An investigation of rail bearing reliability under real conditions of use

• Published in: Engineering failure analysis 2003-12, Vol.10 (6), p.745-758
• Main Authors: Ferreira, J.L.A, Balthazar, J.C, Araujo, A.P.N
• Format: Article
• Language: English
• Subjects: Bearing failures; Fatigue; Life prediction; Railway engineering; Reliability analysis
• ISSN: 1350-6307; 1873-1961
• DOI: 10.1016/S1350-6307(02)00052-3

The objective of the present work is to carry out a performance analysis of roller bearings used in railway ore transportation wagons. Data from 47,000 failed bearings divided in seven groups, were used to determine the failure distribution. Six groups corresponded to operation in the years 1985–1988, 1990, 1991 and one group included all the others. The failure distribution in bearings was described by the three parameter Weibull distribution. To conduct the statistical analysis it was necessary to estimate these parameters by the nonlinear regression method, because it does not require the whole group of failed bearings, and direct inference on the sample. A Monte Carlo technique was used to validate the procedure. Analytically two estimate methodologies were considered. Methodology I is a well established technique while Methodology II is an alternative approach proposed by the authors, which was inspired by a procedure used to design rail shafts. Studying the bearing failure behavior led to the following observations: based on sample inference techniques, an excessive variation was observed in the parameters that characterize the failure distribution. A dispersion of the order of 40 or 25% was found in the minimum and nominal lives when evaluated by the direct inference method or by the nonlinear regression method, respectively, and a dispersion of 17% calculated by nonlinear regression was observed for the shape parameter, β ̂ . In spite of this dispersion, it was found that the minimum lives estimated by direct inference were conservative by a factor of four, when compared to analytical methods, while estimates of the nominal life were shown to be similar to its smallest value observed in the sample groups.