Geomagnetic micro-pulsation automatic detection via deep leaning approach guided with discrete wavelet transform

Abstract Ultra-low frequency (ULF) signals in the geomagnetic records are important indicators for many phenomena; therefore identification of such signals is an important issue. Automatic identification of these ULF signals is not an easy target because of their small magnitudes. Through this study in hand, two algorithms are proposed to automatically detect these micro-pulsations. The first algorithm uses the multi-level components (details) of the discrete wavelet transform (DWT) instead of the original geomagnetic record. The vector of the maximum values of the cross-correlation between the record and an arbitrary chosen ULF pattern in the same frequency range is a good indicator for the existence of these micro-pulsations. The second algorithm is based on convolutional neural network (CNN) framework guided with the multi-resolution-analysis (MRA) of the DWT. Preprocessing the geomagnetic records using the MRA of DWT to produce the fifth and the sixth details to be the input to the deep CNN topology, highly improved the accuracy to approach 91.11%. In addition, deep learning based algorithm showed better results than the DWT based algorithm in light of all the performance metrics.