Drone Radio Fingerprint ID
Dynamic Knowledge Distillation Wireless RFF-LLM Framework for UAV Individual Identification
The integration of Unmanned Aerial Vehicles (UAVs) into the global low-altitude economy presents a complex duality of transformative economic potential and profound asymmetrical security risks. As commercial and consumer drone adoption accelerates, the existing mechanisms for airspace management and security enforcement face obsolescence. The research document UAV Individual Identification via Distilled RF Fingerprints-Based LLM in ISAC Networks (arXiv 2508.12597v2) introduces a sophisticated architectural response to the critical challenge of authenticating non-cooperative aerial systems. This proposed framework represents a paradigm shift from reliance on digital, easily spoofable identifiers such as Media Access Control (MAC) addresses or cooperative Remote ID broadcasts. Instead, the authors advocate for and demonstrate a method leveraging the immutable physical layer (PHY) imperfections inherent in transmitter hardware—specifically Radio Frequency Fingerprinting (RFF)—to assign a unique, unclonable identity to individual UAVs.
Originating from the National Mobile Communications Research Laboratory at Southeast University in Nanjing, China, this research aligns with a broader geopolitical and technological trend toward granular control of the electromagnetic spectrum and the physical airspace it governs. The proposed system engineers a hybrid artificial intelligence framework that integrates a Large Language Model (LLM) based on a modified GPT-2 architecture acting as a “teacher” network and a lightweight Lite-HRNet functioning as a “student” network. The transfer of feature extraction capabilities between these two entities is governed by a novel dynamic distillation process controlled by a Proximal Policy Optimization (PPO) agent. This algorithmic innovation addresses a persistent inefficiency in deep learning model compression, where static distillation parameters often trap lightweight models in local optima, thereby degrading their operational performance.
The claimed performance metrics for this system are substantial and demand rigorous scrutiny. The authors report an identification accuracy of 98.38 percent across a test group of 20 discrete commercial UAVs, achieved while maintaining a remarkably lightweight computational footprint of 0.15 million parameters and an inference latency of merely 2.74 milliseconds. These specifications suggest a high degree of suitability for deployment at the network edge, specifically within the computational constraints of next-generation Integrated Sensing and Communication (ISAC) base stations. If validated in the field, such a system would enable cellular networks to function as ubiquitous, passive radar grids capable of identifying specific drone units without the target’s consent or cooperation.
However, a forensic analysis of the system’s experimental design and validation methodology reveals critical operational limitations that temper these theoretical successes. The reliance on a dataset (DRFF-R1) constructed exclusively from signals captured on Channel 149 (5.745 GHz) introduces a potentially fatal vulnerability regarding real-world applicability against frequency-hopping protocols like DJI’s OcuSync. Furthermore, the explicit assumption of a hovering state with no Doppler effect artificially simplifies the electromagnetic environment, raising significant doubts about the system’s robustness against high-velocity, maneuvering targets. While the PPO-driven distillation strategy is theoretically sound and technically advanced, the system’s operational readiness serves currently as a proof of concept rather than a deployable defense asset.
The Low-Altitude Security Imperative
The domain of low-altitude airspace—typically defined as the region below 400 feet (120 meters) above ground level—has historically been an unregulated void. The rapid democratization of UAV technology has transformed this void into a congested and contested operational theater. The security challenges here are multifaceted and evolving. Malicious actors utilize drones for illegal reconnaissance of sensitive facilities, delivery of contraband into correctional institutions, and potentially for kinetic terrorist attacks using modified commercial off-the-shelf (COTS) platforms. In these scenarios, the primary asset of the attacker is anonymity.
Traditional airspace surveillance relies heavily on cooperative identification technologies. Systems like ADS-B (Automatic Dependent Surveillance-Broadcast) or the newly mandated Remote ID protocols require the aircraft to actively and truthfully broadcast its identity, location, and velocity. However, a determined adversary operating a non-cooperative drone can easily disable these broadcasts or spoof the data packets. Consequently, reliance on digital ID protocols leaves a gaping vulnerability in the security posture of critical infrastructure.
