Predicting remaining useful life, method by method.

What an RUL prediction actually is

Prognostics, the discipline RUL prediction belongs to, is defined in ISO 13381-1 as the estimation of the time to failure and the risk of existing or future failure modes. The distinction from diagnostics matters: diagnostics tells you what is wrong with the machine now, prognostics tells you how long before it stops doing its job. Diagnostics looks backward at damage already done; prognostics runs that damage forward in time.

A serious RUL prediction is a distribution, not a number. "This bearing fails in 42 days" is false precision; "between 30 and 60 days with 90% confidence" is something a planner can act on. The width of that interval is information in itself: it tells you how urgently you need better data, and it typically narrows as the asset degrades and evidence accumulates.

A prediction is also never final. Every new vibration reading, temperature trace, or oil sample is evidence, and the estimate should update as it arrives. RUL prediction is a continuous process that tracks the asset, not a report that gets filed.

Physics-based models: simulate the damage

Physics-based methods model the degradation mechanism itself with first-principles equations. Fatigue cracks are the textbook case: Paris' law relates how fast a crack grows to the stress cycles the component sees. Given the current damage state and the expected future loading, the model integrates the damage forward until it reaches the failure threshold. The time that takes is the RUL. The same logic applies to bearing spall propagation, wear, creep, and corrosion, wherever the mechanism has an established mathematical description.

The strengths follow directly. A physics-based model needs little or no failure history, because the equations already encode how the failure unfolds. It extrapolates to operating conditions never seen in the data, because the physics holds there too. And its predictions arrive in engineering quantities (crack length, stress, remaining cycles) that an engineer can audit against the actual machine.

The costs are equally real. Each failure mode needs its own model, built and validated by someone who understands the mechanism, and the model's parameters must be identified for the specific asset. Hardest of all, the internal damage state is rarely observable directly: nobody can see the crack inside an operating gearbox. Estimating that hidden state from indirect measurements is a substantial part of the work, and errors in it propagate into the prediction.

Statistical degradation models: extrapolate the trend

Statistical methods skip the mechanism and model the symptom. A health indicator is built from condition-monitoring data (vibration energy, temperature, particle counts in oil), its evolution over time is modeled as a random process, and the RUL is the time until that process first crosses a failure threshold. Wiener processes with drift and gamma processes for strictly accumulating damage are the standard tools, and the review by Si and colleagues remains the field's reference map of this territory.

The appeal is that uncertainty comes built in. Because the model is a random process, the time to threshold is naturally a probability distribution, exactly the form a maintenance decision needs. Data requirements are moderate: degradation trajectories are enough, and complete failures are not strictly required to fit the trend, though some failure evidence is needed to place the threshold credibly.

The method's weight rests on two assumptions. First, that the health indicator actually tracks the damage; an indicator that responds to load changes or sensor drift rather than degradation produces confident nonsense. Second, that the future degrades like the past. A change in operating regime, a different product mix, or an intervening repair breaks the extrapolation, and the model has no physics to fall back on when it does.

Machine learning: map data to lifetime

Machine learning methods learn the mapping from sensor data to remaining life directly from examples. Deep learning removed the need to hand-design features: recurrent and convolutional networks ingest raw sensor sequences and output an RUL estimate. Most published results are benchmarked on NASA's C-MAPSS dataset, a set of simulated turbofan run-to-failure trajectories that has become the field's standard test bed since Saxena and colleagues introduced it in 2008.

That benchmark choice reveals the central problem. Supervised learning needs many examples of the journey from health to failure, and critical industrial equipment is maintained precisely so that journey never completes. The field benchmarks on simulated engines because real run-to-failure histories are rare. Lei and colleagues' systematic review of machinery prognostics makes the point plainly: most published methods are validated on laboratory or simulated data, and industrial deployments are far scarcer than the publication volume suggests.

Two further limits matter for expensive decisions. Pure data-driven models extrapolate poorly, and give no warning when they leave the conditions they were trained on. And their predictions arrive without engineering reasons, which makes them hard to act on when the action is shutting down a production line. Our guide on physics-informed neural networks covers both failure patterns in detail.

The four families, compared

The trade-offs line up cleanly once you ask what each family needs before it can work:

Directional comparison drawn from the prognostics reviews cited below. Individual implementations vary.

Read as a decision aid: if the failure physics is established and failures are rare, start physics-based or hybrid. If a trustworthy health indicator exists and conditions are stable, statistical methods are efficient and honest about uncertainty. Pure machine learning earns its keep on large fleets of inexpensive, identical components where run-to-failure data genuinely accumulates. The more critical the asset, the higher the bar for explainability, and the more the balance tilts toward methods whose predictions can be checked.

Hybrid methods: physics plus learning

Hybrid methods pair the two things each side does best. The physics supplies structure: the form of the degradation law, conservation constraints, the bounds of what the asset can physically do. The learning fills in what the physics leaves out: unmodeled effects, per-asset quirks, the gap between a clean equation and a dirty machine. Prognostics reviews, including Lei and colleagues', point toward these combinations as the field's direction of travel, because they attack the data problem and the trust problem at the same time.

Physics-informed neural networks are the most prominent recent form of the idea: the governing equations are built into the model's training objective, so predictions that violate the physics are penalized before they ever reach an operator. The practical effect on RUL work is that the model can reason about a failure mode it has never seen an example of, because the physics of that failure is encoded even where the data is absent.

For an operator of critical equipment, the hybrid family resolves the dilemma the other three each leave open: it produces remaining-life estimates without demanding years of failure history, keeps those estimates inside what is physically possible, and expresses them in terms an engineer can verify. The price is upfront engineering effort, which is why it is the natural choice for assets whose failures are expensive, and overkill for assets whose failures are not.

How Vermilion estimates remaining useful life

Vermilion sits squarely in the hybrid family. Its Large Reasoning Model reasons against first-principles digital twins of the equipment it watches (bearing dynamics, gear-mesh harmonics, fluid-film stability), with constitutive parameters fitted per asset under known physical priors. Degradation is tracked as a physical state, and remaining-life estimates come from running that state forward under the asset's expected duty, with the uncertainty carried through.

The consequences are the ones this guide has been describing. Vermilion produces useful remaining-life estimates without years of labeled failure history, refuses estimates that would violate the physics, and states each one in engineering terms (loads, temperatures, remaining margin) that a reliability team can verify against the machine before committing a maintenance window to it.