A UAV Recognition Model Transfer Method Based on Standard Radio Frequency Calibration Signals

By introducing standard radio frequency calibration signals and hierarchical fine-tuning of neural network topology into the UAV recognition model, the problem caused by hardware differences in cross-device deployment is solved, and the high-precision model can be effectively migrated and quickly updated at low-computing-power edge devices.

CN123070239APending Publication Date: 2026-11-10CHINA ELECTRONICS IND ENG CO LTD +2
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611213149.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-11
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies face challenges when deploying drone recognition models across devices, including decreased recognition accuracy due to hardware differences, increased false alarm rates, data acquisition barriers, limitations in computing resources, and high human resource requirements, making it difficult to achieve effective migration of high-precision models.

Method used

By introducing a standard radio frequency calibration signal as a physical benchmark and combining it with a specific neural network hierarchical fine-tuning topology, offline benchmark construction and online migration fine-tuning are performed to quantify and eliminate hardware distortion, thereby enabling cross-device migration of the model.

Benefits of technology

The high-precision UAV recognition model has been successfully deployed on edge devices with low computing power, reducing computing resource consumption and labor costs, shortening model update time, and improving recognition accuracy and applicable hardware range.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123070239A_ABST
    Figure CN123070239A_ABST
Patent Text Reader

Abstract

This invention discloses a method for transferring a UAV recognition model based on standard radio frequency calibration signals. The transfer method includes an overall process comprising an offline benchmark construction stage and an online transfer fine-tuning stage. The offline benchmark construction stage specifically includes: step S1.1: standard signal transmission and source domain benchmark feature solidification; the online transfer fine-tuning stage specifically includes: step S2.1: target domain calibration signal acquisition; step S2.2: initial feature extraction of the target domain; step S2.3: construction of hardware distortion residual loss; step S2.4: directional fine-tuning topology based on residual loss, constructing the computational graph topology of the fine-tuning stage on the target reconnaissance device, and executing a network hierarchical update strategy; step S2.5: model solidification and deployment. By introducing a standard radio frequency calibration device as a physical benchmark source and combining it with a neural network hierarchical fine-tuning topology with a specific structure, cross-device model transfer is achieved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of radio reconnaissance and counter-drone technology, and more particularly to a drone identification model transfer method based on standard radio frequency calibration signals. Background Technology

[0002] In the field of radio reconnaissance and counter-drone operations, deep learning-based drone radio frequency signal type identification technology has become mainstream. Its basic technical chain is as follows: reconnaissance equipment receives the image transmission or telemetry and control radio frequency signals from the drone via an antenna. These signals are then processed by down-conversion and analog-to-digital conversion at the radio frequency front-end, outputting baseband discrete IQ data (a complex sequence containing in-phase components I and quadrature components Q). This IQ data is then input into a pre-trained deep learning model (such as an improved network model based on one-dimensional convolution or two-dimensional time-frequency graph convolution) to extract high-dimensional features of the signal, thereby outputting the specific drone type classification result.

[0003] However, in actual cross-device deployments, the aforementioned basic links face inherent hardware differences. Different models of reconnaissance equipment inevitably exhibit physical distortions in their RF front-end hardware links (including low-noise amplifiers, mixer oscillators, bandpass filters, etc.), such as local oscillator frequency offset, I / Q quadrature imbalance, nonlinear distortion, and differences in filter amplitude and phase response. These hardware distortions are directly superimposed on the baseband I / Q data, causing a sharp drop in recognition accuracy or a significant increase in false alarm rate when a high-precision recognition model trained on the source reconnaissance equipment is directly transferred to the target reconnaissance equipment due to the shift in input data distribution. To address this model failure problem caused by cross-device hardware distortions, the industry has developed corresponding model transfer and adaptive alignment techniques.

[0004] The following inherent defects exist in signal processing mechanisms, dataset construction structures, computational resource allocation, and job execution processes: At the signal processing mechanism level, the lack of a physical reference leads to the failure of feature decoupling and poor model robustness against interference. The difference in IQ data when the source reconnaissance equipment and the target reconnaissance equipment collect the same UAV is purely due to distortion caused by the physical differences in the RF front-end hardware of the two devices. Existing technology directly inputs UAV signals collected by the target equipment in the field, mixed with real environmental noise, into the network, attempting to smooth out the feature distribution with pure data-driven mathematical algorithms. Due to the lack of a pure physical reference signal, the algorithm cannot distinguish between "hardware physical distortion features" and "real electromagnetic environment noise features" in mathematical space, and is prone to mistakenly learning the environmental background noise unique to the target equipment as the inherent feature of the UAV, resulting in an extremely high false alarm rate in complex electromagnetic environments after the model is transferred.

