Prognostics and Remaining Useful Life

Section 01

From forecasting to prognostics

The first monograph in this series turned a sensor into ordered numbers. The second learned to forecast those numbers forward. This one asks the question that maintenance engineers actually lose sleep over: given everything the machine has told us so far, how much longer will it run before it fails? That answer is the remaining useful life, or RUL, and computing it well is the discipline of prognostics.

The core move is simple to state. Distil the machine's condition into a single number that rises (or falls) steadily as the machine wears, the health indicator. Decide the value of that number at which the machine is considered failed, the failure threshold. Then forecast the health indicator forward and read off where it crosses the threshold. The gap between now and that crossing is the RUL. Prognostics, in other words, is forecasting with a finish line.

Third in a series

This page assumes the vocabulary of its companions. Time Series as Sensory Data built features, residuals and the Kalman filter; Forecasting Time Series built point forecasts, prediction intervals and conformal coverage. Here those threads converge: a health indicator is a feature, its forecast is a RUL, and its interval is the difference between a confident maintenance plan and a guess.

What makes prognostics distinct from ordinary forecasting is the asymmetry of being wrong. Forecast tomorrow's temperature two degrees high or low and the cost is symmetric. Forecast a turbine's failure two weeks late and you have a catastrophe; two weeks early and you have wasted life and an unnecessary teardown. Every method here is shadowed by that asymmetry, which is why the uncertainty around a RUL matters at least as much as the RUL itself.

Section 02

Building a health indicator

Before you can forecast a failure, you need something that actually trends toward it. A raw vibration or temperature channel rarely does; it is dominated by the operating point, not the slow march of damage. The art of constructing a health indicator (HI) is to combine features, often many of them, into one curve with three desirable properties.

Monotonicity

The HI should move in one direction as damage accumulates, not wander up and down. Wear does not heal, so a good indicator does not either.

Trendability

The HI should change by enough over a unit's life to be measurable above the noise. A flat indicator carries no prognostic information.

Prognosability

Units should fail at similar HI values, so a single threshold is meaningful across the fleet rather than unit by unit.

In practice the HI is often a fusion: a weighted blend of RMS vibration, kurtosis, specific bearing-fault band energies, temperatures and currents, sometimes learned by a small model trained to score these three properties. The companion pages give you the raw materials; the skill here is choosing and combining them so the resulting curve is a clean, climbing thing a forecaster can grab onto. Everything that follows assumes you have done this work and hands you a usable HI.

Section 03

Degradation models

The most direct approach fits a parametric curve to the health indicator so far and extrapolates it to the threshold. The shape you assume matters enormously, because you are betting the future on it. Three forms cover most wear: a linear model for steady erosion, an exponential model for damage that feeds on itself (a crack that grows faster as it lengthens), and a power-law model for the slow-then-sudden curve typical of fatigue. Picking the wrong shape early, when little of the curve is visible, is the largest single source of RUL error.

linear h(t) = a + b·t
exponential h(t) = a·e^b t
power-law h(t) = a·t^b → solve h(t)=H_fail for the crossing time

The simulation fits the model form you choose to the observed portion of a genuinely accelerating degradation, then extrapolates to the failure threshold. Reveal more of the curve with the slider and watch the estimate converge; switch the model form and watch a mismatched shape miss the true failure, sometimes badly.

FIGURE 1 · fit a degradation model, extrapolate to failure · interactive

model form: linear exponential power-law

Observed so far 55%

True curvature 2.2

Measurement noise 0.04

Rust dots: the health indicator observed up to now. Teal curve: the chosen model fitted to those dots and extrapolated. Dashed rust line: the failure threshold. The grey marker on the axis is the true failure cycle. When the model shape matches the true curvature the crossing lands near it; mismatch it and the forecast drifts early or late.

Section 04

Tracking the state with a Kalman filter

Fitting a fixed curve to all the data treats every past point as equally relevant, but degradation rates change, and the freshest data should weigh most. A state-space view fixes this. Model the hidden state as the current health level and its current rate of change, let both drift a little each cycle, and use the Kalman filter from the first monograph to track them as measurements arrive. The estimate adapts continuously, and crucially it carries a covariance: a running statement of how uncertain it is.

