Detection and Classification of Drones Using Radars, AI, and Full-Wave Electromagnetic CAD Tool

Abstract

Detection and classification of drones have become crucial due to their potential usage in illicit activities. Radar systems can provide a promising solution to this needed task when combined with machine learning (ML) and artificial intelligence (AI) models. Radar datasets that contain drone information are needed to train AI models. Generating radar datasets that contain drone information is one of the most important challenges in this application as it is expensive and time-consuming. In addition, such datasets are limited to the radar used, the background environment, and drone types. In this chapter, full-wave electromagnetic (EM) and computer-aided design (CAD) tools are proposed for use to generate radar datasets that contain drone information. The proposed method overcomes this prevailing challenge in the field of radar detection and classification of drones. Furthermore, drones are widely classified using their range-Doppler information, which depends on their mechanical motions. The impact of the control systems of four different drones on their range-Doppler signatures is examined using a full-wave EM CAD tool. Finally, we demonstrate how we advance state-of-the-art literature on the detection and classification of drones utilizing radar systems, a mechanical control-based machine learning (MCML) algorithm is used to classify the four unmanned air vehicles (UAVs).

1. Introduction

Unmanned air vehicles (UAVs), also known as drones, have become easily accessible worldwide in which they are possibly be used in many terrorist attacks and illegal activities [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]. This requires having systems to detect and classify drones at a distance to have time to take actions if needed. Radar-based systems are preferred in comparison with optical, acoustic, and RF-based systems due to their advantages in this application [11, 12, 13, 14, 15, 16]. Radar systems work night and day, and in bad weather conditions, they can detect several drones at a time, track autonomous drones, and classify them when combined with ML models [11, 12, 13, 14, 15, 16].

Classification of radar targets is widely achieved through the generation of range-Doppler images and micro-Dopler signatures [17, 18, 19]. Generating radar datasets that contain UAV information using real measurements costs a lot and wastes time. In addition, these datasets are limited to the used radars, UAV types, and the location and surroundings in which these measurements are made [20]. In this chapter, we propose to use a full-wave EM CAD tool, such as Ansys high-frequency structure simulator (HFSS) [21], to generate radar datasets the contain UAV information for the purpose of training ML algorithms [22, 23, 24, 25]. Traditionally, full-wave EM CAD tools are used for designing and simulating high-frequency stationary electronic products. The proposed method using Ansys HFSS SBR+ solver [21, 22, 23, 24, 25] can be used to move drones and perform time-based full-wave analysis. For example, Figure 1a–c illustrate the DJI S900 hexacopter UAV, Ansys HFSS model of this hexacopter, and three-time stamps showing the rotation of its blades in Ansys HFSS, respectively.

Figure 1.

(a) The DJI S900 UAV [26], (b) Ansys HFSS model for the DJI S900 UAV, and (c) three-time stamps for the modeled DJI S900 UAV showing its blades’ rotation.

2. Datasets generation

Six different UAVs are modeled using the full-wave EM tool, the six UAVs are a fixed-wing UAV, a helicopter UAV, 3 quadcopter UAVs, and a hexacopter UAV. They are modeled according to the Albatross drone [27, 28], the Black Eagle 50 drone [29], the DJI FPV drone [30], the MD4-1000 drone [31, 32], the Phantom 3 standard drone [33], and the DJI S900 drone [26], respectively. The specifications of these drones are shown in Table 1. A W-band frequency modulated continuous wave (FMCW) radar is used to generate the datasets required for this work for its high resolution. The radar parameters used in this work are shown in Table 2, these parameters are chosen for simplicity.

Type	UAV	Dimensions (m)
Fixed wing	Albatross	0.74 × 0.2 × 0.15
Helicopter	Black Eagle 50	2.65 × 0.56, blade 3.75
Quadcopter A	DJI FPV	0.178 × 0.232 × 0.127
Quadcopter B	MD4-1000	1.136 × 1.730 × 0.495
Quadcopter C	Phantom 3	Diagonal 0.35 and blade 0.24
Hexacopter	DJI S900	Diagonal 0.9 and arm 0.358

Table 1.

Dimensions of the six UAVs.

Quantity	Symbol	Value
Center frequency	f0	77 GHz
Bandwidth	BW	300 MHz
Range resolution	ΔR	0.5 m
Velocity resolution	ΔV	0.4 m/s
Maximum range	Rmax	60 m

Table 2.

Radar parameters.

The six UAVs are modeled to pitch forward to the radar from a distance of 50 m/s with 5 m/s velocity, this distance is chosen to match the state-of-the-art literature using mmWave radars for this application [34, 35, 36, 37, 38]. Figure 2 shows the Ansys HFSS simulation setup used to generate the required datasets for this work, while Figure 3a–f shows the range-Doppler maps obtained through Ansys HFSS for the Albatross, Black Eagle 50, DJI FPV, MD4 1000, DJI Phantom 3 standard, and DJI S900 UAVs, respectively.

Figure 2.

Ansys HFSS test setup.

Figure 3.

Range-Doppler image for the (a) Albatross UAV, (b) Black Eagle 50 UAV, (c) DJI FPV UAV, (d) MD4 1000 UAV, (e) DJI Phantom 3 standard UAV, and (f) DJI S900 UAV.

