← Back to Autonomy
Monograph № 5 · On Prognostics

Remaining
Useful Life

How much longer will it last? A hands-on tour of the four families of methods engineers reach for — from simple threshold crossings to Gaussian processes with honest confidence bounds.

By Majid Mazouchi

Why the question is hard

Give an engineer a running component — a bearing, a battery, a power transistor, a jet engine — and ask: how much longer will it work? The answer matters a great deal. Too conservative and you replace parts that still had life in them, wasting money and spare inventory. Too aggressive and you discover your mistake when the component fails unannounced, with consequences that range from costly downtime to loss of life. The gap between these two mistakes is where the entire field of prognostics lives.

The question is hard for reasons that compound. First, the component is still running — you cannot observe its failure time, only its current condition. Second, no two units are identical: manufacturing variation, operating conditions, and environmental stresses push each one along its own private aging curve. Third, your prediction has to project forward through an uncertain future — future load, future temperature, future vibration, none of it known. Fourth, your training data is a graveyard: every run-to-failure example you have comes from a unit that is dead. Learning to predict a future that hasn't happened yet from a past that already has is the core tension of the whole discipline.

What engineers call Remaining Useful Life (RUL) is the residual quantity: the time (or cycles, or operating hours) between now and the moment the component is declared unfit for duty. Tools for estimating it fall into a handful of families, each with a different answer to the question of how to learn from dead units what the living ones are about to do.

Key idea

Every RUL method is a bet about what aging looks like. Some bet the signal itself tells you (threshold crossings). Some bet history repeats (similarity matching). Some bet patterns in windows (feature regression). Some bet in time (sequence models). The right bet depends on your data.

Degradation and its threshold

The simplest model of aging is a degradation signal — a scalar summary of component health that drifts monotonically away from its healthy baseline over time. Bearing vibration RMS climbs as the raceway fatigues. Battery capacity falls as electrochemistry erodes. Transistor on-state resistance creeps up as bond wires stress. Pick the right scalar and its trajectory usually has the shape of a hockey stick: flat for most of life, then bending sharply upward (or downward) near the end. A failure threshold — a regulatory limit, a spec, a practical cutoff — marks the moment the unit is retired.

HealthyCrossed thresholdThreshold
Fifteen nominally identical units running to failure — same design, same duty cycle, same environment. Each one ages differently because the underlying degradation process is stochastic. Slide the failure threshold and watch the population's lifetime distribution shift.

The lifetime distribution at the bottom is the central object of population prognostics. Its mean tells you expected life; its standard deviation tells you how tight the design is; its tail tells you how many units fail early. But notice that none of this tells you the RUL of your particular unit, which is what you actually want. For that, you need to look at this unit's signal — which is where the more sophisticated methods earn their keep.

The similarity-based approach

The first honest data-driven answer to "how much longer" is the oldest trick in the forecaster's book: find units in your historical records that looked like yours when they were at the same age, and see what happened to them next. This is the similarity-based or trajectory-matching approach, and it has the great virtue of requiring no assumptions about the shape of the degradation curve. You store a library of run-to-failure trajectories from retired units. You match the observed portion of a live unit against the equivalent early portion of each library member using a distance metric (usually L2 on the trajectories, after alignment). The K closest neighbors cast votes: their individual remaining-life values get averaged, and the average is your predicted RUL.

The weakness of this method is exactly its strength. It assumes nothing, so it learns nothing about why units age — it just hopes history rhymes. If your library covers only one operating regime and your live unit sees a different one, the method will confidently return the wrong answer. But if your library is rich and your operating conditions stationary, similarity methods frequently beat more elaborate models.

Matching to history

A small library of historical degradation trajectories (gray), each belonging to a unit that ran to failure long ago. A new unit (dark) is running right now; we've observed it up to t_current. The method finds the K library members whose early-life behavior most closely matches what we've seen so far and highlights them. Each of those K has a known remaining life beyond t_current; their average is our predicted RUL.

LibraryNeighborsLive unit
Slide the current time forward: early on, many library trajectories are plausible matches and the RUL estimate is noisy; later, as more of the live unit's history is observed, neighbors converge and the prediction sharpens. This is the fundamental trajectory of any prognostics method — uncertainty shrinks as evidence accumulates.

Windowed features & ML regression

The second family generalizes the first. Instead of matching full trajectories, summarize each window of recent signal values into a handful of features — mean, standard deviation, trend, peak, kurtosis, frequency-band energies, whatever you suspect carries information about health — and train a supervised regressor to map those features directly to RUL. The regressor can be anything: linear regression, random forest, gradient boosting, kernel methods, shallow neural networks. Training data comes from historical run-to-failure trajectories: for each timestep in each historical unit, you know the features from its window and you know its true RUL at that moment, so you have a labeled training example. Train the regressor on the pooled set, then apply it to live windows.

