Successful tracking of hypersonic flight by U.S. HBTSS satellites is a big deal

Why successful tracking of hypersonic flight by U.S. HBTSS satellites is a crucial development ?

According to the Missile Defense Agency., the Defense Department’s advanced missile tracking satellites logged their first views of a hypersonic flight test. Hypersonic and Ballistic Tracking Space Sensor (HBTSS) satellites were able to track a hypersonic system. MDA didn’t disclose the date of the flight, which took off from Wallops Island in Virginia. The agency said in a June 14 statement, “Initial reports show the sensors successfully collected data after launch. MDA will continue to assess flight data over the next several weeks.” 

Viewers may note, that the Missile Defense Agency (MDA) and Space Development Agency (SDA) are actively working on components of a hypersonic missile defense system to protect against hypersonic weapons and other emerging missile threats. This includes developing the tracking and transport layers of the Proliferated Warfighter Space Architecture (PWSA) and various interceptor programs. In this video, Defense Updates analyzes why successful tracking of hypersonic flight by US HBTSS satellites is a crucial development ?

 #defenseupdates #hypersonicmissile #hypersonicweapon 

Chapters: 00:11 INTRODUCTION 02:15 HYPERSONIC WEAPON 04:09 TRACKING HYPERSONIC THREATS 06:49 AID IN HYPERSONIC WEAPON TESTING  06:49 ANALYSIS

 Summary

Successful Tracking of Hypersonic Flight by US HBTSS Satellites: A Crucial Development 

The successful tracking of a hypersonic flight test by the US HBTSS satellites is a significant milestone in the development of hypersonic missile defense systems. Hypersonic weapons, capable of flying at speeds greater than Mach 5, pose a significant threat to modern air defenses due to their extreme speed and maneuverability. The US DoD has recognized the urgency of developing effective defenses against these threats, stating that existing terrestrial- and space-based sensor architectures are insufficient to detect and track hypersonic weapons. The HBTSS satellites are part of the tracking layer of the Proliferated Warfighter Space Architecture (PWSA), designed to provide global indications, warning, tracking, and targeting of advanced missile threats, including hypersonic missile systems. With the successful tracking of the hypersonic flight test, the US is one step closer to developing a comprehensive missile defense system against these emerging threats.

Revolution at Mach 10: NASA-Backed UVa Researchers Control Hypersonic ScramJets with Light

Dual-mode scramjet engine flowfield characteristics

Revolution at Mach 10: NASA-Backed Hypersonic Jets Poised to Transform Space Travel

Summary

This article discusses recent advancements in hypersonic jet technology for space travel, focusing on research conducted at the University of Virginia with NASA funding. Here are the key points:

1. Researchers are exploring hypersonic jets as a potential alternative to rocket-based space travel, aiming for aircraft-like spacecraft that can take off and land horizontally.

2. The study, published in Aerospace Science and Technology, demonstrates for the first time that airflow in supersonic combusting jet engines (scramjets) can be controlled using an optical sensor.

3. The team achieved adaptive control of a scramjet engine, which is a significant breakthrough for hypersonic propulsion.

4. The research builds on NASA's earlier X-43A and X-51 Waverider projects, which proved the concept of hypersonic flight but lacked reliable engine control.

5. Optical sensors could potentially replace traditional pressure sensors, offering faster response times and more comprehensive data about engine conditions.

6. The team used optical emission spectroscopy to analyze the flame within the engine, providing information about the engine's state that pressure sensors cannot capture.

7. The wind tunnel demonstration showed that engine control can be both predictive and adaptive, smoothly transitioning between scramjet and ramjet functioning.

8. This technology could lead to more efficient, safer, and reusable single-stage-to-orbit aircraft for space travel.

9. While more research is needed, the findings suggest that optical sensors could play a crucial role in controlling future hypersonic vehicles.

10. The ultimate goal is to develop aircraft-like space vehicles that combine cost-efficiency, safety, and reusability, potentially transforming space travel in the coming decades.

Based on the article, here are the researchers involved, their institutions, and related published work:

Researchers and Institution:
1. Christopher Goyne - Professor and Director of the UVA Aerospace Research Laboratory, University of Virginia School of Engineering and Applied Science
2. Chloe Dedic - Associate Professor, University of Virginia School of Engineering and Applied Science
3. Max Chern - Doctoral student, University of Virginia
4. Andrew Wanchek - Former graduate student, University of Virginia
5. Laurie Elkowitz - Doctoral student, University of Virginia
6. Robert Rockwell - Senior scientist, University of Virginia

All researchers are affiliated with the University of Virginia School of Engineering and Applied Science. The main research discussed in this article was published in the journal Aerospace Science and Technology in June 2024. The specific paper is:

Title: "Control of a dual-mode scramjet flow path utilizing optical emission spectroscopy"
Authors: Max Y. Chern, Andrew J. Wanchek, Laurie Elkowitz, Robert D. Rockwell, Chloe E. Dedic and Christopher P. Goyne
Publication Date: 18 April 2024
DOI: 10.1016/j.ast.2024.109144

The article mentions that this work was supported by a NASA ULI (University Leadership Initiative) grant led by Purdue University, suggesting collaboration with other institutions. This article describes an experimental study on controlling the shock train leading edge (STLE) location in a dual-mode scramjet (DMSJ) engine using optical emission spectroscopy (OES) sensors. Here are the key points:

1. The study demonstrates closed-loop control of STLE location using OES sensor feedback for the first time.

2. OES sensors measure light emitted by excited chemical species (OH* and C2*) in the combustor to estimate and control STLE location.

3. The research compares OES sensor feedback to traditional wall pressure sensors for STLE control.

4. Two control methods were tested and compared:
   - Proportional-Integral (PI) controller
   - Characteristic Model-Based All-Coefficient Adaptive Control (ACAC)

5. Experiments were conducted at the University of Virginia Supersonic Combustion Facility, simulating conditions for a Mach 5 flight vehicle.

6. OES sensors provided smoother response compared to discrete wall pressure measurements.

7. Both PI and ACAC controllers performed similarly under nominal conditions and with introduced plant changes.

8. ACAC tended to command the fuel valve more efficiently than the PI controller.

9. The study demonstrates the potential of using OES sensors for STLE control in scramjet engines, which could improve performance and prevent unstart events.

10. This research contributes to the development of hypersonic propulsion systems for military applications and reusable launch vehicles.

The article emphasizes the novelty of using OES sensors for STLE control and the first experimental demonstration of adaptive control using OES feedback in a scramjet combustor.

Sensor Comparison Metrics

Based on the article, the main metrics used to compare STLE control using OES sensors versus traditional pressure sensors were:

1. Smoothness of response: The article mentions that utilizing an OES sensor for feedback "provided a smoother response than when utilizing discrete wall-based pressure measurements that tended to discretize the estimated STLE location."

2. Bias in STLE location: When using pressure measurements (PRM) for feedback, a bias in STLE location was noticed. The algorithm favored locations near pressure measurements, causing a difference between PRM and OES sensor estimated STLE locations.

3. Controller performance: The study compared how the controllers (PI and ACAC) performed using OES feedback versus pressure sensor feedback. This included looking at their ability to maintain and adjust STLE location according to reference points.

4. Sampling rate: The article notes that one key advantage of OES sensors is "the faster sampling rate that an optical sensor can provide" compared to traditional pressure sensors.

5. Direct measurement: OES sensors were noted to provide "direct measurements of the combustion process itself rather than its effect," which is what pressure sensors measure.

6. Response to plant changes: The study examined how the control systems using different sensors responded to introduced changes, such as mass addition downstream of the combustor and reduced fuel valve response.

7. Fuel valve efficiency: The article mentions that when using OES sensors, the ACAC controller "tended to command the fuel valve more efficiently" compared to the PI controller.

While the article doesn't provide specific quantitative metrics, these qualitative comparisons were used to evaluate the performance of OES sensors against traditional pressure sensors for STLE control. 

Control Algorithm Comparison

Based on the information provided in the article, the comparison between the Proportional-Integral (PI) and Characteristic Model-Based All-Coefficient Adaptive Control (ACAC) controllers was primarily qualitative. However, we can infer some metrics that were likely used to compare their performance:

1. Response to nominal operation: The article states that "ACAC and PI controllers behave similarly both during nominal operation."

2. Response to plant changes: The controllers were compared when changes were introduced, such as "mass addition downstream of the combustor and a reduction of fuel valve response."

3. Fuel valve efficiency: The article mentions that "the ACAC controller tended to command the fuel valve more efficiently."

4. Ability to maintain/track STLE location: This was likely assessed based on how well each controller maintained the desired STLE location during different duty cycles.

5. Stability during transient processes: While not explicitly compared, this was mentioned as an advantage of the ACAC controller.

The article doesn't provide specific scoring methods for these metrics. The comparison appears to be based on observational analysis rather than a formal scoring system. The researchers likely used their "expert judgment" to assess the relative performance of the controllers based on these criteria.

sciencedirect.com

 

scitechdaily.com

Eric Williamson, University of Virginia School of Engineering and Applied Science

---

This is an artist’s depiction of a Hyper-X research vehicle under scramjet power in free-flight following separation from its booster rocket. New research into hypersonic jets may transform space travel by making scramjet engines more reliable and efficient, leading to aircraft-like spacecraft. Credit: NASA

Wind Tunnel Study Reveals Hypersonic Jet Engine Flow Can Be Controlled Optically

Researchers at the University of Virginia are exploring the potential of hypersonic jets for space travel, using innovations in engine control and sensing techniques. The work, supported by NASA, aims to enhance scramjet performance through adaptive control systems and optical sensors, potentially leading to safer, more efficient space access vehicles that function like aircraft.

The Future of Space Travel: Hypersonic Jets

What if the future of space travel were to look less like Space-X’s rocket-based Starship and more like NASA’s “Hyper-X,” the hypersonic jet plane that, 20 years ago this year, flew faster than any other aircraft before or since?

In 2004, NASA’s final X-43A unmanned prototype tests were a milestone in the latest era of jet development — the leap from ramjets to faster, more efficient scramjets. The last test, in November of that year, clocked a world-record speed only a rocket could have achieved previously: Mach 10. The speed equates to 10 times the speed of sound.

NASA culled a lot of useful data from the tests, as did the Air Force six years later in similar tests on the X-51 Waverider, before the prototypes careened into the ocean.

Although hypersonic proof of concept was successful, the technology was far from operational. The challenge was achieving engine control, because the tech was based on decades-old sensor approaches.

NASA’s B-52B launch aircraft cruises to a test range over the Pacific Ocean carrying the third and final X-43A vehicle, attached to a Pegasus rocket, on November 16, 2004. Credit: NASA / Carla Thomas

Breakthroughs in Hypersonic Engine Control

This month, however, brought some hope for potential successors to the X-plane series.

