Developing RFML Intuition: An Automatic Modulation Classification Architecture Case Study

Abstract

The application of machine learning to Automatic Modulation Classification (AMC) has typically used transfer learning from architectures found in the image classification domain. This work examines deviations from the image classification architectures by drawing from traditional expert feature systems within the AMC domain. Two types of ‘expert architectures’ are contrasted against the traditional image processing architectures; the first utilizes a more traditional one-versus-all binary classification with decision fusion approach, while the second inherits a hierarchical decision tree structure that leverages expert knowledge of the classes. When compared with a typical image processing architecture there are marginal classifier performance gains associated with the structures taken from expert AMC systems; however, the expert architectures allow for greater intuition, adaptability, and future-proofing in general.