Maintenance Reliability Through Regenerative Cycles

From Repair Cycles to Long-Run Performance

A repairable system repeatedly operates, fails, and returns to service. Its long-run performance depends on both parts of that cycle: a system with long operating periods may still have poor availability if repairs take a long time, while a system with brief repairs may incur frequent failure costs. Regenerative cycles provide a direct way to calculate availability, failure frequency, and average cost from the behavior of one representative cycle.

Work on a probability space \((\Omega,\mathcal{F},\mathbb{P})\). Let \(Y_t\in\{0,1\}\), \(t\ge0\), be the equipment status, where \(Y_t=1\) means operating and \(Y_t=0\) means under repair. Assume the system starts at time \(0\) just as it enters service. Cycle \(k\) consists of an operating duration \(U_k\), followed by a repair duration \(D_k\), after which the next cycle starts. The cycle length is \(C_k=U_k+D_k\). Assume the pairs \((U_k,D_k)\), \(k\ge1\), are independent and identically distributed, \(C_k>0\) almost surely, and \(0<\mathbb{E}C_1<\infty\). Dependence between \(U_k\) and \(D_k\) within the same cycle is allowed. Each cycle contains exactly one failure, at the end of its operating period.

Write \(N_f(t)\) this way so it is distinct from a cycle count: a cycle is complete only after repair, whereas a failure occurs at the end of operation. The two counts differ by at most one, so they have the same long-run rate when that rate exists. That fact is also justified below.

The Cycle-Level Formulas

The renewal-reward theorem states that for independent, identically distributed cycles with positive finite mean length and integrable total reward per cycle, the cumulative reward per unit time converges almost surely to expected reward per cycle divided by expected cycle length. For availability, the reward is operating time within a cycle, \(U_k\); for failure frequency, it is one failure per cycle. The availability reward is integrable because \(0\le U_k\le C_k\) and \(\mathbb{E}C_k<\infty\).

Proof. Let \(S_n=\sum_{k=1}^n C_k\) be the time at which cycle \(n\) is complete, and let \(M(t)=\max\{n:S_n\le t\}\) be the number of completed cycles by time \(t\). The renewal-reward theorem applied to cycle rewards \(U_k\) gives the long-run operating-time fraction \(\mathbb{E}U_1/\mathbb{E}C_1\). Applied to reward \(1\) per completed cycle, it gives \(M(t)/t\to1/\mathbb{E}C_1\) almost surely. By time \(t\), failures have occurred in the completed cycles and possibly in the current, incomplete cycle; thus \(M(t)\le N_f(t)\le M(t)+1\). Dividing by \(t\) shows \(N_f(t)/t\) has the same limit as \(M(t)/t\). Finally, linearity of expectation gives \(\mathbb{E}C_1=\mathbb{E}U_1+\mathbb{E}D_1\), proving both formulas. \(\square\)

The ratios use expected cycle quantities, not an unweighted average of each cycle’s operating fraction. The denominator is calendar time: both operation and repair contribute to cycle length. In particular, the failure frequency is not the reciprocal of mean operating time unless repair duration is zero.

Proof. Since \(\mathbb{E}U_1>0\) and \(\mathbb{E}D_1\ge0\), availability is \(\mathbb{E}U_1/(\mathbb{E}U_1+\mathbb{E}D_1)\), with positive denominator. Multiplying \(A>a\) by that denominator and dividing by the positive number \(a\mathbb{E}U_1\) preserves the inequality and gives \(\mathbb{E}D_1/\mathbb{E}U_1<(1-a)/a\). Reversing these algebraic steps proves the converse. The strict inequality in the conclusion corresponds to exceeding, not merely meeting, the threshold. \(\square\)

Worked Maintenance Calculations

Scope and Common Pitfalls

The calculation depends on genuine regenerative cycles: after each repair, the next cycle must have the same joint distribution of operating and repair durations and be independent of earlier cycles. If a repair does not restore the equipment to a state with this property, the cycle formulas cannot be used without modifying the model. A finite mean cycle length is also essential to the stated renewal-reward limits. For example, if operating duration is always \(1\) but repair durations have an infinite mean, then \(\mathbb{E}C_1\) is infinite and the displayed finite-mean formulas do not apply; one must not substitute an infinite value and treat the resulting expression as a justified rate.

Independence of operating and repair duration within a cycle is not required for the formulas: only the expected operating reward and expected total cycle length enter. By contrast, independence and identical distribution across cycles are part of the regenerative assumption used by the renewal-reward theorem. If cycle laws change over time, a ratio computed from one cycle’s means need not describe a long-run limit. For instance, in a deterministic sequence with \(U_k=1\) and \(D_k=k\), after \(n\) complete cycles the total operating time is \(n\), while total time is \(n+n(n+1)/2\); their ratio tends to zero, and the cycles are not identically distributed.

Three checks help prevent common errors. First, include repair time in the calendar-time denominator. Second, use one failure reward per cycle only when the model really has one failure per cycle. Third, divide expected cycle reward by expected cycle length; do not average cycle-specific ratios unless cycle lengths are equal or a separate argument justifies that calculation. Availability is a long-run time fraction, not automatically the probability of being up at an arbitrary time from a specified initial condition; the regenerative limit describes the long-run fraction under the hypotheses stated above.