International Journal for Innovation Education and Research ISSN 2411-2933 01-02-2022 International Educative Research Foundation and Publisher © 2022 pg. 178 Mechanical Failures Detection by Vibration Analysis in Rotary Machines Using Wavelet and Artificial Neural Network in the Gera Maranhão Plant. Érico Fernando da Costa Vale Engenheiro Eletricista, Usina Termelétrica Gera Maranhão, Gerência de planejamento e controle Maranhão, Brazil. Link do Lattes: http://lattes.cnpq.br/5001592417786652 ORCID: https://orcid.org/0000-0001-6924-276X E-mail: erico@geramaranhao.com.br Antonio Carlos Duarte Ricciotti Professor, Department of Electrical Engineering, Federal University of Rondonia, Rondonia, Brazil. ORCID: https://orcid.org/0000-0002-4986-6601 Email: acdricciotti@unir.br Viviane Barrozo da Silva Professor, Department of Electrical Engineering, Federal University of Rondonia, Rondonia, Brazil. ORCID: https://orcid.org/0000-0002-1948-1532 Email: viviane-barrozo@unir.br Abstract This article presents aspects of a tool to assist in predictive maintenance based on vibration analysis in rotating machines using wavelet transform and artificial neural networks. The work analyzed the experimental results of applying a methodology based on the combination of the discrete wavelet transform using a Gaussian window with an artificial neural network for condition monitoring of three- phase induction motors. This approach consisted of simulating faulty and flawless signals using software developed in LabVIEW, their processing, appropriate choice of signals, establishing statistical measures of the chosen signs, and forming the input vectors presented to the artificial neural network. The input vectors are constituted based on statistical measures involving measures of central tendency (mean and centroid), measures of dispersion (RMS value and standard deviation), and a measure of asymmetry (Kurtosis). The most promising configuration was the Multiple Perceptron Layer (MPL) network with four hidden layers containing 256 neurons. Such network showed satisfactory performance for both mechanical failures, with a correct range of around 97%. These results proved to be very effective for detecting mechanical failures, thus being an auxiliary instrument in predictive maintenance.