As part of a new NASA-funded study, University of Virginia School of Engineering and Applied Science researchers published data in the June issue of the journal Aerospace Science and Technology that showed for the first time that airflow in supersonic combusting jet engines can be controlled by an optical sensor. The finding could lead to more efficient stabilization of hypersonic jet aircraft.

In addition, the researchers achieved adaptive control of a scramjet engine, representing another first for hypersonic propulsion. Adaptive engine control systems respond to changes in dynamics to keep the system’s overall performance optimal.

“One of our national aerospace priorities since the 1960s has been to build single-stage-to-orbit aircraft that fly into space from horizontal takeoff like a traditional aircraft and land on the ground like a traditional aircraft,” said professor Christopher Goyne, director of the UVA Aerospace Research Laboratory, where the research took place.

“Currently, the most state-of-the-art craft is the SpaceX Starship. It has two stages, with vertical launch and landing. But to optimize safety, convenience, and reusability, the aerospace community would like to build something more like a 737.”

Doctoral student Max Chern takes a closer look at the wind tunnel setup where University of Virginia School of Engineering and Applied Science researchers demonstrated that control of a dual-mode scramjet engine is possible with an optical sensor. Credit: Wende Whitman, UVA Engineering

Goyne and his co-investigator, Chloe Dedic, a UVA Engineering associate professor, believe optical sensors could be a big part of the control equation.

“It seemed logical to us that if an aircraft operates at hypersonic speeds of Mach 5 and higher, that it might be preferable to embed sensors that work closer to the speed of light than the speed of sound,” Goyne said.

Additional members of the team were doctoral student Max Chern, who served as the paper’s first author, as well as former graduate student Andrew Wanchek, doctoral student Laurie Elkowitz and UVA senior scientist Robert Rockwell. The work was supported by a NASA ULI grant led by Purdue University.

Enhancing Scramjet Engine Performance

NASA has long sought to prevent something that can occur in scramjet engines called “unstart.” The term indicates a sudden change in airflow. The name derives from a specialized testing facility called a supersonic wind tunnel, where a “start” means the wind has reached the desired supersonic conditions.

UVA has several supersonic wind tunnels, including the UVA Supersonic Combustion Facility, which can simulate engine conditions for a hypersonic vehicle traveling at five times the speed of sound.

“We can run test conditions for hours, allowing us to experiment with new flow sensors and control approaches on a realistic engine geometry,” Dedic said.

Goyne explained that “scramjets,” short for supersonic combustion ramjets, build on ramjet technology that has been in common use for years.

This computational fluid dynamics image from the original Hyper-X tests shows the engine operating at Mach 7. Credit: NASA

Ramjets essentially “ram” air into the engine using the forward motion of the aircraft to generate the temperatures and pressures needed to burn fuel. They operate in a range of about Mach 3 to Mach 6. As the inlet at the front of the craft narrows, the internal air velocity slows down to subsonic speeds in a ramjet combustion engine. The plane itself, however, does not.

Scramjets are a little different, though. While they are also “air-breathing” and have the same basic setup, they need to maintain that super-fast airflow through the engine to reach hypersonic speeds.

“If something happens within the hypersonic engine, and subsonic conditions are suddenly created, it’s an unstart,” Goyne said. “Thrust will suddenly decrease, and it may be difficult at that point to restart the inlet.”

Testing a Dual-Mode Scramjet Engine

Currently, like ramjets, scramjet engines need a step-up to get them to a speed where they can intake enough oxygen to operate. That may include a ride attached to the underside of a carrier aircraft as well as a rocket boost.

The latest innovation is a dual-mode scramjet combustor, which was the type of engine the UVA-led project tested. The dual engine starts in ramjet mode at lower Mach numbers, then shifts into receiving full supersonic airflow in the combustion chamber at speeds exceeding Mach 5.

Preventing unstart as the engine makes that transition is crucial.

Christopher Goyne, professor and director of the UVA Aerospace Research Laboratory, and Chloe Dedic, associate professor. Credit: Wende Whitman, UVA Engineering

Incoming wind interacts with the inlet walls in the form of a series of shock waves known as a “shock train.” Traditionally, the leading edge of those waves, which can be destructive to the aircraft’s integrity, have been controlled by pressure sensors. The machine can adjust, for example, by relocating the position of the shock train.

But where the leading edge of the shock train resides can change quickly if flight disturbances alter mid-air dynamics. The shock train can pressurize the inlet, creating the conditions for unstart.

So, “If you are sensing at the speed of sound, yet the engine processes are moving faster than the speed of sound, you don’t have very much response time,” Goyne said.

He and his collaborators wondered if a pending unstart could be predicted by observing properties of the engine’s flame instead.

Sensing the Spectrum of a Flame

The team decided to use an optical emission spectroscopy sensor for the feedback needed to control the shock train leading edge.

No longer limited to information obtained at the engine’s walls, as pressure sensors are, the optical sensor can identify subtle changes both inside the engine and within the flow path. The tool analyzes the amount of light emitted by a source — in this case, the reacting gases within the scramjet combustor — as well as other factors, such as the flame’s location and spectral content.

“The light emitted by the flame within the engine is due to relaxation of molecular species that are excited during combustion processes,” explained Elkowitz, one of the doctoral students. “Different species emit light at different energies, or colors, offering new information about the engine’s state that is not captured by pressure sensors.”

Current UVA Engineering mechanical and aerospace doctoral students Laurie Elkowitz and Max Chern were among the influential members of the team. Credit: Wende Whitman, UVA Engineering

The team’s wind tunnel demonstration showed that the engine control can be both predictive and adaptive, smoothly transitioning between scramjet and ramjet functioning.

The wind tunnel test, in fact, was the world’s first proof that adaptive control in these types of dual-function engines can be achieved with optical sensors.

“We were very excited to demonstrate the role optical sensors may play in the control of future hypersonic vehicles,” first author Chern said. “We are continuing to test sensor configurations as we work toward a prototype that optimizes package volume and weight for flight environments.”

Building Toward the Future

While much more work remains to be done, optical sensors may be a component of the future Goyne believes will be realized in his lifetime: plane-like travel to space and back.

Dual-mode scramjets would still require a boost of some sort to get the aircraft to at least Mach 4. But there would be the additional safety of not relying exclusively on rocket technology, which requires highly flammable fuel to be carried alongside large amounts of chemical oxidizer to combust the fuel.

That decreased weight would allow more room for passengers and payload.

Such an all-in-one aircraft, which would glide back to Earth like the space shuttles once did, might even provide the ideal combination of cost-efficiency, safety and reusability.

“I think it’s possible, yeah,” Goyne said. “While the commercial space industry has been able to lower costs through some reusability, they haven’t yet captured the aircraft-like operations. Our findings could potentially build on the storied history of Hyper-X and make its space access safer than current rocket-based technology.”

Reference: “Control of a dual-mode scramjet flow path utilizing optical emission spectroscopy” by Max Y. Chern, Andrew J. Wanchek, Laurie Elkowitz, Robert D. Rockwell, Chloe E. Dedic and Christopher P. Goyne, 18 April 2024, Aerospace Science and Technology.
DOI: 10.1016/j.ast.2024.109144

 

Control of a dual-mode scramjet flow path utilizing optical emission spectroscopy

J. Peeters

---

https://doi.org/10.1016/j.ast.2024.109144

Get rights and content

Abstract

Shock train leading edge (STLE) location control within a Dual-Mode Scramjet (DMSJ) flow path was demonstrated using an optical emission spectroscopy (OES) sensor for control feedback. Emission from electronically excited chemical species, OH^⁎ and C${}_2^{⁎}$, was observed within the combustor and used for feedback to control the STLE within the DMSJ isolator. An optical emission sensor was used to experimentally demonstrate STLE control using a Proportional-Integral (PI) controller. Feedback using this sensor was compared the traditional approach of using wall pressure sensors. Utilizing an OES sensor for feedback proved to be an effective method to estimate and control the STLE location and provided a smoother response than when utilizing discrete wall-based pressure measurements that tended to discretize the estimated STLE location. Characteristic Model-Based All-Coefficient Adaptive Control (ACAC) was also implemented and compared to the PI controller response using the OES signal as feedback. Plant changes were introduced in the form of mass addition downstream of the combustor and a reduction of fuel valve response. Experiments showed that the ACAC and PI controllers behave similarly both during nominal operation and with plant changes, though the ACAC controller tended to command the fuel valve more efficiently. This paper presents the first experimental demonstration of closed-loop control of the STLE location within a scramjet flow path when using OES sensor feedback. This is also the first demonstration of adaptive control of STLE location utilizing OES sensor feedback.

Access through your organization

Check access to the full text by signing in through your organization.

Access through your institution

Introduction

Scramjet technology has the potential to propel hypersonic military systems and reusable launch vehicles for responsive space access. Concepts such as the NASA X-43A have been developed to demonstrate the practical use of scramjets [1], [2]. A special type of scramjet, known as a dual-mode scramjet (DMSJ), makes ample use of an isolator upstream of the combustor. The isolator reduces the incoming flow to subsonic conditions via a shock train. This extends the operational range of the engine by allowing it to operate in a ramjet mode. Therefore, between about Mach 3 and 6, a DMSJ typically operates with a precombustion shock train in the isolator and with subsonic fuel-air mixing and combustion in the combustor [3].

The isolator shock train works to match the scramjet inlet pressure with the higher downstream combustor pressure. The location of the shock train leading edge (STLE) is indicative of engine conditions and is sensitive to inlet, and combustor conditions [4]. As the combustor pressure increases relative to the inlet pressure, the shock train increases in length. The STLE can travel forward into the scramjet inlet, causing subsonic flow in the inlet, during an event called unstart. Unstart is detrimental to a flight vehicle as it greatly reduces mass flow into the engine, which in turn, reduces thrust and increases drag [5]. Unstart prevention is required for hypersonic air-breathing engines as an unexpected unstart event could lead to the loss of the vehicle. A common, passive method to reduce the risk of unstart is to increase the length of the isolator, but this also increases the overall engine length and weight. Alternatively, active control of the STLE offers another approach for maximizing the operational envelope and the performance of the engine.