This approach sits in a sweet spot. It's more flexible than similarity matching (can represent nonlinear relationships between features and RUL), but more constrained than end-to-end deep learning (you choose the features, so you inject domain knowledge). It works well when the degradation process has a clear signature in a few well-chosen scalar features.

A model that counts down

Below is a single unit's degradation trajectory with a sliding window of recent history. Slide the current time and watch three things happen together: the window's position on the signal, the features extracted from it (mean, std, slope), and the RUL prediction those features produce. The lower panel compares predicted RUL to true RUL across the entire life — this is what validation curves for feature-based RUL models look like in practice.

A linear regressor fit to [mean, std, slope, max] of the window predicts RUL. Early in life, all windows look boring and the prediction sits near the population-average expected RUL. As degradation kicks in, feature values move rapidly and the regressor tracks the countdown. The gap between dashed (true) and solid (predicted) tells you how accurate the model is at each life stage.

Sequence models: LSTMs & Transformers

The next step is to let the model choose its own features. Instead of hand-engineering summaries of a window, feed the raw windowed time series into a sequence model — a Long Short-Term Memory network (LSTM), a 1D convolutional network, or more recently a Transformer — and let it learn both the feature extraction and the RUL regression end-to-end. The loss function is the same as before: squared error between predicted and true RUL, often with asymmetric weighting to penalize late predictions more than early ones (an unannounced failure costs more than an unnecessary replacement).

Sequence models shine when the degradation process has temporal structure that hand-engineered features miss: subtle changes in the frequency content of vibration, gradual shifts in the shape of current waveforms, early-warning patterns that span many timesteps and look like noise from any fixed window. They are harder to train than feature-based methods (more hyperparameters, more data needed, more prone to overfitting), but they set the state of the art on public benchmarks like the NASA CMAPSS turbofan engine dataset, and they are what you reach for when simpler models plateau below your accuracy target.

A contemporary RUL sequence model typically looks like: a normalization layer, a few 1D conv or LSTM layers for temporal feature extraction, optional self-attention if the window is long, a small MLP head mapping the pooled features to a single RUL number. Dropout and weight decay for regularization, Adam as the optimizer, early stopping on a held-out set of units (not of timesteps — more on that below). The work is less in architecture and more in data preparation and validation protocol.

Gaussian Processes & honest uncertainty

The previous three families give you a point estimate of RUL. For many decisions that isn't enough. Whether to replace a blade in a jet engine, whether to restrict the duty cycle of a motor, whether to pull a battery pack out of service — these decisions depend on the distribution of possible remaining lives, not just its mean. An asset manager choosing between "replace now" and "run ten more days" needs to know the probability of failure in those ten days, not just the expected life. This is the domain of Gaussian Process regression, and more generally of probabilistic prognostics.

A Gaussian process models the degradation trajectory as a random function with a prior encoded by a mean function (capturing the average trend) and a covariance function (capturing how smoothly the degradation evolves). Conditioning the prior on observed data gives a posterior over future trajectories: not a single prediction, but a Gaussian distribution at every future time. Project that posterior forward to when it crosses the failure threshold and you get a distribution of RUL values — a mean, a median, a 5th percentile, an upper bound on optimism. The figure below uses Bayesian linear regression with a quadratic basis, which produces the same qualitative behavior as a GP with quadratic mean function: uncertainty narrow where you have data, widening where you extrapolate.

Confidence bands that mean something

A single unit's degradation trajectory, observed up to t_current. The model fits the observations and extrapolates forward, producing a posterior mean (the gold curve) and a 95% confidence band (shaded). Where the band meets the failure threshold, we have a distribution over RUL values, shown as the PDF at the bottom. As you slide the current time later, more data arrives, the band narrows, and the RUL distribution sharpens.

ObservationsMean95% bandThreshold
Early on, the extrapolation fan is very wide — reflecting that a quadratic fit on sparse early data leaves the future genuinely uncertain. As observations accumulate, the posterior mean curves to match the accelerating degradation and the band tightens around the true trajectory. The RUL distribution at the bottom sharpens from a broad shape into a narrow one. This is probabilistic prognostics in one figure.

