Deep Learning for Modulation Classification: Signal Features in Performance Analysis

An increasing trend towards making the modulation classification (MC) algorithms better-suited for real world implementation in Cognitive Radio (CR) equipment, can be seen as recent works adopt novel agile deep learning models and datasets which include different kinds of signal impairments. Considering the large range of studies in the field, a unifying examination of the common state-of-the-art methods will benefit the future developments by providing a starting point for comparison of their efficiency for recognition of various communication signals. The purpose of this study is to provide a comparative analysis which gives clarity on the ways in which the dataset's type and contents (statistical features of the signals as well as the presence of noise) influences the classification performance. Thus, a more systematic approach to choosing the appropriate input data and a suitable classifier model can be employed in future developments.