[0005] At the dataset construction level, there is a strong reliance on real target samples, creating an insurmountable data acquisition barrier. Existing domain adversarial training technologies heavily depend on the target device actually receiving real drone signals aligned with the source device type in the field. This presents an unsolvable structural flaw in real-world engineering. On one hand, certain drone models are rarely seen in the target device's deployment area. The source device may have collected data during its development phase, but the target device cannot generate similar data when it is refitted. On the other hand, drone model distribution has strong regional attributes, with different common models in different regions. If the target device is deployed across regions, the data set it can collect in the field will inevitably be severely mismatched with the historical data of the source device. This results in the domain adversarial network being unable to complete alignment training due to the lack of target domain samples corresponding to the model.

[0006] At the computational resource allocation level, the network architecture is too large, hindering the deployment of high-precision models to low-computing-power edge devices. Existing domain adversarial schemes must temporarily connect a large-scale "domain discriminator" network in parallel to the existing identification backbone network, and conduct dual-network adversarial games through a gradient inversion layer. This computational topology consumes enormous computing resources and several hours of iteration time. Although high-performance servers can meet this computing power requirement, it is precisely this enormous computing power consumption that prevents high-precision UAV identification models, which could be trained on high-performance platforms, from being smoothly migrated and deployed to the numerous low-computing-power edge reconnaissance devices (such as handheld direction finders and UAV-borne pods) on the front lines, greatly limiting the applicable hardware range of high-precision models.

[0007] At the operational implementation level, there is a lack of standardized closed-loop processes, resulting in extremely high human resource barriers and a period of operational vulnerability. Existing technologies have turned "cross-device migration" into an extremely cumbersome manual intervention project. Domain adversarial training involves the dynamic adjustment of the source and target domain data ratios and repeated trial and error of gradient inversion layer hyperparameters. This process requires the on-site intervention of senior algorithm engineers with a strong theoretical foundation, resulting in extremely high technical barriers and limiting the independent replacement of target devices. At the same time, when the target device needs to update its model, it must go through a lengthy process of "downtime waiting for field data collection, manual data transmission, several hours of offline fine-tuning by experts, and then the release of the updated model firmware," preventing the target device from being immediately put into practical use and generating high sunk costs in time and manpower. Summary of the Invention

[0008] In view of the above problems, the present invention is proposed to provide a UAV identification model transfer method based on standard radio frequency calibration signals to overcome or at least partially solve the above problems.

[0009] According to one aspect of the present invention, a method for transferring a UAV identification model based on a standard radio frequency calibration signal is provided, the method comprising: The overall process includes: an offline benchmark building phase and an online migration fine-tuning phase; The offline benchmark construction phase specifically includes: Step S1.1: Standard signal transmission and source domain reference characteristics are solidified; The online migration fine-tuning phase specifically includes: Step S2.1: Acquisition of target domain calibration signals; Step S2.2: Initial feature extraction of the target domain; Step S2.3: Construct the hardware distortion residual loss; Step S2.4: Based on residual loss, perform directional fine-tuning of the topology, construct the computational graph topology for the fine-tuning stage on the target reconnaissance equipment, and execute the network hierarchical update strategy; Step S2.5: Model solidification and deployment.

[0010] Optionally, step S1.1: standard signal transmission and source domain reference feature solidification specifically includes: Place the standard radio frequency calibration device at the antenna port of the source reconnaissance equipment; The control calibration device transmits a preset typical radio frequency signal for the UAV; The source detection equipment acquires the signal through its radio frequency front end, and outputs a baseband IQ data stream after down-conversion processing. The baseband IQ data stream is truncated to obtain a set of IQ data segments of fixed length, and a source domain calibration IQ dataset Ds is generated. The source domain calibration IQ dataset Ds is input into the pre-trained initial recognition model, and the corresponding source domain calibration feature set Fs is output through the feature extraction backbone network of the initial recognition model. The source domain calibration IQ dataset Ds and the source domain calibration feature set Fs are packaged as a hardware feature benchmark file and pre-stored in the storage medium of the source reconnaissance device.

