additive non-Gaussian alpha-stable noise, correntropy, cyclic correntropy function, cyclostationary descriptors, fractional lower-order cyclic autocorrelation function, digital modulations, impulsive noise, automatic modulation recognition, machine learning, logistic regression.

The rapid growth of applications supported by wireless communications systems drives the search for new communications systems that allow to efficiently explore the frequency spectrum, such as systems based on cognitive radio. The cognitive radio can be defined as an intelligent communications system, capable of adapting autonomously to the communication channel, through the reconfiguration of its operating parameters. An important property of cognitive radios is the ability to automatically recognize the type of modulation employed in an RF signal, thus enabling interoperability between systems, improving spectral efficiency, or even enabling electronic surveillance (in military application contexts) ). This attribute is known as automatic modulation classification (AMC). Among the AMC techniques that characterize the state-of-the-art, are those that are based on the detection of patterns obtained from the analysis of second order cycle stationary characteristics. Although very widespread, these techniques are unable to recognize some types of digital modulations, such as high-order M-QAM and M-PSK modulations. On the other hand, the higher order cycle stationary analysis techniques, used to extract singular descriptors of these modulations, have a very high computational cost and are only suitable for communication environments with AWGN noise. Although the AWGN noise model is widely used to characterize wireless communication channels, there are several practical scenarios that are better modeled by non-Gaussian distributions, such as HF communication, whose environment presents a strong contamination by impulsive noise. Recently, two new cyclostationary analysis functions, the lower order fractional cyclic autocorrelation function (FLOCAF), and the cyclic current correlation function (CCF), were defined and evaluated for the purpose of spectral sensing in impulsive environments, being the spectral sensing a less complex problem in relation to the automatic classification of modulations. In fact, knowing that there was no satisfactory solution in the literature for the automatic classification of high-order modulations in channels with impulsive noise, this problem was addressed in this work. In this work, automatic modulation classification architectures are developed based on the FLOCAF and CCF cyclostationary functions, combined with decision tree classification and logistic regression techniques. The architectures were developed for the recognition of BPSK, QPSK, 8-QAM, 16-QAM, and 32-QAM digital modulations, and evaluated in different contexts of alpha-stable additive noise contamination. The results showed that all architectures were able to operate in impulsive environments, however, architectures based on CCF were the most efficient.