The authors of the analyzed document explicitly position their research as a direct countermeasure to these threats. They cite incidents of privacy infringement and illegal surveillance as primary motivators for developing a more robust identification method. The strategic intent is to strip the veil of anonymity from these devices by exploiting the one characteristic they cannot easily alter—the physical interaction of their internal hardware components with the laws of physics. This shifts the security paradigm from “what the drone says it is” (a digital claim) to “what the drone actually is” (a physical reality).
Integrated Sensing and Communication (ISAC) Networks
The research is deeply embedded within the context of 5G-Advanced and the emerging 6G telecommunications standards, specifically focusing on Integrated Sensing and Communication (ISAC). ISAC represents a technological convergence where cellular base stations evolve beyond their traditional role as mere data relays. In an ISAC architecture, base stations function simultaneously as communication hubs and as active or passive radar sensors.
The authors propose embedding their RFF-LLM identification model directly into the edge computing infrastructure of these ISAC networks. This architectural choice is significant because it implies a future surveillance state where the cellular network itself becomes a ubiquitous, city-wide sensor grid. Every base station becomes a listening post capable of detecting, tracking, and identifying specific drone hardware in its vicinity. The intent is to create a seamless, wide-area mesh of identification nodes that requires no additional, expensive radar hardware. By leveraging the existing and ubiquitous telecommunications infrastructure, the system aims to achieve persistent airspace dominance and granular situational awareness in urban canyons where traditional radar often fails due to occlusion.
Theoretical Foundations of Radio Frequency Fingerprinting
Physics of Hardware Impairments
To critically evaluate the proposed system, it is essential to understand the underlying mechanism of Radio Frequency Fingerprinting (RFF). This technique does not involve decoding the encrypted payload of the drone’s control link or video stream. Instead, it analyzes the subtle analog distortions introduced into the signal by the transmitter’s hardware chain. These distortions arise from microscopic manufacturing tolerances and material impurities that are economically and physically impossible to eliminate in consumer-grade electronics.
The document highlights two primary sources of these fingerprints which serve as the foundation for their feature extraction algorithms. The first is Carrier Frequency Offset (CFO). This phenomenon occurs when the local oscillator (LO) of the transmitter does not oscillate at the exact intended frequency. Due to the unique crystalline structure of the quartz used in oscillators, every device exhibits a constant, unique frequency deviation. Even two drones of the exact same model, coming off the same assembly line, will have oscillators that vibrate at infinitesimally different frequencies.
The second primary source is I/Q Imbalance. In modern direct-conversion transmitters, the signal is split into In-phase (I) and Quadrature (Q) components which must be perfectly matched in amplitude and separated by exactly 90 degrees in phase. Hardware imperfections in the mixers, amplifiers, and filters lead to gain imbalance (\alpha) and phase imbalance (\varphi). The system model provided in the document describes the received signal s(t) as a convolution of the channel H(t) and the baseband signal b(t), plus environmental noise. The baseband signal b(t) explicitly incorporates these imperfections. The core theoretical premise is that while the channel H(t) is mutable and changes with the environment, the hardware parameters (\alpha and \varphi) remain intrinsic to the device. The fundamental challenge for any AI model in this domain is to mathematically disentangle the variable channel effects from the immutable hardware fingerprints.
The Problem of Feature Extraction
Historically, RFF identification relied on manual feature engineering. Engineers would design algorithms to calculate statistical moments, kurtosis, spectral flatness, or fractal dimensions of the signal. The authors of the paper correctly argue that these traditional methods introduce excessive computational overhead and lack robustness in complex, noisy environments. Manual features are often brittle—they work well in the lab but fail when the signal is subjected to the multipath fading and interference of the real world.
