semantic features. This accurately enhances the effective features corresponding to the pilot signal and suppresses redundant noise features. While ensuring the model's lightweight characteristics and without increasing computational overhead, it significantly improves the feature representation capability of weak pilot signals, effectively addressing signal detection errors in low signal-to-noise ratio scenarios. By performing frequency offset alignment and feature position registration on the enhanced feature map, and combining dynamic weighted fusion of local feature ranges to achieve bidirectional cross-scale feature interaction, it preserves subtle frequency domain differences in pilot signals, significantly improving the ability to finely distinguish pilot signals from different UAV models using the same protocol. Furthermore, this invention uses a preset loss function to complete iterative model training, ensuring overall model convergence and detection stability. This embodiment of the invention can significantly improve the recognition accuracy and robustness of UAV pilot signals in complex electromagnetic environments while maintaining model lightweightness, meeting the real-time detection needs of embedded edge devices.

[0032] In one embodiment, the feature extraction unit includes a first feature extraction branch, a second feature extraction branch, and a third feature extraction branch; both the second and third feature extraction branches include a C2f module, and the third feature extraction branch includes an SPPF module; the output of each C2f module and SPPF module is connected to a frequency domain block local channel attention module; the C2f module is used to perform multi-branch residual feature extraction on the input features and output the corresponding initial feature map; the SPPF module is used to perform multi-scale pooling feature extraction on the input features and output the corresponding initial feature map; the frequency domain block local channel attention module is used to divide the initial feature map into local frequency blocks, complete the channel weight calculation within the block, and optimize the channel weights according to the back gradient corresponding to the high-level semantic features output by the feature fusion unit during model training, and output the enhanced feature map.

[0033] In this embodiment, the C2f module employs a multi-branch residual structure, dividing the input features into two parallel paths. One path performs deep feature extraction through multiple Bottleneck units, while the other path directly performs cross-stage connections. Finally, the features from the two paths are concatenated and fused. The SPPF module uses multiple serial max-pooling layers of the same size to concatenate pooling results at different scales. The frequency domain block-based local channel attention module first divides the initial feature map along the frequency axis into multiple non-overlapping continuous local frequency blocks. Within each local frequency block, global average pooling is used to obtain channel dimension statistics. The size of the one-dimensional convolutional kernel is adaptively adjusted according to the number of channels to capture local channel dependencies. Then, the final channel weights are dynamically adjusted by combining the P5 high-level semantic features output by the feature fusion unit, and the weights are multiplied element-wise with the original feature map. This structure, while maintaining lightweight design, can obtain richer gradient flow information and multi-scale global receptive fields. It accurately enhances the key features of weak UAV pilot signals with almost no increase in computational overhead, suppresses redundant channels corresponding to background noise and interference signals, and provides a highly discriminative feature foundation for subsequent feature fusion.

[0034] In one embodiment, the first feature extraction branch is used to extract fine-grained local features of the input time-frequency map and output high-resolution shallow enhancement features; the second feature extraction branch includes at least two cascaded sub-branches, which are used to extract mesoscale structural features at different resolutions and output mesoscale enhancement features at different resolutions; the third feature extraction branch includes at least two cascaded sub-branches, which are used to extract global semantic features at different resolutions and output high-level strong semantic enhancement features at different resolutions.

[0035] In this embodiment, the high-resolution shallow enhancement features output by the first feature extraction branch correspond to 1 / 8 resolution (P3) of the input time-frequency map. The two sub-branches of the second feature extraction branch output mesoscale enhancement features corresponding to 1 / 16 (P4) and 1 / 32 (P5) resolutions of the input time-frequency map, respectively. The two sub-branches of the third feature extraction branch output high-level strong semantic enhancement features corresponding to 1 / 64 (P6) and 1 / 128 (P7) resolutions of the input time-frequency map, respectively. Before entering the feature fusion unit, the initial feature maps output by all the branches mentioned above undergo a unified frequency-domain block local channel attention module for enhancement. This module first divides the initial feature map output by each branch into multiple non-overlapping continuous local frequency blocks along the frequency axis, ensuring each block corresponds to a physically meaningful continuous spectral range. Within each local frequency block, global average pooling is used to obtain the channel dimension statistics for that frequency band. The size of the one-dimensional convolutional kernel is adaptively adjusted based on the number of channels to capture local channel dependencies, generating initial channel attention weights. During model training, the module receives the backpropagation gradient corresponding to the P5 high-level semantic features output by the feature fusion unit. The initial weights are dynamically corrected using the globally fused semantic information, ultimately generating optimal channel weights that match the characteristics of the current input signal. These weights are then multiplied element-wise with the original feature map to output the enhanced single-scale feature map. After this processing, the three branches together form five enhanced multi-scale feature maps (P3-P7) at different resolutions, which are then fed into the feature fusion unit.