Many different active control methods for preventing unstart have been proposed including boundary layer suction [6], mass addition via vortex generator jets [7], plasma actuation [8], back pressure regulation via ramp actuation downstream of the combustor [9], [10], and active fueling control [11], [12], [13], [14]. An advantage to using active fueling control for unstart prevention is that it can largely use the existing fuel system, reducing complexity of the hardware design and fabrication. Controlling fuel as a method of controlling STLE location is possible because of the pressure balance taking place within the engine. As the fueling rate increases, increased heat release in the combustor leads to an increase in combustor pressure, which moves the STLE forward in the isolator. The opposite occurs when the fueling rate is decreased. Coupling of the combustion reaction with the overall engine state and STLE location enables the utilization of combustor sensors as feedback for STLE estimation and control.

Pressure sensors are typically used to measure combustor state, but recent studies have shown benefits of using optical sensors [14], [15]. Optical emission spectroscopy (OES) provides a spectrally-resolved measurement of the light emitted following the formation of electronically-excited species in a chemical reaction. During ethylene combustion, or any hydrocarbon combustion process, electronically excited OH^⁎, CH^⁎ and C${}_2^{⁎}$ are formed as intermediate species [16]. The relative intensities of light emitted from these species have been shown to be correlated with local fueling equivalence ratio (ER) in partially premixed flames [17]. Such an established correlation can be used to control ER or STLE location in a scramjet, through closed-loop control approaches. Fig. 1 visualizes how OES measurements may be used to control the STLE in a scramjet in such a control loop. It depicts two different measured emission spectra that correspond to two different STLE locations, at a different equivalence ratios in the scramjet combustor. As the ER increases, the spectra measured from the flame chemiluminescence changes. This corresponds to a change in shock train length and STLE location. The key advantages of using OES for sensing of engine parameters is the ability to obtain direct measurements of the combustion process itself rather than its effect, i.e. isolator pressure rise, as well as the faster sampling rate that an optical sensor can provide. Previous work has applied this concept to DMSJ control by using the ratio of integrated C${}_2^{⁎}$ and OH^⁎ emission intensities to control fuel pressure and, hence, combustor global ER [14]. Although Refs. [14] and [15] establish a link between the OES signal at a specific location and global ER, and STLE control using OES has been proposed in previous work. [18], the concept has not yet been experimentally demonstrated.

If inlet conditions of a scramjet remain constant, a correlation between the combustor state, as estimated by an OES sensor, and STLE location may be established. However, this is not always the case during flight. Many factors can cause inlet conditions to change during flight, such as changes in altitude, weather [19], and Angle-of-Attack (AoA) of the vehicle [20], [21]. These factors will cause the STLE behavior to deviate from the baseline condition and may lead to unfavorable controller responses. Other changes, such as hardware malfunctions (i.e. a reduction in valve response), engine geometry change due to thermal expansion, or changes in the optical transmission of windows will likewise cause a change in the plant of the control system. Ultimately, an engine control system on a flight vehicle may use multiple sensor types for feedback. The goal of this work is to demonstrate that the relationship between OES sensor measurements and STLE location is sufficient for closed loop control of the STLE and to explore the advantages of using OES sensors over traditional pressure measurement.

An adaptive controller may be required to mitigate issues that arise from uncertainties in scramjet operability and performance of a real flight system. Therefore, a Model Reference Adaptive Controller (MRAC) was examined in this study; an approach called “Characteristic Model-Based All-Coefficient Adaptive Control (ACAC).” [22] MRAC is typically deployed on plants where uncertainties can be in the form of degradation, uncertain plant dynamics, or other unmodeled phenomenon [23]. Direct MRAC utilizes adaptive parameters to adjust controller gains and provide a defined system performance, whereas indirect MRAC utilizes adaptive parameters to model the system dynamics, and are used directly in the control law. The form of MRAC used in this work, ACAC, is an indirect method developed by Wu and Xie [24] and has been applied to many engineering plants [22], including the control of a high-speed rotor supported on magnetic bearings [25] and simulated altitude control for the X-34 launch vehicle [26]. ACAC was selected for use in the current study due to its relatively simple implementation and its stability during transient processes [24]. ACAC utilizes a plant estimation method known as characteristic modeling to provide a real-time estimation of the plant response. This estimation is then implemented directly into the golden section adaptive control law to provide a robustly stable system, even during transient processes [24].

Other control laws can be implemented to aid in the performance of the ACAC controller while taking advantage of its stability properties. Maintaining/tracking, derivative, and integral control laws have been employed with ACAC to provide good closed-loop performance for engineering systems [22]. Although ACAC has been around since the 1990's, there is little experimental results published using the controller. This study provides an experimental analysis of this controller when compared to the more standard proportional-integral (PI) controller. Due to the uncertain and non-linear nature of DMSJ operation and link to combustor emission, ACAC is a good candidate for the control of the DMSJ flowpath using the OES sensors as feedback due to its adaptive qualities.

This study builds on the work of Ref. [18] where the possibility of utilizing OES for STLE control was explored through simulation. The present study has two objectives: (1) demonstrate experimental closed-loop control of the STLE location in a scramjet isolator flow path utilizing OES feedback, and (2) explore adaptive control in comparison to a standard PI control method for this system. This is the first demonstration of closed-loop control of the STLE location within a scramjet isolator flow path utilizing OES sensor feedback. This is also the first time adaptive control was experimentally demonstrated utilizing OES sensor feedback from a scramjet combustor.

In this paper, the experimental facility setup is described in two parts: subsection 2.1 provides details of the facility conditions, instrumentation, actuators, and control hardware, and subsection 2.2 provides details on the method of optical emission sensing and its synchronization with the control system. The control setup is then described in four parts: subsection 3.1 explains the method by which pressure measurements within the isolator were used to estimate the STLE, subsection 3.2 explains how these estimations were used in conjunction with OES measurements to provide calibrations for the control loop, subsection 3.3 explains the closed loop control loop that was implemented, and finally, subsection 3.4 explains the control law of the ACAC controller. The results and discussion section begins with a brief outline of the test conditions and duty cycles that were examined and then the results are examined in two parts: subsection 4.1, which discusses the performance of the controller using OES feedback in comparison to pressure sensors, and subsection 4.2, which discusses the performance of the ACAC controller. Finally, a conclusion summarizes the results and impacts of this work.

Section snippets

Facility and instrumentation

The experiments in this paper were conducted at the University of Virginia Supersonic Combustion Facility (UVASCF). The facility consists of an electrically heated, continuous-flow, direct-connect wind tunnel which was used to provide air at 1200 K and 300 kPa total conditions to a DMSJ flow path through a Mach 2 nozzle. This condition simulates the engine inflow Mach number and enthalpy of a flight vehicle at approximately Mach 5 [27], [28]. The facility conditions are measured and recorded on

Control setup

This section details the methods by which closed loop STLE control was achieved. This begins with a description of how the STLE was detected, followed by the method by which the OES signal was calibrated to the STLE. Details are then provided on how the control loop was implemented, followed by the control law of the ACAC controller.

Results and discussion

In order to fully examine control effectiveness, two duty cycles with step changes of reference STLE location was exercised with each controller: the “Constant Step Size” duty cycle with steps of ±5 x/H between -20 and -35 x/H, and the “Differing Step Size” duty cycle with steps of different sizes between -20 and -35 x/H. The time between the step changes was 3 seconds to allow the STLE location to stabilize at each location. Each duty cycle was repeated at least twice for each controller at

Conclusion

The control of the STLE location using OES sensor feedback was experimentally demonstrated and was shown to be a viable approach. Utilizing the OES sensor as feedback provided a comparable response to using the PRM as feedback. Bias in STLE location was noticed during control with the PRM, though, as the algorithm favored locations near pressure measurements. This bias caused a difference between PRM and OES sensor estimated STLE location. Introducing flow imaging of the isolator flow path was

CRediT authorship contribution statement

Max Y. Chern: Data curation, Investigation, Writing – original draft, Writing – review & editing. Andrew J. Wanchek: Writing – original draft, Software, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Laurie Elkowitz: Methodology, Formal analysis, Data curation, Validation, Writing – original draft. Robert D. Rockwell: Data curation, Investigation, Supervision, Validation. Chloe E. Dedic: Methodology, Investigation, Funding acquisition, Conceptualization, Project

Declaration of Competing Interest

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Andrew Wanchek reports financial support was provided by NASA Aeronautics Research Mission Directorate. Max Chern reports financial support was provided by NASA Aeronautics Research Mission Directorate. Laurie Elkowitz reports financial support was provided by National Defense Science and Engineering Graduate Fellowship. If there are other authors, they declare

Acknowledgements

The authors would like to thank Jack Donnellan and Joe Fritch from GE Aerospace for their support. This work was financially supported by NASA's Space Technology Research Grants Program (NASA ULI Grant #80NSSC21M0069 P00001). L. Elkowitz was supported by the National Defense Science and Engineering Graduate (NDSEG) Fellowship.

References (35)

• et al.

  Mechanisms of C₂* and CH* formation in a hydrogen-oxygen flame containing hydrocarbon traces

  Symp. (Int.) Combust.

  (1971)

• N. Li

  Response of shock train to fluctuating angle of attack in a scramjet inlet-isolator

  Acta Astronaut.

  (2022)

• K. Xu et al.

  Mechanism of shock train rapid motion induced by variation of attack angle

  Acta Astronaut.

  (2017)

• L. Marshall et al.

  A chief engineer's view of the NASA X-43A scramjet flight test

• T.P. Jones et al.

  Evaluation of the X-43A scramjet engine controller performance by Monte Carlo technique

• C. Segal

  The Scramjet Engine: Processes and Characteristics

  (2011)

• D.B. Le et al.

  Shock train leading-edge detection in a dual-mode scramjet

  (2008)

• J.L. Wagner et al.

  Experimental investigation of unstart in an inlet/isolator model in Mach 5 flow

  AIAA J.

  (2009)

• K. Kang et al.

  Fast-acting boundary-layer suction control of unstarting flows in an ethylene-fueled dual-mode scramjet

  AIAA J.

  (2021)

• A. Valdivia et al.

  Active control of supersonic inlet unstart using vortex generator jets

S. Im et al.

The manipulation of an unstarting supersonic flow by plasma actuator

J. Phys. D, Appl. Phys.

(2012)

L. Vanstone et al.

Closed-loop control of unstart in a Mach 1.8 isolator

(2020)

J.R. Hutzel

Scramjet Isolator Modeling and Control Dissertation

(2011)

R. Rockwell et al.

Simulated shock train control using an all-coefficient adaptive control approach

J.M. Donbar

Shock train position control in an axisymmetric scramjet combustor flowpath

A. Goel et al.

Retrospective cost adaptive control of unstart in a model scramjet combustor

AIAA J.

(2018)

L. Elkowitz et al.

Dual-mode scramjet control using optical emission sensors

Appl. Opt.

(2024)

Cited by (0)

¹

Department of Mechanical and Aerospace Engineering.