Deep learning automates this feature extraction, allowing neural networks to learn the most relevant characteristics from the raw data. However, deep learning models face the “black box” dilemma. A neural network might inadvertently learn to identify a drone based on the background noise of the environment where it was recorded (a spurious correlation) rather than the device’s actual fingerprint. To remediate this, the authors utilize a massive, pre-trained Large Language Model (modified GPT-2) as a “Teacher.” The hypothesis is that the LLM’s capacity to understand complex sequences allows it to guide the feature extraction process, forcing the student model to learn more robust, generalized representations of the signal structure rather than memorizing environmental artifacts.
Architectural Deconstruction The Teacher Network
Adapting GPT-2 for the RF Domain
The decision to utilize a GPT-2 architecture for Radio Frequency (RF) analysis is a bold and somewhat unconventional design choice. GPT-2 is fundamentally a sequence-to-sequence model designed for natural language processing (NLP). It excels at predicting the next token in a sequence of text by attending to long-range dependencies within the data. The authors repurpose this text-generation engine to process electromagnetic signals, treating the time-frequency spectrogram of a radio wave as a “language” with its own syntax and grammar.
However, a direct application of GPT-2 to RF data is insufficient. Standard GPT-2 utilizes absolute positional encoding, which tags each piece of data with its fixed position in the sequence (e.g., “this is the 5th word”). The authors argue that this is suboptimal for Time-Frequency spectra generated by Short-Time Fourier Transform (STFT). Wireless signals are non-stationary; their features shift over time due to modulation schemes and channel effects. A specific signal feature, such as a protocol preamble, might arrive slightly earlier or later depending on synchronization. Absolute positioning might confuse the model if the feature does not appear at the exact expected index.
To remediate this limitation, the authors replace the standard positional embedding with a Bidirectional Long Short-Time Memory (BiLSTM) network. The BiLSTM processes the spectral sequence in both forward and backward directions, creating a dynamic, content-aware temporal embedding. This allows the subsequent Transformer layers to focus on the relationship between signal components rather than their rigid positions. This modification demonstrates a nuanced understanding of signal processing, acknowledging that RF data possesses a temporal fluidity that text data does not always exhibit. The BiLSTM allows the model to understand the signal’s context regardless of slight temporal shifts, effectively synchronizing the “understanding” of the neural network with the incoming radio waves.
Architectural Deconstruction The Student and Distillation
Lite-HRNet The Edge-Native Student
While the RFF-LLM teacher is powerful, it is computationally obese. The GPT-2 architecture involves millions of parameters and requires substantial GPU memory, making it impractical for deployment on a standard cellular base station or a portable edge device where power and latency constraints are stringent. The authors address this deployment gap by distilling the intelligence of the teacher into a Lite-HRNet (Lightweight High-Resolution Network).
Lite-HRNet was originally designed for human pose estimation in computer vision. Its key architectural strength is its ability to maintain high-resolution representations throughout the network depth. Most convolutional neural networks (like ResNet or UNet) aggressively down-sample the input image to a low-resolution “bottleneck” to save computation, and then up-sample it back. In the context of RF spectrograms, this down-sampling can be destructive. RF fingerprints often manifest as subtle, high-frequency micro-features within the spectrogram—tiny aberrations in the frequency or phase that occupy very few pixels. Aggressive down-sampling could obliterate the very “fingerprints” the system is trying to detect.
The choice of Lite-HRNet suggests that the authors prioritize the preservation of these fine spectral details. The student model achieves a parameter count of merely 0.15 million. This is orders of magnitude smaller than the teacher network and significantly smaller than competing models like ResNet (0.74 million parameters). Despite this massive reduction in size, the student model reportedly retains 98.38 percent accuracy. This high compression ratio with minimal loss of accuracy validates the efficacy of the distillation process, proving that the essential “knowledge” of the fingerprint can be represented much more compactly than the raw ability to extract it.
