Predictive maintenance models compared: rules, anomaly detection and remaining useful life

Maintenance teams choose between four broad strategies. Reactive maintenance repairs after failure. Preventive maintenance replaces or services on a calendar or usage count. Condition-based maintenance acts when a measured condition crosses a limit. Predictive maintenance goes one step further and estimates how the condition will develop, so work can be planned before the limit is reached.

Every predictive approach depends on one physical fact: the failure mode must show a detectable precursor some time before functional failure. Reliability engineers call the gap between the first detectable sign and the failure the P-F interval. If that interval is shorter than the time it takes to get parts, people and a maintenance window together, no model will help, however accurate it is on paper.

So the comparison below starts from two questions per failure mode: is there a measurable precursor, and how much labeled history exists for it?

A bearing temperature limit from the manufacturer, a differential-pressure limit across a filter or a motor current ceiling are rules with physics behind them. They are cheap, auditable and easy for a technician to act on, and a model should be judged on whether it adds lead time beyond the rule.

Statistical baselines improve on fixed limits by asking what is normal for the current operating mode. A pump at partial load vibrates differently from one at full load. Segmenting history by mode (load band, speed, product grade, ambient range) and applying statistical process control within each segment removes many of the false alarms that make fixed limits unpopular, at the cost of re-establishing baselines after overhauls and control retuning.

Unsupervised methods learn the structure of normal multivariate behavior and score how far new data departs from it: principal component analysis with reconstruction error, isolation forests, one-class models and autoencoders are common choices. Most organizations have years of sensor history but very few cleanly recorded failures, which is their appeal.

Their limit is equally clear. An anomaly score says that something is different, not that a component is degrading. Sensor faults, process changes and a new product grade all look anomalous. That makes triage part of the system design: each anomaly needs contribution plots showing which signals moved, and an engineer's verdict recorded against it. Those verdicts become the labels that, over time, allow a supervised model to be trained.

A supervised classifier learns to recognize the signature of a specific failure mode in a window before it occurs. It needs enough labeled examples of that mode, on assets similar enough to share a signature, plus reliable timestamps. Work-order records rarely provide this directly: failure codes are often generic, dates record when work was done rather than when degradation began, and many failures are fixed during other work without being coded at all.

Remaining useful life models estimate time to failure from a degradation path. Survival models, such as Weibull or proportional hazards models, are often the better statistical framing because they use censored units, those removed or repaired before failing, instead of discarding them. Either way, the output should be a distribution, and the maintenance decision should be tied to a probability of failure within the planning horizon rather than to a single date.

Two families of international standards give useful structure. ISO 17359, Condition monitoring and diagnostics of machines: General guidelines, sets out how to build a condition monitoring program, from identifying equipment and failure modes to choosing measurements and alert criteria¹. ISO 13374, Condition monitoring and diagnostics of machines: Data processing, communication and presentation, describes the processing chain from acquiring data to generating advice².

MIMOSA's OSA-CBM specification is an implementation of the ISO 13374 functional specification and organizes that chain into six functional blocks³: data acquisition, data manipulation, state detection, health assessment, prognostic assessment and advisory generation. Mapping each approach in the table to those blocks is a quick way to see that a rule covers state detection only, while a remaining useful life model reaches into prognostics and still needs an advisory step people can act on.