View full text

© 2024 Elsevier Masson SAS. All rights reserved.

 

Robust Consumer Drone RF Link Detection and Classification in Low SNR Environments

Robust Low-Cost Drone Detection and Classification in Low SNR Environments

Electrical Engineering and Systems Science > Signal Processing

[Submitted on 26 Jun 2024]

Stefan Glüge, Matthias Nyfeler, Ahmad Aghaebrahimian, Nicola Ramagnano, Christof Schüpbach

The proliferation of drones, or unmanned aerial vehicles (UAVs), has raised significant safety concerns due to their potential misuse in activities such as espionage, smuggling, and infrastructure disruption. This paper addresses the critical need for effective drone detection and classification systems that operate independently of UAV cooperation.

We evaluate various convolutional neural networks (CNNs) for their ability to detect and classify drones using spectrogram data derived from consecutive Fourier transforms of signal components. The focus is on model robustness in low signal-to-noise ratio (SNR) environments, which is critical for real-world applications.

A comprehensive dataset is provided to support future model development. In addition, we demonstrate a low-cost drone detection system using a standard computer, software-defined radio (SDR) and antenna, validated through real-world field testing. On our development dataset, all models consistently achieved an average balanced classification accuracy of >= 85% at SNR > -12dB.

In the field test, these models achieved an average balance accuracy of > 80%, depending on transmitter distance and antenna direction. Our contributions include: a publicly available dataset for model development, a comparative analysis of CNN for drone detection under low SNR conditions, and the deployment and field evaluation of a practical, low-cost detection system.

Comments:	11 pages, submitted to IEEE Open Journal of Signal Processing
Subjects:	Signal Processing (eess.SP); Machine Learning (cs.LG)
Cite as:	arXiv:2406.18624 [eess.SP]
	(or arXiv:2406.18624v1 [eess.SP] for this version)

Submission history

From: Stefan Glüge [view email]
[v1] Wed, 26 Jun 2024 12:50:55 UTC (16,275 KB)

Summary

The paper focuses on radio frequency (RF) signals from several types of consumer and hobbyist drones. Specifically:

• 1. The development dataset included signals from 6 drones and 4 remote controllers:
•    - DJI Phantom 4 Pro drone and its remote control
•    - Futaba T7C remote control
•    - Futaba T14SG remote control and R7008SB receiver
•    - Graupner mx-16 remote control and GR-16 receiver
•    - Taranis ACCST X8R Receiver
•    - Turnigy 9X remote control
• 2. The signals were recorded in the 2.4 GHz ISM band, which is commonly used by consumer drones for communication between the drone and its remote control.
• 3. The paper mentions that these signals occur in short bursts, typically 1.3-2 ms long, with repetition periods ranging from about 60 to 600 ms depending on the specific drone model.
• 4. The signals were recorded at a sampling frequency of 56 MHz initially, then downsampled to 14 MHz for processing.
• 5. For the field test, they used a slightly different set of drones/controllers, including:
•    - DJI Phantom Pro 4 drone and remote
•    - Futaba T14 remote control
•    - Futaba T7 remote control
•    - FrySky Taranis Q X7 remote control
•    - Turnigy Evolution remote control

The focus was on the RF signals emitted by these consumer-grade drones and their remote controls, as detecting these signals can indicate the presence of a drone even when it's not visible. This paper presents research on detecting and classifying drones using radio frequency (RF) signals and convolutional neural networks (CNNs). Key points include:

• 1. The authors developed CNN models to detect and classify drones using spectrogram data from RF signals.
• 2. They focused on model robustness in low signal-to-noise ratio (SNR) environments, which is important for real-world applications.
• 3. A comprehensive dataset was created and made publicly available to support future research.
• 4. The authors implemented a low-cost drone detection system using standard computer hardware, software-defined radio, and an antenna.
• 5. In lab tests, the models achieved ≥85% balanced accuracy at SNRs above -12 dB.
• 6. Field tests showed >80% average balanced accuracy, varying based on transmitter distance and antenna direction.
• 7. The simplest model (VGG11 BN) performed as well as more complex models.
• 8. Most misclassifications occurred between noise and drone signals, rather than between different drone types.
• 9. The system could reliably detect drones up to 670 meters away in real-world conditions.
• 10. Limitations included potential interference in field tests and the inability to detect multiple imultaneous transmitters.

The research demonstrates the feasibility of using CNNs for drone detection with RF signals in challenging real-world conditions, while also providing resources for further research in this area.

The paper primarily focused on using variations of the Visual Geometry Group (VGG) CNN architecture. Specifically:

1. VGG11 BN
2. VGG13 BN
3. VGG16 BN
4. VGG19 BN

Key points about these models:

- The "BN" suffix indicates that these versions include batch normalization layers after the convolutions.

- The main idea of the VGG architecture is to use multiple layers of small (3x3) convolutional filters instead of larger ones.

- The number in each model name (11, 13, 16, 19) refers to the number of layers with weights in the network.

- For the dense classification layer, they used 256 linear units followed by 7 linear units at the output (one unit per class).

- The models were trained on 2D spectrogram data derived from the RF signals.

- Interestingly, the authors found no significant performance advantage in using the more complex models (like VGG19 BN) over the simplest model (VGG11 BN) for this specific task.

- The VGG11 BN model, being the least complex, required the least number of training epochs to achieve optimal performance on the validation set.

The authors chose these VGG variants due to their proven effectiveness in image classification tasks, adapting them to work with the 2D spectrogram representation of the RF signals. They focused on comparing the performance of these different VGG variants rather than exploring other types of CNN architectures.

• Corresponding author: Stefan Glüge (email: stefan.gluege@zhaw.ch).

  Robust Low-Cost Drone Detection and Classification in Low SNR Environments

  Stefan Glüge${}^{\text{1}}$ Matthias Nyfeler${}^{\text{1}}$ Ahmad Aghaebrahimian${}^{\text{1}}$
  Nicola Ramagnano${}^{\text{2}}$ and Christof Schüpbach${}^{\text{3}}$ Fellow IEEE Institute of Computational Life Sciences, Zurich University of Applied Sciences, 8820 Wädenswil, Switzerland Institute for Communication Systems, Eastern Switzerland University of Applied Sciences, 8640 Rapperswil-Jona, Switzerland Armasuisse Science + Technology, 3602 Thun, Switzerland

  Abstract

  The proliferation of drones, or unmanned aerial vehicles (UAVs), has raised significant safety concerns due to their potential misuse in activities such as espionage, smuggling, and infrastructure disruption. This paper addresses the critical need for effective drone detection and classification systems that operate independently of UAV cooperation. We evaluate various convolutional neural networks (CNNs) for their ability to detect and classify drones using spectrogram data derived from consecutive Fourier transforms of signal components. The focus is on model robustness in low signal-to-noise ratio (SNR) environments, which is critical for real-world applications. A comprehensive dataset is provided to support future model development. In addition, we demonstrate a low-cost drone detection system using a standard computer, software-defined radio (SDR) and antenna, validated through real-world field testing. On our development dataset, all models consistently achieved an average balanced classification accuracy of $\ge 85\%$ at SNR $>-12$ dB. In the field test, these models achieved an average balance accuracy of $>80\%$, depending on transmitter distance and antenna direction. Our contributions include: a publicly available dataset for model development, a comparative analysis of CNN for drone detection under low SNR conditions, and the deployment and field evaluation of a practical, low-cost detection system.

  {IEEEkeywords}

  Deep neural networks, Robustness, Signal detection, Unmanned aerial vehicles

  1 INTRODUCTION

  \IEEEPARstart

  Drones, or civil UAVs, have evolved from hobby toys to commercial systems with many applications. In particular, mini/amateur drones have become ubiquitous. With the proliferation of these low-cost, small and easy-to-fly drones, safety issues have became more pressing (e.g. spying, transfer of illegal or dangerous goods, disruption of infrastructure, assault). Although regulations and technical solutions (such as transponder systems) are in place to safely integrate UAVs into the airspace, detection and classification systems that do not rely on the cooperation of the UAV are necessary. Various technologies such as audio, video, radar, or radio frequency (RF) scanners have been proposed for this task [1].

  In this paper, we evaluate different CNNs for drone detection and classification using the spectrogram data computed with consecutive Fourier transforms for the real and imaginary parts of the signal. To facilitate future model development, we make the dataset publicly available. In terms of performance, we focus on the robustness of the models to low SNRs, as this is the most relevant aspect for a real-world application of the system. Furthermore, we evaluate a low-cost drone detection system consisting of a standard computer, SDR, and antenna in a real-world field test.

  Our contributions can therefore be summarised as follows:

  • • We provide the dataset used to develop the model. Together with the code to load and transform the data, it can be easily used for future model development.
  • • We compare different CNNs using $2$D spectrogram data for detection and classification of drones based on their RF signals under challenging conditions, i.e. low SNRs down to $-20$ dB.
  • • We visualise the model embeddings to understand how the model clusters and separates different classes, to identify potential overlaps or ambiguities, and to examine the hierarchical relationships within the learned features.
  • • We implement the models in a low-cost detection system and evaluate them in a field test.

  1.1 RELATED WORK

  A literature review on drone detection methods based on deep learning (DL) is given in [1] and [2]. Both works reflect the state of the art in $2024$. Different DL algorithms are discussed with respect to the techniques used to detect drones based on visual, radar, acoustic, and RF signals. Given these general overviews, we briefly summarise recent work based on RF data, with a particular focus on the data side of the problem to motivate our work.

  With the advent of DL-based methods, the data used to train models became the cornerstone of any detection system. Table 1 provides an overview of openly available datasets of RF drone signals. The DroneRF dataset [3] is one of the first openly available datasets. It contains RF time series data from three drones in four flight modes (i.e. on, hovering, flying, video recording) recorded by two universal software radio peripheral (USRP) SDR transceivers [4]. The dataset is widely used and enabled follow-up work with different approaches to classification systems, i.e. DL-based [5, 6], focused on pre-processing and combining signals from two frequency bands [7], genetic algorithm-based heterogeneous integrated k-nearest neighbour [8], and hierarchical reinforcement learning-based [9]. In general, the classification accuracies reported in the papers on the DroneRF dataset are close to $100\%$. Specifically, [4], [5], and [6] report an average accuracy of $99.7\%$, $100\%$, and $99.98\%$, respectively, to detect the presence of a drone. There is therefore an obvious need for a harder, more realistic dataset.