PPO Dynamic Distillation
The most technically sophisticated component of the framework—and its primary claim to novelty—is the dynamic knowledge distillation strategy. In standard knowledge distillation, a “temperature” parameter (T or \tau) controls how much the teacher’s knowledge is “softened” before being passed to the student. A high temperature flattens the probability distribution, revealing information about the “wrong” classes (known as “dark knowledge”), while a low temperature sharpens the focus on the correct class.
Typically, \tau is a fixed hyperparameter set before training begins. The authors identify this as a flaw, noting that a fixed temperature can trap the student model in local optima. A student might need broad guidance early in training (high temperature) and sharp correction later (low temperature), or vice versa depending on the difficulty of the specific samples. To counter this rigidity, they employ Proximal Policy Optimization (PPO), a reinforcement learning algorithm.
In this framework, the PPO agent views the training process as a game. The State (s_t) is the current accuracy and convergence rate of the student network. The Action (a_t) available to the agent is adjusting the temperature \tau. The Reward (R_t) is defined as the improvement in validation accuracy and the reduction in loss functions. This creates a self-tuning feedback loop. When the student is struggling to differentiate between similar drones, the PPO agent might increase \tau to provide richer, softer guidance from the teacher. When the student is confident, the agent might decrease \tau to sharpen the decision boundaries. This adaptive mechanism is likely the primary driver behind the system’s ability to outperform static baselines, as it customizes the learning curriculum in real-time based on the student’s progress.
Experimental Methodology and Data Forensics
The DRFF-R1 Dataset
The credibility of any machine learning model is inextricably bound to the quality and representativeness of its training data. It is in this domain that the proposed system exhibits its most significant vulnerabilities. The authors constructed the DRFF-R1 dataset using a USRP B210 receiver to capture signals from 20 drones across 7 distinct commercial models. While the effort to create and release a public dataset is commendable given the scarcity of such resources in the RF security community, the specific methodology used for data collection introduces severe limitations that compromise the system’s real-world utility.
The Channel 149 Fallacy
The document explicitly states that signal collection occurred on Channel 149. Channel 149 corresponds to a center frequency of 5.745 GHz in the U-NII-3 band. This detail is critical because the majority of modern commercial drones, particularly the DJI models (Mavic, Mini, Phantom) included in the study, utilize proprietary transmission protocols like OcuSync or Lightbridge.
These advanced protocols utilize Frequency Hopping Spread Spectrum (FHSS) for the control link (uplink) and sophisticated Orthogonal Frequency-Division Multiplexing (OFDM) schemes for the video downlink. FHSS systems rapidly switch carrier frequencies across the available band—hopping between 2.4 GHz and 5.8 GHz, or across multiple channels within the 5.8 GHz band—to avoid interference and mitigate jamming. A drone using OcuSync does not stay on Channel 149; it might transmit a packet on Channel 149, the next on Channel 157, and the next on Channel 161, all within milliseconds.
By restricting data collection to a single fixed channel, the authors have introduced a fundamental sampling bias. This implies one of two scenarios occurred. First, the researchers may have manually forced the drones into a “fixed-frequency mode” using developer tools or hardware modifications. While this allows for clean, continuous data collection, it fundamentally alters the RF signature. A radio transmitting continuously on a fixed frequency exhibits different thermal characteristics and power amplifier non-linearities than one that is rapidly switching frequencies. The “fingerprint” learned in this artificial static mode may not match the “fingerprint” of a drone operating in its normal, hopping state. Second, they may have simply recorded opportunistically whenever the hopping sequence happened to land on Channel 149. This would result in discontinuous data bursts and fails to capture the transient responses associated with the frequency switching itself—transients that are often the most fingerprint-rich portion of the signal. Consequently, a model trained on this dataset risks catastrophic failure when deployed against a drone operating in its default configuration. The system would be effectively “blind” to the drone for the vast majority of time it is transmitting on channels other than 149.
