IEEE - Aerospace and Electronic Systems - June 2022 - 34

Cognitive Radio for Aeronautical Mobile Telemetry: A Machine Learning-Based Approach
set of values. Frequently used classifiers include k-nearest
neighbors (KNNs), support-vector machine (SVM), linear
discriminant analysis (LDA), decision tree (DT), and random
forest. Useful regression models include linear, quadratic,
and logistic regression, SVMs, and Gaussian process regression.
In addition, deep learning algorithms such as convolutional
neural networks are more complex but exhibit higher
performance thanmany traditional models.
Many of the common machine learning classifiers have
Figure 2.
Comparison of the PSDs of the IRIG 106 modulations.
spectral efficiency by a factor of 3 relative to PCM/FM. In
2004, the Telemetry Group (TG) ofthe Range Commander's
Council (RCC) adopted SOQPSK-TG as the modulation to
meet the first goal. SOQPSK-TG is a partial response CPM
with a constrained ternary alphabet, modulation index h ¼
0.5, and a frequency pulse spanning eight bit times. ARTM
CPM was adopted in 2007 to meet the second goal. ARTM
CPM is a partial response CPM with quaternary alphabet,
two modulations indexes h ¼ {4/16, 5/16} applied in an
alternating fashion, and a pulse shape spanning three symbol
times. The PSDs ofthe modulations are plotted in Figure 2.
MACHINE LEARNING FOR MODULATION RECOGNITION
AND BITRATE ESTIMATION
Machine learning algorithms are highly effective for finding
patterns in data and predicting responses to new input. Many
statistical machine learning algorithms are supervised learning
methods comprising two steps: training on input data
with known target values, and then using the trained model to
make predictions on new data. Algorithms can perform both
classification on categories and regression on a continuous
been applied to the problem of automatic modulation recognition.
The features used for the classification include spectral
features and statistical time domain features. In [11], the performance
ofSVM, KNN, and naı¨ve Bayes classifiers to recognize
2-, 4-, 16-, and 64-ary QAM was evaluated, using
higher order cumulants as the features for classification.
Using SVM classifiers on higher order statistical features is a
frequent technique for automatic modulation recognition
[12], [13]. In [14], fourth-order statistical features are used
with an artificial neural network to classify one of the six
modulation types (three analog modulations and three digital
modulations). Increasing in complexity, the process in [15]
involves a convolutional neural network followed by a long
short-termmemory network to extract features from raw data
and classify among 11 different modulation types (three analog
modulations and eight digital modulations).
Spectral features for modulation detection have also been
explored. Exploring a less complex method of classification
for cognitive radio applications in [16], the maximum value of
the PSD is used as a metric to distinguish between 2ASK,
BPSK, QPSK, 16QAM, 2FSK, and 4FSK, but accuracy is not
high at SNRs below 15 dB. In [17], several instantaneous statistical
features are usedwith a DT to recognize ASKand FSK
modulations, resulting in correct classification above 7 dB.
In summary, machine learning algorithms can be trained
and testedwith a wide variety ofinputs obtained by processing
the received signal samples. Some ofthe processing is sophisticated
with significant computational complexity. Because
the modulation type and the bitrate are unknown, our experiments
began with simple PSD estimates. The PSD estimates
with known bitrate andmodulation type were used in the training
phase. In the testing phase, the PSD estimates were used.
The results are described in the following two sections.
TRAINING
Figure 3.
Experimental configuration used to capture the data used for
training and evaluation.
34
The training data were generated using a multimode telemetry
transmitter connected to an ADALM Pluto SDR [18].
The experimental setup is illustrated in Figure 3. The transmitter
is a flight-certified model and represents the current
state of the art in producing the CPMs used in aeronautical
mobile telemetry. For the training phase, the ADALM Pluto
SDR is used to convert the continuous-time RF waveform to
baseband I/Q samples. Because ADALMPluto SDR is inexpensive,
it does not have the high-quality RF front-end
IEEE A&E SYSTEMS MAGAZINE
JUNE 2022
IEEE - Aerospace and Electronic Systems - June 2022

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - June 2022
