No single technique among predictive maintenance models gives the best remaining useful life (RUL) estimate everywhere. The right choice depends on three facts about your failure history: how many real failures you have seen, how much of the fleet is censored (still running, so its failure time is unknown), and how long and consistent each unit’s monitoring record is. Rich historian data with almost no labelled failures points to anomaly detection. A censored fleet points to survival models such as Cox proportional hazards or Weibull accelerated failure time (AFT). Very few failures point to Weibull priors, hierarchical Bayesian pooling across similar assets, or physics-informed hybrids. An RUL estimate without a stated level of confidence is of limited use in mission-critical applications.
Choosing predictive maintenance models by failure count, censoring and history length
The table below gives a starting point. It gives each regime the model family to try first and a known way it breaks.
| Your data | Model family to start with | What breaks it |
|---|---|---|
| Long historian or SCADA record, few or no labelled failures | A normal-behaviour model or autoencoder trained on healthy data, raising anomaly alerts ahead of any RUL estimate | Training data contaminated by pre-failure behaviour |
| No failures yet, right-censored run times, a trusted historical Weibull shape | Weibayes (Weibull with an assumed shape parameter) | A historical shape that does not fit this population |
| Some failures, many still-running units, sensor covariates | Cox proportional hazards, Weibull AFT | Proportional hazards violated, or censoring linked to failure risk |
| Few failures per unit type, but a fleet of related types | Hierarchical Bayesian model with fleet-informed priors | Negative transfer when the pooled units are not really alike |
| Few failures, but a credible physics model of the failure mode | Physics-informed or hybrid physics plus data-driven model | An incomplete physics model or poor parameter calibration |
| Enough samples, Cox assumptions not met | Parametric-mixture or deep survival models | Too few samples for the extra flexibility |
Check the quality and quantity of the data, and what the business expects from the output, before choosing a model. For RUL work that means counting real failure events (not records), writing down what “failed” means for this component, and naming the decision the output has to support, whether that is a ranking, an alarm or a probability by a date.
Matching the RUL technique to the failure history you actually have
The modelling regime is set by the data, not by the asset class, so a pump fleet, a bearing population and a set of process probes can land in different rows of the table. Practitioners choose between physical, data-driven and hybrid approaches according to the physics of the problem, the data available and what the solution must deliver. The subsections below take each common situation in turn.
When is a simple anomaly-detection model on historian data enough?
A label-free anomaly detector is enough when the job is early warning rather than a time-to-failure estimate, and when labelled failures are too scarce to train a supervised RUL model. Anomaly detection asks a different question: it flags deviations from normal operation that may indicate an emerging fault, and does not try to estimate remaining operating time. Normal-behaviour modelling on SCADA data, reviewed in Wind Energy Science in 2023, trains on healthy periods and treats an abnormally large or trending prediction error as a sign that a component may be failing.
The wind-turbine pipelines in that review are built from existing SCADA data, pooled across turbines into one model per signal per farm to keep the maintenance burden acceptable to operators. The review favours the elastic net as the main model because it is simple, transparent, handles many correlated predictors and works with limited training data. More complex models gave only small improvements, showed no advantage for rotor brush or generator bearing failures, and were more sensitive to operating states rare in the training data. On NASA’s simulated CMAPSS turbofan data, an unsupervised LSTM autoencoder trained only on healthy cycles reached recall of at least 0.725 with specificity above 0.96, despite anomalous windows making up only 0.8% to 1.3% of the data.
Two conditions apply, though. Someone has to own the rule for what counts as healthy training data: the review excludes data from the four months before a failure and the month after it. Thresholds also have to be tuned, and tuning with only a handful of failure examples is hard. You need more than anomaly detection once the decision asks for a date or a probability by a date, which is where survival and Bayesian models come in.
Sparse failure labels with long censored run times
When most units have run for a long time without failing, survival analysis is the natural fit, because it models time to an event that may be only partly observed. The still-running units carry information rather than noise, and how to use them is covered in its own section later in this article.