[0011] Optionally, step S2.1: target domain calibration signal acquisition specifically includes: When it is necessary to migrate and deploy the initial identification model to the target reconnaissance equipment, deploy the same standard radio frequency calibration device near the antenna port of the target reconnaissance equipment and transmit a typical radio frequency signal with the same parameters as in step S1.1; The target reconnaissance equipment acquires the signal through its own radio frequency front-end, performs down-conversion processing, and generates a target domain calibration IQ dataset D according to the same truncation rules as in step S1.1. t .

[0012] Optionally, step S2.2: initial feature extraction of the target domain specifically includes: On the computing platform of the target reconnaissance equipment, an initial identification model with the exact same structure as the source equipment is loaded; The target domain calibration IQ dataset is input into the model, and the signal is allowed to propagate unidirectionally forward through the feature extraction backbone network. Without performing any backpropagation calculations, the target domain calibration feature set F is extracted. t .

[0013] Optionally, step S2.3: constructing the hardware distortion residual loss specifically includes: On the computing platform of the target reconnaissance equipment, the pre-stored source domain calibration feature set is read; Since the sources of both the input source domain calibration feature set and the target domain calibration feature set are known standard signals, the statistical differences between them in the high-dimensional feature space are physically equivalent to the differences in the physical distortion of the radio frequency front-end hardware between the source reconnaissance equipment and the target reconnaissance equipment.

[0014] Optionally, the method for calculating the statistical difference between the two in the high-dimensional feature space includes: Step S2.3.1: Perform feature compression, respectively for F s and F t Global average pooling is performed along the feature channel dimension to compress the 3D feature map into a one-dimensional global mean vector μ. s and μ t ; Step S2.3.2: Perform distance measurement, call the cosine similarity calculation module, and calculate μ. s and μ t The cosine distance between them is defined as the hardware distortion residual loss L. calib .

[0015] Optionally, step S2.4: Based on residual loss, the directional fine-tuning topology is constructed on the target reconnaissance equipment to build the computational graph topology for the fine-tuning stage, and the network hierarchical update strategy is executed, specifically including: Step 2.4.1, bottom layer freeze truncation: In the backpropagation path of the computation graph, insert the gradient stopping operator to absolutely freeze the network weight parameters of the first N layers of the feature extraction backbone network in the initial recognition model, and force the gradient to be set to zero. N is a positive integer and is the bottom convolutional layer for extracting basic time-frequency texture features. Step 2.4.2, Top-level directional update: Construct the total loss function L total =L calib Backpropagation updates are enabled only for the last few layers of the feature extraction backbone network and the top-level classification network; L is minimized using the optimizer. calib The mean vector μ of the forced target domain feature distribution t The mean vector μ of the source domain baseline characteristic distribution in the mathematical space.s Align them closely.

[0016] Optionally, step S2.5: model solidification and deployment specifically includes: Monitoring total loss function L total The numerical change of the graph is interrupted when it decreases and converges to a preset threshold range. Extract the network layer states containing the updated weight parameters from the current computation graph, serialize and save them, and generate a target recognition model file adapted to the target reconnaissance equipment. The temporary computation graph is destroyed, and the target recognition model file is loaded into the inference engine of the target reconnaissance equipment to take over the subsequent real-time UAV model recognition task.

[0017] This invention provides a method for transferring a UAV identification model based on standard radio frequency calibration signals. The transfer method includes an overall process comprising an offline benchmark construction stage and an online transfer fine-tuning stage. The offline benchmark construction stage specifically includes: step S1.1: standard signal transmission and source domain benchmark feature solidification; the online transfer fine-tuning stage specifically includes: step S2.1: target domain calibration signal acquisition; step S2.2: initial target domain feature extraction; step S2.3: constructing hardware distortion residual loss; step S2.4: based on the residual loss, directional fine-tuning topology, constructing the computational graph topology of the fine-tuning stage on the target reconnaissance device, and executing a network hierarchical update strategy; step S2.5: model solidification and deployment. By introducing a standard radio frequency calibration device as a physical benchmark source and combining it with a neural network hierarchical fine-tuning topology with a specific structure, cross-device model transfer is achieved.

[0018] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating a UAV identification model transfer method based on standard radio frequency calibration signals, provided as an embodiment of the present invention; Figure 2 The core process and network topology diagram provided for embodiments of the present invention. Detailed Implementation

[0021] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0022] The terms "comprising" and "having," and any variations thereof, in the specification, embodiments, claims, and drawings of this invention are intended to cover non-exclusive inclusion, such as including a series of steps or units.

