Vous êtes sur la page 1sur 2

Application of Machine Learning to Antenna

Design and Radar Signal Processing: A Review


Youngwook Kim
Electrical and Computer Engineering, California State University, Fresno, 2320 E. San Ramon Ave, Fresno, CA93740

Abstract — Recently, machine learning has received a great and target classification, based on radar imagery, are
deal of attention thanks to its success in a range of applications addressed.
in engineering science, medicine, and economics. The new
technology has also been gaining popularity in electromagnetics.
In this paper, the application of machine learning algorithms, 2. Machine Learning Algorithms
for antenna design and radar signal processing, is reviewed.
Index Terms — Machine learning, Deep learning, Antenna When the analytical model of a system is not available,
optimization, Radar signal processing, Deep convolution neural while input and output data can be measured or simulated,
networks, Transfer learning, Classification. machine learning can be an alternate tool to explain the
system via mathematical equations. The mathematical model
includes numerous parameters that can be optimized to
1. Introduction approximate the system transfer function. The determination
of the parameter is referred to as a training process that
Machine learning is one of the rapidly emerging
searches for parameters that make the model best fit the data.
disciplines that can be widely applied in the fields of
Among machine learning algorithms, the neural network
engineering, science, medicine, and economics, to name just is one of the most famous mathematical structures to have
a few. The area is a subset of artificial intelligence that uses become a foundation of the data-driven model [6]. The
computational statistics to find a mathematical model neural network consists of several layers, each of which is
describing input and output data. The constructed composed of several perceptrons. The perceptrons of one
mathematical model, referred to as a data-driven archetype, layer are connected with those of another layer via weighting
can offer a substitute for an analytical counterpart. Because factors. In the perceptron, a non-linear function, known as
the data-driven model can interpolate output based on the activation function, is embedded to enable the nonlinear
unknown input, it can solve regression problems. With description of the system. The structure of the neural
classification problems sharing the same fundamental basis network is shown below.
as regression, machine learning has been used intensively for
classification problems such as image recognition, speech
recognition, and target classification [1].
Recently, the application of machine learning has also
been extended to electromagnetics (EM). Thanks to the
advancement of machine learning algorithms, especially
deep-learning technology that enables the modeling of high- Fig. 1. Structure of neural networks.
level abstractions in data, many antenna parameter
optimization and radar target classification problems could Conventionally, a single inner layer has been used because
be revisited. Antenna optimization, which requires high this has been proven sufficient to approximate any non-linear
computational complexity, can be addressed by machine systems. In addition, multiple layers induce more parameters
learning algorithms to reduce the time cost involved. Rather to be found, meaning that the computational complexity and
than employing a computationally expensive EM simulator, limited memory size are the main restrictions on its
machine learning, such as that found in artificial neural feasibility.
networks, can substitute the EM simulator following training
[2]. In addition, target classification, based on radar
measurement, has been a significant topic of machine 3. Deep Learning
learning for defense and surveillance purposes for decades The difference between general machine learning and deep
[3–5]. learning is that the latter uses multiple layers to enable more
In this paper, we review the applications of machine powerful abstractions and generalization. The conventional
learning in the EM field, in particular antenna design and use of neural networks requires a prior feature extraction
radar target classification. The basic concepts of machine process to reduce the networks’ computational burden. If the
learning and deep learning are discussed, and their features need to be extracted inside the neural networks, they
applications for antenna geometric parameter optimization will require a large network size with powerful training
capability. Thus, the chance of successful training was not
high, considering a practical computational power and One of the bottlenecks of using deep learning for radar
memory size. Recently, advances in computer power and imagery is the lack of training data. Unlike with optical
larger memory sizes have enabled the successful training of camera images, high costs are involved in obtaining the radar
large networks. In deep learning, feature extraction and images, meaning that the number of images is not sufficient
classification/generalization can be performed in a single to train large networks. To overcome this issue, transfer
network, leading to remarkable improvement. Various deep learning has been suggested, which has shown great
learning technologies have seen success, including deep improvement [4]. By borrowing the pre-trained
neural networks, deep convolutional neural networks convolutional neural networks, such as AlexNet and VGG16,
(DCNN), and deep recursive neural networks (DRNN), to only the last stage, the fully connected layer, needs to be re-
trained.
name a few. In particular, DCNN has obtained great
popularity in image classification, while DRNN has shown
its capability in time-domain signal processing.