Very few or no failures: start from a Weibull prior
With few or no failures, a Weibull analysis that takes the shape parameter from history (Weibayes) still gives lower confidence bounds on percentiles, survival probabilities and failure probabilities, provided the data are right-censored and the historical value is accurate. With no failures at all, Minitab’s guidance notes that a 50% lower bound is commonly used as the estimate of the scale parameter. A structured-effect neural network study on RUL likewise reports that the Weibull distribution has proved effective even with few observed lifetimes.
That sits awkwardly next to a DataStatPro survival-analysis tutorial, which gives 50 to 100 events as a heuristic for reliable parametric estimation. The two are consistent once you see where the information comes from: Weibayes fixes the shape from prior knowledge, so the sparse data only has to inform the scale. The result is only as good as that historical shape, so document its origin and test how sensitive the bounds are to it.
How can a hierarchical model borrow strength across similar assets without overstating confidence?
A hierarchical Bayesian model fits each asset group its own parameters while tying those parameters to a shared fleet-level distribution, so groups with little data borrow statistical strength from data-rich ones. Bull et al. (2023) report that this partial pooling raised the predictive log-likelihood of hazard models for a truck fleet by 15% and 13%, and cut the standard deviation of interpretable survival parameters by 70% on average. A study of industrial machinery fleets in Ghana reports a hierarchical Weibull model narrowing the 95% credible interval for one under-sampled fleet’s mean time to failure to 720 to 1,210 hours. Jia, Papaioannou and Straub report the same pattern in model-based prognostics: bounds from historical data alone are large and narrow significantly once current data are included.
The gain is largest for sparse groups and smallest for data-rich ones. That is also where the risk sits: if the assumptions about which parameters to tie are wrong, the model can suffer negative transfer. Bull et al. found the likelihood failed to improve over single-task models for some alternator tasks, and note that a much larger dataset would be needed to tell whether negative transfer had occurred. Their extended model gives each component type its own discrepancy term while the linear-component parameters stay shared across all tasks.
To avoid overstating confidence, hold out data per subgroup, compare the pooled model against independent per-group models on that held-out data, and tie only parameters that engineering knowledge says should be shared. Bull et al. encode that knowledge as weakly informative priors, and suggest uninformative priors on the variance parameters that control cross-task correlation when such knowledge is not available.
Should physics-of-failure or vendor component-life models fill the gap?
Yes, where a credible physics model of the failure mode exists, because physics-informed machine learning keeps predictions physically consistent in exactly the regimes where data are sparse. Arias Chao et al. (2020) fused a physics-based performance model with deep learning on a simulated fleet of nine turbofan engines and extended the prediction horizon by nearly 127% over purely data-driven approaches. When they shrank the training set, the hybrid’s RMSE rose by 2% against 17% for the purely data-driven model. Physics can enter in several ways: as simulated training data, as constraints in the model design, or as a fatigue law such as Paris’ law built into the model with a learned correction for effects it misses. In a hierarchical Bayesian setting, the Paris’ law parameters themselves can be inferred from fleet data.
The same idea scales down. One structured-effect neural network encodes the known lifetime-distribution shape in its structure instead of relying on vast input data. ISO 13381-1:2015 makes the underlying point for prognostics generally: forecasting fault progression needs foreknowledge of the likely failure modes and their relationship to operating conditions, and may need the physics of those modes.
The limits are real. Physics-based models are often incomplete, and the hybrid’s performance depends on how well the physics parameters are calibrated, although Arias Chao et al. found it still beat the data-driven baseline under noisy calibration. A vendor’s rated component life is best treated as a prior on the scale of the life distribution, not as a model of your duty cycle, and your own data should be allowed to move it.
pH and DO probes: does an aircraft RUL pattern transfer?
