OptimalTREND™ Wiki / Detecting failures
Method
Detecting failures
How OptimalTREND™ turns raw signals into a warning with a date on it: learn the baseline, alarm on the right features, estimate the life remaining, and point at the cause.
The method
- Connect to existing sensors, historian or SCADA.
- Learn each asset normal envelope across many signals and operating states.
- Detect multivariate deviations and score their severity.
- Estimate remaining useful life from the degradation trend.
- Alert with the likely failure mode, and learn from the outcome.
Alarm strategies
Alarming well is the hard part of detection: early enough to act, specific enough to diagnose, and quiet enough to be trusted. Mature condition monitoring layers several alarm types rather than relying on one.
| Overall, broadband level | The trended overall, velocity in mm/s RMS assessed against ISO 20816 zones A to D, or acceleration and enveloped g. Simple and robust, but late: a defect frequency can grow many times over before it moves the overall. |
|---|---|
| Standards and vendor limits | Absolute limits set by machine class, power and mounting (ISO 20816), or by the OEM. Usually two-tier: an Alert to plan and investigate, and a Danger or Trip to protect. Permanent protection systems follow API 670. |
| Statistical baseline | Limits learned from the machine itself, typically the mean plus two or three standard deviations, set per operating state so a load or speed change does not read as a fault. Self-tunes to each asset, but needs a clean learning period. |
| Spectral band alarms | Alarm bands placed on the frequencies that matter for that machine, 1x and 2x, the bearing defect frequencies, gear-mesh and their sidebands, so a rise is attributed to a cause, not just noticed. Far earlier than the overall. |
| Envelope and PeakVue bands | Thresholds on the enveloped, demodulated signal, tuned to impacting. The earliest reliable warning of rolling-element bearing and gear-tooth defects, at stage 1 to 2. |
| Spectral mask, adaptive | A learned baseline spectrum with a tolerance envelope, or mask, laid over it. Anything that pokes through the mask alarms, which suits complex spectra where fixed bands are hard to place. |
| Rate-of-change | Alarms on the slope of the trend, not just its level, so an accelerating degradation is caught while the absolute value is still moderate. |
| Model-based, multivariate | An expected-value model across vibration, temperature, load and speed flags the residual when reality departs from the model. It normalises for operating condition and catches the combined early signature a single tag misses. This is the analytics layer. |
Keeping alarms trustworthy. Early alarming is worthless if it cries wolf. Good practice gates alarms to the running-and-loaded state, suppresses them through start-up and transients, requires persistence over several readings before raising, latches and asks for acknowledgement, and prioritises by asset criticality. False positives, not missed faults, are what kill a monitoring programme.
OptimalTREND™ combines these: it learns the normal envelope of each asset per operating state, derives the band and envelope alarms from the asset kinematics, watches the multivariate residual and the trend slope, scores severity so attention goes to the few that matter, and learns from each confirmed outcome. The result is an alarm that is early, attributed to a likely failure mode, and trusted.
Failure rate, Weibull and availability
How an item fails sets the strategy. Weibull shape parameter β classifies it:
| β < 1 | Decreasing failure rate, infant mortality. The answer is burn-in and quality, not a scheduled overhaul. |
|---|---|
| β = 1 | Constant failure rate, random. R(t) = e−t/η. On-condition detection is the only real lever. |
| β > 1 | Increasing failure rate, wear-out. A useful life exists, so age-based and condition-based both apply. |
Reliability follows R(t) = e−(t/η)^β, and steady-state availability is MTBF ÷ (MTBF + MTTR). Predictive maintenance moves availability by cutting the unplanned share of MTTR and by catching random and wear-out failures inside their P-F interval.
Remaining useful life
Remaining useful life is the time from now until the asset can no longer do its job, the point F on the DIPF curve, not the point P where the defect first became visible. Estimating it is prognostics, and the honest output is a window with a confidence, not a single date.
| Trend extrapolation, data-driven | Fit the health feature and project it to the failure threshold. Linear while degradation is slow, exponential or power-law as it accelerates toward F. The workhorse, and the basis of the estimate below. |
|---|---|
| Physics of failure, model-based | Project a known degradation mechanism, crack growth, bearing spall growth, wear or fatigue, using its physics. Accurate where the mechanism and the loads are well understood. |
| Similarity and machine learning | Match the current degradation path against historical run-to-failure trajectories for the same asset type, and read the life left from the closest analogues. |
| Reliability statistics, Weibull | A population prior from the Weibull life. Useful as a sanity check and a starting point, but it is a population figure, not an individual prediction, so it never replaces the measured trend. |
Three things shape a usable RUL. It depends on the chosen failure threshold, so a conservative limit buys earlier action. Its confidence band is wide far out and tightens as the defect advances and F nears. And the decision it serves is simple: is the net P-F interval, the life left minus the time to detect, plan and act, long enough to raise the work, get the parts and take the outage. Consistent with diagnosis, not prediction, OptimalTREND™ estimates RUL from the developing trend of the asset itself and updates it on every reading, rather than quoting a population average.
Once a health feature is trending to a limit, a first-order estimate of the life left is how far it is above the failure threshold divided by its rate of change:
RUL ≈ (Current health − Failure threshold) ÷ Degradation rate
Worked example: health at 80, threshold 30, falling 5 a month, gives (80 − 30) ÷ 5 = 10 months. Real prognostics fit the degradation curve, usually exponential or power-law as it approaches F, and carry a confidence band; the linear estimate is the intuition and the first alert, not the final word.
Try it
Set an asset current health and how fast it is degrading to see the first-order warning window.
Failure threshold fixed at 30. Green above 6 months of warning, amber 2 to 6, red under 2.
Warning window
The ring shows current health; the number is the first-order remaining useful life.