Algorithmic Feature Selection and Dimensionality Reduction in Signal Classification Tasks

This paper presents a research endeavour addressing the recognition of acoustic emission signals, aiming to enhance their utilisation in non-destructive defectoscopy and machining process control. The classification task can be accomplished through two approaches: representing signals using a suitable attribute set, or directly passing the signals in their entirety to the classification algorithm. Our primary focus was on the meticulous selection of methods and tools for automating the extraction of a comprehensive set of features from the signals, followed by dimensionality reduction techniques. Subsequently, we conducted a comprehensive performance evaluation by comparing various classifiers applied to the low-dimensional projections. Lastly, we put the feature based classification approach to the test with direct signal classification employing convolutional neural networks.