4. Application to Antenna Design


(a)
Because antenna design is involved with a
computationally expensive EM simulator, the simulation
time ranges from a few seconds to a few days, depending on
the size of the antenna, operating frequency, and
computational power. In particular, because the antenna
needs to be optimized to maximize the performance while
reducing the size, rapid optimization techniques are highly
(b)
desirable. Rather than using an EM simulator, machine Fig. 2. (a) DCNN for micro-Doppler processing and (b) DCNN for SAR
learning algorithms can work as an alternative to the general classification [9].
optimization techniques such as genetic algorithm, particle
swarm optimization, and simulated annealing. For example,
6. Conclusion
the relationship between the size of the patch antenna and its
resonant frequency is trained by neural networks, allowing In this paper, we briefly reviewed the applications of deep
the size to be immediately determined through the trained learning to the applied EM. Greater advances in machine
machine [2]. Another approach is to estimate the parameter learning algorithms would benefit other areas, including EM.
for the model of the antenna parameter using the machine Although a distinct improvement could have been found in
learning technique. The antenna characteristics, which can be the radar image classification, it is expected that its use could
expressed by a mathematical model, can be determined if the augment the antenna design area, such as helping in the
model parameters are known [7]. Array antenna designs have design of conventional antennae by non-experienced
also been addressed by neural networks [8]. engineers or identifying new antenna structures.

5. Application to Radar Signal Processing References


One of the purposes of radar signal processing is to [1] G. James, D. Witten, T. Hastie, and R. Tibshirani, An Introduction to
identify and recognize the measured target. The latter’s Statistical Learning (with applications in R), Springer, 2013.
[2] M. Singhal and G. Saini, “Optimization of antenna parameters using
features, such as RCS, range profile, and radar imagery, offer artificial neural network: A review,” International Journal of
a possibility of target recognition. Conventionally, the Computer Trends and Technology, vol. 44, 2017.
classifiers based on machine learning algorithms, such as [3] Y. Kim and T. Moon, “Human detection and activity classification
neural networks and support vector machines, have been based on micro-Dopplers using deep convolutional neural
networks,” IEEE Geoscience and Remote Sensing Letters, vol. 13, pp.
used for target recognition problems. In particular, DCNN 2–8, Jan. 2016.
has recently garnered significant attention in image [4] J. Park, J. Rios, T. Moon, and Y. Kim, “Micro-Doppler based
classification problems, with its success encouraging its use classification of human activities on water via transfer learning of
for radar imagery. Convolutional filters extract features and convolutional neural networks,” Sensors, vol. 16, pp. 19–90, Nov.
2016.
the fully connected layers construct a class boundary in a [5] S. Chen and H. Wang, “SAR target recognition based on deep
single neural network. As such, DCNN has been applied to learning,” International Conference on Data Science and Advanced
micro-Doppler signatures. Human activities, hand gestures, Analytics (DSAA), Nov. 2014.
aquatic activities, drones, and vehicles are measured using [6] C. Bishop, Neural Network for Pattern Recognition, Clarendon Press,
Jan. 1996.
radar, and the corresponding spectrogram is classified using [7] Y. Kim, S. Keely, J. Ghosh, and H. Ling, “Application of artificial
DCNN. Another notable advance was made in synthetic neural networks to broadband antenna design based on a parametric
aperture radar (SAR) images for automatic target recognition. frequency model,” IEEE Transactions on Antennas and Propagation,
This not only allowed the SAR image classification to be vol. 55, pp. 669–674, Mar. 2007.
[8] A. Rawata, R. N. Yadavb, and S. C. Shrivastavac, “Neural network
improved, the whole process can now be handled by DCNN applications in smart antenna arrays: A review,” International
rather than the traditional routine of detection, discrimination, Journal of Electronics and Communications, Vol. 66, May 2012.
and classification. [9] C. Danilla, “Convolutional neural networks for contextual de-noising
and classification of SAR images,” M.S. Thesis, University of
Twente, 2017.