[0036] By dividing the feature extraction unit into three functionally distinct branches and outputting five enhanced feature maps of different resolutions, it is possible to comprehensively cover UAV pilot signals of different intensities and durations. High-resolution features are specifically responsible for detecting short-term, narrow-band, and weak pilot signals, while mesoscale features are responsible for detecting signals of normal intensity, and high-level features are responsible for detecting strong signals. This effectively avoids feature loss of weak pilot signals during multiple rounds of downsampling.

[0037] In one embodiment, frequency offset alignment and feature position registration are performed on the enhanced multi-scale feature map. Then, bidirectional cross-scale feature interaction is performed on the aligned feature map from top to bottom and bottom to top. During the interaction, learnable non-negative weights are assigned to the local feature range for dynamic weighted fusion, and the multi-scale fused feature map is output. This includes: completing frequency offset alignment and feature position registration by calculating the cross-correlation value of adjacent scale features in the frequency dimension, so that the frequency position of the same pilot signal in the feature maps of different scales remains consistent; performing top-down feature transfer on the aligned feature map, upsampling the high-level features to the same resolution as the adjacent low-level features, and fusing them with the low-level features of the corresponding scale; performing bottom-up feature transfer on the fused feature map, downsampling the low-level features to the same resolution as the adjacent high-level features, and fusing them with the high-level features of the corresponding scale; and assigning learnable non-negative weights to the local feature range in each feature fusion process, and outputting the multi-scale fused feature map after completing the dynamic weighted fusion.

[0038] In this embodiment, frequency offset alignment and feature position registration are first performed on the enhanced P3-P7 feature maps: by calculating the cross-correlation value of adjacent scale features in the frequency dimension, the frequency offset of the same pilot signal in the feature maps of different scales is estimated, and the feature position deviation is corrected to ensure that the features of the same pilot signal can be accurately superimposed during cross-scale fusion, avoiding the dilution of small frequency offset information during the fusion process. After completing frequency offset alignment and feature position registration, the feature maps enter a bidirectional cross-scale feature interaction process from top to bottom and bottom to top: first, a complete top-down transmission is performed, upsampling the high-level strong semantic features of P7 by 2 times step by step, and fusing them with the low-level fine-grained features of P6, P5, P4, and P3 in sequence; then, a complete bottom-up transmission is performed, downsampling the low-level detailed features of P3 by 2 times step by step, and fusing them with the high-level semantic features of P4, P5, P6, and P7 in sequence. In the feature overlay process of each improved BiFPN fusion node, a dynamic weighted fusion mechanism based on local feature ranges is adopted: using a 3×3 pixel region within the feature map as the basic computational unit, an independent learnable non-negative weight is assigned to each local range (the non-negativity of the weights is ensured by the ReLU activation function). The feature values ​​of each input feature corresponding to the local range are multiplied by the weights, summed, and normalized. Then, feature optimization is completed through 3×3 convolution, BN normalization, and SiLU activation. With almost no increase in computational overhead, this significantly improves the feature representation capability of weak pilot signals and the discrimination accuracy of similar signals, providing a highly discriminative feature foundation for the accurate localization and classification of subsequent detection heads.

[0039] In one embodiment, performing bounding box regression and classification prediction on the multi-scale fused feature map to output the pilot signal recognition result includes: inputting multi-scale fused feature maps of different resolutions into the corresponding scale detection head; each detection head performing bounding box regression and classification prediction on the input multi-scale fused feature map and outputting the preliminary detection result of the corresponding scale; and post-processing the preliminary detection results of all scales to obtain the pilot signal recognition result.

[0040] In this embodiment, the feature fusion module outputs three multi-scale fused feature maps with different resolutions, corresponding to 1 / 8, 1 / 16, and 1 / 32 resolutions of the input time-frequency spectrum, respectively. Each fused feature map is input to an independent detection head. The post-processing uses a non-maximum suppression algorithm to remove redundant detection boxes with high overlap, retaining the detection results with the highest confidence. By setting corresponding detection heads for fused feature maps of different resolutions, each detection head can focus on detecting pilot signals at its corresponding scale, improving the targeting and accuracy of detection. Unified post-processing of the preliminary detection results at all scales can integrate the outputs of different detection heads to obtain comprehensive and accurate final recognition results, reducing the missed detection and false detection rates.

