Predictive Maintenance Without the Hype: Which Failures Can Be Forecast?

The failure-pattern reality: most breakdowns aren't age-related

The single most useful correction to predictive-maintenance hype is understanding how equipment actually fails over time. Reliability research that began with United Airlines engineers Stan Nowlan and Howard Heap in their 1978 Reliability-Centered Maintenance report — and was later repeated in naval and industrial studies — found that most components do not wear out predictably with age. Only a small share of failures follow the classic bathtub curve; in the civil-aviation study roughly four percent of components did, and only about 11 to 20 percent of failures across the main studies proved strongly age-related.

In practice, most reliability engineers work with a simpler summary: only about 15–20% of failures can be forecast from age or running hours, while the other 80–85% appear random with respect to time. That does not mean they have no cause — misalignment, contamination, moisture, operator error and external shocks are all real causes — it means their timing cannot be predicted from a calendar. The consequence is that time-based overhauls alone will never prevent most breakdowns, and an analytics model that only counts operating hours will miss most of the value.

• Infant mortality: a burst of failures right after installation or overhaul.
• The chance or random period: a relatively constant failure rate for most of the service life.
• The wear-out zone: a sharp rise for components that genuinely age or cycle.

What condition monitoring can actually forecast

What can be forecast is degradation — a physical process that begins to change measurable condition well before the machine stops working. This is the potential-to-functional-failure (P-F) interval concept used in reliability-centered maintenance. Vibration analysis, oil and debris analysis, thermography, ultrasound and current monitoring detect the moment a defect becomes detectable (point P), leaving a window before functional failure (point F) in which you can plan a repair rather than react to a breakdown.

Concretely, the classes that lend themselves to prediction include bearing and bushing wear, shaft misalignment and unbalance, gear fatigue, impeller and pump wear, filter and screen clogging, lubrication breakdown visible in oil samples, insulation and winding degradation in motors, and gradual capacity loss in batteries. Each leaves a signature — rising vibration, rising temperature, metal particles in oil, rising current — that a sensor or an inspector can catch in time.

The failures prediction cannot catch

Prediction fails for failures that reach functional failure without a usable advance window, or whose cause is external and instantaneous. Operator error, debris or oversized material ingested into a machine, lightning and power surges, shipping or installation damage, and sudden single-point electronic failures typically offer no detectable P-F interval at all. A model cannot sound an early warning about an event that has no warning.

For these, the honest strategy is not more analytics but protection, redundancy and cause removal: guarding and interlocks, surge protection, spares and redundancy for single points of failure, and root-cause analysis that eliminates the triggering condition. Predictive maintenance that only flags these after the fact — or that mislabels a sensor fault as a real machine fault — wastes money and erodes trust.

How forecasts are made — and why they degrade

Where a measurable degradation process exists, three families of methods estimate remaining useful life (RUL). Purely data-driven models learn patterns from sensor histories; physics-based models simulate wear, fatigue and stress from first principles; hybrid physics-informed models combine both. Data-driven models are easier to deploy but are only as good as their training data, extrapolate poorly to never-seen conditions, and can produce physically impossible answers. Physics-based models generalize better but demand detailed knowledge of each component.

A recurring practical limit is that forecasts get less reliable the further they reach: accuracy that is high for tomorrow degrades sharply for next week. On top of that, few plants have clean full-lifecycle failure data, because machines are repaired before they fail. Treat any quoted "percent accurate" figure as context-specific rather than a promise.

Making predictions you can act on: false alarms, data and trust

The biggest operational killer of predictive maintenance is not the algorithm but the way alarms are used. Route-based monitoring only sees a machine at the moment of inspection, so a machine can fail between rounds; continuous sensors close that gap but can drown teams in noise. When many warnings turn out to be false positives, or cannot be acted on because crews are busy with today's breakdowns, technicians stop trusting the tool — a pattern sometimes called the predictive-AI paradox.

Good practice includes sizing the inspection interval as a fraction of the P-F interval so a symptom is seen several times before failure, checking that sensors actually sit where the signal exists (poor placement can make a fifth of them useless), confirming a reading before acting, and ensuring every confirmed prediction feeds a work order and a root-cause review. Detection alone does not remove the cause: unless the failure mode is eliminated, the same random fault can recur.

• Set measurement frequency to capture several points inside the P-F interval.
• Confirm a warning with a repeat measurement before stopping the line.
• Route every confirmed prediction to a work order and root-cause analysis.
• Avoid the "data graveyard": a sensor that is not connected or analyzed delivers nothing.

Where prediction belongs in your maintenance mix

Prediction is one tool in a portfolio, not a replacement for strategy. For strongly age- and cycle-driven parts that are cheap to replace, scheduled time-based maintenance may still be the most cost-effective answer. For degradation that offers a usable P-F window and where downtime is expensive, condition-based and predictive maintenance earns its keep. For high-consequence failures with no warning, invest in design, protection and root-cause elimination.

The screener below is a practical way to classify your own failure modes before you buy sensors and models. Most sites find that only a minority of assets justify full continuous prediction; the rest are better served by inspections, scheduled replacement, protection, or well-managed run-to-failure.