The Doppler Omission
The experimental setup explicitly assumes a hovering UAV and a stationary receiver, leading the authors to conclude that there is “no Doppler effect” and no influence from the receiver’s own motion. This is a theoretical simplification that severely undermines the operational utility of the model.
In a real-world security scenario, a malicious drone is rarely hovering statically. It is likely traversing the airspace at speed—potentially 15 to 20 meters per second for consumer drones, and significantly faster for fixed-wing variants. This motion induces Doppler shifts that compress or stretch the signal waveform. Furthermore, the rapid rotation of the drone’s propellers generates micro-Doppler effects that modulate the signal. These micro-Doppler signatures are often used by radar systems to identify drone types, but in the context of transmitter fingerprinting, they act as noise that can obscure the subtle impairments of the oscillator and amplifier.
By training on zero-Doppler data, the RFF-LLM is learning a feature set that exists only in a laboratory vacuum. When presented with a signal from a moving drone, the temporal features extracted by the BiLSTM would be warped by Doppler contraction or dilation. The model, having never seen these distortions, would likely fail to correlate the warped signal with the static fingerprint it learned. The 98.38 percent accuracy metric is, therefore, valid only for hovering targets, rendering the system potentially ineffective against kinetic threats or drones executing complex flight paths.
Performance Benchmarking and Comparative Analysis
Efficiency Metrics
The authors provide a comparative analysis of their Lite-HRNet-KD model against several standard deep learning architectures: ResNet, LMSC, and ShuffleNet-v2. The data reveals a stark contrast in efficiency and performance.
The ResNet model, a heavy industry standard, achieved an accuracy of 91.85 percent. However, it required 0.74 million parameters to do so. While its inference latency was identical to the proposed model at 2.74 ms, its lower accuracy and higher memory footprint make it less ideal for resource-constrained edge devices.
The LMSC (Lightweight Multi-Scale Convolutional) model, designed specifically for efficiency, failed to deliver on accuracy, achieving only 82.87 percent. Despite being a “lightweight” architecture, it paradoxically utilized the highest number of parameters at 1.34 million and had a slightly slower latency of 2.78 ms. This suggests that the LMSC architecture struggles to capture the subtle complexities of RF fingerprints without significant parameter bloat.
ShuffleNet-v2, another popular efficient architecture, performed well with 94.73 percent accuracy. It used 0.94 million parameters and had a latency of 2.83 ms. This makes it the closest competitor to the proposed system.
The proposed Lite-HRNet-KD model, however, outperformed all benchmarks. It achieved the highest accuracy at 98.38 percent while using the absolute fewest parameters—0.15 million. This is a reduction in model size of approximately 84 percent compared to ShuffleNet-v2 and nearly 80 percent compared to ResNet, with no penalty in latency (2.74 ms). This demonstrates the power of the dynamic distillation process; the student model effectively “punched above its weight” by learning the most critical features from the teacher, allowing it to discard the redundant parameters that bloated the other models.
Operational Latency
The inference latency of 2.74 ms is a critical specification. In a security context, decisions must be made in near-real-time. A drone flying at 20 meters per second covers 5 centimeters in 2.74 milliseconds. This processing speed allows the system to generate hundreds of identification estimates per second, enabling it to track a target continuously and smooth out occasional misclassifications through voting mechanisms. If the latency were in the order of hundreds of milliseconds, the system would struggle to keep up with a fast-moving target or to scan multiple channels effectively. The low latency confirms the suitability of Lite-HRNet for real-time edge deployment.
Vulnerability, Adversarial, and Ethical Analysis
Susceptibility to Signal Replay
The authors claim that RFF is “exceptionally resistant to malicious imitation” because it is based on hardware imperfections that are difficult to physically clone. While it is true that manufacturing a transmitter with the exact same silicon defects as a target device is nearly impossible, the system remains vulnerable to Signal Replay Attacks.