  Consequently, [10] investigate the detection and classification of drones in the presence of Bluetooth and Wi-Fi signals. Their system used a multi-stage detector to distinguish drone signals from the background noise and interfering signals. Once a signal was identified as a drone signal, it was classified using machine learning (ML) techniques. The detection performance of the proposed system was evaluated for different SNRs. The corresponding recordings ($17$ drone controls from eight different manufacturers) are openly available [11]. Unfortunately, the Bluetooth/Wi-Fi noise is not part of the dataset. Ozturk et al. [12] used the dataset to further investigate the classification of RF fingerprints at low SNRs by adding white Gaussian noise to the raw data. Using a CNN, they achieved classification accuracies ranging from $92\%$ to $100\%$ for SNR $\in [-10,30]$dB.

  The openly available DroneDetect dataset [13] was created by Swinney and Woods [14]. It contains raw in-phase and quadrature (IQ) data recorded with a BladeRF SDR. Seven drone models were recorded in three different flight modes (on, hovering, flying). Measurements were also repeated with different types of noise, such as interference from a Bluetooth speaker, a Wi-Fi hotspot, and simultaneous Bluetooth and Wi-Fi interference. The dataset does not include measurements without drones, which would be necessary to evaluate a drone detection system. The results in [14] show that Bluetooth signals are more likely to interfere with detection and classification accuracy than Wi-Fi signals. Overall, frequency domain features extracted from a CNN were shown to be more robust than time domain features in the presence of interference. In [15] the drone signals from the DroneDetect dataset were augmented with Gaussian noise and SDR recorded background noise. Hence, the proposed approach could be evaluated regrading its capability to detect drones. They trained a CNN end-to-end on the raw IQ data and report an accuracy of $99\%$ for detection and between $72\%$ and $94\%$ for classification.

  The Cardinal RF dataset [16] consists of the raw time series data from six drones + controller, two Wi-Fi and two Bluetooth devices. Based on this dataset, Medaiyese et al. [17] proposed a semi-supervised framework for UAV detection using wavelet analysis. Accuracy between $86\%$ and $97\%$ was achieved at SNRs of $30$ dB and $18$ dB, while it dropped to chance level for SNRs below $10$ dB to $6$ dB. In addition, [18] investigated different wavelet transforms for the feature extraction from the RF signals. Using the wavelet scattering transform from the steady state of the RF signals at $30$ dB SNR to train SqueezeNet [19], they achieved an accuracy of $98.9\%$ at $10$ dB SNR.

  In our previous work [20], we created the noisy drone RF signals dataset¹¹https://www.kaggle.com/datasets/sgluege/noisy-drone-rf-signal-classification from six drones and four remote controllers. It consists of non-overlapping signal vectors of $16384$ samples, corresponding to $\approx 1.2$ ms at $14$ MHz. We added Labnoise (Bluetooth, Wi-Fi, Amplifier) and Gaussian noise to the dataset and mixed it with the drone signals with SNR $\in [-20,30]$ dB. Using IQ data and spectrogram data to train different CNNs, we found an advantage in favour of the $2$D spectrogram representation of the data. There was no performance difference at SNR $\ge 0$ dB but a major improvement in the balanced accuracy at low SNR levels, i.e. $84.2$% on the spectrogram data compared to $41.3$% on the IQ data at $-12$ dB SNR.

  Recently, [21] proposed an anchor-free object detector based on keypoints for drone RF signal spectograms. They also proposed an adversarial learning-based data adaptation method to generate domain independent and domain aligned features. Given five different types of drones, they report a mean average precision of $97.36\%$, which drops to $\approx 55\%$ when adding Gaussian noise with $-25$ dB SNR. The raw data used in their work is available²²https://www.kaggle.com/datasets/zhaoericry/drone-rf-dataset, but yet, unfortunately not usable without any further documentation.

  Table 1: Overview on openly available drone RF datasets.

  Dataset	Year	Datatype	UAV	Noise	Size
  DroneRF [3]	2019	Raw Amplitude	$3$ drones + $3$ controller	Background RF activities	$3.75$ GB
  Drone remote controller RF signal dataset [11]	2020	Raw Amplitude	$17$ controller	none	$124$ GB
  DroneDetect dataset [13]	2020	Raw IQ	$7$ drones + $7$ controller	Bluetooth, Wi-Fi devices	$66$ GB
  Cardinal RF [16]	2022	Raw Amplitude	$6$ drones + $6$ controller	Bluetooth, Wi-Fi	$65$ GB
  Noisy drone RF signals [20]	2023	Pre-processed IQ and Spectrogram	$6$ drones + $4$ controller	Bluetooth, Wi-Fi, Gauss	$23$ GB

  1.2 MOTIVATION

  As we have seen in other fields, such as computer vision, the success of DL can be attributed to: (a) high-capacity models; (b) increased computational power; and (c) the availability of large amounts of labelled data [22]. Thus, given the large amount of available raw RF signals (cf. Tab. 1) we promote the idea of open and reusable data, to facilitate model development and model comparison.

  With the noisy drone RF signals dataset [20], we have provided a first ready-to-use dataset to enable rapid model development, without the need for any data preparation. Furthermore, the dataset contains samples that can be considered as “hard” in terms of noise, i.e. Bluetooth + Wi-Fi + Gaussian noise at very low SNRs, and allows a direct comparison with the published results.

  While the models proposed in [20] performed reasonably well in the training/lab setting, we found it difficult to transfer their performance to practical application. The reason was the choice of rather short signal vectors of $16384$ samples, corresponding to $\approx 1.2$ ms at $14$ MHz. Since the drone signals occur in short bursts of $\approx 1.3\text{–}2$ ms with a repetition period of $\approx 60\text{–}600$ ms, our continuously running classifier predicts a drone whenever a burst occurs and noise during the repetition period of the signal. Therefore, in order to provide a stable and reliable classification per every second, one would need an additional “layer” to pool the classifier outputs given every $1.2$ ms.

  In the present work, we follow a data-centric approach and simply increase the length of the input signal to $\approx 75$ ms to train a classifier in an end-to-end manner. Again, we provide the data used for model development in the hope that it will inspire others to develop better models.

  In the next section, we briefly describe the data collection and preprocessing procedure. Section 3 describes the model architectures and their training/validation method. In addition, we describe the setup of a low-cost drone detection system and of the field test. The resulting performance metrics are presented in Section 4 and are further discussed in Section 5.

  2 MATERIALS

  We used the raw RF signals from the drones that were collected in [20]. Nevertheless, we briefly describe the data acquisition process again to provide a complete picture of the development from the raw RF signal to the deployment of a detection system within a single manuscript.

  2.1 DATA ACQUISITION

  The drone’s remote control and, if present, the drone itself were placed in an anechoic chamber to record the raw RF signal without interference for at least one minute. The signals were received by a log-periodic antenna and sampled and stored by an Ettus Research USRP B210, see Fig. 1. In the static measurement, the respective signals of the remote control (TX) alone or with the drone (RX) were measured. In the dynamic measurement, one person at a time was inside the anechoic chamber and operated the remote control (TX) to generate a signal that is as close to reality as possible. All signals were recorded at a sampling frequency of $56$ MHz (highest possible real-time bandwidth). All drone models and recording parameters are listed in Tab. 2, including both uplink and downlink signals.

  

  Figure 1: Recording of drone signals in the anechoic chamber. A DJI Phantom 4 Pro drone with the DJI Phantom GL300F remote control.

  Table 2: Transmitters and receivers recorded in the development dataset and their respective class labels. Additionally, we show the center frequency (GHz), the channel spacing (MHz), the burst duration (ms), and the repetition period of the respective signals (ms).

  Transmitter	Receiver	Label	Center Freq. (GHz)	Spacing (MHz)	Duration (ms)	Repetition (ms)
  DJI Phantom GL300F	DJI Phantom 4 Pro	DJI	$2.44175$	$1.7$	$2.18$	$630$
  Futaba T7C	-	FutabaT7	$2.44175$	$2$	$1.7$	$288$
  Futaba T14SG	Futaba R7008SB	FutabeT14	$2.44175$	$3.1$	$1.4$	$330$
  Graupner mx-16	Graupner GR-16	Graupner	$2.44175$	$1$	$1.9/3.7$	$750$
  Bluetooth/Wi-Fi Noise	-	Noise	$2.44175$
  Taranis ACCST	X8R Receiver	Taranis	$2.440$	$1.5$	$3.1/4.4$	$420$
  Turnigy 9X	-	Turnigy	$2.445$	$2$	$1.3$	$61$, $120$-$2900$ a
  a The repetition period of the Turnigy transmitter is not static. First bursts were observed after $61$ ms, the following signal bursts were observed in the interval $[120,2900]$ ms

  We also recorded three types of noise and interference. First, Bluetooth/Wi-Fi noise was recorded using the hardware setup described above. Measurements were taken in a public and busy university building. In this open recording setup, we had no control over the exact number or types of active Bluetooth/Wi-Fi devices and the actual traffic in progress.

  Second, artificial white Gaussian noise was used, and third, receiver noise was recorded for $30$ seconds from the USRP at various gain settings ($[30,70]$ db in steps of $10$ dB) without the antenna attached. This should prevent the final model from misclassifying quantisation noise in the absence of a signal, especially at low gain settings.

  2.2 DATA PREPARATION

  To reduce memory consumption and computational effort, we reduced the bandwidth of the signals by downsampling from $56$ MHz to $14$ MHz using the SciPy [23] signal.decimate function with an $8$th order Chebyshev type I filter.

  The drone signals occur in short bursts with some low power gain or background noise in between (cf. Tab. 2). We divided the signals into non-overlapping vectors of $1048576$ samples ($74.9$ ms) and only vectors containing a burst, or at least a partial burst, were used for the development dataset. This was achieved by applying an energy threshold. As the recordings were made in an echo-free chamber, the signal burst is always clearly visible. Hence, we only used vectors that contained a portion of the signal whose energy was above the threshold, which was arbitrarily set at $0.001$ of the average energy of the entire recording.

  The selected drone signal vectors $x$ with $i\in \{1,\mathrm{\dots }k\}$ were normalised to a carrier power of $1$ per sample, i.e. only the part of the signal vector containing drone bursts was considered for the power calculation ($m$ samples out of $k$). This was achieved by identifying the bursts as those samples where a smoothed energy was above a threshold. The signal vectors $x$ are thus normalised by

  $$\widehat{x}(i)=x(i)/\sqrt{\frac{1}{m}\sum _i{|x(i)|}^2}.$$

  (1)

  Noise vectors (Bluetooth, Wi-Fi, Amplifier, Gauss) $n$ with samples $i\in \{1,\mathrm{\dots }k\}$ were normalised to a mean power of $1$ with

  $$\widehat{n}(i)=n(i)/\sqrt{\frac{1}{k}\sum _i{|n(i)|}^2}.$$

  (2)

  Finally, the normalised drone signal vectors were mixed with the normalised noise vectors by

  $$\widehat{y}(i)=\frac{\left(\sqrt{k}\cdot \widehat{x}(i)+\widehat{n}(i)\right)}{\sqrt{k+1}},\text{ with }k={10}^{\frac{\text{SNR}}{10}},$$

  (3)

  to generate the noisy drone signal vectors $\widehat{y}$ at different SNRs.

