Physics-informed prognostics, and where the failure estimate comes from

Diagnostics identifies what is happening now and why. Prognostics estimates how long before a defined condition is reached. They are frequently sold together and they are different problems with different evidence requirements.

The difference that matters is verifiability. A diagnostic claim can be checked by inspection: open the machine and see whether the named mechanism is there. A prognostic claim can only be checked by waiting, and by then the decision it informed has already been taken.

That asymmetry is why prognostics attracts more overstatement than any other part of this field, and why the discipline around it has to come from method rather than from results.

Physics-based prognostics models the degradation mechanism itself: a crack growth law, a wear rate, a fouling accumulation. Where the mechanism is well characterised and the loading is known, this projects from a described process rather than a fitted curve, which is a better basis outside the data. It is not a guarantee, and the projection still has to be validated over the range it is used in. Where the mechanism is not well characterised, which is most industrial cases, it is a confident guess with equations attached.

Data-driven prognostics learns from degradation histories, and it works best where a population has run to failure, which in practice means aerospace, automotive and anything else with a large fleet and instrumented failures. Censored histories, accelerated test data, transfer learning and simulation can substitute in part. On a plant with eight of something and no failure history, what remains is thin.

Hybrid prognostics uses a physical model of the degradation with data-driven correction of its parameters from the specific asset. It is the one of the three that usually suits industrial energy assets, because it needs neither a fully characterised mechanism nor a fleet of failures.

A statistical projection extends an observed trend. That is defensible while the future resembles the past, and on a renewable-coupled asset the future routinely does not: the resource varies year to year, the tariff structure changes, the operating strategy is still being tuned.

A physical model projects under stated conditions instead of under continuity. Given this duty, this temperature history and this degradation law, the state reaches the criterion at approximately this time. Change the duty assumption and the answer changes, which is a feature rather than a weakness, because duty is exactly what a plant can decide.

That makes the output usable in a way a trend extension is not. It supports the question an operator actually asks, which is what happens if we run it differently, rather than only the question of what happens if nothing changes.

The same reasoning underlies converting a measured trend into a replacement date, which is set out in remaining useful life reporting for lenders.

Every prognostic carries at least three uncertainties: in the current state estimate, in the degradation model, and in the assumed future loading. They compound, and the compounded range is usually wide enough that a single number misrepresents it.

A defensible output is a distribution or a range with a stated confidence, the criterion it was computed against, the duty assumption it used, and the date it was produced. Five items, and their absence is the fastest way to tell a marketing figure from an engineering one.

There is a commercial temptation in the other direction, because a range is harder to put on a slide than a date. It is worth resisting for a practical reason as well as an honest one: the first time a single-number prediction is visibly wrong, everything else the system says loses credibility with the people who have to act on it.

The hard question for any prognostic claim is what it was validated against, and the honest answers are limited.

Seeded fault testing, where a known degradation is introduced deliberately in a test rig, gives a controlled check of whether the method detects and projects the mechanism. It does not prove behaviour on a plant.

Historical backtesting against events that did occur, with the model given only data from before each event, is the strongest available evidence and requires a failure history most sites do not have.

Consistency checking is the weakest but most available: does the estimate remain stable as new data arrives, or does it swing on every update. A prognostic that changes its answer by months each week is reporting noise.

Whatever is claimed, the sample matters more than the accuracy figure. An accuracy quoted without the number of events behind it is uninterpretable, and the number is often very small.

The end-of-life criterion, stated explicitly. It is an input rather than an output, and changing it changes the answer more than the model does.

The uncertainty range, and what it is a range over.

The duty assumption, and whether alternatives were run.

The validation basis, with the number of events.

The failure modes the system does not attempt to predict. Some mechanisms are effectively instantaneous, belong to protection systems, and claiming to predict them is one of the reliable markers of a vendor overselling.

And what happens when the asset moves outside the range the model was calibrated on, because on a variable-duty plant it will, and a system without an answer to that is going to be confidently wrong at exactly the wrong time.