[0023] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0024] Example 1 A method for transferring a drone identification model based on standard radio frequency calibration signals, the method comprising: an offline benchmark construction stage and an online transfer fine-tuning stage; The offline reference construction stage specifically includes: Step S1.1: Standard signal transmission and source domain reference feature solidification; The online migration fine-tuning stage specifically includes: Step S2.1: Target domain calibration signal acquisition; Step S2.2: Initial feature extraction of the target domain; Step S2.3: Construction of hardware distortion residual loss; Step S2.4: Based on the residual loss, directional fine-tuning topology, constructing the computational graph topology of the fine-tuning stage on the target reconnaissance equipment, and executing the network layered update strategy; Step S2.5: Model solidification and deployment.

[0025] This invention provides a cross-device UAV identification model transfer method based on standard radio frequency calibration signals. This method introduces a standard radio frequency calibration device as a physical reference source and combines it with a neural network of a specific structure for hierarchical fine-tuning topology to achieve cross-device model transfer.

[0026] The overall process is divided into an offline baseline building phase and an online migration fine-tuning phase, with the specific steps as follows: Phase 1: Offline Benchmark Building Phase Step 1.1: Standard Signal Transmission and Source Domain Benchmark Feature Consolidation. Place the standard RF calibration device near the antenna port of the source reconnaissance equipment (i.e., the equipment used in the original training model). Control the calibration device to transmit a preset typical UAV RF signal (e.g., an OFDM signal with a center frequency of 5.8 GHz and a bandwidth of 20 MHz). The source reconnaissance equipment acquires this signal through its RF front-end, and outputs a baseband IQ data stream after down-conversion processing. Truncate this data stream to obtain a fixed-length (e.g., 1024 sampling points) set of IQ data segments, generating the source domain calibration IQ dataset D. s D s The input is fed into a pre-trained initial recognition model, and through the feature extraction backbone network of this model, the corresponding source domain calibration feature set F is output. s D s and F s The package is stored as a "hardware characteristic baseline file" in the storage medium of the source reconnaissance device.

[0027] Phase Two: Online Migration Fine-tuning Phase Step 2.1: Target Domain Calibration Signal Acquisition. When the initial identification model described above needs to be migrated and deployed to the target reconnaissance equipment, a standard RF calibration device is deployed near the antenna port of the target reconnaissance equipment, transmitting a typical RF signal with the same parameters as in Step 1.1. The target reconnaissance equipment acquires this signal through its own RF front-end, performs down-conversion processing, and generates a target domain calibration IQ dataset D according to the same truncation rules as in Step 1.1. t .

[0028] Step 2.2: Initial Feature Extraction of the Target Domain. On the computing platform of the target reconnaissance equipment, an initial recognition model with the exact same structure as the source equipment is loaded. The target domain calibration IQ dataset is input into this model, allowing only unidirectional forward propagation of the signal through the feature extraction backbone network. Without performing any backpropagation calculations, the target domain calibration feature set F is extracted. t .

[0029] Step 2.3: Construct the hardware distortion residual loss. On the computing platform of the target reconnaissance equipment, read the pre-stored source domain calibration feature set. Since the sources of both the input source domain calibration feature set and the target domain calibration feature set are known standard signals, the statistical difference between them in the high-dimensional feature space is physically equivalent to the difference in physical distortion of the RF front-end hardware between the source reconnaissance equipment and the target reconnaissance equipment. The system executes the following two layers of computational logic to quantify this difference: Step 2.3.1 (Feature Compression): For F respectively s and F t Global average pooling is performed along the feature channel dimension to compress the 3D feature map into a one-dimensional global mean vector μ.s and μ t ; Step 2.3.2 (Distance Metric): Call the cosine similarity calculation module to calculate μ. s and μ t The cosine distance between them is defined as the hardware distortion residual loss L. calib .

[0030] Step 2.4: Targeted Fine-tuning of Topology Based on Residual Loss. Construct the computational graph topology for the fine-tuning phase on the target reconnaissance equipment and execute the following network hierarchical update strategy: Step 2.4.1 (Bottom-level Freeze Truncation): Insert a gradient stopping operator in the backpropagation path of the computation graph to absolutely freeze the network weight parameters of the first N layers (N is a positive integer, usually the bottom convolutional layer that extracts basic time-frequency texture features) of the feature extraction backbone network in the initial recognition model, and force its gradient to be set to zero. Step 2.4.2 (Top-level Directed Update): Construct the total loss function L total =L calib (In an optional embodiment, if a very small number of real labeled samples collected by the target device are combined, then L is constructed) total =L task + λ˙L calib L task The cross-entropy classification loss is used, where λ is a preset weight coefficient. Backpropagation updates are only enabled for the last few layers of the feature extraction backbone network and the top-level classification network. The optimizer minimizes L... calib The mean vector μ of the forced target domain feature distribution t The mean vector μ of the source domain baseline characteristic distribution in the mathematical space. s Align them closely.

