Desenvolvimento de classificadores fuzzy dedicados à manutenção de locomotivas, baseado nas técnicas de Manutenção Centrada na Confiabilidade

Abstract

Maintenance, in its new era, is based on the inclusion of new strategies, which aim to accurately assess the vital parameters of certain equipment, enabling dedicated interventions in order to optimize their reliability indexes. We highlight the methodologies attributed to the aid of decision making, aiming at the optimization of resources according to the reliability required for each production process, such as maintenance centered on reliability (RCM), cause and effect analysis (FMEA) and analysis root cause (RCA). The aim of this study was to unify the concepts of RCA, FMEA and RCM in a single structure, in order to establish criteria for the development of an intelligent algorithm, dedicated to the detection of certain failure modes installed in a group of equipment. For this, two stages were used. The first consisted of selecting the most representative and likely occurring disturbances in a group composed of 47 diesel-electric locomotives, based on FMEA and RCA methodologies. In all, 32 failure modes were identified, and through a selection process, based on the RCM method, four were designated for the development of an intelligent system dedicated to their detection. The second step, consisted of the elaboration of a classifier based on the fuzzy C-means clustering techniques, aimed at detecting and classifying these occurrences. The development of the algorithm was based on six stages: data acquisition, data pre-processing, feature extraction, training, validation, testing and classification. Data acquisition was performed using embedded equipment installed in the analysis group, designed to collect 26 different operational parameters. Pre-processing was responsible for reducing the dimensionality of the database via the selection of useful parameters. Characteristic extraction was used to define characteristics relevant to the representativeness of each failure mode. The training and validation stage was applied to create a model that best represents the study information. And the testing step, consisted of testing the efficiency of the clusters for a later classification of the failure modes by means of labels and thresholds granted to them. The algorithm was successful for classification in 3 of the 4 failure modes selected, presenting an accuracy level of 92%. The use of such methodologies, in order to establish criteria for the development of the algorithm, provided the project with a certain logical basis, converting the large amount of information provided by the productive environment, often inaccurate if analyzed in isolation, in determining fundamentals for its construction. It is concluded that maintenance methodologies have high functionality, when attributed to the development of algorithms dedicated to equipment reliability.