[0041] In one embodiment, the detection head adopts a decoupled detection head structure, including mutually independent bounding box regression branches and classification branches, which respectively complete the position prediction and category prediction of the pilot signal.

[0042] In this embodiment, the decoupled detection head structure completely separates the bounding box regression task and the classification task. The bounding box regression branch learns the target's positional offset information through convolutional layers, while the classification branch learns the target's category feature information through convolutional layers. The two branches share the input multi-scale fused feature map, but their parameters are updated independently. Using a decoupled detection head structure avoids mutual interference between the bounding box regression and classification tasks, enabling the model to simultaneously optimize localization accuracy and classification accuracy. In particular, it improves the convergence accuracy of bounding box regression for pilot signals under low signal-to-noise ratio conditions and reduces localization bias.

[0043] In one embodiment, the time-frequency spectra in the training samples are labeled with the bounding box location information and category information of the UAV pilot signal.

[0044] In this embodiment, the time-frequency spectra in the training samples are obtained through self-experimentation, self-screening, and self-annotation. Each image includes a different number and degree of blur of pilot signal stripes. All pilot signals are labeled with the coordinates of the upper left and lower right corners of a rectangular bounding box, as well as uniform "pilot" category information. Using time-frequency spectra labeled with bounding box position and category information as training samples provides the model with a clear learning objective, enabling the model to learn not only the category features of the pilot signals but also the positional features of the pilot signals in the time-frequency spectra, achieving end-to-end accurate localization and recognition of pilot signals.

[0045] In one embodiment, the loss function includes bounding box regression loss and classification focusing loss; the classification focusing loss is configured with a structure-aware factor and a multimodal modulation factor, the structure-aware factor is used to assign a higher loss weight to the boundary region corresponding to the pilot signal than to the center region, and the multimodal modulation factor is used to match the corresponding loss adjustment coefficient according to different sample types.

[0046] In this embodiment, the pilot signal appears as a narrow strip extending along the time axis in the time-frequency spectrum. Its discriminative information is mainly concentrated in the boundary region of the signal. Accurate detection of the boundary region is crucial for distinguishing between different models of the same protocol and overlapping pilot signals. The structure-aware factor assigns 1.5-2 times the loss weight of the central region to pixels within a certain distance from the boundary within the bounding box, guiding the model to focus on learning the boundary features of the pilot signal and improving boundary localization accuracy. The multimodal modulation factor is used to address the differentiated learning problem of different types of difficult samples: this factor divides the training samples into five types: simple background samples, regular intensity pilot samples, low signal-to-noise ratio weak pilot samples, overlapping pilot samples, and blurred boundary samples, assigning an independent learnable loss adjustment coefficient to each sample type. For simple background samples and regular intensity pilot samples, the modulation factor approaches 0, significantly reducing their loss contribution; for low signal-to-noise ratio weak pilot samples, overlapping pilot samples, and blurred boundary samples, the modulation factor approaches 1, fully preserving their loss value and guiding the model to focus on learning these difficult samples that are hard to distinguish.

[0047] In one specific embodiment, such as Figure 2As shown, a schematic diagram of the BEF-YOLOv8 architecture is provided, which consists of three sequentially connected modules: Backbone (feature extraction), Neck (multi-scale feature fusion), and Head (detection output). In the feature extraction stage, the input time-frequency spectrum is sequentially processed through two convolutional layers, four C2f modules, and one SPPF module for multi-scale feature extraction, generating five initial feature maps of different resolutions, P3 to P7. The outputs of each C2f module and SPPF module are connected in series with a frequency domain block-based local channel attention module. During model training, the channel weights of each attention module are optimized based on the backpropagation gradient corresponding to the P5 high-level semantic features output by the neck fusion module, resulting in an enhanced multi-scale feature map. In the feature fusion stage, the enhanced feature maps P3 to P7 are input into the neck region. First, frequency offset alignment and feature location registration are performed on features at adjacent scales. Then, four locally weighted improved BiFPN modules, based on frequency offset alignment and feature location registration, sequentially complete bidirectional cross-scale feature interaction from top to bottom and bottom to top. At each fusion node, independent learnable non-negative weights are assigned to each feature local range as the basic unit for dynamic weighting, ultimately generating multi-scale fused feature maps of scales P3, P4, and P5. In the detection output stage, the fused feature maps of scales P3, P4, and P5 are input into the corresponding decoupled detection heads. Each detection head contains independent bounding box regression and classification branches, respectively completing the location and category determination of the pilot signal. Simultaneously, each detection head uses a loss function that focuses on difficult samples, guiding the model to prioritize low signal-to-noise ratio and partially overlapping difficult pilot samples. The improved network achieves systematic optimization from feature extraction to loss calculation, significantly improving the recognition accuracy and robustness of UAV pilot signals in complex electromagnetic environments while maintaining lightweight design.

