Similarity-based methods for machine diagnosis

• Ano de defesa: 2018
• Autor(a) principal: Ribeiro, Felipe Moreira Lopes
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal do Rio de Janeiro; Brasil; Instituto Alberto Luiz Coimbra de Pós-Graduação e Pesquisa de Engenharia; Programa de Pós-Graduação em Engenharia Elétrica; UFRJ
• Palavras-chave em Português: Engenharia elétrica; Diagnóstico; Detecção de falhas de máquinas; Aprendizado de máquina; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA
• Link de acesso: http://hdl.handle.net/11422/11671

This work presents a data-driven condition-based maintenance system based on similarity-based modeling (SBM) for automatic machinery fault diagnosis. The proposed system provides information about the equipment current state (degree of anomaly), and returns a set of exemplars that can be employed to describe the current state in a sparse fashion, which can be examined by the operator to assess a decision to be made. The system is modular and data-agnostic, enabling its use in different equipment and data sources with small modifications. The main contributions of this work are: the extensive study of the proposition and use of multiclass SBM on different databases, either as a stand-alone classification method or in combination with an off-the-shelf classifier; novel methods for selecting prototypes for the SBM models; the use of new similarity functions; and a new production-ready fault detection service. These contributions achieved the goal of increasing the SBM models performance in a fault classification scenario while reducing its computational complexity. The proposed system was evaluated in three different databases, achieving higher or similar performance when compared with previous works on the same database. Comparisons with other methods are shown for the recently developed Machinery Fault Database (MaFaulDa) and for the Case Western Reserve University (CWRU) bearing database. The proposed techniques increase the generalization power of the similarity model and of the associated classifier, having accuracies of 98.5% on MaFaulDa and 98.9% on CWRU database. These results indicate that the proposed approach based on SBM is worth further investigation.