From Frequency Content to Signal Dynamics Using DNNs

This study developed a novel method for analyzing and decomposing signal into its main dynamics small large timescales. Our proposal is based on decoupled hybrid system of convolutional recurrent neural networks that uses as inputs the power spectrum spectrogram given signal, giving output dynamic behavior. We define classification predicted using previously known characterized through training signals: periodic, quasi-periodic, aperiodic, chaotic, randomness.We created synthetic dataset comprising more than 50 signals from different categories. For real-world dataset, we used photoplethysmographic 40 students obtained Spanish medical study. tested system’s performance in real biological synthetical signals, obtaining noteworthy results. All results are evaluated qualitatively quantitatively. Still, novelty lack similar works, cannot compare reliably rigorously our with other at least can retrieve exposed this work three key ideas: DNN-based solutions capable learning generalizing behavior signals; learned correctly to distinguish between reference provided find some unidirectional similarities aperiodicity cases; PPG reveal seem exhibit multi-dynamic changes depending timescale, being quasi-periodically dominant short-term aperiodically long-term.