Notice something important: even the narrow band at 85% observation still has meaningful width. Honest uncertainty never collapses to a point estimate, and that is a feature, not a bug. The question "am I 95% confident this unit will last another 20 cycles?" has a clean answer here — it's whether the 5th percentile of the RUL distribution exceeds 20. No point estimate, however accurate, gives you that answer.

Which tool when?

MethodStrengthsWeaknessesReach for it when…
Threshold crossingDead simple; interpretable; no model to train.Single-unit RUL is just extrapolated threshold crossing, no uncertainty.The degradation signal is clean and the threshold is well-defined.
Similarity-basedNo distributional assumptions; works with small libraries; easy to explain.Fails under operating-condition shift; performance caps at neighbor quality.You have run-to-failure histories from the same regime as the live unit.
Feature-based MLInjects domain knowledge via features; robust; fast training.Only as good as the features; misses temporal dependencies within a window.Known degradation physics gives you strong candidate features.
Sequence / deepLearns features end-to-end; captures subtle temporal patterns; strong on benchmarks.Needs lots of data; hyperparameter-heavy; poor out-of-distribution behavior.Feature-based methods plateau and you have hundreds of run-to-failure traces.
Cox / survivalHandles censoring natively; interpretable hazard ratios; well-developed theory.Proportional-hazards assumption often violated; requires careful validation.Population-level RUL with strong covariate effects matters more than per-unit curves.
Gaussian ProcessCalibrated uncertainty; smooth extrapolation; robust with small data.O(n³) scaling; kernel & mean-function choice matters; calibration needs care.Your decisions depend on RUL distributions, not just point estimates.

Production systems often combine several. A common layered architecture: use a Gaussian process for short-horizon trajectory extrapolation, feed GP features into a feature-based regressor for RUL, and use a separate Cox model for population-level alerting when covariates shift. The right answer is usually not a single tool but a small ensemble that exposes where its members disagree — because disagreement is where you should look first when something unexpected happens.

Where these methods earn their keep

Battery State-of-Health & RUL

Capacity fade and internal-resistance growth over charge-discharge cycles. EV packs, grid storage, consumer electronics. Sequence models on cycle-level features have become standard; GP layers provide uncertainty for warranty and second-life decisions.

Bearing & rotating-machinery prognostics

Vibration RMS, kurtosis, and envelope-spectrum indicators as degradation signals. Classic domain for similarity-based and feature-based methods; deep learning has made steady gains on public bearing datasets (FEMTO, IMS).

Aircraft engine prognostics

Turbofan engine RUL from multivariate sensor streams. The NASA CMAPSS benchmark has been the proving ground for nearly every published RUL method; LSTM and CNN-LSTM architectures dominate current leaderboards.

Power electronics & inverters

IGBT and MOSFET aging through bond-wire lift-off, solder fatigue, gate-oxide wearout. On-state resistance and thermal impedance as degradation features. Feature-based ML is the workhorse; GP uncertainty matters for safety-critical automotive qualification.

Semiconductor fab equipment

Chamber-pressure sensors, RF-power monitors, and optical-emission spectroscopy predicting equipment downtime. High-volume, high-value maintenance decisions where even small RUL improvements justify substantial modeling effort.

Motor winding & insulation

Partial-discharge monitoring, dielectric dissipation, temperature rise histories. Long-tail lifetimes where survival analysis dominates; RUL prognostics come into play under accelerated-aging test regimes.

Pipeline & civil infrastructure

Corrosion coupon data, acoustic emission, strain gauges on bridges and towers. Low sampling rate, extreme variability in environments — Gaussian processes and physics-hybrid models outperform pure black-box ML here.

Medical device lifecycling

Pacemaker battery drain, pump-wear cycles, implant fatigue. Rigorous uncertainty quantification is not optional; probabilistic prognostics with calibrated intervals is the regulatory bar.

