Abstract

In addressing temporal dependencies within data, specifically in signal analysis, the integration of Deep Neural Networks (DNN) has demonstrated notable improvements when coupled with a preprocessing stage designed for extracting implicit information. In this context, the widely adopted Wavelet Transform (WT) has garnered attention for its remarkable results. However, inherent challenges, such as the imperative definition of parameters for optimal information extraction across diverse scales and resolutions, as well as the prerequisite batch conversion of signals prior to network training, underscore the need for innovative solutions. In response to these challenges, the main contribution of this manuscript is a novel DNN architecture to replace the preprocessing phase. This architecture produces output characteristics resembling those derived from WT, preventing the necessity for a preceding batch execution. Our contribution not only stands as an independent solution but also seamlessly integrates with other modeling techniques, eliminating the prerequisite for the upfront execution of any wavelet transformations. To assess its performance, our methodology undergoes rigorous evaluation against DNNs in classifying signals from real-world applications. Our findings indicate the promising potential of end-to- end schemes in advancing signal analysis applications.