Neural Networks for Periodic Signals in Time and Frequency Domains – The Advantage of a Spectral Stride CNN

In many fields, as exemplified by machine condition monitoring, the use of neural networks (NN) for the processing of signals is attracting more attention than the use of established knowledge-based approaches. While for decades researchers have focussed on designing and refining signal processing techniques based on their a priori understanding of the properties of signals, in the last few years this approach has been challenged by a posteriori justified artificial-intelligence (AI) techniques. The inability to provide reasons for the choice of NN structures has been accompanied by a lack of understanding of what the NN does to reach a certain output, which significantly compromises the trust in this approach. The importance of the NN input domain of representations (e.g., signal representation in time and/or frequency domains) is well understood. However, the effects of the choice of input domain on the network architecture and its overall effectiveness remain unexplored for the case of periodic signals. Therefore, the approach of interpretable NN will need to be further examined in terms of the role of the input to the NN. In this paper, the authors aim at applying a signal-processing-guided analysis to neural networks operating on both time and frequency domain representations of periodic signals. The scope of this work will include NN based on convolutional and fully connected layers. The paper will analyse how the role of these layers will change depending on the domain of the input signal and propose an effective solution to frequency-domain analysis consisting of a CNN with stride equal to kernel size.