[0031] Step 2.5: Model Consolidation and Deployment. Monitor the total loss function L. total The numerical change of the target weights is monitored, and when it decreases and converges to a preset threshold range, the fine-tuning of the computation graph is interrupted. The network layer states containing the updated weight parameters are extracted from the current computation graph, serialized and saved, and a target recognition model file adapted to the target reconnaissance equipment is generated. The temporary computation graph is destroyed, and the target recognition model file is loaded into the inference engine of the target reconnaissance equipment to take over the subsequent real-time UAV type recognition task.

[0032] like Figure 2 As shown in the figure, all modules, data flows, and operators are labeled in Chinese, and are divided into three vertical regions from left to right: the physical signal layer, the feature extraction layer, and the residual alignment layer. Physical signal layer: contains standard radio frequency calibration devices. The same standard physical signal is collected by heterogeneous devices of source reconnaissance equipment and target reconnaissance equipment to form source domain calibration IQ dataset Ds and target domain calibration IQ dataset Dt.

[0033] Feature extraction layer: It contains two feature extraction backbone networks, which output tensor source domain calibration feature set Fs and target domain calibration feature set Ft, respectively.

[0034] Residual Alignment Layer: The source domain calibration feature set Fs and the target domain calibration feature set Ft first enter the "global average pooling" processing box, and are respectively reduced in dimensionality to output a one-dimensional vector mean vector μ. s and mean vector μ t The result is then fed into the cosine distance calculation module to obtain the hardware distortion residual loss L. calib L calib By bypassing the frozen layers (first N layers, gradient set to zero) of the feature extraction backbone network, and connecting only to the unfrozen layers and the top-level classification network, this allows only the top-level network to perform the feature extraction based on L. calib Perform directional fine-tuning during backpropagation.

[0035] Example 2 Example: Cross-device migration and deployment of 50 types of UAV models based on the 1435B-V calibration source This embodiment details the complete engineering process of standardizing and migrating a complex identification model capable of recognizing at least 50 UAV models across reconnaissance equipment with different hardware specifications. Those skilled in the art will understand that the specific models, quantities, and network layers described below are merely preferred parameters for implementing the present invention and do not constitute a limitation on the scope of protection.

[0036] 1. System hardware and initial model baseline configuration Standard RF calibration device: using Siyi Technology 1435B-V portable vector signal generator.

[0037] Initial recognition model architecture: adopts a deep one-dimensional convolutional neural network (1D-CNN), which includes 12 convolutional layers (Conv1 to Conv12, where Conv1 to Conv7 are the bottom basic feature extraction layers, and Conv8 to Conv12 are the high-level semantic abstraction layers) and 1 fully connected classification layer (the number of output nodes is 50, corresponding to 50 drone models).

[0038] Calibration signal library configuration: The 1435B-V comes pre-installed with various RF signal waveform files simulating actual drone models, with a focus on image transmission / telemetry signals from mainstream DJI (such as Mavic 3, Mini 3, etc.) and Autel (such as EVO II, etc.). Users can import specific drone RF packet capture waveforms of interest to them as supplementary calibration sources through the 1435B-V.

[0039] 2. Phase One: Detailed Execution of the Offline Benchmark Building Phase Step 1.1 Execution (Standard Signal Transmission and Source Domain Reference Feature Fixing): Place the 1435B-V in an open area with a clean electromagnetic environment and no interference signals. Control the 1435B-V to sequentially transmit pre-installed signals from DJI, Autel, and other actual aircraft models (covering the 2.4GHz and 5.8GHz frequency bands). The source reconnaissance equipment directly collects the radiated signal through its own antenna and outputs a baseband IQ data stream after down-conversion. The system truncates the IQ data stream according to a fixed time window (1024 sampling points) to generate a source domain calibration IQ dataset D containing features of various aircraft models. s D s The input is fed into a pre-trained 50-class recognition model, and unidirectional forward propagation extracts the source domain calibration feature set F from the Conv12 output. s D s and F s Packaged as a "hardware feature baseline file" and encrypted and stored in the storage medium of the source device.