That covariance is the prize. To forecast the RUL, project the level and its growing uncertainty forward; the threshold crossing is no longer a single cycle but a distribution of plausible crossings. The simulation runs a constant-velocity Kalman filter on the health indicator and projects forward, drawing the widening uncertainty cone and the resulting spread of RUL. Trust the measurements more or the model more and watch the cone tighten or flare.

FIGURE 2 · Kalman state-space degradation & RUL spread · interactive

Observed so far 55%

Process noise (Q) 0.020

Measurement noise (R) 0.05

Rust dots: observed health. Teal: the Kalman estimate and its forward projection. The shaded cone is the projected uncertainty; where its centre meets the threshold is the median RUL, and the band where the cone straddles the threshold is the RUL interval. High process noise makes the filter nimble but the cone wide; low process noise tightens the cone at the risk of lagging a real change in rate.

Section 05

Particle filters for nonlinear wear

The Kalman filter is optimal when everything is linear and the noise is Gaussian. Real degradation is often neither: exponential crack growth, multiplicative noise, hard bounds. The particle filter handles all of it by brute, honest sampling. Represent the belief about the hidden state not as a tidy mean and covariance but as a swarm of hundreds of particles, each a candidate state with a weight. Each cycle, push every particle forward through the (possibly nonlinear) degradation model, reweight them by how well they explain the new measurement, and resample so that good hypotheses multiply and bad ones die.

To forecast the RUL, you do something no curve fit can: take the surviving particle cloud and run every particle forward to the threshold, collecting one failure time from each. The histogram of those times is the RUL distribution, with no Gaussian assumption imposed on it, and it can be skewed or even multi-modal in ways that match reality. The simulation runs a genuine sampling-importance-resampling particle filter on exponential degradation. Watch the cloud concentrate as evidence arrives and the RUL histogram sharpen.

FIGURE 3 · particle filter & the RUL distribution · interactive

Observed so far 55%

Process noise 0.06

Measurement noise 0.05

Top: the health indicator (rust dots) with the particle cloud's projected trajectories fanning toward the threshold in faint teal. Bottom: the RUL histogram built by running each particle to the threshold. Notice it is not symmetric; the long right tail is the small chance the machine limps along, exactly the asymmetry a maintenance plan must respect.

Section 06

Bayesian degradation updating

The particle filter tracked a belief by brute sampling. When the degradation is closer to linear, there is a far cheaper way to carry uncertainty that gives the same honest spread with a clean closed form: Bayesian updating of the degradation parameters. Start with a prior over the slope and intercept of the health trend, encoding what you believe before this unit has said much, then let each new measurement sharpen that belief into a posterior.

The payoff is a predictive band that is wide when evidence is thin and tightens as the curve reveals itself, and which extrapolates to the threshold as a genuine RUL distribution rather than a single line. The simulation runs Bayesian linear regression on the health indicator, drawing the prior's vague forecast, the posterior's sharpened one, and the RUL interval that falls out. Weaken the prior and early forecasts swing wildly on little data; strengthen it and the model leans on belief until the measurements earn the right to overrule it.

FIGURE 4 · prior to posterior, with a predictive RUL band · interactive

Observed so far 50%

Prior strength 1.0

Measurement noise 0.05

Rust dots: observed health. Faint teal band: the prior predictive, before this unit's data is trusted. Solid teal with shaded band: the posterior predictive mean and its uncertainty, extrapolated to the threshold. The point where the band crosses the threshold gives the RUL and its interval. A strong prior keeps the early forecast steady but stubborn; a weak one is nimble but jumpy.

Section 07

Similarity-based RUL

Sometimes you do not need to model the physics at all, because you have history: a library of units that ran to failure, each leaving a full degradation trajectory. The similarity-based approach treats RUL as a matching problem. Take the partial trajectory of the unit you care about, slide it against every complete trajectory in the library, and find the ones whose shape it most resembles so far. Each close match has a known remaining life from that alignment point; blend those, weighted by similarity, and you have a data-driven RUL with no assumed curve.

This is the approach behind many strong results on run-to-failure benchmarks, because it captures whatever shape the real degradation takes, including kinks and plateaus a parametric model would never guess. Its weakness is its appetite: it needs a library of genuine failures, which are precisely the rare and expensive events. The simulation matches a developing trajectory against a library of run-to-failure curves and forms a similarity-weighted RUL estimate.

