Application of machine learning methods for the classification of vibroacoustic signals measured at the moment of transformer energization with varying degrees of core and winding defect complexity