When a predictive-maintenance model says a jet engine has "between 40 and 90 cycles left, with 90% confidence," the maintenance crew cares far more about whether that promise holds in the last 30 cycles of the engine's life than in the first 200. Conformal prediction, a family of methods that wraps any forecaster and guarantees its intervals contain the truth a chosen fraction of the time without distributional assumptions, already reaches about 90% coverage on turbofan remaining-useful-life (RUL) benchmarks. That guarantee is averaged over the whole life of the engine. Engines spend most of their lives in a long, benign stable phase and only a short time in the non-linear, accelerating wear phase before failure. An interval method can therefore meet its 90% target overall while badly missing in exactly the window where a wrong answer causes an unplanned shutdown.
This note proposes calibrating RUL intervals separately for each band of true remaining life, so that coverage is guaranteed to hold within the near-failure band itself, not just on average. The technical difficulty is that true remaining life is the unknown being predicted, so the test point cannot simply be sorted into a calibration group. We resolve this by inverting the test: a candidate RUL value is kept in the interval only if it conforms to the calibration errors of its own life band. Under unit-level exchangeability this gives finite-sample coverage conditional on the life band. We pair it with an asymmetric, one-sided variant aimed at the operationally dangerous error, overestimating remaining life, and test it on C-MAPSS and N-CMAPSS against the strongest existing probabilistic and conformal RUL methods.
Core Claim
Conformal Prediction Intervals for Remaining Useful Lifetime Estimation conformalizes CNN and gradient-boosting RUL predictors with split conformal prediction, conformalized quantile regression (CQR), and a non-exchangeable variant on C-MAPSS. It reports roughly 90% coverage and intervals that narrow near failure. The authors explicitly name stage-conditional coverage as the open problem, and they report no per-stage coverage numbers. The phase structure documented in the degradation literature makes the concern concrete. That structure includes a long stable phase in which damage accumulates silently (Wu), followed by S-curve acceleration (Zeng et al.) or knee points (Pene et al.). Point-like predictors systematically overestimate RUL near the failure boundary (Leão). Narrowing intervals combined with an overestimating centre is precisely the recipe for late-life under-coverage hidden by a good marginal average.
Our claim has two parts. First, the guarantee practitioners need is coverage conditional on the true-RUL band $B_k$:
$$P\big(Y \in C(X) \mid Y \in B_k\big) \ge 1-\alpha \quad \text{for every band } k,$$
where $Y$ is true RUL, $X$ is the sensor history up to the current cycle, and $C(X)$ is the interval. Second, this guarantee is achievable at modest width cost by label-conditional calibration: the interval is the union over bands of $\{y \in B_k : s(X,y) \le \hat q_k\}$. Here $s$ is a CQR-style score from a quantile forecaster, and $\hat q_k$ is the conformal quantile of calibration scores whose true RUL lies in $B_k$. Validity follows from the same rank argument as ordinary split conformal prediction, applied within the band. The argument needs calibration and test examples with $Y \in B_k$ to be exchangeable. Cycles of the same engine are strongly dependent, so we enforce exchangeability across engines by drawing one cycle per band per calibration engine.
The novelty is not group-conditional conformal prediction in general. Kandinsky Conformal Prediction and Identifying Homogeneous and Interpretable Groups for Conformal Prediction supply that machinery and the evidence that marginal wrappers fail on hard subgroups, but they condition on observable groups. The novelty is conditioning on a group defined by the target itself, in a regime where observable proxies for that group are biased in exactly the dangerous direction.
Research Program
The base forecaster supplies quantile estimates. We reuse the structured state-space model with simultaneous quantile regression of A Quantile Regression Approach for Remaining Useful Life Estimation with State Space Models. That model trains S4-family sequence models to output several quantiles at once via pinball loss, but it offers no coverage guarantee. Its quantiles define the CQR score of Conformalized Quantile Regression. We also run a gradient-boosting backbone so results do not hinge on one architecture.
The calibration layer comes in three forms. The two-sided form gives band-conditional coverage of the full interval. The one-sided form calibrates only a lower bound on RUL, using the score $\hat q_{lo}(X) - y$. It addresses the asymmetry raised by Javanmardi and Hüllermeier's call for physics-aware scores: overestimating life late is catastrophic, while underestimating early only costs some conservatism. The third form adds a monotonicity check. Remaining life decreases by exactly one per cycle, so an engine's successive lower bounds can be clipped to be consistent with each other. We will test whether this tightens intervals without breaking validity, using a held-out check rather than a proof, because the clipping couples cycles.
The analysis has four parts. We derive the minimum number of calibration engines per band, which is $\lceil(1-\alpha)(n_k+1)\rceil \le n_k$, meaning at least 9 engines per band at 90%. From this we characterize the band-count versus width trade-off for FD001's 100 run-to-failure training engines. We measure how often the inverted set is not a single interval. We also quantify the price of conditioning, defined as the width increase relative to marginal CQR in each band.