FIGURE 5 · trajectory matching against a failure library · interactive

Observed so far 50%

This unit's wear rate 1.00

Match bandwidth 0.30

Faint grey curves: a library of units that ran to failure. Rust: the current unit, observed only up to now. The library curves most similar to it over the observed region glow teal; their known remaining lives, blended by similarity, give the RUL marker. A tight bandwidth trusts only the closest matches; a loose one averages over more of the library.

Section 08

Discrete health states with an HMM

Not all degradation is a smooth curve. Sometimes the natural description is a handful of regimes, healthy, degraded, critical, with the machine dwelling in one and occasionally jumping to a worse one, and never jumping back without repair. A hidden Markov model captures exactly this: the true health state is hidden and follows a Markov chain, while the sensor emits a noisy observation that only hints at which state the machine is in.

Given the noisy stream, the forward algorithm computes, at every moment, the probability of being in each state, fusing the new observation with the model's beliefs about how states evolve. The simulation generates a true state path and noisy observations, then runs the forward algorithm live. Watch the state-probability bands swap as the machine deteriorates, and notice how heavier emission noise blurs the handover, the discrete analogue of every uncertainty in this guide.

FIGURE 6 · forward-algorithm state probabilities · interactive

Emission noise 0.30

Transition rate 0.04

healthy degraded critical

Top: the noisy health observation, with the true state shaded behind it. Bottom: the forward-algorithm probability of each state, stacked to one. As the machine steps from healthy through degraded to critical, the coloured mass shifts. With low noise the bands are crisp and the handover sharp; with high noise they overlap and the estimate hedges, exactly as it should when the evidence is ambiguous.

Section 09

Honest RUL intervals

A point RUL is a liability. The maintenance decision rests on the lower edge of the interval, the soonest plausible failure, so an interval that is too tight is not a smaller error, it is a hazard. The challenge is that the fancy models do not always produce trustworthy intervals: a particle filter's spread can be over-confident if its noise is mis-specified, and a curve fit's textbook interval assumes a shape that may be wrong.

Conformal prediction, introduced in the forecasting monograph, is the antidote. It wraps any RUL predictor in an interval with a coverage guarantee that holds without trusting the model's assumptions. The recipe is mechanical: on a set of past units, record how far each RUL prediction missed (the nonconformity scores), take the appropriate quantile of those errors, and pad every future prediction by that amount. If you want 90 percent coverage, you use the 90th percentile of historical errors, and you get close to 90 percent coverage on new units, by construction. The simulation tracks a RUL prediction as a unit ages, drawing the conformal band and the alpha-lambda accuracy funnel that prognostics uses as its target.

FIGURE 7 · conformal RUL band & the alpha-lambda funnel · interactive

Target coverage 90%

Accuracy band α 0.20

Prediction noise 0.18

true remaining life predicted RUL conformal band ±α funnel

Grey: the true remaining life, ticking down to zero at failure. Teal: the RUL predicted at each moment, noisy early and sharpening late. Rust band: the conformal interval sized for your chosen coverage. Ochre funnel: the alpha-lambda target, narrowing toward failure, the standard against which a prognostic is graded. A good predictor enters and stays inside the funnel as failure approaches.

Section 10

Operating conditions as covariates

Here is the trap that fells naive prognostics: the same health indicator at the same level can mean wildly different remaining lives depending on what the machine is about to be asked to do. A bearing at 60 percent health has months left if the duty cycle stays gentle and weeks if it is about to be run hot and heavy. Extrapolating the health indicator in a vacuum ignores this, and the error is not random, it is systematic and dangerous.

The fix is to make the operating condition a covariate of the degradation rate, not just a nuisance to be averaged away. The model learns how load, speed and temperature accelerate or slow the wear, and the RUL forecast then takes a future operating profile as an input. This is also what lets prognostics inform operations: ease off the load and you genuinely extend the life, and the forecast shows it. Change the assumed future load below and watch the same present state yield very different remaining lives.

FIGURE 8 · same state, different futures · interactive

Observed so far 50%

Future load factor 1.0

Past load (so far) 1.0

Rust dots: health observed under the past load. From the same present point, three teal futures branch out under light, nominal and heavy future loads, each crossing the threshold at a different cycle. The bold teal curve is your selected future-load scenario. This branching is why a RUL is meaningless without a stated future operating profile.