[0048] To address the issues of weak and ambiguous pilot signals, redundant channel information, and strong noise interference that exists only in local continuous frequency bands, this invention utilizes an improved ECA (Frequency Domain Block Local Channel Attention Module) module. This module leverages local frequency block statistics to achieve precise focusing on the effective frequency region while maintaining a lightweight design, taking advantage of the frequency domain locality characteristics of time-frequency spectra. The ECA module's workflow is as follows: Figure 3 As shown, it includes the following steps: (1) Frequency Domain Adaptive Blocking. Based on the frequency resolution corresponding to the feature map, the feature map is divided into multiple non-overlapping local frequency blocks along the frequency axis (spatial width dimension) according to continuous frequency segments: the P3 layer (1 / 8 resolution) is divided into 16 blocks, the P4 layer (1 / 16 resolution) is divided into 8 blocks, the P5 layer (1 / 32 resolution) is divided into 4 blocks, and the P6-P7 layers are not divided. At the same time, a learnable block size adjustment factor is introduced to automatically adjust the block division granularity according to the signal energy distribution of each frequency block during training, ensuring that the frequency segment where the pilot signal is located is completely contained in a single block.

[0049] (2) Feature gradient feedback fusion. The P5 layer fusion feature (the intermediate scale feature with the richest semantic information) output by the improved BiFPN is downsampled and matched to the same size as the current backbone scale feature. During the model training phase, the channel weights are optimized by the back gradient of this high-level feature to obtain an enhanced feature map that fuses high-level semantic information: ; in To improve the P5 layer fusion features of the BiFPN output, For the first The original feature map of each channel, For the enhanced feature map, This is a downsampling operation.

[0050] (3) Block-based global average pooling. First, the enhanced feature map output from step (2) is processed... The local frequency blocks are divided according to step (1). Let the first... Within the block, the first The feature sub-images of each channel are as follows ,in Let these be the height and width of the block. Perform global average pooling on this subgraph to obtain the local channel statistics for that block: ; in, As a scalar, it represents the first... Within the frequency block, the first Local descriptive statistics for each channel For the first Within the block, the first The feature submap of each channel at position The pixel value.

[0051] (4) Adaptive kernel selection. ECA adaptively determines the size of the one-dimensional convolution kernel through a nonlinear mapping function. : ; in This represents the number of channels in the current feature map. For hyperparameters, , This indicates taking the nearest odd number, a constant. Used for numerical offset calibration to ensure that higher-dimensional channels correspond to a longer interaction range and lower-dimensional channels correspond to a shorter interaction range.

[0052] (5) Local one-dimensional convolution to calculate attention weights. Within each frequency block, the convolution kernel size obtained from step (4) is adaptively solved. Perform one-dimensional convolution operations to model local channel dependency interactions on the aggregated statistical features of channels within the block, and obtain the attention weights of each channel within the block: ; in It is the Sigmoid activation function. For learnable weights, For the first The local neighborhood channel indexes of each channel. In this step, the boundary regions of the one-dimensional convolution operation are padded with zeros to ensure that the convolution kernel can slide and compute normally across the entire channel range.

[0053] (6) Local feature recalibration. The attention weights learned within each frequency block are... The enhanced feature map output by step (2) Element-wise multiplication is used to enhance local features. ; This precisely enhances the characteristic response of the local frequency band where the UAV pilot signal is located, and suppresses the weight of pure noise frequency blocks.