Practical notes from the trenches

  1. Validate by holding out whole units, never timesteps within a unit. A model that scores well on the last 20% of timesteps from each training unit will dramatically overestimate its generalization. The honest protocol: split your units into train/val/test sets, and evaluate the test set as if each unit had never been seen.
  2. Use asymmetric scoring functions. A late RUL prediction (you said 50 cycles, it failed at 30) costs more than an early one. The NASA PHM challenge score function is the canonical asymmetric metric: exp(−d/10)−1 for early, exp(d/13)−1 for late, where d is the prediction error. Train with it if your downstream decision is asymmetric.
  3. Clip the target RUL during training. Early in a unit's life, true RUL can be enormous (hundreds of cycles out). Asking the model to distinguish 300 from 320 cycles hurts more than it helps. Clip labels to a ceiling (e.g., 125 in CMAPSS) so the model focuses on the portion of life where RUL matters.
  4. Operating-condition drift will wreck your similarity matches. If your library units all ran at 1000 rpm and your live unit is at 1500 rpm, matching on raw signal amplitudes is a disaster. Normalize by operating regime first, or condition your similarity metric on operating-point covariates.
  5. Uncertainty quantification is not a feature — it's the deliverable. The point estimate is useful for reporting. The distribution is what a maintenance scheduler actually consumes. Always report at least a 90% prediction interval, and calibrate it on held-out data (Prediction Interval Coverage Probability, PICP).
  6. Beware the hockey-stick overfit. Real degradation curves are flat for 80% of life, then bend sharply. Models that minimize global MSE can over-smooth the knee and predict reasonable averages while missing the critical end-of-life rapid transition. Check your residuals as a function of true RUL, not just overall.
  7. Feature engineering is still king for small datasets. If you have 20 run-to-failure traces, a LSTM will overfit; a random forest on well-chosen features will generalize. The rule of thumb: deep learning starts winning around 100+ units with rich multivariate sensing. Below that, domain-informed features with a tree ensemble or GP is the safer bet.
  8. Monitor RUL predictions against actual survivals in production. Deployment is an experiment. Log every prediction with its timestamp; when a unit fails or is retired, compute the realized RUL and compare. A drift in RUL accuracy is often the earliest signal that your operating conditions have shifted away from your training distribution.
  9. Physics-informed models beat pure black-box models when physics is available. If you know the degradation follows a Paris-law crack growth, an Arrhenius temperature dependence, or an electrochemical fade pattern, bake that into the model as a structured backbone with ML for the residuals. You get interpretable extrapolation and much better small-data behavior than a pure data-driven fit.
  10. Don't confuse RUL estimation with anomaly detection. RUL assumes the degradation mode is known and the end is approaching along a predictable trajectory. Anomaly detection catches unexpected events that don't fit the training distribution at all. You need both — anomaly detection as a gatekeeper, RUL for the expected wear path.

References & further reading

  1. Saxena, A., Goebel, K., Simon, D., & Eklund, N. (2008). Damage Propagation Modeling for Aircraft Engine Run-to-Failure Simulation. International Conference on Prognostics and Health Management. Introduces the CMAPSS turbofan benchmark.
  2. Wang, T., Yu, J., Siegel, D., & Lee, J. (2008). A Similarity-Based Prognostics Approach for Remaining Useful Life Estimation of Engineered Systems. PHM 2008.
  3. Heimes, F. O. (2008). Recurrent Neural Networks for Remaining Useful Life Estimation. PHM 2008. Won the original CMAPSS challenge.
  4. Sateesh Babu, G., Zhao, P., & Li, X.-L. (2016). Deep Convolutional Neural Network Based Regression Approach for Estimation of Remaining Useful Life. DASFAA.
  5. Zheng, S., Ristovski, K., Farahat, A., & Gupta, C. (2017). Long Short-Term Memory Network for Remaining Useful Life Estimation. PHM 2017. Reference LSTM baseline.
  6. Li, X., Ding, Q., & Sun, J.-Q. (2018). Remaining Useful Life Estimation in Prognostics Using Deep Convolution Neural Networks. Reliability Engineering & System Safety, 172, 1–11.
  7. Ellefsen, A. L., Bjørlykhaug, E., Æsøy, V., Ushakov, S., & Zhang, H. (2019). Remaining Useful Life Predictions for Turbofan Engine Degradation Using Semi-Supervised Deep Architecture. Reliability Engineering & System Safety, 183, 240–251.
  8. Rasmussen, C. E., & Williams, C. K. I. (2006). Gaussian Processes for Machine Learning. MIT Press. Freely available online.
  9. 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, 213, 1–14.
  10. Lei, Y., Li, N., Guo, L., Li, N., Yan, T., & Lin, J. (2018). Machinery Health Prognostics: A Systematic Review from Data Acquisition to RUL Prediction. Mechanical Systems and Signal Processing, 104, 799–834.
  11. Fink, O., Wang, Q., Svensen, M., Dersin, P., Lee, W.-J., & Ducoffe, M. (2020). Potential, Challenges and Future Directions for Deep Learning in Prognostics and Health Management Applications. Engineering Applications of Artificial Intelligence, 92.
  12. Severson, K. A. et al. (2019). Data-Driven Prediction of Battery Cycle Life Before Capacity Degradation. Nature Energy, 4, 383–391. Influential paper on early-life battery RUL.