Prognostics
& Health Management

Forward in time.

A bearing has a crack. It is small. The machine is running. The question that matters is not only is something wrong — diagnostics answered that — but how long do I have? Weeks? Hours? Until the next shift? Prognostics is the discipline of answering that question, quantitatively, with honest uncertainty.

Prognostics and Health Management (PHM) sits where diagnostics becomes a planning tool. A diagnostic system catches a fault the moment it becomes distinguishable from noise. A prognostic system reads the fault's trajectory and projects forward — combining what it has seen with what it knows about how this kind of fault tends to grow — to estimate when the asset will reach the end of its useful life.

The central object is the health indicator (HI), a scalar that captures the asset's current condition, and its evolution — degradation — over time. The central output is Remaining Useful Life (RUL): a distribution, ideally, over the time until the health indicator crosses a failure threshold.

Two distinctions worth making up front. Diagnostics vs. prognostics: diagnostics is a present-tense claim about the system’s state; prognostics is a future-tense claim, necessarily uncertain. And point estimate vs. distribution: a prognostic that outputs “42 hours” without a confidence band is almost useless. A prognostic that outputs “RUL is log-normal with median 42 h and 90% CI [18, 97]” lets a maintenance planner actually decide something.

Remaining Useful Life — a first-passage time.

RUL is the time from “now” until the asset first crosses a failure threshold. Every prognostic method, whatever its machinery, is ultimately trying to estimate this distribution.

Three families dominate in practice. They differ in whether they rely on a physics model, a population of historical units, or raw pattern recognition.

2.1Particle filter — physics plus Bayes

The particle filter works when you have a stochastic degradation model and noisy measurements. At each time step, it carries a cloud of particles — each one a hypothesis about the system’s current hidden state. Particles are propagated forward by the degradation model. When a measurement arrives, particles are weighted by how well they explain it, and the cloud is resampled to concentrate near the likely state.

To estimate RUL, each particle is extrapolated further into the future (without new measurements) until it crosses the failure threshold. The collected crossing times form an empirical RUL distribution.

2.2Similarity-based — the library of histories

Sometimes you have no physics model but you do have a library of run-to-failure histories from similar units. For the current asset, you find the historical units whose trajectory most closely resembles the one you are observing, and use their remaining lives to estimate yours.

This is simple, surprisingly effective on real data, and scales beautifully as the library grows. It assumes the current unit will age like the library did — an assumption that needs checking when populations drift.

2.3Deep learning — learn the mapping

Given enough labeled run-to-failure data, deep networks can map raw sensor windows directly to RUL. CNN and LSTM architectures dominate the public benchmarks (C-MAPSS, PRONOSTIA, PHM Society challenges). Transformers and temporal convolutional networks have joined the fray more recently.

How they compare in practice

• Particle filterPrincipled uncertainty, requires a physics or stochastic model, computationally heavy for large N.
• Similarity-basedMinimal modeling, needs a run-to-failure library, honest uncertainty via nearest-neighbor spread.
• Deep learningHigh performance when data is abundant and stationary, uncertainty estimation is an open research problem.

Degradation modeling — the shape of getting old.

Before predicting when an asset will fail, you need a model of how it ages. Three stochastic processes dominate the literature, each capturing a different character of aging.

• Wiener processDamage accumulates as a drift plus Brownian noise. Can fluctuate up and down — appropriate when the health indicator is noisy or reversible in the short term. First-passage-time to a threshold is inverse-Gaussian distributed, giving a closed-form RUL.
• Gamma processDamage arrives in strictly positive random increments. Paths are monotonically increasing and often jumpy. The right model for wear, crack growth, corrosion — anything that cannot spontaneously heal.
• Inverse Gaussian processA generalization with a specific moment structure; useful when the data shows heavier right tails in its failure times than Gamma can capture.

Health indicators — squeezing many sensors into one scalar.

A machine is a forest of sensors. Vibration RMS, bearing temperature, motor current harmonics, oil particle counts, acoustic envelopes. None of them, alone, tells the whole story of aging. A health indicator is a single scalar derived from the raw sensor set that captures “how healthy is this thing, overall.”

Construction is part craft, part mathematics. The common recipes:

• Weighted sumEngineering-assigned weights on normalized sensors. Transparent, tunable, common in production.
• PCA projectionProject onto the first principal component of the sensor matrix during degradation. Data-driven, captures the dominant degradation direction.
• Autoencoder latentTrain a neural network on healthy data; use the reconstruction error or a latent-space distance as HI.
• Mahalanobis distanceDistance from the current sensor vector to the healthy-distribution centroid, scaled by covariance.

The key property of a good health indicator is monotonicity: it should trend consistently from “healthy” to “failed” over the asset’s life, with as little back-and-forth as possible. A non-monotonic HI makes RUL estimation brittle and maintenance decisions confusing.

Uncertainty — often more important than the estimate.

A prognostic that says “RUL = 50 hours” is not actionable. A prognostic that says “RUL is 50 hours ± 3 hours with 95% confidence” schedules a maintenance visit. A prognostic that says “RUL is 50 hours ± 40 hours with 95% confidence” schedules a second measurement. The width of the interval is the content.

Two kinds of uncertainty must be tracked. Aleatoric uncertainty is the irreducible randomness of the degradation process — even perfect knowledge of the current state cannot predict the future Brownian noise. Epistemic uncertainty is the modeler’s own ignorance — uncertain parameters, mismatched models, insufficient observations. Aleatoric uncertainty narrows as you approach the threshold; epistemic uncertainty narrows as you acquire more data.

The practical output is a reliability function: the probability that the asset will still be operating at time τ in the future. This is the object a maintenance planner actually uses.

Condition-based and predictive maintenance.

The economic point of PHM. A health indicator that no one reads and an RUL distribution that no one acts on is a research project, not an engineering system. Maintenance strategies are how PHM pays for itself.

Four archetypes span the landscape:

• ReactiveRun to failure. Cheapest until the failure occurs; then it is the most expensive, with unplanned downtime and possible collateral damage.
• Scheduled (time-based)Maintain on a fixed calendar — every 5,000 hours, every 12 months. Prevents most failures but wastes remaining life on everything replaced early.
• Condition-based (CBM)Maintain when the current health indicator crosses a warning threshold. Uses diagnostics but not forecasting. Good improvement over scheduled.
• Predictive (prognostic-driven, PdM)Maintain when the RUL distribution predicts failure within a planning horizon. Consumes maximum useful life and minimizes surprise failures.