Quasi-Stationarity of Discrete-Time Markov Chains with Drift to Infinity

We consider a discrete-time Markov chain on the non-negative integers with drift to infinity and study the limiting behavior of the state probabilities conditioned on not having left state 0 for the last time. Using a transformation, we obtain a dual Markov chain with an absorbing state such that absorption occurs with probability 1. We prove that the state probabilities of the original chain conditioned on not having left state 0 for the last time are equal to the state probabilities of its dual conditioned on non-absorption. This allows us to establish the simultaneous existence, and then equivalence, of their limiting conditional distributions. Although a limiting conditional distribution for the dual chain is always a quasi-stationary distribution in the usual sense, a similar statement is not possible for the original chain.     

Adaptive age replacement strategies based on nonparametric predictive inference

We consider an age replacement problem using nonparametric predictive inference (NPI) for the lifetime of a future unit. Based on n observed failure times, NPI provides lower and upper bounds for the survival function for a future lifetime X_n+1, which are lower and upper survival functions in the theory of interval probability, and which lead to upper and lower cost functions, respectively, for age replacement based on the renewal reward theorem. Optimal age replacement times for X_n+1 follow by minimizing these cost functions. Although the renewal reward theorem implicitly assumes that the corresponding optimal strategy will be used for a long period, we study the effect on this strategy when the observed value for X_n+1, which is either an observed failure time or a right-censored observation, becomes available. This is possible due to the fully adaptive nature of our nonparametric approach, and the next optimal strategy will be for X_n+2. We compare the optimal strategies for X_n+1 and X_n+2 both analytically and via simulation studies. Our NPI-based approach is fully adaptive to the data, to which it adds only few structural assumptions. We discuss the possible use of this approach, and indeed the wider importance of the conclusions of this study to situations where one wishes to combine the statistical aspects of estimating a lifetime distribution with the more traditional operational research approach of determining optimal replacement strategies for lifetime distributions that are assumed to be known.     

Nonparametric adaptive opportunity-based age replacement strategies

We consider opportunity-based age replacement (OAR) using nonparametric predictive inference (NPI) for the time to failure of a future unit. Based on n observed failure times, NPI provides lower and upper bounds for the survival function for the time to failure X_n+1 of a future unit which lead to upper and lower cost functions, respectively, for OAR based on the renewal reward theorem. Optimal OAR strategies for unit n+1 follow by minimizing these cost functions. Following this strategy, unit n+1 is correctively replaced upon failure, or preventively replaced upon the first opportunity after the optimal OAR threshold. We study the effect of this replacement information for unit n+1 on the optimal OAR strategy for unit n+2. We illustrate our method with examples and a simulation study. Our method is fully adaptive to available data, providing an alternative to the classical approach where the probability distribution of a unit's time to failure is assumed to be known. We discuss the possible use of our method and compare it with the classical approach, where we conclude that in most situations our adaptive method performs very well, but that counter-intuitive results can occur.     

Quasi-stationary Distributions for a Class of Discrete-time Markov Chains

This paper is concerned with the circumstances under which a discrete-time absorbing Markov chain has a quasi-stationary distribution. We showed in a previous paper that a pure birth-death process with an absorbing bottom state has a quasi-stationary distribution—actually an infinite family of quasi-stationary distributions— if and only if absorption is certain and the chain is geometrically transient. If we widen the setting by allowing absorption in one step (killing) from any state, the two conditions are still necessary, but no longer sufficient. We show that the birth–death-type of behaviour prevails as long as the number of states in which killing can occur is finite. But if there are infinitely many such states, and if the chain is geometrically transient and absorption certain, then there may be 0, 1, or infinitely many quasi-stationary distributions. Examples of each type of behaviour are presented. We also survey and supplement the theory of quasi-stationary distributions for discrete-time Markov chains in general.      

Bayesian reliability demonstration with multiple independent tasks

^** Corresponding author. Email: frank.coolen{at}durham.ac.uk We consider optimal testing of a system in order to demonstratereliability with regard to its use in a process after testing,where the system has to function for different types of tasks,which we assume to be independent. We explicitly assume thattesting reveals zero failures. The optimal numbers of tasksto be tested are derived by optimisation of a cost criterion,taking into account the costs of testing and the costs of failuresin the process after testing, assuming that such failures arenot catastrophic to the system. Cost and time constraints ontesting are also included in the analysis. We focus on studyof the optimal numbers of tests for different types of tasks,depending on the arrival rate of tasks in the process and thecosts involved. We briefly compare the results of this studywith optimal test numbers in a similar setting, but with analternative optimality criterion which is more suitable in caseof catastrophic failures, as presented elsewhere. For thesetwo different optimality criteria, the optimal numbers to betested depend similarly on the costs of testing per type andon the arrival rates of tasks in the process after testing.     

Non-parametric predictive reliability demonstration for failure-free periods

We derive sample sizes for testing as required for reliabilitydemonstration, using non-parametric predictive inference forthe predicted number of future failures on the basis of testinformation. We assume that tests lead to zero failures, asis typical, for example, for high-reliability testing. Optimizationis in terms of failure-free periods for a process after testing,and we also consider total expected cost minimization, withparticular attention to deterministic and Poisson processes.We show that, perhaps surprisingly, the deterministic case isa worst-case scenario with regard to number of tests needed.     

Condition monitoring: a new perspective

We present a novel approach to support preventive replacement decisions based on condition monitoring. It is assumed that the condition of a technical unit is measured continuously, with the measurement indicating in which of k≥2 states the unit is. A new unit is in state S₁, and failure occurs the moment a unit leaves state S _k . A unit will only go from state S _j to state S_j+1, and these immediately observed transitions occur at random times. At such transition moments a decision is required on preventive replacement of the unit, which would require preparation time. Such a decision needs to be based on inferences about the total residual time till failure. We present a method for such inferences that relies on minimal model assumptions added to data on state transitions. We illustrate the approach via simulated examples, and discuss possible use of the method and related aspects.     

Opportunity-based age replacement with a one-cycle criterion

In this paper age replacement (AR) and opportunity-based age replacement (OAR) for a unit are considered, based on a one-cycle criterion, both for a known and unknown lifetime distribution. In the literature, AR and OAR strategies are mostly based on a renewal criterion, but in particular when the lifetime distribution is not known and data of the process are used to update the lifetime distribution, the renewal criterion is less appropriate and the one-cycle criterion becomes an attractive alternative. Conditions are determined for the existence of an optimal replacement age T^* in an AR model and optimal threshold age T_opp^* in an OAR model, using a one-cycle criterion and a known lifetime distribution. In the optimal threshold age T_opp^*, the corresponding minimal expected costs per unit time are equal to the expected costs per unit time in an AR model. It is also shown that for a lifetime distribution with increasing hazard rate, the optimal threshold age is smaller than the optimal replacement age. For unknown lifetime distribution, AR and OAR strategies are considered within a nonparametric predictive inferential (NPI) framework. The relationship between the NPI-based expected costs per unit time in an OAR model and those in an AR model is investigated. A small simulation study is presented to illustrate this NPI approach.