If an adversary captures the raw I/Q signal of a legitimate, authorized drone (a “whitelist” drone), they can store this high-fidelity recording. Later, using a high-end Software Defined Radio (SDR) capable of high sample rates and bit depths, they could replay this recorded signal. If the RFF system relies heavily on the preamble or specific synchronization sequences where the fingerprint is most prominent, the system might be fooled by the replay. The high-fidelity replay would contain the CFO and I/Q imbalance of the original authorized drone. The document does not discuss defenses against this, such as challenge-response protocols or integrating timestamp/sequence number validity checks (which would require decoding the packet, contradicting the “physical layer only” approach).
Adversarial Perturbations
The reliance on deep neural networks opens the door to adversarial attacks. Research in AI security has demonstrated that adding minute, human-imperceptible noise patterns to an input can cause a model to misclassify with high confidence. In the RF domain, an intelligent adversary could modulate their transmission with a specific noise pattern designed to trigger a misclassification in the Lite-HRNet. Because the student model is so lightweight (0.15M parameters), it may have a simplified decision boundary that is easier to exploit than a larger model. An attacker could theoretically mask a malicious drone as a benign bird or a different, authorized drone model by injecting this “adversarial noise” into their transmission.
Privacy and the Surveillance State
The provenance of this research—Southeast University’s National Mobile Communications Research Laboratory—carries significant geopolitical weight. This institution is a key node in the telecommunications research infrastructure, often contributing to national standards for 5G and 6G. The explicit focus on “UAV individual identification” aligns with a broader state interest in granular airspace control and the management of the “Internet of Drones.”
While the stated intent is security and accountability (countering terrorism and privacy invasion), the capability to identify specific individual drones creates a powerful tool for ubiquitous surveillance. It allows for the tracking of specific users, not just device types. If deployed across a city-wide ISAC network, this system would enable the historical logging of flight paths for every individual drone operator. It essentially creates a “license plate reader” system for the sky. This raises profound privacy concerns regarding the warrantless tracking of journalists, activists, hobbyists, and commercial operators. Unlike a license plate, which is visible only when the car is on a public road, an RF fingerprint is broadcast omnidirectionally and can be detected through walls and over great distances. The passive nature of the system means operators would have no knowledge that they are being tracked.
Strategic Recommendations and Future Directions
Expanding the Data Horizon
For this system to transition from a theoretical framework to a deployable security asset, the experimental design must be radically expanded. Future datasets must include signals collected from drones in rapid flight to incorporate Doppler effects. The training pipeline must include “Doppler augmentation”—mathematically simulating velocity-induced shifts on static data—to robustify the model against moving targets. Without this, the model is operationally useless against any drone that is not hovering perfectly still.
Solving the Frequency Agility Problem
The “Channel 149” limitation must be addressed. A deployable receiver architecture needs to track frequency-hopping patterns. This could involve a wideband receiver capable of monitoring the entire 2.4 GHz and 5.8 GHz ISM bands simultaneously, capturing the signal regardless of its hop channel. Alternatively, the receiver could be integrated with a “frequency follower” that detects the hop sequence and retunes the radio in real-time. The model effectively needs to be trained to recognize the fingerprint across the entire frequency spectrum, as hardware imperfections like I/Q imbalance often vary as a function of frequency (a fingerprint at 5.7 GHz might look different than the fingerprint of the same drone at 5.8 GHz).
Adversarial Hardening
The training phase should include adversarial training, where the model is exposed to perturbed signals and taught to resist them. This would increase the model’s resilience against intelligent jamming and spoofing attempts.
The system described in UAV Individual Identification via Distilled RF Fingerprints-Based LLM in ISAC Networks represents a sophisticated convergence of wireless engineering and modern artificial intelligence. The use of a GPT-2 based teacher and a PPO-distilled student model demonstrates a cutting-edge approach to automated feature extraction and model compression. The reported accuracy of 98.38 percent and latency of 2.74 ms set a new benchmark for lightweight edge identification, validating the potential of LLM-inspired architectures in the physical layer.