[0040] 3. Phase Two: Detailed Execution of the Online Migration Fine-tuning Phase In practical applications, the solution of this invention is perfectly compatible with the following two typical cross-device deployment scenarios: Scenario A: The source and target devices have comparable performance (both equipped with training GPUs). In this scenario, the target device has strong backpropagation computing capabilities and can perform complete network topology fine-tuning.

[0041] Step 2.1 (Target Domain Calibration Signal Acquisition): The same 1435B-V is moved to the field where the target reconnaissance equipment is located, and radiates the same calibration signal as in Step 1.1 under clean electromagnetic environment conditions. The target reconnaissance equipment acquires this radiated signal through its radio frequency front end, and after down-conversion processing, generates the target domain calibration IQ dataset Dt according to the same truncation rule as in Step 1.1 (1024 sampling points).

[0042] Step 2.2 Execution (Initial Feature Extraction of the Target Domain): On the computing platform of the target reconnaissance equipment, load an initial recognition model with the exact same structure as the source equipment. Input the target domain calibration IQ dataset into this model, allowing the signal to propagate unidirectionally forward through the feature extraction backbone network. Without performing any backpropagation calculations, extract the target domain calibration feature set F. t .

[0043] Step 2.3 Execution (Constructing Hardware Distortion Residual Loss): On the computing platform of the target reconnaissance equipment, read the pre-stored source domain calibration feature set F. s Because of input Fs With F t The sources of both are known standard signals. The statistical difference between them in the high-dimensional feature space is physically equivalent to the difference in physical distortion of the radio frequency front-end hardware between the source reconnaissance equipment and the target reconnaissance equipment. The system executes the following two layers of computational logic to quantify this difference: Step 2.3.1 Execution (Feature Compression): Perform feature compression on F respectively. s and F t Global average pooling is performed along the feature channel dimension to compress the 3D feature map into a one-dimensional global mean vector μ. s and μ t ; Step 2.3.2 Execution (Distance Metric): Call the cosine similarity calculation module to calculate μ. s and μ t The cosine distance between them is defined as the hardware distortion residual loss L. calib .

[0044] Step 2.4 Execution (Targeted Fine-tuning Topology Based on Residual Loss): Construct the computational graph topology for the fine-tuning phase on the target reconnaissance equipment and execute the following network layering update strategy: Step 2.4.1 Execution (Bottom-level Freeze-Truncation): In the backpropagation path of the computation graph, insert the gradient stopping operator to absolutely freeze the network weight parameters of the first 7 layers (i.e., Conv1 to Conv7) of the feature extraction backbone network in the initial recognition model, and force its gradient to be set to zero. Step 2.4.2 Execution (Top-Level Directed Update): Construct the total loss function L total =L calib Backpropagation updates are enabled only for the last 5 layers (Conv8 to Conv12) of the feature extraction backbone network and the top-level classification network. The optimizer minimizes L... calib, The mean vector μ of the forced target domain feature distribution t The mean vector μ of the source domain baseline characteristic distribution in the mathematical space. s Align them closely.

[0045] Step 2.5 Execution (Model Consolidation and Deployment): Monitor the total loss function L total The numerical change of the target weights is monitored, and when it decreases and converges to a preset threshold range, the fine-tuning of the computation graph is interrupted. The network layer states containing the updated weight parameters are extracted from the current computation graph, serialized and saved, and a target recognition model file adapted to the target reconnaissance equipment is generated. The temporary computation graph is destroyed, and the target recognition model file is loaded into the inference engine of the target reconnaissance equipment to take over the subsequent real-time UAV type recognition task.

[0046] Scenario B: The source device has powerful performance, while the target reconnaissance device has average performance (only equipped with an embedded edge computing chip). In this scenario, the target edge device is unable to handle the gradient calculation for backpropagation due to limitations in computing power and memory. An offline fine-tuning strategy of edge-cloud collaboration (decoupling of acquisition and computation) is adopted.

[0047] Step 2.1 Execution (Target Domain Calibration Signal Acquisition): The target reconnaissance equipment is placed in a clean electromagnetic environment, using 1435B-V radiated signals. The target reconnaissance equipment uses its radio frequency front-end to acquire signals, and after down-conversion processing, generates a target domain calibration IQ dataset D according to the same truncation rules as in Step 1.1 (1024 sampling points). t The target device does not perform any model training tasks, but only transfers D... t Export as the raw data file.

