Prediction failure in electric motors bearings using vibration signals and Long Short Term-memory neural networks

Resumo

Bearing failures modify the vibration regime of electric motors. The acquisition and analysis of these
signals may provide important information about the operating condition of these components. In this context, the
use of failure prediction techniques can ensure the rolling bearings will be always in good operating condition,

ensuring the production processes continuity and avoiding accidents. This paper investigates the subject of fail-
ure prediction in rolling bearings from vibration signals using Long Short-Term Memory (LSTM) networks. The

experimental was carried out on the vibration signals from the data set IMC. The method consisted of building 6

models in different training settings by using either raw data or 13 statistical descriptors in time domain. Perfor-
mance evaluation was accomplished by means of accuracy, precision, sensibility, sensitivity and F1-Score. The

best result (92% of accuracy, 94% of precision, 86% of sensibility, 94% of specificity and 89% of F1-score), indi-
cates the use of LSTM aiming to predicting failures in rolling bearings can improve the reliability of production

systems, by anticipating preventive actions and reducing the need for corrective maintenance.