Pulse and Signal Data Classification Using Conventional and Few-Shot Machine Learning

Abstract

Signal detection is a key component in a radar system; however, signals are often muddled with noise and interference, which can make singling out the pure signals difficult. Also, signals are often interleaved with other signals, which makes it difficult to tell from first glance where a signal starts and ends. This paper will focus on classifying pulses and signals that were generated in MATLAB using few-shot machine learning and conventional machine learning techniques. The signals will be filtered using the Hilbert transform, and the envelope will be taken in order for the data to be used to train machine learning models. The few-shot learning method used in this study involves meta-learning and utilizes an algorithm that was adapted to handle data rather than images. Specifically, the models will be trained using pure time domain data, and the validation accuracies of each model will be compared to see which technique fares best when using minimal data. The trained models will then be used to try to classify a test set and observe if they correctly classify whether a given sample of data represents a pulse or a signal. In the second portion of this experiment, the data will also be labeled based on the number of pulses or signals present in the given sample, using the same methodology but with eight classes instead of two. These results will be compared to see not only how the models fare against one another, but also how having a larger number of classes with a certain attribute to identify affects the accuracy as well.