Detecção de falhas de curto-circuito em motores de indução trifásicos usando classificadores baseados em protótipos

• Ano de defesa: 2019
• Autor(a) principal: Sousa, Diego Perdigão
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Teleinformática; Motores; Localização de falhas (Engenharia); Prototype-based learning; Three-phase induction motor; Failures detection
• Link de acesso: http://www.repositorio.ufc.br/handle/riufc/49906

Three-phase induction motors are one of the most important equipment of modern industry. However, in many situations, these equipment are subject to inappropriate conditions such as in environments with high temperatures and humidity, abrupt variations of load above specified, excessive vibrations, among others. These conditions make motors more susceptible to various failures, whether external or internal, which are obviously undesirable in industrial processes. In this context, the predictive maintenance plays a relevant role, where the detection and correct diagnosis of failures in a timely manner leads to increasing the useful life of the motor and, consequently, to the reduction of costs with production stoppage due to corrective maintenance. Considering these factors, this dissertation proposes a methodology for detecting short-circuit failures in three-phase induction motors, which involves prototypes-based algorithms. To this end, both unsupervised techniques - such as the K-means and supervised algorithm, such as the LVQ (Learning Vector Quantization) family classifiers are used. The methodology starts with the seeking of the optimal number of prototypes from the unsu- pervised analysis of clusters and techniques clustering validation. Then, the prototypes that were found are used in the supervised training of various classifiers of the LVQ family. The influence that each type of clustering validation criterion exerts on the various LVQ classifiers implemented is deeply evaluated. In particular, the GRLVQ (Generalized Relevance Learning Vector Quantization) classifier obtained the best results where it presented a maximum classifica- tion rate of 98.3%, with the Dunn and Silhouette criteria standing out as the most efficient in determining the optimal quantity of prototypes.