A convolutional neural network (CNN) model is used to classify the six UAVs. The architecture of the CNN model showing the number, types, and dimensions of layers used in this work is illustrated in Figure 4. The CNN model is based on the DopplerNet CNN model [39], a max pooling layer is added to this model to decrease its complexity and avoid overfitting. The CNN model is applied to a dataset that contains a total of 1200 range-Doppler maps, 200 range-Doppler maps for each UAV. The classification result is found to exceed 97% as shown in Figure 5. The findings of this work demonstrate how accurate the proposed method can be used to generate radar datasets that contain UAV information and training ML models on them.

Figure 4.

The CNN model architecture.

Figure 5.

The confusion matrix for the CNN model.

3. Mechanical control-based machine learning

UAVs have five basic motions, which are hovering, pitching, throttling, rolling, and yawing. To perform each motion, the speeds of a UAV’s rotors have to be changed, which affects its Doppler signature [23]. In the state-of-the-art literature, ML models are trained on datasets that contain a UAV’s hovering and pitching motions only, and they are tested on the same datasets [36, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50]. This leads to a degradation on ML accuracy if these trained models were tested using different datasets that contain other motions [23]. As radar signature of UAVs is highly dependent on the mechanical control information of them [23]. In this section, the impact of the mechanical control information of four UAVs on ML accuracy is investigated, and a mechanical control-based machine learning (MCML) algorithm is proposed to overcome this effect [23]. Figure 6a and b shows the traditional algorithm used for the state-of-the-art literature and the MCML algorithm, respectively [23].

Figure 6.

(a) Traditional algorithm and (b) MCML algorithm.

To perform this investigation, the full-wave EM simulator is used to model the different motions of the four UAVs, a helicopter, a hexacopter, and two different quadcopters, Figure 7a–d show the throttle, pitch, roll, and yaw motions, respectively [23]. As an example, the range-Doppler maps for the hexacopter UAV for the different motions are shown in Figure 8.

Figure 7.

Ansys HFSS simulation setups for different motions. (a) Throttle, (b) pitch, (c) roll, and (d) yaw.

Figure 8.

Range-Doppler maps for the hexacopter UAV at different motions.

Two different datasets, containing the range-Doppler maps for the four UAVs, are generated: the first dataset contain the pitch and hover motions only to match the state-of-the-art literature, and the second dataset contain all the basic motions for the UAVs. The CNN model shown in Figure 4 is trained on the first dataset and is tested on the same dataset as done in the state-of-the-art literature, and then it is tested on the second dataset that contains different motions to investigate the impact of the mechanical control information of the four UAVs. The result of this investigation is summarized in Figure 9a and b, which illustrates how the CNN model failed to classify the four UAVs when it was tested on another dataset that has different motions. Subsequently, the MCML algorithm [23] is applied, yielding an accuracy of 92.5% as shown in Figure 10. The MCML method gets over the loss in classification accuracy that occurs if the mechanical control information of UAVs is ignored.

Figure 9.

The confusion matrix of the CNN model when it is tested on: (a) same dataset and (b) different dataset.

Figure 10.

The confusion matrix of the CNN model when applying the MCML method.

4. Countering radar deception

Using numerical simulations to generate radar drones datasets facilitate the process for UAV and radar researchers to design and test appropriate radar systems for the detection and classification of drones. Furthermore, the proposed method can be used to decrease the effects of the attempts taken to deceive radar systems. Some of these several attempts that are not limited to modifying a UAV’s body/blades, stealth-coating to reduce a UAV’s radar cross section (RCS) area, equipping a standard UAV with explosives, and training birds to hide a UAV from radar detection. All these examples and more can be modeled using the proposed method to generate radar drones datasets and train ML models on them. For example, Figures 11 and 12 and show the DJI S900 hexacpoter equipped by dynamite, and an octocopter is hidden by a group of birds, respectively. Radar signature of these cases can be generated to be studied, and in addition, signal processing techniques and ML models can be developed to counter these attempts.

Figure 11.

Ansys HFSS model for the DJI S900 hexacopter UAV with dynamite attached.

Figure 12.

Ansys HFSS model for a hidden octocopter by a group of trained birds.

5. Conclusions

In this chapter, full-wave electromagnetic CAD tools are used to generate radar UAV’s datasets to train machine learning (ML) models. The method provides accurate results for range and Doppler information of UAVs. The accuracy of a CNN model used in this work is found to exceed 97%. The proposed method presents a paradigm shift in how machine learning experts think about the application of radar classification of UAVs. The effect of the mechanical control information of UAVs on machine learning accuracy is explored using a full-wave electromagnetic CAD tool. The Doppler information is found to be highly dependent on mechanical control information of UAVs. A MCML method gets over the loss in classification accuracy that occurs if the mechanical control information of UAVs is ignored. The accuracy of the MCML algorithm is found to exceed 90% compared with the state-of-the-art literature in the application of radar detection and classification of UAVs.

Acknowledgments

This work was supported in part by NSERC. We would like to acknowledge CMC microsystems and Ansys for providing licenses for the CAD tools used throughout this work.

Conflict of interest

The authors declare no conflict of interest.

Abbreviations

ML	Machine learning
AI	Artificial intelligence
EM	Electromagnetic
CAD	Computer-aided design
FMCW	Frequency modulated continuous wave
MCML	Mechanical control-based machine learning
UAV	Unmanned air vehicle
HFSS	High-frequency structure simulator
BW	Bandwidth
CNN	Convolutional neural network
RCS	Radar cross section