Classificador por votação baseado em otimização por enxame de partículas relativísticas para a detecção de falhas simples e combinadas em máquinas elétricas rotativas

Abstract

With the advance in the process of automating industries, various methodologies for identifying faults in rotating electrical machines are being proposed. A fault in this equipment leads to losses in productivity, increased costs and the risk of accidents on the part of operators. This paper proposes a new methodology for diagnosing both simple and compound faults in rotating electrical machines. To implement this new approach, a weighted voting hybrid classifier is created that combines the predictions of the following algorithms: multilayer perceptron neural network, support vector machine, k-nearest neighbors, random forest, gradient boosting and lightGBM. In addition, relativistic particle swarm optimization is used to find the weights for each classifier. The algorithm is applied to three databases that contain some of the main types of faults that occur in these devices, including unbalance, misalignment, bearing faults and short-circuit windings. The mechanical faults identified in this work are isolated and combined. The results show the robustness of the proposed technique, which achieved accuracy results of 88.68% for the first database, 100% for the second and 98.95% for the third. The results, when compared with other studies in the literature, were fair.