  2.3 DEVELOPMENT DATASET

  To facilitate future model development, we provide our resulting dataset³³https://www.kaggle.com/datasets/sgluege/noisy-drone-rf-signal-classification-v2 along with a code example⁴⁴https://github.com/sgluege/noisy-drone-rf-signal-classification-v2 to load and inspect the data. The dataset consists of the non-overlapping signal vectors of $2^{20}$ samples, corresponding to $\approx 74.9$ ms at $14$ MHz.

  As described in Sec. 2.2, the drone signals were mixed with noise. More specifically, $50\%$ of the drone signals were mixed with Labnoise (Bluetooth + Wi-Fi + Amplifier) and $50\%$ with Gaussian noise. In addition, we created a separate noise class by mixing Labnoise and Gaussian noise in all possible combinations (i.e., Labnoise $+$ Labnoise, Labnoise $+$ Gaussian noise, Gaussian noise $+$ Labnoise, and Gaussian noise $+$ Gaussian noise). For the drone signal classes, as for the noise class, the number of samples for each SNR level was evenly distributed over the interval of SNRs $\in [-20,30]$ dB in steps of $2$ dB, i.e., $679$-$685$ samples per SNR level. The resulting number of samples per class is given in Tab. 3.

  Table 3: Number of samples in the different classes in the development dataset.

  Class	DJI	FutabaT14	FutabaT7	Graupner	Taranis	Turnigy	Noise
  #samples	$1280$	$3472$	$801$	$801$	$1663$	$855$	$8872$

  In our previous work [20] we found an advantage in using the spectrogram representation of the data compared to the IQ representation, especially at low SNRs levels. Therefore, we transform the raw IQ signals by computing the spectrum of each sample with consecutive Fourier transforms with non-overlapping segments of length $1024$ for the real and imaginary parts of the signal. That is, the two IQ signal vectors ($[2\times 2^{20}]$) are represented as two matrices ($[2\times 1024\times 1024]$). Fig. 2 shows four samples of the dataset at different SNRs. Note that we have plotted the log power spectrogram of the complex spectrum ${\widehat{y}}_{\mathrm{fft}}$ as

  $${\log }_{10}|{\widehat{y}}_{\mathrm{fft}}|={\log }_{10}(\sqrt{\mathrm{Re}{({\widehat{y}}_{\mathrm{fft}})}^2+\mathrm{Im}{({\widehat{y}}_{\mathrm{fft}})}^2})$$

  (4)

  

  (a) FutabaT14 at SNR $26$ dB

  

  (b) DJI at SNR $6$ dB

  

  (c) Taranais at SNR $-4$ dB

  

  (d) Noise at SNR $22$ dB

  Figure 2: Log power spectrogram and IQ data samples from the development dataset at different SNRs (2(a)-2(d))

  2.4 DETECTION SYSTEM PROTOTYPE

  For field use, a system based on a mobile computer was used as shown in Fig. 3 and illustrated in Fig. 4. The RF signals were received using a directional left-hand circularly polarised antenna (H&S SPA 2400/70/9/0/CP). The antenna gain of $8.5$ dBi and the front-to-back ratio of $20$ dB helped to increase the detection range and to attenuate the unwanted interferers in the opposite direction. Circular polarisation has been chosen to eliminate the alignment problem as the transmitting antennas have a linear polarisation. The USRP B210 was used to down-convert and digitise the RF signal at a sampling rate of $14$ Msps. On the mobile computer, the GNU Radio program collected the baseband IQ samples in batches of one second and send one batch at a time to our PyTorch model, which classified the signal. To speed up the computations in the model we utilised an Nvidia GPU in computer. The classification results were then visualised in real time in a dedicated GUI.

  

  
  Figure 3: Block diagram of the mobile drone detection system.

  

  Figure 4: Detection prototype at the Zurich Lake in Rapperswil.

  3 METHODS

  3.1 MODEL ARCHITECTURE AND TRAINING

  As in [20] we chose the Visual Geometry Group (VGG) CNN architecture [24]. The main idea of this architecture is to use multiple layers of small ($3\times 3$) convolutional filters instead of larger ones. This is intended to increase the depth and expressiveness of the network, while reducing the number of parameters. There are several variants of this architecture, which differ in the number of convolutional layers ($11$ and $19$, respectively). We used a variant with a batch normalisation [25] layer after the convolutions, denoted as VGG11_BN to VGG19_BN. For the dense classification layer, we used $256$ linear units followed by $7$ linear units at the output (one unit per class).

  A stratified $5$-fold train-validation-test split was used as follows. In each fold, we trained a network using $80\%$ and $20\%$ of the available samples of each class for training and testing, respectively. Repeating the stratified split five times ensures that each sample was in the test set once in each experiment. Within the training set, $20\%$ of the samples were used as the validation set during training.

  Model training was performed for $200$ epochs with a batch size of $8$. The PyTorch [26] implementation of the Adam algorithm [27] was used with a learning rate of $0.005$, betas $(0.9,0.999)$ and weight decay of $0$.

  3.2 MODEL EVALUATION

  During training, the model was evaluated on the validation set after each epoch. If the balanced accuracy on the validation set increased, it was saved. After training, the model with the highest balanced accuracy on the validation set was evaluated on the withheld test data. The performance of the models on the test data was accessed in terms of classification accuracy and balanced accuracy.

  As accuracy simply measures the proportion of correct predictions out of the total number of observations, it can be misleading for unbalanced datasets. In our case, the noise class is over-represented in the dataset (cf. Tab. 3). Therefor, we also report the balanced accuracy, which is defined as the average of the recall obtained for each class, i.e. it gives equal weight to each class regardless of how frequent or rare it is.

  3.3 VISUALISATION OF MODEL EMBEDDINGS

  Despite their effectiveness, CNNs are often criticised for being “black boxes”. Understanding the feature representations, or embeddings, learned by the CNN helps to demystify these models and provide some understanding of their capabilities and limitations. In general, embeddings are high-dimensional vectors generated by the intermediate layers that capture essential patterns from the input data.

  In our case, we chose the least complex VGG11_BN model to visualise its embeddings. When inferencing the test data, we collected the activations at the last dense classification layer, which consists of $256$ units. Given $3549$ test samples, this results in a $256\times 3549$ matrix. Using t-distributed Stochastic Neighbor Embedding (t-SNE) [28] and Uniform Manifold Approximation and Projection (UMAP) [29] as dimensionality reduction techniques, we project these high-dimensional embeddings into a lower-dimensional space, creating interpretable visualisations that reveal the model’s internal data representations.

  Our goals were to understand how the model clusters and separates different classes, to identify potential overlaps or ambiguities, and to examine the hierarchical relationships within the learned features.

  3.4 DETECTION SYSTEM FIELD TEST

  We conducted a field test of the detection system in Rapperswil at the Zurich Lake. The drone detection prototype was placed on the shore (cf. Fig. 4) in line of sight of a wooden boardwalk across the lake, with no buildings to interfere with the signals. The transmitters were mounted on a $2.5$ m long wooden pole. The signals from the transmitters were recorded (and classified in real time) at four positions along the walkway at approximately $110$ m, $340$ m, $560$ m and $670$ m from the detection system. Figure 5 shows an overview of the experimental setup. At each recording position, we measured with the directional antenna at three different angles, i.e. at $0^{\circ }$ – facing the drones and/or remote controls, at ${90}^{\circ }$ – perpendicular to the direction of the transmitters, and at ${180}^{\circ }$ – in the opposite direction. Directing the antenna in the opposite direction should result in $\approx 20$ dB attenuation of the radio signals.

  

  Figure 5: Experimental measurement setup at the Zurich Lake in Rapperswil. One can see the four recording positions along the wooden walkway and the detection system positioned at the lake side. Further, recordings were done at different angels of the directional antenna indicated by the arrows at the detection system.

  Table 4 lists the drones and/or remote controls used in the field test. Note that the Graupner drone and remote control are part of the development dataset (cf. Tab. 2), but were not measured in the field experiment. We assume that no other drones were present during the measurements, so recordings where none of our transmitters were used are labelled as “Noise”.

  Table 4: Drones and/or remotes used in the field test

  Class	Drone/remote control
  DJI	DJI Phantom Pro 4 drone and remote
  FutabaT14	Futaba T14 remote control
  FutabaT7	Futaba T14 remote control
  Taranis	FrySky Taranis Q X7 remote control
  Turnigy	Turnigy Evolution remote control

  For each transmitter, distance, and angle, $20$ to $30$ s, or approximately $300$ spectrograms were live classified and recorded. The resulting number of samples for each class, distance, and angle are shown in Tab. 5.

  Table 5: Number of samples (#samples) for each class, distance and antenna direction (angle) recorded in the field test. Recordings at $0$ m distance have no active transmitter and were therefore labelled “Noise”

  class	#samples	Distance [m]	#samples	Angle [∘]	#samples
  DJI	$4900$	$0$	$2597$	$0$	$9110$
  FutabaT14	$5701$	$110$	$6208$	$90$	$9076$
  FutabaT7	$5086$	$340$	$6305$	$180$	$9226$
  Noise	$2597$	$560$	$6358$
  Taranis	$5094$	$670$	$5942$
  Turnigy	$4032$

  4 RESULTS

  4.1 CLASSIFICATION PERFORMANCE ON THE DEVELOPMENT DATASET

  Table 6 shows the general mean $\pm$ standard deviation of accuracy and balanced accuracy on the test data of the development dataset (cf. Sec. 2.3), obtained in the $5$-fold cross-validation of the different models.

  There is no meaningful difference in performance between the models, even when the model complexity increases from VGG11_BN to VGG19_BN. The number of epochs for training (#epochs) shows when the highest balanced accuracy was reached on the validation set. It can be seen that the least complex model, VGG11_BN, required the least number of epochs compared to the more complex models. However, the resulting classification performance is the same.

  Table 6: Mean $\pm$ standard deviation of the accuracy (Acc.) and the balanced accuracy (balanced Acc.) obtained in $5$-fold cross-validation of the different models on the test data of the development dataset. An indication of the model training time is given with the mean $\pm$ standard deviation of the number of training epochs (#epochs), i.e. when the highest balanced accuracy on the validation set was reached. The number of trainable parameters (#params) indicates the complexity of the model.

