Bearing faults classification of induction motor using advanced deep learning techniques

Abstract

This study investigates the application of advanced neural network models for bearing fault detection using vibration and current signals. Bearing faults in induction motors pose significant challenges to industrial operations, often leading to unexpected downtimes and increased main-tenance costs. The study explores the performance of Artificia lNeura lNetwork s(ANN) ,1D- Convolutional Neural Networks (1D-CNN), 1D-CNN with multi-kernel sizes, and Long Short-Term Memory (LSTM) models. Findings indicate that 1D-CNN and its multi-kernel size variant outper-form other models, achieving accuracies up to 99.95% under various load conditions for vibration data. The 1D-CNN multi-kernel size model’s ability to capture diverse features through different kernel sizes proved advantageous, reflectin g asignifica ntimproveme ntov erprevio usmethodologies that relied on extensive preprocessing.For the current signal dataset,Our recent finding ssurpas sall prior results, particularly in variable speed operation, where our work marks a pioneering effort. In our current signal dataset, the pinnacle of accuracy, reaching 99.88%, was attained through the application of the 1D-CNN model with the variable load operation dataset. This remarkable success highlights the effectivenes so fmergin g1D-CN Nwit hVariationa lMod eDecompositio n(VMD) ,en- abling the proficien tdecompositio no fsignal san dresolutio no fboundar yeffec ts tohand leintricate fault patterns. Despite encountering greater complexities in variable speed operation, our models persevered and achieved commendable accuracies. Notably, the 1D-CNN model achieved an accuracy of up to 99.36%. These results highlights the significan tachievemen tmad ei nterm so fdiagnosi si ninduction motors.