Baselines
The naive reference is plain split conformal prediction with a fixed absolute-residual margin, run on the same backbone. It is the cheapest method with a valid marginal guarantee, and it shows how badly a constant width fails across phases.
The strongest existing conformal RUL method is marginal CQR as in Javanmardi and Hüllermeier, reimplemented with our backbones, splits, and calibration engines so that the only difference is the calibration rule. Their reported figure is approximately 90% marginal coverage at the 90% target on C-MAPSS. No per-band number was found on Lacuna. To beat it, our method must keep per-band coverage within sampling error of 90% in the final band, while marginal CQR falls below, at a final-band width penalty we will report.
The strongest uncalibrated probabilistic baselines are raw SQR-SSM quantiles (Frizzo et al.), which report an FD001 median RMSE of 37.96 for S4 against 31.46 for a Transformer, and the deep sigma point process of Uncertainty-aware Remaining Useful Life predictor. On N-CMAPSS that model reports an NLL of 3.10 against 4.26 for MC Dropout, with RMSEs of 7.38 and 7.31 respectively. We read off its 5% and 95% predictive quantiles and score them on the same per-band coverage metric. All learned components share the same hyperparameter search budget, tuned on training engines only.
The strongest time-series conformal competitor is SPCI (Sequential Predictive Conformal Inference for Time Series), applied per engine online. Sun and Yu report that reactive methods of this kind drop to 80% or below after change points.
The key ablation is Mondrian calibration on predicted-RUL bands, which is identical to our method except that test points are routed by the forecaster's median rather than by inverting over candidate values. This isolates the label-conditional ingredient.
Why not just…?
Why not just condition on predicted RUL or a health indicator, which is observable? This is the real threat. If the forecaster's band assignment were accurate near failure, it would match our method at lower complexity. Our concern is that predicted-band conditioning guarantees coverage among points the model believes are near failure. Engines whose wear has accelerated but whose predicted RUL still looks large are routed into the stable band and inherit its errors, which is the lag documented by Leão and Cheng. The ablation settles the question directly. If predicted-band Mondrian calibration achieves final-true-band coverage within one standard error of ours, we will report that the simpler method suffices and our contribution reduces to the evaluation protocol.
Why not just use a better probabilistic backbone with marginal CQR, since intervals already narrow near failure? It is possible that strong quantile backbones already cover the last band. In that case the gap Javanmardi and Hüllermeier raised would be empty on these benchmarks, and our per-band audit would show it. That would be a useful negative result, and the experiment is designed to detect it rather than assume the opposite.
Why not just lower α until the worst phase is covered? Raising nominal coverage everywhere also covers the late band eventually, but it widens the long stable phase for no benefit. We compare at matched worst-band coverage and report the total width over an engine's life. Our method wins only if it is tighter at that matched level.
Why not use anticipatory switching-state calibration like CPTC? CPTC keeps per-regime score pools driven by a discrete switching model and was tested on cyclic regimes, not on irreversible, absorbing wear. It also requires an online warm-up within each engine's short life. We include it where its switching model can be fitted, but we expect its regime predictor to suffer the same routing problem as predicted-band calibration.
Proposed Experiments
We use C-MAPSS FD001 and FD002, and N-CMAPSS with Unit 14 as the covariate-shift stress test. Per-cycle evaluation needs true RUL at every cycle, and the official C-MAPSS test set gives it only at truncation, so we split run-to-failure training engines at the engine level into training, calibration, and test groups. We repeat this over 20 random engine splits and evaluate on uncapped true RUL. Metrics are coverage per true-RUL band, with emphasis on worst-band and final-band coverage; the lower-bound violation rate in the last 20 cycles; mean width per band; the fraction of non-contiguous sets; and marginal coverage as a sanity check.
The decisive checks are three. First, does marginal CQR under-cover the final band? If not, the premise fails on these benchmarks. Second, does our method restore final-band coverage where predicted-band Mondrian calibration does not? Third, is the width penalty in the final band smaller than that of uniform α-lowering at matched worst-band coverage? On N-CMAPSS Unit 14 the exchangeability assumption is deliberately broken, and we report the resulting coverage loss honestly rather than claim a guarantee there.
References
- Conformal Prediction Intervals for Remaining Useful Lifetime Estimation
- Effective Latent Representation for Prediction of Remaining Useful Life
- Fault Diagnosis and Condition Division Criterion of DC Gas Insulating Equipment Based on SF6 Partial Discharge Decomposition Characteristics
- Aging modeling and lifetime prediction of a proton exchange membrane fuel cell using an extended Kalman filter
- Failure Prognostics of a Hydraulic Pump Using Kalman Filter
- Kandinsky Conformal Prediction: Beyond Class- and Covariate-Conditional Coverage
- Identifying Homogeneous and Interpretable Groups for Conformal Prediction
- A Quantile Regression Approach for Remaining Useful Life Estimation with State Space Models
- Conformalized Quantile Regression
- Uncertainty-aware Remaining Useful Life predictor
- Sequential Predictive Conformal Inference for Time Series
- Conformal Prediction for Time-series Forecasting with Change Points