Baseband Modulation Classification using Incremental Learning
Todd Morehouse, Niloofar Rahimi, Ming Shao, Ruolin Zhou · 2020
Automatic modulation classification (AMC) has been used in channel estimation to better understand the active users and signals. Traditional methods used complex and computationally expensive feature extraction and algorithms to perform AMC. More recently machine learning (ML) has been used to provide high accuracy AMC using methods such as convolutional neural network (CNN), residual neural network (ResNet), and long short-term memory (LSTM). These methods typically cannot integrate new modulation types to be classified without needing to be retrained with the entire dataset. One challenge that has not been addressed so far is the growing number of data sets and the limited capacity of computational systems to tackle this challenge. Incremental Learning (IL) algorithms introduce methods to train networks on new data, without forgetting knowledge learned from previous data. IL enables a network to learn new classes without requiring the entire dataset from previous classes. This work proposes an incremental learning classification model on signal constellation plots that is capable of adding classes incrementally to the model.