IEEE - Aerospace and Electronic Systems - November 2019 - 53

Huizing et al.

Figure 7.
Denoising adversarial auto-encoder applied to spectrograms corrupted with synthetic noise.

GANomaly network, as shown in Figure 4, except that the
separate encoder is omitted and the objective function for
training of the auto-encoder only consists of the mean
square error between noise corrupted versions of the spectrograms and the uncorrupted spectrograms.
Figure 7 shows two examples of measured microDoppler spectrograms of a T-REX 550 helicopter with a
high SNR in the first column. The second column shows
the same spectrograms to which synthetic Gaussian noise
has been added to lower the SNR. The third column shows
the reconstructed spectrograms using a denoising adversarial auto-encoder. Although some details of the microDoppler modulations are distorted, it is still possible to
recognize the micro-Doppler contribution due to the tail
rotor and, up to a certain extent, the blade tip contribution.
In a second experiment, the denoising adversarial
auto-encoder has been tested with measured spectrograms
that are characterized by a lower SNR. As apparent from
the reconstructed spectrograms shown in Figure 8, a signal
to background ratio enhancement in the order of 20 dB is
achieved by the denoising auto-encoder. Furthermore,
information associated to tail rotor signature and blade tip
rotation is partially restored. Although the SNR of the
spectrograms seems to be significantly enhanced by the
denoising adversarial auto-encoder, the impact of this preprocessing on the accuracy of a target classifier, such as a
CNN still has to be investigated.

imbalance in the training set used in this paper
(see Table 1) is approximately a factor of two. This imbalance can have a negative impact on the convergence of
the classifier during training and the overall accuracy [20].
There are several ways to mitigate the effect of an imbalanced training set. The first category concerns methods
that leave the training set intact and modify the training
procedure or the classifier to deal with the imbalance. The
second category involves changes to the training set itself
by generating data for classes that are underrepresented.
In this section, a generative network called InfoGAN has
been investigated for the generation of realistic training
data for underrepresented target classes [21]. When compared with a standard GAN, an InfoGAN has the advantage that it learns in an unsupervised way, interpretable
and disentangled representations of challenging datasets.
Figure 9 shows the architecture of an InfoGAN network

ADVERSARIAL TRAINING FOR SPECTROGRAM
GENERATION
Training sets for target classifiers based on machine learning are often characterized by an imbalance in the number
of examples for different target classes. For example, the
NOVEMBER 2019

Figure 8.
Denoising adversarial auto-encoder applied to measured low SNR
spectrograms.

IEEE A&E SYSTEMS MAGAZINE

IEEE - Aerospace and Electronic Systems - November 2019

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - November 2019