However, the system in its current iteration functions as a “fair-weather” sensor. Its reliance on a static, single-channel dataset renders it brittle against the dynamic, frequency-hopping, high-velocity reality of modern drone operations. The assumption of zero Doppler and the restriction to Channel 149 are simplifications that strip the problem of its most difficult real-world variables. The technology holds immense promise for airspace security—offering a way to identify threats that evade digital detection—but that promise is currently bounded by the artificial constraints of its experimental design. Significant expansion in data diversity, environmental complexity, and adversarial robustness is required before this framework can be trusted to secure the skies against determined, non-cooperative actors.
‐‐————————
anonymous_sanaa_dif_dox.txt
# 🚀 DRFF-R1 Dataset Description
## Language Selection
– [English](README.en.md)
– [ä¸æ–‡](README.md)
## 1. Equipment Introduction
### 1. Drones and Flight Controllers
This dataset uses DJI drones and their flight control systems, covering 20 drones across 7 models:
– 1 × Mavic Mini
– 1 × Mavic Air
– 1 × Mavic 3
– 3 × Mini SE
– 3 × Mini 3 Pro
– 5 × Mavic Air 2
– 6 × Mavic Air 2S
### 2. RF Signal Receiver
The RF signal reception device is the USRP-B210 from Ettus Research. Key specifications:
– Frequency range: 70 MHz – 6 GHz
– 2 TX / 2 RX channels
– Maximum instantaneous bandwidth: 56 MHz
– Signal gain: 0-90 dB
– Maximum I/Q sampling rate: 61.44 MS/s
### 3. Data Cable
The data cable connecting the host and USRP is an E164571-KS AWM 2725, supporting up to 5 GB/s transfer rate with latency <10 ms.
### 4. Host Computer
The host runs Windows 11 with an Intel(R) Core(TM) i5-13450HX CPU, 16 GB RAM, and a 2 TB portable HDD for storage.
A GNU Radio environment is installed to drive the USRP for signal acquisition and data I/O.
## 2. Dataset
### Naming Convention
Taking `mini_3pro_1_0.mat` as an example:
– `mini_3pro`: Drone model
– `1`: The 1st drone of this model
– `0`: Collection distance
– `0`: Baseline signal in an anechoic chamber (no EMI)
– `10`, `30`, `50`, `70`, `90`: Hovering distances (meters)
### Dataset Content
Each `.mat` file contains:
– 140 million In-phase (I) and Quadrature (Q) signal samples
– Sampling frequency (`Fs`)
– Center frequency (`CenterFreq`)
For detailed collection workflow, refer to `collect.py`.
### Data Collection Strategy
During the actual signal acquisition process, we simultaneously collected RF signals from real environments and reference signals with low channel-effect interference. Specifically:
– Control the drone to hover at different altitudes (10m/30m/50m/70m/90m) while keeping both the receiver and flight controller stationary, collecting RF signals without Doppler effects in the actual environment.
– To eliminate multipath effects caused by environmental reflections in real scenarios, we placed the USRP receiver, normally connected drone, and flight controller together inside microwave-absorbing foam for signal acquisition (the foam effectively absorbs electromagnetic waves in the 5-6 GHz band with reflection attenuation >30 dB).
📌 **Tips:**
When using the dataset:
– Signals collected under microwave-absorbing foam wrapping (marked as 0m distance data files) can serve as reference signals, which are unaffected by multipath and Doppler effects.
– You can superimpose different channel models (e.g., Rayleigh fading, multipath delay) on this reference signal through simulation to construct training datasets for complex scenarios.
– RF signals from real environments (marked with distances ≥10m) can directly validate models in real-world scenarios.
## 3. Collection Workflow
Two operators are required: Operator A controls the drone, Operator B manages the host.
1. **Operator A**:
– Powers on the drone near the host and places it in standby mode, connecting to a calibrated remote controller or mobile app.
– Maintains drone hovering at specified distances (anechoic chamber, 10m, 30m, 50m, 70m, 90m), with real-time distances monitored via the app.