[0054] Bidirectional Feature Pyramid Network (BiFPN) is an efficient fusion structure for multi-scale feature interactions, such as... Figure 4 As shown, a comparative diagram of the workflows of FPN, PANet, and improved BiFPN is provided. Addressing the issues of traditional BiFPN's global weighting failing to suppress local electromagnetic pulse noise and lacking frequency offset alignment and feature position registration mechanisms leading to the loss of minute frequency offset information, this invention employs a local dynamic weighted fusion module (improved BiFPN) with frequency offset alignment and feature position registration to fuse multi-scale feature maps enhanced by ECA in the frequency domain. This module addresses the characteristics of minute frequency offsets in pilot signals and localized point-like noise distribution in time-frequency spectra by adding sub-pixel-level frequency offset alignment, feature position registration, and local region weighting steps to the bidirectional fusion process. The core steps are as follows: First, define the multi-scale features, assuming the multi-scale features output by the backbone network are... Each The resolution represents the input image. The feature map. Among them It is a high-resolution shallow feature, responsible for the detection of short-time, narrow-band, and weak pilot signals; , It is a mesoscale feature, responsible for detecting conventional intensity signals; , For high-level strong semantic features, it is responsible for strong signal detection. Subsequently, the improved BiFPN constructs bidirectional paths from top to bottom and bottom to top, enabling bidirectional complementary interaction between features of different scales.

[0055] (1) Subpixel-level frequency offset alignment and feature position registration preprocessing Before cross-scale fusion, frequency offset estimation and alignment are performed on the low-resolution feature maps: the sub-pixel-level frequency offset of the pilot signal is estimated by calculating the cross-correlation coefficient of adjacent scale feature maps in the frequency dimension (vertical axis). ; ; in, Let be the cross-correlation function in the frequency dimension, representing the frequency offset as The frequency matching degree between high-level and low-level feature maps. This is the upsampled high-level feature map. For lower-level feature maps in the same layer, select the one with the highest cross-correlation. This serves as the estimated offset frequency. Then, the upsampled high-level feature map is subpixel shifted and aligned along the frequency direction to ensure that the frequency positions of the same pilot signal in feature maps of different scales precisely coincide.

[0056] ① Top-down path After upsampling the aligned high-level strong semantic features, they are then dynamically weighted and fused with the mid-level features. Initialization ,for : ; Upsample( ) is 2x upsampling, , 3 for each 3. Learnable weights for local regions (guaranteed non-negative via ReLU). To prevent division by zero, use superscript. td This represents intermediate variables in the top-down fusion path.

[0057] ② Bottom-up path After downsampling the aligned low-level fine-grained features, they are then dynamically fused with the high-level features locally. Initialization ,for : ; in, Original features of this layer Weight, for td intermediate features of the stage Weight, To downsample low-level features Weight, For the fusion features fed into the detection head, superscript out This represents the final output variable of the bottom-up fusion path. This process enables each layer of features to simultaneously acquire high-level semantic information and low-level detailed information, achieving full-scale information complementarity.

[0058] (2) Scale-differentiated fusion weight limit To address the characteristic that feature maps at different scales correspond to different frequency resolutions, this invention further designs scale-differentiated fusion weighting coefficients to achieve precise division of labor among the functions of each scale: P3-P4 layers (high-resolution layers): highest frequency resolution (approximately) 1kHz (per pixel), with a focus on preserving high-frequency narrowband pilot details and limiting the weights for high-level feature fusion. The upper limit is The specific possible values ​​are 0.3 ,Right now: ; in The function constrains the weights to [ 0 , Within the range, ensure that low-level fine-grained features dominate the fusion process, and avoid high-level semantic features diluting the subtle frequency domain structure of the pilot signal.

[0059] Layers P6-P7 (low-resolution layers): lowest frequency resolution (approximately) 16kHz (per pixel), focusing on suppressing broadband noise interference and limiting the weights of low-level feature fusion. The upper limit is Specifically, the possible value is 0.3: ; This constraint ensures that high-level semantic features dominate the fusion process, effectively filtering out random noise interference from lower layers.

[0060] P5 layer (intermediate layer): No weight restrictions are set; it serves as a transition and balance between high and low layer features. Finally, after each fusion, the data is processed... 3×3 Convolution, BN normalization, and SiLU activation optimize the features, making the fused features smoother and more discriminative. ; This invention applies an improved BiFPN to the YOLOv8 neck network to achieve multi-scale feature fusion from P3 to P7. On one hand, it enhances high-resolution shallow features of P3 through a bidirectional path, ensuring that weak features such as edges and textures of weak pilot signals from UAVs are not overwhelmed by higher-level semantics, significantly improving the perception and utilization of features for small and weak targets. On the other hand, the dynamic weighted fusion mechanism allows the model to autonomously assign high weights to signal features, effectively suppressing interference from electromagnetic noise and other interference signals. Frequency-domain block ECA and the improved BiFPN form a channel-scale two-dimensional collaborative enhancement: frequency-domain block ECA completes accurate selection of channel-dimensional features, while the improved BiFPN achieves deep aggregation of scale and spatial-dimensional features. The combined optimization of these two technologies improves signal feature perception and detection performance.

