Repairable system analysis in presence of covariates and random effects

https://doi.org/10.1016/j.ress.2014.04.009Get rights and content

Highlights

  • We describe the failure process of buses powertrain system subject to heterogeneity.

  • Heterogeneity due to different service types is explained by a covariate.

  • Random effect is modeled through a joint pdf on failure process parameters.

  • The powertrain reliability under new future operating conditions is estimated.

Abstract

This paper aims to model the failure pattern of repairable systems in presence of explained and unexplained heterogeneity. The failure pattern of each system is described by a Power Law Process. Part of the heterogeneity among the patterns is explained through the use of a covariate, and the residual unexplained heterogeneity (random effects) is modeled via a joint probability distribution on the PLP parameters. The proposed approach is applied to a real set of failure time data of powertrain systems mounted on 33 buses employed in urban and suburban routes. Moreover, the joint probability distribution on the PLP parameters estimated from the data is used as an informative prior to make Bayesian inference on the future failure process of a generic system belonging to the same population and employed in an urban or suburban route under randomly chosen working conditions.

Introduction

Most of the papers on the reliability analysis of repairable systems deal with the homogeneous case although homogeneous populations and homogeneous operating conditions can be hardly found in practice [1]. Indeed, in many cases a substantial heterogeneity is observed among the failure patterns of apparently identical repairable systems, due to the presence of differences among units characteristics and/or differences in operating and environmental conditions. Part of such heterogeneity can be frequently explained by the presence of assignable causes (often called explanatory variables or covariates). The residual part (often called the unobserved heterogeneity or random effects) remains rather unexplained, since it is caused by unassignable (unidentified or not recorded) factors. Refs. [1], [2], [3], [4], [5], [6], [7] considered, among others, heterogeneity in non-homogeneous Poisson processes, whereas Lindqvist et al. [8] discussed heterogeneity in the more general case of trend-renewal processes.

An interesting example of the joint presence of explained heterogeneity and random effects is provided by the powertrain system of a fleet of buses employed in urban and suburban routes of Naples, Italy. Routes differ among themselves mainly in the traffic conditions: suburban service is characterized by noticeable thinner traffic, with a more regular use of the engine and a less frequent use of the gearbox with respect to the urban service, so that service type is expected to have some significant effect on the powertrain reliability. However, the large variability observed in the failure patterns (depicted in Fig. 1, Fig. 2, for urban and suburban service, respectively) is only partially explained by service type. Other factors which potentially affect the reliability of the system (inherent characteristics of each system and/or not recorded unit specific service characteristics, such as the mean slope of the roads, the mean payload, etc.) leave unexplained part of the existent heterogeneity.

This paper introduces and studies a method to analyze the failure pattern of the powertrain systems in presence of identified and unidentified sources of heterogeneity. In particular, the failure pattern of each individual system is described by a Power Law Process (PLP) [9], [10], that is, the non-homogeneous Poisson process with failure intensityλ(t)=βα(tα)β1,α,β>0

The heterogeneity among the patterns is then modeled through the combined use of:

  • (a)

    a joint probability distribution g(α,β) on the PLP parameters α and β that models the random effects, and

  • (b)

    a dummy covariate x, that defines service type (urban or suburban service) and acts only on the distribution of the scale parameter α, via the covariate parameter q.

The proposed joint distribution g(α,β) is indexed by five hyperparameters θ=(a,b, c,d,y) in order to allow the first two marginal moments of α and β and the cross moment E{αβ} to be assessed each independently of the others.

The hyperparameters θ and the covariate parameter q, are estimated via two different maximum likelihood procedures: (a) the herein-called 2-step procedure, which first estimates the PLP parameters α and β of each bus and then estimates θ and q by treating the obtained ML estimates of α and β as they were the true (unit specific) values of the considered unknown parameters; and (b) the herein-called 1-step procedure, which estimates θ and q directly from the observed failure data. The hypotheses of presence of explained heterogeneity and random effects against the null hypothesis of common parameters across all the systems are checked via the likelihood ratio testing procedure.

Based on the ML estimates of θ and q, a joint distribution on the PLP parameters of a generic powertrain system belonging to the same population and employed in an urban or suburban route under randomly chosen working conditions is formulated. This informative distribution is then used in a Bayesian framework to make inference on the failure process of a future system, in conjunction with its early failure data. Finally, to make the potential users more confident in the effectiveness of the proposed approach, a comparison between the results obtained adopting the proposed informative Bayesian procedure and those obtained without using prior information closes the work.

Section snippets

Powertrain data and preliminary heterogeneity analyses

The application refers to the failure data of the powertrain system of m=33 buses built by IVECO and Breda Menarinibus, put into service in the early months of 1999 and observed until December 31, 2004. The main design characteristics (weight, size, maximum load capacity) of the IVECO and Breda Menarinibus buses were nearly identical. Again, all the buses mounted an identical copy of the powertrain (that includes a FIAT engine, a transmission with ZF gearbox, driveshafts, and differentials).

The

Joint distribution on the PLP parameters

In order to model the unidentified heterogeneity among buses undergoing a same service type (urban or suburban) the PLP parameters α and β are assumed to be random variables which are given the following joint densityg(α,β)=badcΓ(a)Γ(c)βaycβαcβ+1exp(bβ)exp[d(yα)β]

The proposed distribution (2), which is a reparameterized form of the joint density previously proposed by Guida and Pulcini [12] in a Bayesian estimation context, is indexed by five parameters θ=(a,b,c,d,y), so that the first two

Modeling and estimating the assignable heterogeneity

The analyses performed in Section 3 confirm what it was observed at the end of Section 2: the random scale parameter α is strongly affected by service type, which seems to act on α through a multiplicative factor (the two coefficients of variation of α are quite similar). On the contrary, the random shape parameter β is not appreciably affected by the type of service. Thus, a regression model is used, which assumes that the type of service, modeled via a dummy covariate z, acts only on the

Bayes inference on future system

The joint distribution g(α0,β) can be used in a Bayesian framework as an informative prior density on the PLP parameter for a generic powertrain system belonging to the same population and employed in an urban or suburban route under randomly chosen working conditions. This informative distribution can be used in conjunction with the early failure data of a future system to make inference on its failure process.

Let TF denote the operating time of a future system and let τ1<τ2<<τnFTF be the

Conclusions

The model proposed in this paper showed to be able to describe the heterogeneity in the intensity function of the powertrain system of a fleet of 33 buses employed in urban or suburban service. Part of the existing heterogeneity is explained by a covariate representing service type, while the remaining part of the heterogeneity, due to different working conditions, is modeled through a joint probability distribution on the parameters which index the adopted stochastic process, namely the PLP.

References (14)

There are more references available in the full text version of this article.

Cited by (0)

View full text