Whether aircraft-component RUL patterns transfer to single-use bioreactor probes is unproven, and it should not be assumed. What does carry over is the structure of the problem. RUL is defined as the time before health falls below a defined failure threshold, so for a probe the first job is to define failure, for example calibration drift beyond the tolerance your process accepts. The RULSurv authors warn that whenever a model is applied to a new application, the event detector and the predictive model have to be re-evaluated on a sufficient hold-out set of units. For probes with short service lives and many replacements, a Weibull or survival model on calibration-drift events, pooled across probe types where they are genuinely alike, is a reasonable first model; the degradation physics of an aircraft bearing is not a reason to choose it.
Cox proportional hazards or a deep survival model?
Use Cox proportional hazards when the proportional-hazards assumption is plausible and data are modest, and move to a model without that assumption when the check fails or the data are large and non-linear. A benchmark study, “Beyond Cox Models”, found no reason to go beyond Cox regression when linearity and proportional hazards hold, and found Cox remarkably robust at small sample sizes even when they do not. On its synthetic non-linear dataset, Cox was best at 300 samples, and non-linear methods overtook it only at 600 to 1,200 samples depending on the method.
The check comes before the commitment. Under proportional hazards, the Schoenfeld residuals for each covariate should be uncorrelated with time, and the Grambsch-Therneau test formalises that. The lifelines library exposes this as a check_assumptions() method. A covariate that fails can be handled by stratification, which gives each stratum its own baseline hazard while sharing the regression coefficients. Covariates that change over time, such as vibration or temperature trends, enter a modified Cox model in long format with start and stop times per interval. That model has a cost: predicting forward requires covariate values beyond the observed times, which you do not have.
The RULSurv authors note that proportional hazards is reasonable under artificially accelerated degradation but unlikely to hold in an industrial setting, where time-to-event distributions have dependencies at different time scales. An accelerated failure time (AFT) model, which puts multiplicative effects on time rather than on the hazard, is the more intuitive choice when the assumption fails. Deep Survival Machines drops the proportional-hazards assumption entirely using a mixture of parametric distributions, and its parametric form makes it more robust than non-parametric or deep alternatives where data are sparse, such as long event horizons. A more complex model needs enough samples before it improves on Cox.
Which metrics should gate acceptance?
Gate acceptance on a censoring-aware concordance index and on the integrated Brier score together, evaluated on held-out data. The concordance index is a ranking metric and says nothing about calibration, so a model can rank units correctly while its probabilities are wrong. Harrell’s version becomes too optimistic as censoring increases, so the scikit-survival documentation suggests the inverse-probability-of-censoring-weighted (IPCW) version for moderate to high censoring. Where hazards are not proportional, survival curves can cross and Antolini’s C-index is the appropriate generalisation.
The integrated Brier score extends mean squared error to right-censored data and summarises calibration across all time points as one number. In the scikit-survival worked example, a Random Survival Forest and a Cox model scored equally on discrimination while Cox was clearly better calibrated. Split the test set with stratification so it does not end up holding only censored units. We would derive the pass thresholds from what a missed failure or an unnecessary intervention costs, because criticality and mitigation cost should drive the acceptable error trade-off.
Can still-running units train an RUL model instead of waiting for more failures?
Yes: right-censored units (those that have not failed yet) belong in a survival model as censored observations, and dropping them both wastes data and biases the estimate. Ignoring censored observations leads to overestimating failure risk. The excluded units are typically the longest-surviving ones, so training only on failed units skews the failure-time distribution towards early failure. The RULSurv study found that a LASSO model trained only on uncensored bearings gave poorer accuracy than censoring-aware survival models across most datasets and censoring levels.
The mechanism is simple. A censored unit tells you only that failure happens after its last observed time, a lower bound on its event time, and survival methods build the likelihood so that each censored and uncensored observation contributes correctly. In practice each unit needs a duration, an event flag (1 for failed, 0 for censored) and its covariates. Cox, AFT and Aalen’s additive models all accept that format. Whether your telemetry supports Cox or AFT depends on its covariates: static features fit either, while trends that change over time need the long-format, time-varying Cox setup with its prediction limits. Feature engineering may need rework if the telemetry was never designed for time-to-event modelling.