[0061] FocalLoss automatically reduces the contribution of simple samples to the training process, guiding the model to focus on learning difficult samples. This invention replaces the original YOLOv8 classification loss function with a structure-aware multimodal difficult sample focusing loss (improved FocalLoss), forming a full-link collaborative optimization with improved frequency domain block ECA, frequency offset alignment, and improved feature location registration BiFPN. The formula is as follows: ; in: For the corresponding real category t At that time, the predicted probability output by the model, This is the positive and negative sample balancing factor.

[0062] Structure-aware factor: Based on the prior knowledge of the strip structure of the pilot signal, pixels in the boundary region are assigned. 1.5-2 The loss weight is multiplied by a factor of 1, and the base weight is assigned to the pixels in the central region. 1 The boundary region is defined as the region at a distance from the edge of the pilot signal bounding box. 3 The area within one pixel is key to distinguishing different models and overlapping signals.

[0063] Multimodal modulation factor: The modulation factor is automatically adjusted based on the sample type to achieve differentiated optimization for different challenging samples, such as low signal-to-noise ratio samples (predicted probability is...). 0.3-0.7 ): =3 Partially overlapping samples ( IoU And the predicted probability is 0.3-0.7 ): =2.5 Boundary region samples: =2 Simple samples (predicted probabilities) > 0.9 or<0.1 )Pick =1 When a sample satisfies multiple conditions simultaneously, the largest one is taken. The value ensures that the modulation effect of difficult samples is prioritized.

[0064] This loss function combines the time-frequency representation characteristics of UAV pilot signals to achieve differentiated adaptive optimization. Multiple control factors work together to improve model training accuracy and stability: First, for weak pilot signal features in complex electromagnetic environments, a multi-modal modulation factor adaptively adjusts the loss contribution of samples with different recognition difficulties, avoiding the problem of simple features dominating training in conventional loss functions, and prompting the model to continuously learn weak and easily submerged effective signal features. Second, for the problem of blurred target boundaries caused by overlapping UAV signals, a structure-aware factor applies differentiated weight constraints to the target boundary region, strengthening the model's ability to learn pilot signal edge details and effectively improving the accuracy of boundary localization and differentiation of overlapping signals. Third, for the problem of uneven sample distribution with a low proportion of effective pilot target features and a high proportion of background redundancy, a positive and negative sample balancing factor balances the gradient contributions of positive and negative samples, suppressing the interference of massive background redundancy features on the training process, ensuring stable convergence of the model training process, and accurately focusing on the effective target features of UAV pilot signals.

[0065] To guide the frequency-domain block ECA attention module, frequency offset alignment, and feature location registration improved BiFPN module to focus on the channels and regions where difficult samples are located, this invention further introduces an attention gradient enhancement mechanism: during backpropagation, the gradient corresponding to the difficult sample is multiplied by an enhancement coefficient, explicitly amplifying the gradient signal in the key region, thus solving the problem of ineffective learning by the lower-level modules due to gradient decay. The specific formula is as follows: ; in For structure-aware gradient enhancement coefficients (pixels in the boundary region are taken as follows): 1.5 Take from other areas 1 ), Gradient enhancement coefficients for difficult samples (for low signal-to-noise ratio samples) 2 Other samples were taken 1 When the sample is also a low signal-to-noise ratio sample in the boundary region, the gradient enhancement coefficient is: 1.5×2=3 The weighted synergy with the loss layer strengthens the learning signal in key regions. This mechanism explicitly enhances the gradient signal in key regions and difficult samples, solving the problem of ineffective learning of previous modules caused by gradient decay, without affecting the forward inference speed.

[0066] It should be understood that, although Figure 1The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages: firstly, the sub-steps and stages do not need to be completed at the same time and can be executed in different time periods; secondly, the order of the sub-steps themselves is not limited, and they can be alternated and interspersed with other steps, or parts of the sub-steps or stages of other steps.