Section 11

The population view: survival & hazard

Everything so far has tracked one unit's signal. There is a complementary tradition that asks a different question: across a whole population of similar machines, how is the time-to-failure distributed, regardless of any individual's sensor trace? This is survival analysis, and its workhorse is the Weibull distribution, prized because a single shape parameter captures the qualitative character of the failure.

The shape parameter k tells the story. Below one, the hazard rate, the instantaneous risk of failing given survival so far, decreases with age: infant mortality, where survivors of early defects grow more reliable. Equal to one, the hazard is constant: failures arrive at random, memoryless. Above one, the hazard increases: wear-out, the regime most condition monitoring lives in. Stitch the three together and you get the classic bathtub curve of reliability. The simulation draws the survival and hazard curves for a Weibull you control, with the operative regime named.

survival S(t) = e^{−(t/λ)^k} · hazard h(t) = (k/λ)(t/λ)^k−1 · MTTF = λ·Γ(1 + 1/k)

FIGURE 9 · Weibull survival and hazard · interactive

Shape k 2.2

Scale λ (char. life) 120

survival S(t) hazard h(t)

Teal: the fraction of the population still running at each age. Rust: the hazard rate, the risk of failing right now given survival so far. Push the shape below one and the hazard falls with age (infant mortality); set it to one and the hazard is flat (random failure); raise it and the hazard climbs (wear-out). The readout names the regime and reports the mean time to failure.

Section 12

Evaluating a prognostic

Because late predictions are dangerous and early ones merely wasteful, prognostics is not graded with the symmetric metrics of ordinary forecasting. A few specialized measures encode the asymmetry and the importance of timing.

Prognostic horizon

How early before failure the prediction first becomes and stays accurate. Earlier is better, because it buys time to act.

Alpha-lambda accuracy

Whether the prediction sits inside a shrinking band (plus or minus alpha) around the truth at chosen fractions of life. The shrinking funnel seen with conformal RUL intervals.

Asymmetric cost

Scoring that penalizes late predictions far more than early ones, matching the real consequences of an unplanned failure.

Convergence

Whether the estimate steadily homes in on the truth as failure nears, rather than wandering or jumping.

The honest test, as always, is on units the model never saw, and ideally many of them, because a single run-to-failure curve tells you almost nothing about variance. A prognostic that looks brilliant on one bearing and falls apart on the next has learned that bearing, not the failure mode.

Section 13

Optimizing the maintenance decision

Every method in this guide has been building toward a single act: deciding when to intervene. A RUL distribution is not the answer, it is the input to the answer. The decision turns on an asymmetry stated plainly. Replace too early and you throw away good life and pay for an unnecessary teardown, the preventive cost. Wait too long and the machine fails in service, the corrective cost, which is usually far larger once you count secondary damage, downtime and safety.

The principled choice minimizes the long-run cost per unit time. For a chosen replacement age, the expected cost blends the corrective cost weighted by the chance of failing first with the preventive cost weighted by the chance of surviving to the planned replacement; divide by the expected time the unit actually runs and you have a cost rate to minimize. The simulation sweeps the replacement age over a Weibull failure distribution and finds the optimum. Crank the cost ratio up and the optimal replacement marches earlier, buying safety at the price of wasted life; this single curve is what the whole pipeline exists to produce.

cost rate g(t_r) = [ C_c·F(t_r) + C_p·S(t_r) ] ÷ ∫₀^t_r S(t) dt · choose t_r to minimize it

FIGURE 10 · the cost-rate curve and the optimal replacement age · interactive

Corrective / preventive cost 8×

Failure-time mean 120

Failure spread (shape k) 2.8

failure-time density cost rate optimal replacement

Teal: the distribution of when the machine fails. Rust: the long-run cost per unit time as a function of the age at which you preventively replace. Its minimum, the ochre line, is the optimal replacement age. Run-to-failure sits at the far right where you pay the corrective cost often; replace too early and you pay the preventive cost too frequently. Raise the cost ratio and the optimum slides left, trading life for safety.

Section 14

The PHM pipeline

Stand all three monographs end to end and the full prognostics and health management loop appears. Sensors produce signals; signals become features; features fuse into a health indicator; a detector watches the present while a forecaster watches the future; and a RUL with its interval feeds a maintenance decision, which acts back on the machine.