  Model	Acc.	balanced Acc.	#epochs	#params
  VGG11_BN	$0.944\pm 0.005$	$0.932\pm 0.002$	$66.4\pm 24.4$	$9.36\cdot {10}^6$
  VGG13_BN	$0.947\pm 0.003$	$0.935\pm 0.003$	$138.6\pm 46.4$	$9.54\cdot {10}^6$
  VGG16_BN	$0.947\pm 0.006$	$0.937\pm 0.005$	$101.8\pm 41.5$	$14.86\cdot {10}^6$
  VGG19_BN	$0.952\pm 0.006$	$0.939\pm 0.008$	$98.2\pm 45.8$	$20.17\cdot {10}^6$

  Figure 6 shows the resulting $5$-fold mean balanced accuracy over SNRs $\in [-20,30]$dB in $2$ dB steps. Note that we do not show the standard deviation to keep the plot readable. In general, we observe a drastic degradation in performance from $-12$ dB down to near chance level at $-20$ dB.

  The vast majority of misclassifications occurred between noise and drones and not between different types of drones. Figure 7 illustrates this fact. It shows the confusion matrix for the VGGG11_BN model for a single validation on the test data for the samples with $-14$ dB SNR.

  

  Figure 6: Mean balanced accuracy obtained in the $5$-fold cross-validation of the different models on the test set of the development dataset over the SNRs levels.

  

  Figure 7: Confusion matrix of the outputs of the VGG11_BN model on a single fold for the samples at $-14$ dB SNR from the test data. The average balanced accuracy is $0.71$.

  4.2 EMBEDDING SPACE VISUALISATION

  Figure 8 shows the 2D t-SNE visualisation of the VGG11_BN embeddings of $3549$ test samples from the development dataset. It can be seen that each class forms a separate cluster. While the different drone signal clusters are rather small and dense, the noise cluster takes up most of the embedding space and even forms several sub-clusters. This is most likely due to the variety of the signals used in the noise class, i.e. Bluetooth and Wi-Fi signals plus Gaussian noise.

  We used t-SNE for dimensionality reduction because of its ability to preserve local structure within the high-dimensional embedding space. Furthermore, t-SNE has been widely adopted in the ML community and has a well-established track record for high-dimensional data visualisation. However, it is sensitive to hyperparameters such as perplexity and requires some tuning, i.e. different parameters can lead to considerable different results.

  It can be argued that UMAP would be a better choice due to its balanced preservation of local and global structure together with its robustness to hyperparameters. Therefore, we created a web application⁵⁵https://visvgg11bndronerfembeddings.streamlit.app that allows users to test and compare both approaches with different hyperparameters.

  

  Figure 8: 2D t-SNE visualisation of the VGG11_BN embeddings of $3549$ test samples from the development dataset. The hyperparameters for t-SNE were: metric “eucleidean”, number of iterations $1000$, perplexity $30$ and method for gradient approximation “barnes_hut”

  4.3 CLASSIFICATION PERFORMANCE IN THE FIELD TEST

  For each model architecture, we performed $5$-fold cross-validation on the development dataset (cf. Sec. 3.1), resulting in five trained models per architecture. Thus, we also evaluated all five trained models on the field test data. We report the balanced accuracy $\pm$ standard deviation for each model architecture for the complete field test dataset averaged over all directions and distances in Tab. LABEL:tab:field_test_acc.

  Table 7: Mean $\pm$ standard deviation of the balanced accuracy (balanced Acc.) of the complete field test recordings for the different models.

  Model	balanced Acc.
  VGG11_BN	$0.792\pm 0.022$
  VGG13_BN	$0.807\pm 0.011$
  VGG16_BN	$0.811\pm 0.009$
  VGG19_BN	$0.806\pm 0.016$

  As observed on the development dataset (cf. Tab. 6), there is no meaningful difference in performance between the model architectures. We therefore focus on VGG11_BN, the simplest model trained, in the more detailed analysis of the field test results.

  A live system should trigger an alarm when a drone is present. Therefore, the question of whether the signal is from a drone at all is more important than predicting the correct type of drone. Therefore, we also evaluated the models in terms of a binary problem with two classes “Drone” (for all six classes of drones in the development dataset) and “Noise”.

  Table LABEL:tab:field_test_acc_type_direction shows that the accuracies were highly depend on the class. Our models generalise well to the drones in the dataset, with the exception of the DJI. The dependence on direction is not as strong as expected. Orienting the antenna 180^∘ away from the transmitter reduces the signal power by about $20$ dB, resulting in lower SNR and lower classification accuracy. However, as the transmitters were still quite close to the antenna, the effect is not pronounced. As we have seen on the development dataset in Fig. 6, there is a clear drop in accuracy once the SNR is below $-12$ dB. Apparently we were still above this threshold, regardless of the direction of the antenna.

  Table 8: Mean balanced accuracy $\pm$ standard deviation of the VGG11_BN models on the field test recordings for the different classes for each direction ($0^{\circ }$, ${90}^{\circ }$ and ${180}^{\circ }$). The upper part shows the accuracies for the classification problem (seven classes) and the lower part the accuracies for the detection problem “Drone” or “Noise”

  Class	$0^{\circ }$	${90}^{\circ }$	${180}^{\circ }$
  DJI	$0.623\pm 0.080$	$0.624\pm 0.035$	$0.540\pm 0.051$
  FutabaT14	$0.716\pm 0.101$	$0.984\pm 0.011$	$0.911\pm 0.042$
  FutabaT7	$0.724\pm 0.041$	$0.737\pm 0.034$	$0.698\pm 0.059$
  Noise	$0.554\pm 0.038$	$0.924\pm 0.026$	$0.833\pm 0.068$
  Taranis	$0.936\pm 0.008$	$0.879\pm 0.002$	$0.858\pm 0.002$
  Turnigy	$0.899\pm 0.129$	$0.958\pm 0.027$	$0.962\pm 0.024$
  Drone	$0.958\pm 0.011$	$0.859\pm 0.014$	$0.847\pm 0.017$
  Noise	$0.554\pm 0.038$	$0.924\pm 0.026$	$0.833\pm 0.068$

  What may be surprising is the low accuracy on the signals with no active transmitter, labelled as “Noise”, in the direction of the lake ($0^{\circ }$). Given the uncontrolled nature of a field test, it could well be that there a drone was actually flying on the other side of the $2.3$ km wide lake. This could explain the false positives we observed in that direction.

  Table LABEL:tab:field_test_acc_distance shows the average balanced accuracy of the VGG11_BN models on the field test data collected at different distances for each antenna direction. There is a slight decrease in accuracy with distance. However, the longest distance of $670$ m appears to be too short to be a problem for the system. Unfortunately, this was the longest distance within line-of-sight that could be recorded at this location.

  Table 9: Mean balanced accuracy $\pm$ standard deviation of the VGG11_BN models on the field test data with active transmitters collected at different distances for each antenna direction ($0^{\circ }$, ${90}^{\circ }$ and ${180}^{\circ }$). The upper part shows the accuracies for the classification problem (seven classes) and the lower part the accuracies for the detection problem “Drone” or “Noise”

  Classification
  Distance (m)	$0^{\circ }$	${90}^{\circ }$	${180}^{\circ }$
  110	$0.852\pm 0.044$	$0.838\pm 0.024$	$0.786\pm 0.044$
  340	$0.815\pm 0.078$	$0.916\pm 0.017$	$0.828\pm 0.031$
  560	$0.730\pm 0.063$	$0.796\pm 0.007$	$0.764\pm 0.017$
  670	$0.708\pm 0.068$	$0.805\pm 0.008$	$0.777\pm 0.007$
  Detection
  Distance (m)	$0^{\circ }$	${90}^{\circ }$	${180}^{\circ }$
  110	$0.955\pm 0.007$	$0.851\pm 0.025$	$0.849\pm 0.018$
  340	$0.986\pm 0.008$	$0.940\pm 0.010$	$0.913\pm 0.018$
  560	$0.960\pm 0.011$	$0.820\pm 0.013$	$0.807\pm 0.022$
  670	$0.925\pm 0.021$	$0.826\pm 0.013$	$0.823\pm 0.014$

  Figure 9 shows the confusion matrix for the outputs of the VGG11_BN model of a single fold on the field test data. As with the development dataset (cf. Fig. 7), most of the confusion is between noise and drones rather than between different types of drones.

  

  Figure 9: Confusion matrix of the outputs of the VGG11_BN model on a single fold for the samples from the field test data. The average balanced accuracy is $0.80$.

  5 DISCUSSION

  We were able to show that a standard CNN, trained on drone RF signals recorded in a controlled laboratory environment and artificially augmented with noise, generalised well to the more challenging conditions of a real-world field test.

  The drone detection system consisted of rather simple and low budget hardware (consumer grade notebook with GPU + SDR). Recording parameters such as sampling frequency, length of input vectors, etc. were set to enable real-time detection with the limited amount of memory and computing power. This means that data acquisition, pre-processing and model inference did not take longer than the signal being processed ($\approx 74.9$ ms per sample in our case).

  Obviously, the VGG models were able to learn the relevant features for the drone classification from the complex spectrograms of the RF signal. In this respect, we did not find any advantage for the use of more complex models, such as VGG19_BN, over the least complex model, VGG11_BN (cf. Tabs. 6 and LABEL:tab:field_test_acc).

  Furthermore, we have seen that the misclassifications mainly occur between the noise class and the drones, and not between the different drones themselves (cf. Figs. 7 and 9). This is particularly relevant for the application of drone detection systems in security sensitive areas. The first priority is to detect any kind of UAV, regardless of its type.

  Based on our experience and results, we see the following limitations of our work. The field test showed that the models can be used and work reliably (cf. Tab. LABEL:tab:field_test_acc_type_direction). However, it is the nature of a field test that the level of interference from WiFi/Bluetooth noise and the possible presence of other drones cannot be fully controlled. Furthermore, due to the limited space/distance between the transmitter and receiver in our field test setup, we were not able to clearly demonstrate the effect of free space attenuation on detection performance (cf. Tab. LABEL:tab:field_test_acc_distance).

  Regarding the use of simple CNNs as classifiers, it is not possible to reliably predict whether multiple transmitters are present. In that case, an object detection approach on the spectrogams could provide a more fine-grained prediction, see for example the works [30, 31] and [21]. Nevertheless, the current approach will still detect a drone if one or more are present.

  We have only tested a limited set of VGG architectures. It remains to be seen whether more recent architectures, such as the pre-trained Vision Transformer [32], generalise as well or better. We hope that our development dataset will inspire others to further optimise the model side of the problem and perhaps find a model architecture with better performance.