Three conditions decide whether this works. First, censoring has to be non-informative, meaning the reason a unit is censored is independent of its true failure time. Preventive maintenance itself creates right-censored data, and units pulled because they look unhealthy break that independence. Second, a censored unit is not a missing record: units with unknown status or time are missing data and should be handled as such. Replacement for a different failure mode is a competing risk, and a Kaplan-Meier estimate that treats it as ordinary censoring overestimates the cumulative incidence of the failure you care about. Third, volume matters: in a PHM Society study, ordinal-regression models on small datasets gained too little from a handful of censored units to improve performance.
Why a calibrated probability-of-failure curve beats a single RUL number
A probability-of-failure curve per asset, checked for calibration, gives a failure probability at every time, while a point RUL ignores the uncertainty in the estimate.
As part of AI engineering for manufacturing and automotive, TechnoLynx would turn the failure history, censored units included, into per-unit failure probabilities validated against held-out failures and fed into maintenance scheduling, starting by framing the history against the decision it must support.
The model family matters. Haider et al. (JMLR, 2020) point out that risk scores from a Cox model do not provide survival probabilities, which is why they argue for individual survival distribution models that give a probability at every time. The RULSurv survival model produces curves guaranteed to decrease monotonically over time and gives a probabilistic RUL rather than a point estimate. Bayesian survival models add credible intervals around each unit’s curve. A commonly reported RUL, the median of the survival curve, describes when a population of similar units fails on average only if the curves are D-calibrated.
How do you know the uncertainty band is trustworthy?
Test it. D-calibration checks whether observed failures fall evenly across the predicted survival-probability deciles, about 10% per decile, using a Pearson chi-squared test. Coverage calibration checks whether a p% credible interval contains the outcome p% of the time. In the RULSurv study, the Bayesian model passed coverage calibration. Conformal prediction offers a different route: it wraps any point predictor and guarantees, under mild assumptions, that the interval covers the true RUL at a level you choose. Candès, Lei and Ren extend this to right-censored survival data where censoring is independent of failure time, with calibrated lower predictive bounds, and a low bound signals either high risk or high uncertainty from data scarcity.
Coverage alone is not enough: a trivially wide interval achieves high coverage and is useless for planning. Interval width depends on how precise the underlying predictor is, so calibration and accuracy work together. Building audit-ready evidence around these checks is the subject of our page on calibrating and governing probability-of-failure models.
What to do when the budget deadline arrives first
Validation is itself hard: acting on a failure prediction removes the chance of checking whether it was right. A model moved to a new application has to be re-evaluated, event detector included. Turning the probabilities into a shutdown sequence and work orders is a separate scheduling problem from building the model.
How much monitoring history is enough before acting on the estimates?
No source gives a firm minimum, and the honest answer is that held-out calibration decides, not the calendar. Rules of thumb exist. The DataStatPro tutorial cites at least 20 to 30 events for stable Kaplan-Meier curves, 10 events per predictor for a Cox model, and 50 to 100 events for parametric models. The same tutorial calls ten events per parameter a historical heuristic, not a universal adequacy rule, and says adequacy depends on parameters, event timing, censoring, shrinkage and validation strategy. Observed events, not total units, set the requirement, and the median survival time cannot be estimated if fewer than half the units have failed.
Anomaly detectors are trained on healthy operating data rather than on failure events. The 2023 SCADA review trained on about six months of healthy data per turbine; cutting that to two months hurt rare operating states such as long cool-downs more than steady-state behaviour. For Bayesian models, uncertainty narrows as each unit’s own data accumulate. Early on, Jia, Papaioannou and Straub found the true RUL sometimes lay outside the 95% credible interval.
The practical question is therefore when the held-out calibration checks start passing for your fleet and your failure definition. Until they do, the probabilities can still rank units for inspection, but they should not justify stretching intervals on safety-relevant equipment.
FAQ
How is predictive maintenance different from preventive maintenance?
Both rely on estimates of remaining useful life: preventive maintenance bases those estimates on human knowledge, while predictive maintenance uses data-driven models.