[0048] Step 2.2 Execution (Initial Feature Extraction of Target Domain): Extract the dataset D exported from the target device. t Copy or transfer the dataset to a powerful backend server. On the backend server's computing platform, load the initial recognition model and process the dataset D. t The input signal is fed into the model, allowing it to propagate unidirectionally forward through the feature extraction backbone network. Without performing any backpropagation calculations, the target domain calibration feature set F is extracted. t .

[0049] Step 2.3 Execution (Constructing Hardware Distortion Residual Loss): On the computing platform of the backend server, read the pre-stored source domain calibration feature set F. s The system executes the following two layers of computational logic to quantify the differences in hardware physical distortion between the source and target devices: Step 2.3.1 Execution (Feature Compression): Perform feature compression on F respectively. s and F t Global average pooling is performed along the feature channel dimension to compress the 3D feature map into a one-dimensional global mean vector μ. s and μ t ; Step 2.3.2 Execution (Distance Metric): Call the cosine similarity calculation module to calculate μ. s and μ t The cosine distance between them is defined as the hardware distortion residual loss L. calib .

[0050] Step 2.4 Execution (Targeted Fine-tuning Topology Based on Residual Loss): Construct the computation graph topology for the fine-tuning phase on the backend server, and execute an extremely lightweight network layering update strategy to address the performance bottlenecks of edge devices. Step 2.4.1 Execution (Bottom-level Freeze-Truncation): In the backpropagation path of the computation graph, insert the gradient stopping operator to absolutely freeze the network weight parameters of the first 11 layers (i.e., Conv1 to Conv11) of the feature extraction backbone network in the initial recognition model, and force its gradient to be set to zero. Step 2.4.2 Execution (Top-Level Directed Update): Construct the total loss function L total =L calib Backpropagation updates are enabled only for the last convolutional layer immediately following the fully connected classification layer (i.e., Conv12) and the top-level classification network. The optimizer minimizes L... calib The mean vector μ of the forced target domain feature distribution t The mean vector μ of the source domain baseline characteristic distribution in the mathematical space. s Align them closely.

[0051] Step 2.5 Execution (Model Consolidation and Deployment): Monitor the total loss function L on the backend server. total After convergence, the fine-tuning computation graph is interrupted. The network layer states containing the updated weight parameters are extracted from the current computation graph, serialized and saved, generating a target recognition model file specifically adapted to this edge target reconnaissance device. Finally, this target recognition model file is copied across devices and loaded into the target reconnaissance device's lightweight inference engine to take over subsequent UAV type recognition tasks.

[0052] Beneficial Effects: This invention provides a cross-device UAV identification model transfer method based on standard radio frequency calibration signals. By introducing known prior standard signals as physical reference anchors, it achieves accurate quantification and elimination of hardware physical distortions; by using standard signals emitted by calibration devices to replace real flight samples, it completely breaks down the scarcity and regional barriers of field data acquisition; by eliminating the massive domain discriminator network and adopting a directional freezing fine-tuning strategy, it greatly reduces computational resource consumption, enabling high-precision models to be successfully deployed to low-computing-power edge reconnaissance devices; by encapsulating complex algorithms into a local automatic fine-tuning closed loop based on physical anchors, it eliminates dependence on expert experience and allows target devices to automatically complete model updates and be immediately deployed in actual combat within minutes of being refitted, completely eliminating the combat capability vacuum period.

[0053] The above specific embodiments further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for transferring UAV identification models based on standard radio frequency calibration signals, characterized in that, The migration method includes: The overall process includes: an offline benchmark building phase and an online migration fine-tuning phase; The offline benchmark construction phase specifically includes: Step S1.1: Standard signal transmission and source domain reference characteristics are solidified; The online migration fine-tuning phase specifically includes: Step S2.1: Acquisition of target domain calibration signals; Step S2.2: Initial feature extraction of the target domain; Step S2.3: Construct the hardware distortion residual loss; Step S2.4: Based on residual loss, perform directional fine-tuning of the topology, construct the computational graph topology for the fine-tuning stage on the target reconnaissance equipment, and execute the network hierarchical update strategy; Step S2.5: Model solidification and deployment.