  Another issue to consider is the occurrence of unknown drones, i.e. drones that are not part of the train set. Examining the embedding space (cf. 4.2) gives a first idea of whether a signal is clearly part of a known dense drone cluster or rather falls into the larger, less dense, noise cluster. We believe that a combination of an unsupervised deep autoencoder approach [33, 34] with an additional classification part (cf. [35]) would allow, first, to provide a stable classification of known samples and, second, to indicate whether a sample is known or rather an anomaly.

  References

  
  [1] N. Al-lQubaydhi, A. Alenezi, T. Alanazi, A. Senyor, N. Alanezi, B. Alotaibi, M. Alotaibi, A. Razaque, and S. Hariri, “Deep learning for unmanned aerial vehicles detection: A review,” Computer Science Review, vol. 51, p. 100614, 2 2024. [Online]. Available: https://linkinghub.elsevier.com/retrieve/pii/S1574013723000813
• [2] M. H. Rahman, M. A. S. Sejan, M. A. Aziz, R. Tabassum, J.-I. Baik, and H.-K. Song, “A comprehensive survey of unmanned aerial vehicles detection and classification using machine learning approach: Challenges, solutions, and future directions,” Remote Sensing, vol. 16, p. 879, 3 2024. [Online]. Available: https://www.mdpi.com/2072-4292/16/5/879
• [3] M. S. Allahham, M. F. Al-Sa’d, A. Al-Ali, A. Mohamed, T. Khattab, and A. Erbad, “Dronerf dataset: A dataset of drones for rf-based detection, classification and identification,” Data in Brief, vol. 26, p. 104313, 10 2019. [Online]. Available: https://linkinghub.elsevier.com/retrieve/pii/S2352340919306675
• [4] M. F. Al-Sa’d, A. Al-Ali, A. Mohamed, T. Khattab, and A. Erbad, “Rf-based drone detection and identification using deep learning approaches: An initiative towards a large open source drone database,” Future Generation Computer Systems, vol. 100, pp. 86–97, 11 2019.
• [5] C. J. Swinney and J. C. Woods, “Unmanned aerial vehicle flight mode classification using convolutional neural network and transfer learning,” in 2020 16th International Computer Engineering Conference (ICENCO), 2020, pp. 83–87.
• [6] Y. Zhang, “Rf-based drone detection using machine learning,” in 2021 2nd International Conference on Computing and Data Science (CDS), 2021, pp. 425–428.
• [7] C. Ge, S. Yang, W. Sun, Y. Luo, and C. Luo, “For rf signal-based uav states recognition, is pre-processing still important at the era of deep learning?” in 2021 7th International Conference on Computer and Communications (ICCC), 2021, pp. 2292–2296.
• [8] Y. Xue, Y. Chang, Y. Zhang, J. Sun, Z. Ji, H. Li, Y. Peng, and J. Zuo, “Uav signal recognition of heterogeneous integrated knn based on genetic algorithm,” Telecommunication Systems, vol. 85, pp. 591–599, 4 2024. [Online]. Available: https://link.springer.com/10.1007/s11235-023-01099-x
• [9] A. AlKhonaini, T. Sheltami, A. Mahmoud, and M. Imam, “Uav detection using reinforcement learning,” Sensors, vol. 24, no. 6, 2024. [Online]. Available: https://www.mdpi.com/1424-8220/24/6/1870
• [10] M. Ezuma, F. Erden, C. K. Anjinappa, O. Ozdemir, and I. Guvenc, “Detection and classification of uavs using rf fingerprints in the presence of wi-fi and bluetooth interference,” IEEE Open Journal of the Communications Society, vol. 1, pp. 60–76, 2020. [Online]. Available: https://ieeexplore.ieee.org/document/8913640/
• [11] ——, “Drone remote controller rf signal dataset,” 2020. [Online]. Available: https://dx.doi.org/10.21227/ss99-8d56
• [12] E. Ozturk, F. Erden, and I. Guvenc, “Rf-based low-snr classification of uavs using convolutional neural networks,” ITU Journal on Future and Evolving Technologies, vol. 2, pp. 39–52, 7 2021. [Online]. Available: https://www.itu.int/pub/S-JNL-VOL2.ISSUE5-2021-A04
• [13] C. J. Swinney and J. C. Woods, “Dronedetect dataset: A radio frequency dataset of unmanned aerial system (uas) signals for machine learning detection & classification,” 2021. [Online]. Available: https://dx.doi.org/10.21227/5jjj-1m32
• [14] ——, “Rf detection and classification of unmanned aerial vehicles in environments with wireless interference,” in 2021 International Conference on Unmanned Aircraft Systems (ICUAS), 2021, pp. 1494–1498.
• [15] S. Kunze and B. Saha, “Drone classification with a convolutional neural network applied to raw iq data,” in 2022 3rd URSI Atlantic and Asia Pacific Radio Science Meeting (AT-AP-RASC), May 2022, pp. 1–4. [Online]. Available: https://ieeexplore.ieee.org/document/9814170/
• [16] O. Medaiyese, M. Ezuma, A. Lauf, and A. Adeniran, “Cardinal rf (cardrf): An outdoor uav/uas/drone rf signals with bluetooth and wifi signals dataset,” 2022. [Online]. Available: https://dx.doi.org/10.21227/1xp7-ge95
• [17] O. O. Medaiyese, M. Ezuma, A. P. Lauf, and A. A. Adeniran, “Hierarchical learning framework for uav detection and identification,” IEEE Journal of Radio Frequency Identification, vol. 6, pp. 176–188, 2022.
• [18] O. O. Medaiyese, M. Ezuma, A. P. Lauf, and I. Guvenc, “Wavelet transform analytics for rf-based uav detection and identification system using machine learning,” Pervasive and Mobile Computing, vol. 82, p. 101569, 6 2022. [Online]. Available: https://linkinghub.elsevier.com/retrieve/pii/S1574119222000219
• [19] F. N. Iandola, M. W. Moskewicz, K. Ashraf, S. Han, W. J. Dally, and K. Keutzer, “Squeezenet: Alexnet-level accuracy with 50x fewer parameters and <1mb model size,” CoRR, vol. abs/1602.07360, 2016. [Online]. Available: http://arxiv.org/abs/1602.07360
• [20] S. Glüge., M. Nyfeler., N. Ramagnano., C. Horn., and C. Schüpbach., “Robust drone detection and classification from radio frequency signals using convolutional neural networks,” in Proceedings of the 15th International Joint Conference on Computational Intelligence - NCTA, INSTICC.   SciTePress, 2023, pp. 496–504.
• [21] R. Zhao, T. Li, Y. Li, Y. Ruan, and R. Zhang, “Anchor-free multi-uav detection and classification using spectrogram,” IEEE Internet of Things Journal, vol. 11, pp. 5259–5272, 2 2024. [Online]. Available: https://ieeexplore.ieee.org/document/10221859/
• [22] C. Sun, A. Shrivastava, S. Singh, and A. Gupta, “Revisiting unreasonable effectiveness of data in deep learning era,” in 2017 IEEE International Conference on Computer Vision (ICCV), 2017, pp. 843–852.
• [23] P. Virtanen, R. Gommers, T. E. Oliphant, M. Haberland, T. Reddy, D. Cournapeau, E. Burovski, P. Peterson, W. Weckesser, J. Bright, S. J. van der Walt, M. Brett, J. Wilson, K. J. Millman, N. Mayorov, A. R. J. Nelson, E. Jones, R. Kern, E. Larson, C. J. Carey, İ. Polat, Y. Feng, E. W. Moore, J. VanderPlas, D. Laxalde, J. Perktold, R. Cimrman, I. Henriksen, E. A. Quintero, C. R. Harris, A. M. Archibald, A. H. Ribeiro, F. Pedregosa, P. van Mulbregt, and SciPy 1.0 Contributors, “SciPy 1.0: Fundamental Algorithms for Scientific Computing in Python,” Nature Methods, vol. 17, pp. 261–272, 2020.
• [24] K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” in 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, Y. Bengio and Y. LeCun, Eds., 2015. [Online]. Available: http://arxiv.org/abs/1409.1556
• [25] S. Ioffe and C. Szegedy, “Batch normalization: accelerating deep network training by reducing internal covariate shift,” in Proceedings of the 32nd International Conference on International Conference on Machine Learning - Volume 37, ser. ICML’15.   JMLR.org, 2015, p. 448–456.
• [26] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, “Pytorch: An imperative style, high-performance deep learning library,” in Advances in Neural Information Processing Systems 32, H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, Eds.   Curran Associates, Inc., 2019, pp. 8024–8035.
• [27] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, Y. Bengio and Y. LeCun, Eds., 2015. [Online]. Available: http://arxiv.org/abs/1412.6980
• [28] L. van der Maaten and G. Hinton, “Visualizing data using t-sne,” Journal of Machine Learning Research, vol. 9, no. 86, pp. 2579–2605, 2008. [Online]. Available: http://jmlr.org/papers/v9/vandermaaten08a.html
• [29] L. McInnes, J. Healy, N. Saul, and L. Großberger, “Umap: Uniform manifold approximation and projection,” Journal of Open Source Software, vol. 3, no. 29, p. 861, 2018. [Online]. Available: https://doi.org/10.21105/joss.00861
• [30] K. N. R. Surya Vara Prasad and V. K. Bhargava, “A classification algorithm for blind uav detection in wideband rf systems,” in 2020 IEEE 92nd Vehicular Technology Conference (VTC2020-Fall), 2020, pp. 1–7.
• [31] S. Basak, S. Rajendran, S. Pollin, and B. Scheers, “Combined rf-based drone detection and classification,” IEEE Transactions on Cognitive Communications and Networking, vol. 8, no. 1, pp. 111–120, 2022.
• [32] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” ICLR, 2021.
• [33] S. Lu and R. Li, DAC–Deep Autoencoder-Based Clustering: A General Deep Learning Framework of Representation Learning.   Springer Science and Business Media Deutschland GmbH, 2 2022, vol. 294, pp. 205–216. [Online]. Available: https://link.springer.com/10.1007/978-3-030-82193-7_13
• [34] H. Zhou, J. Bai, Y. Wang, J. Ren, X. Yang, and L. Jiao, “Deep radio signal clustering with interpretability analysis based on saliency map,” Digital Communications and Networks, 1 2023. [Online]. Available: https://linkinghub.elsevier.com/retrieve/pii/S2352864823000238
• [35] E. Pintelas, I. E. Livieris, and P. E. Pintelas, “A convolutional autoencoder topology for classification in high-dimensional noisy image datasets,” Sensors, vol. 21, p. 7731, 11 2021. [Online]. Available: https://www.mdpi.com/1424-8220/21/22/7731