[0067] In one embodiment, a drone pilot signal identification device is provided, comprising: The sample acquisition module is used to acquire training samples; the training samples include time-frequency spectra labeled with UAV pilot signals. The model building module is used to build a pilot signal recognition model; the pilot signal recognition model includes a feature extraction unit, a feature fusion unit, and a detection output unit. The feature enhancement processing module is used to perform multi-scale feature extraction on the input time-frequency spectrum through the feature extraction unit to obtain multiple initial feature maps with different resolutions. Each initial feature map is divided into multiple non-overlapping local frequency blocks according to continuous frequency segments. Adaptive weight allocation of the channel dimension is performed independently within each local frequency block, and the enhanced multi-scale feature map is output. The multi-scale feature fusion module is used to perform frequency offset alignment and feature position registration on the enhanced multi-scale feature map through the feature fusion unit, and then perform bidirectional cross-scale feature interaction from top to bottom and bottom to top on the aligned feature map. During the interaction, learnable non-negative weights are assigned to the local feature range for dynamic weighted fusion, and the multi-scale fused feature map is output. The signal recognition module is used to perform bounding box regression and classification prediction on the multi-scale fused feature map through the detection output unit, and output the recognition result of the pilot signal. The model training module is used to train the pilot signal recognition model using training samples and a pre-set loss function, so as to obtain the trained pilot signal recognition model. The result output module is used to input the time-frequency spectrum corresponding to the UAV communication signal to be identified into the trained pilot signal recognition model, and output the recognition result of the UAV pilot signal in the time-frequency spectrum.

[0068] Specific limitations regarding the UAV pilot signal identification device can be found in the limitations of the UAV pilot signal identification method described above, and will not be repeated here. Each module in the aforementioned UAV pilot signal identification device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0069] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and RAM. The non-volatile storage media stores the operating system and computer programs. The RAM provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for identifying pilot signals of a drone. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0070] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0071] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described above.

[0072] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0073] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for identifying pilot signals of unmanned aerial vehicles (UAVs), characterized in that, The method includes: Obtain training samples; the training samples include time-frequency spectra labeled with UAV pilot signals; A pilot signal recognition model is constructed; the pilot signal recognition model includes a feature extraction unit, a feature fusion unit, and a detection output unit. The feature extraction unit performs multi-scale feature extraction on the input time-frequency spectrum to obtain multiple initial feature maps with different resolutions. Each initial feature map is divided into multiple non-overlapping local frequency blocks according to continuous frequency segments. Within each local frequency block, channel dimension statistics are obtained through global average pooling. Based on the channel dimension statistics, local channel dependencies are captured and initial channel attention weights are generated. The initial channel attention weights are multiplied element-wise with the initial feature map to complete the adaptive weight allocation of the channel dimension. During model training, the channel weights are optimized based on the inverse gradient corresponding to the high-level semantic features in the multi-scale fused feature map output by the feature fusion unit, and the enhanced multi-scale feature map is output. The feature fusion unit performs frequency offset alignment and feature position registration on the enhanced multi-scale feature map. Then, it performs bidirectional cross-scale feature interaction from top to bottom and bottom to top on the aligned feature map. During the interaction, learnable non-negative weights are assigned to the local feature range for dynamic weighted fusion, and a multi-scale fused feature map is output. The detection output unit performs bounding box regression and classification prediction on the multi-scale fused feature map and outputs the recognition result of the pilot signal. Using the training samples, the pilot signal recognition model is trained using a pre-set loss function to obtain the trained pilot signal recognition model; Input the time-frequency spectrum corresponding to the UAV communication signal to be identified into the trained pilot signal recognition model, and output the recognition result of the UAV pilot signal in the time-frequency spectrum; The frequency offset alignment and feature location registration of the enhanced multi-scale feature map include: Frequency offset alignment and feature position registration are completed by calculating the cross-correlation value of adjacent scale features in the frequency dimension. This corrects the pixel offset caused by downsampling in the enhanced multi-scale feature map, so that the pixel position of the same pilot signal in the feature maps of different scales remains consistent.

2. The method according to claim 1, characterized in that, The feature extraction unit includes a first feature extraction branch, a second feature extraction branch, and a third feature extraction branch; the second and third feature extraction branches each include a C2f module, and the third feature extraction branch includes an SPPF module; the output of each C2f module and SPPF module is connected to a frequency domain block local channel attention module. The attention modules corresponding to medium- and high-resolution features enable the frequency domain block strategy, while the attention modules corresponding to low-resolution and high-level features do not perform frequency domain block operations. The C2f module is used to perform multi-branch residual feature extraction on the input features and output the corresponding initial feature map. The SPPF module is used to perform multi-scale pooling feature extraction on the input features and output the corresponding initial feature map. The frequency domain block local channel attention module is used to divide the initial feature map into local frequency blocks, and to complete the channel weight calculation within the block. During model training, the channel weights are dynamically adjusted according to the high-level semantic features in the multi-scale fused feature map output by the feature fusion unit, and the corresponding scale enhanced feature map is output.

