Uncertainty-aware RUL and maintenance decision economics

"My point estimate said 60 cycles. It never mentioned it was a coin flip between 20 and 120. The maintenance planner found out the hard way."

A Newly Humble AI Agent

This section leans on the probabilistic vocabulary of Chapter 4 (predictive distributions, aleatoric versus epistemic uncertainty) and on the calibration and conformal machinery of Chapter 18, which is what makes an RUL interval trustworthy rather than decorative. It takes the point-RUL models of Section 36.4 as given and asks the operational question they leave open: given what we predict and what we know we do not know, what should we do?

From a point to a predictive RUL distribution

A usable prognostic does not output \(\widehat{\text{RUL}}=60\); it outputs a distribution \(p(\text{RUL}\mid x_{1:t})\) over the cycles of life remaining given everything observed so far. Two sources of spread live inside that distribution and they behave differently. Aleatoric uncertainty is irreducible physical variability: two nominally identical pumps fed identical loads still fail at different cycles because of manufacturing scatter and micro-differences in operating history. More data never shrinks it. Epistemic uncertainty is the model's own ignorance: a degradation pattern it has seen only twice, or an operating regime outside its training envelope. This one does shrink with data, and it is the component that spikes precisely when a machine drifts out of distribution, which is exactly when a maintenance mistake is most likely.

The practical ways to obtain the distribution mirror the toolkit of Chapter 18: a deep ensemble of RUL regressors whose disagreement estimates epistemic spread; Monte Carlo dropout as a cheap ensemble surrogate; a model with a heteroscedastic head that predicts both a mean and a variance for aleatoric noise; or a quantile regressor that outputs the 10th, 50th, and 90th RUL percentiles directly. Whatever the source, the object we carry into the decision is the same: a set of RUL samples or quantiles, not a single number.

The maintenance decision as expected-cost minimization

Now attach money. Let \(C_f\) be the cost of a corrective action: an unplanned, in-service failure, including collateral damage, emergency labor, and lost production. Let \(C_p\) be the cost of a preventive action: a planned replacement of a still-working part, wasting whatever life remained. Almost universally \(C_f \gg C_p\), and the whole game is spending a little \(C_p\) to avoid a large \(C_f\) with the right probability. Suppose at the current cycle we may either replace now (cost \(C_p\)) or defer and re-examine after a horizon \(\Delta\) (the inspection interval, or the procurement lead time before a part could even arrive). Deferring risks a failure within that window with probability \(P(\text{RUL}\le\Delta)\), computed from the predictive distribution. Comparing expected costs, deferral is worth it only while

That ratio \(C_p/C_f\) is a critical failure probability: replace as soon as the estimated chance of failing before the next opportunity to act exceeds it. If a preventive swap costs one-twentieth of a failure, you should tolerate up to a 5% failure probability before acting, no more. Notice what sets the threshold: not the RUL midpoint, but the cost asymmetry. A cheap-to-replace, catastrophic-to-fail component (small ratio) triggers early and conservatively; an expensive part with benign failure (large ratio) is run closer to the edge. The predictive distribution enters only through the tail probability \(P(\text{RUL}\le\Delta)\), which is why a well-calibrated lower quantile matters far more than an accurate mean.

For a component we manage over its whole life rather than at one instant, the same ingredients give the classic age-replacement policy from renewal theory: choose a planned replacement age \(\tau\) minimizing long-run cost per unit time,

where \(R(\tau)\) is the reliability (survival) function connecting this directly to the hazard models of Section 36.3. The numerator is the expected cost of one renewal cycle (you either reach \(\tau\) and pay \(C_p\), or fail first and pay \(C_f\)); the denominator is the expected length of that cycle. The code below evaluates both rules on a predictive RUL sample.

import numpy as np
rng = np.random.default_rng(0)

# Predictive RUL distribution at the current cycle, e.g. from a deep ensemble
# or quantile regressor (Chapter 18). Here: a right-skewed posterior in cycles.
rul = rng.lognormal(mean=np.log(60), sigma=0.45, size=50_000)

C_p, C_f = 20.0, 100.0            # preventive vs corrective cost
critical = C_p / C_f             # tolerated failure prob before acting

# (1) Act-now-vs-defer rule over a procurement/inspection horizon of 15 cycles.
horizon = 15
p_fail_soon = np.mean(rul <= horizon)
print(f"critical ratio C_p/C_f = {critical:.2f}")
print(f"P(fail within {horizon} cy) = {p_fail_soon:.3f}  -> "
      f"{'REPLACE NOW' if p_fail_soon > critical else 'defer, re-check later'}")

# (2) Age-replacement: cost-per-cycle-optimal planned replacement age.
def cost_rate(tau):
    R = np.mean(rul > tau)                       # reliability at tau
    exp_cycle_len = np.mean(np.minimum(rul, tau))
    return (C_p * R + C_f * (1 - R)) / exp_cycle_len

taus = np.arange(5, 120)
rates = np.array([cost_rate(t) for t in taus])
tau_star = taus[int(np.argmin(rates))]
print(f"cost-rate-optimal replacement age = {tau_star} cy, "
      f"failure prob by then = {np.mean(rul <= tau_star):.2f}")

Running it, the cost-rate-optimal age lands well below the median RUL: because \(C_f\) is five times \(C_p\), the policy deliberately replaces parts with life still in them to buy down failure risk. Slide \(C_f\) up and \(\tau^\star\) marches earlier; that single knob, not the model, is what a reliability engineer actually tunes.

Calibration and conformal guarantees make the economics trustworthy

The expected-cost rules above are only as honest as the probability \(P(\text{RUL}\le\Delta)\) they consume. If the model's predictive distribution is overconfident, that tail probability is understated, the critical-ratio test fires too late, and the "optimal" policy quietly runs machines to failure. Calibration is therefore not a nicety here; it is load-bearing. The discipline from Chapter 18 applies directly: check that stated RUL intervals have their claimed coverage on held-out, unit-disjoint data (a 90% interval should contain the true RUL 90% of the time), and recalibrate if not.

Conformal prediction is especially attractive for RUL because it wraps any point regressor, ensemble, or foundation-model embedding (Section 36.5) and returns intervals with a finite-sample, distribution-free coverage guarantee, given exchangeable calibration data. The catch, and it is a real one for prognostics, is that degradation trajectories are not exchangeable: a unit late in its life is systematically different from an early one. Naive conformal intervals therefore under-cover in the danger zone, so time-aware and adaptive variants are used, and any coverage claim must be validated on the failure-imminent regime specifically, not just pooled across all cycles.

Finally, calibrated cost-aware RUL is what lets higher-level automation act safely. An operations agent (Chapter 22) that raises work orders, or a fleet scheduler (Section 36.7) that batches replacements across a plant, both consume the expected-cost decision and its uncertainty rather than a raw number; feeding them an uncalibrated point RUL is how automation confidently schedules the wrong thing at scale.

What's Next

In Section 36.7, we scale this single-unit decision to a whole fleet: thousands of assets producing RUL distributions at once, shared spare-parts pools and crew constraints that couple their decisions, and the deployment and monitoring plumbing (drawing on Chapter 69) that keeps a prognostic model calibrated as the fleet, its machines, and its operating conditions all drift over years of service.