An Integrated Toolbox of Time-Frequency Techniques for Preprocessing in AI Networks

Artificial neural networks have recently been widely used in classifying signals with high accuracy and without manual feature extraction procedures. Additionally, simultaneously presenting signals in both time and frequency domains, time-frequency images have been frequently utilized as inputs to train a convolutional neural network. Some typical time-frequency analysis techniques, i.e., the short-time Fourier transform, continuous wavelet transform, Wigner-Ville distribution, and the Hilbert-Huang transform, have been approved to solve complex classification problems in convolutional neural network-based architectures. However, those techniques were considered separately, and one network commonly uses only one method for some types of signals. There is no tool to evaluate which time-frequency techniques are the best for which signals. This paper integrates a time-frequency analysis toolbox to deal with that problem. From then on, it will be an essential preprocessing step in convolutional neural network-based architectures to enhance classification performance.