Earn the health indicator first. No prognostic is better than the curve it forecasts. Spend the effort making a monotonic, trendable indicator before reaching for clever models.
Quantify uncertainty, always. Ship the interval, not just the number, and size it for the asymmetric cost of being late.
Condition on the future. A RUL is a function of the operating profile to come. State the assumed profile, and recompute when the plan changes.
Validate across units. Test on held-out run-to-failure histories, many of them, and report the variance, not just the best case.
Retrain on drift. New failure modes, maintenance actions and fleet changes all shift the degradation. A prognostic is a living model, not a one-time fit.

End of part three, and of the series. From a single sensor sample in Time Series as Sensory Data, through the forecast of its next value in Forecasting Time Series, to the forecast of its failure here, the arc is complete: sensing, predicting, and deciding when to act.

Section 15

Practical notes

Prognostics fails in characteristic ways, and most of them have nothing to do with the sophistication of the model. These are the lessons that decide whether an RUL estimate is trusted on the floor or quietly ignored.

Ship a distribution, not a number. A single RUL figure invites false confidence and is almost always wrong. The interval, how sure you are and how that certainty tightens as failure nears, is the part a maintenance planner can actually use.
Define failure before you predict it. The health-indicator threshold drives every estimate downstream, so set it from physics and consequence, not convenience. Move the threshold and you move every RUL; pretend it is fixed at your peril.
Condition on the operating context. The same wear state has a different future under heavy load than at idle. A trajectory that ignores future duty cycle is forecasting a machine that does not exist.
Score with prognostic metrics. A small RMSE that arrives only after the part has failed is useless. Judge with alpha-lambda accuracy and prognostic horizon, which reward being right early and penalise late confidence.
Respect the scarcity of failures. Most fleets have few full run-to-failure records. Similarity matching, survival models and Bayesian priors all exist to squeeze information out of that scarcity rather than pretending it away.
Accept wide early, demand narrow late. An honest RUL is vague when the unit is young and sharpens with evidence. A model that is confident from the first sample is not calibrated, it is bluffing.
Watch for resets. A repair, a part swap, or a regime change can step the health indicator back. A prognostic that keeps extrapolating the old trend through a maintenance event will be confidently, dangerously wrong.

The one-sentence summary

Remaining useful life is a belief about the future of one specific machine under its specific duty, so anchor it to a defined failure, carry its uncertainty honestly, condition it on how the unit will actually be used, and judge it by whether it warns you in time to act rather than by how close the number looked in hindsight.

Section 16

References and further reading

Lei, Y., et al. (2018). "Machinery health prognostics: a systematic review from data acquisition to RUL prediction." Mechanical Systems and Signal Processing. The standard modern overview of the whole field.
Si, X.-S., Wang, W., Hu, C.-H., & Zhou, D.-H. (2011). "Remaining useful life estimation: a review on the statistical data driven approaches." European Journal of Operational Research. A clear taxonomy of degradation and statistical RUL methods.
Saxena, A., Goebel, K., Simon, D., & Eklund, N. (2008). "Damage propagation modeling for aircraft engine run-to-failure simulation." PHM Conference. The C-MAPSS turbofan benchmark behind much RUL research.
Wang, T., Yu, J., Siegel, D., & Lee, J. (2008). "A similarity-based prognostics approach for remaining useful life estimation." PHM Conference. The origin of trajectory-matching RUL.
Arulampalam, M. S., Maskell, S., Gordon, N., & Clapp, T. (2002). "A tutorial on particle filters for online nonlinear/non-Gaussian Bayesian tracking." IEEE Transactions on Signal Processing. The canonical particle-filter reference.
Simon, D. Optimal State Estimation: Kalman, H-infinity, and Nonlinear Approaches. Wiley. State-space tracking for degradation.
Saxena, A., et al. (2008). "Metrics for evaluating performance of prognostic techniques." PHM Conference. Prognostic horizon, alpha-lambda accuracy and related measures.
Vovk, V., Gammerman, A., & Shafer, G. Algorithmic Learning in a Random World. Springer. The foundations of conformal prediction.
Angelopoulos, A. N., & Bates, S. (2023). "Conformal prediction: a gentle introduction." Foundations and Trends in Machine Learning. A practical modern guide to distribution-free intervals.
Sankararaman, S., & Goebel, K. (2015). "Uncertainty in prognostics and systems health management." International Journal of Prognostics and Health Management. Why and how to quantify RUL uncertainty.

References point to further study; the explanations and simulations above are written and implemented from first principles rather than drawn from any single source.