Monitoring of the electrical equipment as a part of electrical power systems is necessary to ensure the reliability of electrical energy transmission and distribution. Operating modes of power substation transformers and autotransformers are under particular control due to their high functional significance. The most effective method is the real time detection and classification of abnormal operating modes in transformers. Our approach reveals defects of the transformer. The setup consists of a transformer surrounded by three Dialog M-110 microphones connected to a computer via two Orient AU-01SW USB and one Creative Play! (SB1140) RET sound cards. The increase of voltage on one of the transformer windings leads to the increase of the amplitude of the magnetic induction in the transformer core up to the value that corresponds to the deep saturation. We recorded a signal from each microphone to a unique track applying Steinberg Cubase 5 and divided each record into two parts representing different modes, which allowed us to observe different spectra for each mode in each series with embedded "Spectrum Analyzer". We ignored a part of spectra beyond the frequency range of the microphones which was 50 – 16 000 Hz and observed the change in components at 250 Hz and 350 Hz. We witness the increase of harmonics in transformer with defect that miss in a signal, obtained for the normal transformer. The artificial neural networks classify the set of signals. The artificial convolutional neural networks (CNN) method is most popular for sound classification nowadays. Neural network algorithms together with our approach to the study of acoustic signals generated by transformer elements will provide an opportunity to obtain effective methods for timely detection of non-normative operating modes of transformer equipment. The results are applicable for new generation of high-speed automation systems of emergency response in electric power systems.