2. The method for transferring UAV identification models based on standard radio frequency calibration signals according to claim 1, characterized in that, Step S1.1: Standard signal transmission and source domain reference feature solidification specifically includes: Place the standard radio frequency calibration device at the antenna port of the source reconnaissance equipment; The control calibration device transmits a preset typical radio frequency signal for the UAV; The source detection equipment acquires the signal through its radio frequency front end, and outputs a baseband IQ data stream after down-conversion processing. The baseband IQ data stream is truncated to obtain a set of IQ data segments of fixed length, and a source domain calibration IQ dataset Ds is generated. The source domain calibration IQ dataset Ds is input into the pre-trained initial recognition model, and the corresponding source domain calibration feature set Fs is output through the feature extraction backbone network of the initial recognition model. The source domain calibration IQ dataset Ds and the source domain calibration feature set Fs are packaged as a hardware feature benchmark file and pre-stored in the storage medium of the source reconnaissance device.

3. The method for transferring UAV identification models based on standard radio frequency calibration signals according to claim 1, characterized in that, Step S2.1: Target domain calibration signal acquisition specifically includes: When it is necessary to migrate and deploy the initial identification model to the target reconnaissance equipment, deploy the same standard radio frequency calibration device near the antenna port of the target reconnaissance equipment and transmit a typical radio frequency signal with the same parameters as in step S1.1; The target reconnaissance equipment acquires the signal through its own radio frequency front-end, performs down-conversion processing, and generates a target domain calibration IQ dataset D according to the same truncation rules as in step S1.

1. t .

4. The method for transferring UAV identification models based on standard radio frequency calibration signals according to claim 1, characterized in that, Step S2.2: Initial feature extraction of the target domain specifically includes: On the computing platform of the target reconnaissance equipment, an initial identification model with the exact same structure as the source equipment is loaded; The target domain calibration IQ dataset is input into the model, and the signal is allowed to propagate unidirectionally forward through the feature extraction backbone network. Without performing any backpropagation calculations, the target domain calibration feature set F is extracted. t .

5. The UAV identification model transfer method based on standard radio frequency calibration signals according to claim 1, characterized in that, Step S2.3: Constructing the hardware distortion residual loss specifically includes: On the computing platform of the target reconnaissance equipment, the pre-stored source domain calibration feature set is read; Since the sources of both the input source domain calibration feature set and the target domain calibration feature set are known standard signals, the statistical differences between them in the high-dimensional feature space are physically equivalent to the differences in the physical distortion of the radio frequency front-end hardware between the source reconnaissance equipment and the target reconnaissance equipment.

6. The UAV identification model transfer method based on standard radio frequency calibration signals according to claim 5, characterized in that, The methods for calculating the statistical differences between the two in the high-dimensional feature space include: Step S2.3.1: Perform feature compression, respectively for F s and F t Global average pooling is performed along the feature channel dimension to compress the 3D feature map into a one-dimensional global mean vector μ. s and μ t ; Step S2.3.2: Perform distance measurement, call the cosine similarity calculation module, and calculate μ. s and μ t The cosine distance between them is defined as the hardware distortion residual loss L. calib .

7. The UAV recognition model transfer method based on standard radio frequency calibration signals according to claim 1, characterized in that, Step S2.4: Based on residual loss, the topology is fine-tuned by constructing a computational graph topology for the fine-tuning phase on the target reconnaissance equipment, and the network hierarchical update strategy is executed, specifically including: Step 2.4.1, bottom layer freeze truncation: In the backpropagation path of the computation graph, insert the gradient stopping operator to absolutely freeze the network weight parameters of the first N layers of the feature extraction backbone network in the initial recognition model, and force the gradient to be set to zero. N is a positive integer and is the bottom convolutional layer for extracting basic time-frequency texture features. Step 2.4.2, Top-level directional update: Construct the total loss function L total =L calib Backpropagation updates are enabled only for the last few layers of the feature extraction backbone network and the top-level classification network; L is minimized using the optimizer. calib The mean vector μ of the forced target domain feature distribution t The mean vector μ of the source domain baseline characteristic distribution in the mathematical space. s Align them closely.

8. The method for transferring UAV identification models based on standard radio frequency calibration signals according to claim 1, characterized in that, Step S2.5: Model solidification and deployment specifically includes: Monitoring total loss function L total The numerical change of the graph is interrupted when it decreases and converges to a preset threshold range. Extract the network layer states containing the updated weight parameters from the current computation graph, serialize and save them, and generate a target recognition model file adapted to the target reconnaissance equipment. The temporary computation graph is destroyed, and the target recognition model file is loaded into the inference engine of the target reconnaissance equipment to take over the subsequent real-time UAV model recognition task.