3. The method according to claim 2, characterized in that, The first feature extraction branch is used to extract fine-grained local features of the input time-frequency spectrum and output high-resolution shallow enhancement features; The second feature extraction branch includes at least two cascaded sub-branches, which are used to extract mesoscale structural features at different resolutions and output mesoscale enhancement features at different resolutions. The third feature extraction branch includes at least two cascaded sub-branches, which are used to extract global semantic features at different resolutions and output high-level strong semantic enhancement features at different resolutions.

4. The method according to claim 1, characterized in that, A bidirectional cross-scale feature interaction is performed on the aligned feature map, using both top-down and bottom-up methods. During the interaction, learnable non-negative weights are assigned to local feature ranges for dynamic weighted fusion, resulting in a multi-scale fused feature map. Perform top-down feature transfer on the registered feature map, upsample the high-level features to the same resolution as the adjacent low-level features, and fuse them with the low-level features at the corresponding scale; The fused feature map is transferred from bottom to top, and the low-level features are downsampled to the same resolution as the adjacent high-level features and then fused with the high-level features at the corresponding scale. In each feature fusion process, learnable non-negative weights are assigned to the local feature range, and after dynamic weighted fusion, a multi-scale fused feature map is output.

5. The method according to claim 1, characterized in that, Boundary box regression and classification prediction are performed on the multi-scale fused feature map, and the identification results of the output pilot signal include: Input the multi-scale fused feature maps of different resolutions into the detection head of the corresponding scale; Each detection head performs bounding box regression and classification prediction on the input multi-scale fused feature map, and outputs preliminary detection results at the corresponding scale. The preliminary detection results at all scales are post-processed to obtain the pilot signal identification results.

6. The method according to claim 5, characterized in that, The detection head adopts a decoupled detection head structure, including mutually independent bounding box regression branches and classification branches, which respectively complete the position prediction and category prediction of the pilot signal.

7. The method according to claim 1, characterized in that, The time-frequency spectra in the training samples are labeled with the bounding box location information and category information of the UAV pilot signals.

8. The method according to claim 1, characterized in that, The loss function includes bounding box regression loss and classification focusing loss; the classification focusing loss is configured with a structure-aware factor and a multimodal modulation factor, the structure-aware factor is used to assign a higher loss weight to the boundary region corresponding to the pilot signal than to the center region, and the multimodal modulation factor is used to match the corresponding loss adjustment coefficient according to different sample types.

9. A UAV pilot signal identification device applied to the method described in any one of claims 1-8, characterized in that, The device includes: A sample acquisition module is used to acquire training samples; the training samples include time-frequency spectra labeled with UAV pilot signals; The model building module is used to build a pilot signal recognition model; the pilot signal recognition model includes a feature extraction unit, a feature fusion unit, and a detection output unit. The feature enhancement processing module is used to perform multi-scale feature extraction on the input time-frequency spectrum through the feature extraction unit to obtain multiple initial feature maps with different resolutions. Each initial feature map is divided into multiple non-overlapping local frequency blocks according to continuous frequency segments. Adaptive weight allocation of the channel dimension is performed independently within each local frequency block. During model training, the channel weights are optimized according to the inverse gradient corresponding to the high-level semantic features output by the feature fusion unit, and the enhanced multi-scale feature map is output. The multi-scale feature fusion module is used to perform frequency offset alignment and feature position registration on the enhanced multi-scale feature map through the feature fusion unit, and then perform bidirectional cross-scale feature interaction from top to bottom and from bottom to top on the aligned feature map. During the interaction process, learnable non-negative weights are assigned to the local feature range for dynamic weighted fusion, and the multi-scale fused feature map is output. The signal recognition module is used to perform bounding box regression and classification prediction on the multi-scale fused feature map through the detection output unit, and output the recognition result of the pilot signal; The model training module is used to train the pilot signal recognition model using the training samples and a pre-set loss function to obtain the trained pilot signal recognition model. The result output module is used to input the time-frequency spectrum corresponding to the UAV communication signal to be identified into the trained pilot signal recognition model, and output the recognition result of the UAV pilot signal in the time-frequency spectrum.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images