2. **Operator B**:
– Executes Python scripts to connect the host with USRP-B210 via the data cable.
– Configures default parameters:
– Sampling rate: 40 MS/s
– Center frequency: 5.745 GHz (all drones manually set to Channel 149 for consistency across OcuSync 2.0/3.0 variants)
– Gain: 30 dB
– Bandwidth: 20 MHz
– Collects 140 million samples per file, labels files following naming rules, and verifies data integrity.
– Signals Operator A to land the drone after completion, then repeats for other drones.
– **Schematic**:
—
## 4. Data Visualization
### Plotting Amplitude and Time-Frequency Diagrams
To visualize amplitude plots and spectrograms, modify the file path in `plot.py`:
“`matlab
data = load(‘./mini_se_3_0.mat’);
“`
### Mini SE 3
– **I/Q Amplitude Graph and Time-Frequency Graph of data acquired under RF signal wave-absorbing cotton wrapping**
—
import uhd
import numpy as np
from scipy.io import savemat
from datetime import datetime
import time
import os
# ================= 硬件配置参数 =================
center_freq = 5745e6 # 5.745 GHz
sample_rate = 40e6 # 采样率
gain = 30 # 接收增益
rx_antenna = “TX/RX” # B210天线端口
# ================= 采集参数 =====================
total_samples = int(1.4e8) # 1.2亿样本
buffer_size = 4096 # 接收缓冲区大小
# ================= 设备初始化 ===================
def init_usrp():
usrp = uhd.usrp.MultiUSRP(“type=b200”)
usrp.set_rx_rate(sample_rate, 0)
usrp.set_rx_freq(uhd.libpyuhd.types.tune_request(center_freq), 0)
usrp.set_rx_gain(gain, 0)
usrp.set_rx_antenna(rx_antenna, 0)
stream_args = uhd.usrp.StreamArgs(“fc32”, “sc16”)
stream_args.channels = [0]
return usrp, usrp.get_rx_stream(stream_args)
# ================= 主采集流程 ====================
def main():
# 初始化设备
usrp, rx_stream = init_usrp()
save_dir = r”E:\dataset”
os.makedirs(save_dir, exist_ok=True)
full_i = np.zeros(total_samples, dtype=np.float32)
full_q = np.zeros(total_samples, dtype=np.float32)
recv_buffer = np.zeros((1, buffer_size), dtype=np.complex64)
metadata = uhd.types.RXMetadata()
rx_stream.issue_stream_cmd(
uhd.types.StreamCMD(uhd.types.StreamMode.start_cont))
print(f”开始采集，目标样本数：{total_samples}”)
try:
collected = 0
write_index = 0
while collected < total_samples:
num_samps = rx_stream.recv(recv_buffer, metadata)
if num_samps == 0: continue
remaining = total_samples – collected
valid_samps = min(num_samps, remaining)
samples = recv_buffer[0, :valid_samps]
end_index = write_index + valid_samps
full_i[write_index:end_index] = np.real(samples).astype(np.float32)
full_q[write_index:end_index] = np.imag(samples).astype(np.float32)
collected += valid_samps
write_index = end_index
print(f”\r进度：{collected / total_samples * 100:.2f}%”, end=””)
except KeyboardInterrupt:
print(“\n用户中断采集…”)
finally:
rx_stream.issue_stream_cmd(
uhd.types.StreamCMD(uhd.types.StreamMode.stop_cont))
timestamp = datetime.now().strftime(“%Y%m%d_%H%M%S”)
filename = os.path.join(save_dir, f”DRONE_RF_FULL_{timestamp}.mat”)
savemat(filename, {
“RF0_I”: full_i[:collected],
“RF0_Q”: full_q[:collected],
“Fs”: sample_rate,
“CenterFreq”: center_freq
})
print(f”\n已保存文件：{filename} (样本数：{collected})”)
if __name__ == “__main__”:
main()
