Abnormal reading ≠ abnormal process. An out-of-range value can mean the plant is deviating (a real process fault, ISO 14224 PDE) or that the instrument itself has drifted (AIH/AIL). Deciding which is the first, cheapest diagnostic move.
Instruments fail in a few predictable ways: zero drift (offset), span drift (gain error), plus stuck/frozen signals, saturation and noise. Calibration corrects the first two; validation logic catches the rest.
You can catch most bad instruments automatically — range limits, rate-of-change limits, redundancy voting, cross-checks against related signals, and a quality flag on every reading — so a drifting sensor is quarantined, not acted on.
1 · The fork: process or instrument?
When a value goes abnormal, the very first question is whose fault it is. Cheap checks, in order:
- Does anything corroborate it? A real high winding temperature shows up as higher current, more load, a warmer frame. If the temperature jumped but every related signal is flat, suspect the sensor.
- Is the move physically possible? A temperature that rose 40 °C in one second isn’t thermodynamics — it’s a wiring or sensor fault. Rate-of-change is a giveaway.
- Did it step, or drift? A sudden step to a rail (0% or 100%) usually means a broken wire or a dead sensor; a slow creep is classic drift.
- Does a redundant or nearby instrument agree? Two sensors, one story = trust it. Disagreement points straight at the odd one out.
ISO 14224 keeps them separate for a reason. PDE (parameter deviation) is a real process departure; AIH/AIL are the instrument reading high/low. Coding them correctly is what lets your reliability data distinguish “the machine keeps overheating” from “this transmitter keeps drifting” — two very different fixes.
2 · The anatomy of an instrument error
Almost every analogue instrument fault is one of these:
| Error | What happens | Signature |
|---|---|---|
| Zero drift | Whole range shifts up/down by a constant (offset/bias) | Error same at low and high end |
| Span drift | Gain error — reading scales wrong across the range | Error grows toward one end |
| Non-linearity | Reading bows away from true mid-range | Max error in the middle |
| Hysteresis | Different reading rising vs. falling | Loop between up/down sweeps |
| Stuck / frozen | Signal doesn’t move at all | Perfectly flat line (zero variance) |
| Saturation / rail | Pinned at 0% or 100% (or 4 / 20 mA) | Sudden step to a limit |
| Noise | Excess scatter on the signal | High variance, no trend |
The two that calibration fixes — and the two most often behind a slow AIH/AIL — are zero and span. The model below shows how they distort the reading across the range, and when they push it past an abnormal-reading limit.
Interactive — Zero & span error
Live modelDial in a zero offset and a span (gain) error and watch the indicated reading diverge from the true value across the 0–100% range. The dashed band is a ±2% tolerance; step outside it and the instrument is out of calibration.
Indicated vs. true
3 · Why instruments drift
- Sensor ageing — thermocouple decalibration, RTD element drift, strain-gauge creep, electrode fouling.
- Process coating / fouling / plugging — a coated electrode, a blocked impulse line on a pressure transmitter, buildup on a thermowell slowing and biasing the reading.
- Moisture & environment — water ingress, condensation, ground loops, EMI on long signal runs, terminal corrosion.
- Temperature effects on the electronics — ambient swings shifting a transmitter’s zero/span if it’s not temperature-compensated.
- Mechanical — a shifted or damaged thermowell, a loose mount, vibration loosening a connection.
4 · Calibration: zero, span & the paper trail
Calibration re-references the instrument to a traceable standard. The essentials:
- Zero and span are the two adjustments — set the low end, then the gain. A 5-point check (0/25/50/75/100%) exposes non-linearity and hysteresis too.
- As-found / as-left. Always record the error before you adjust (as-found) and after (as-left). The as-found error is the drift since last calibration — the data that sets your interval.
- Traceability. The reference must trace to a national standard (NIST), with its own calibration in date. An out-of-cal reference calibrates nothing.
- Interval. Tighten it for instruments that drift or matter (safety, custody transfer); relax it for stable, non-critical ones. Let the as-found history — not a blanket calendar — set it.
5 · Catching a bad instrument automatically
You don’t want a human eyeballing every tag. Build the checks into the data path:
| Check | Catches |
|---|---|
| Range limits (min/max valid) | Saturation, rail, gross faults |
| Rate-of-change limit | Physically impossible jumps — wiring/sensor faults |
| Flatline / frozen detection | Stuck signal (zero variance over time) |
| Redundancy & voting (e.g. 2oo3) | A single drifted sensor — outvoted by its peers |
| Cross-signal consistency | Temp up but current/load flat → suspect the sensor |
| Quality flag on every reading | Everything above — mark bad/uncertain so downstream logic ignores it |
The quality flag is the keystone. Every reading should carry good / bad / uncertain. A rule engine must never open an alarm on a non-good reading — a dead sensor should never trip a maintenance work order. Validate first; act only on trusted data.
6 · Drift is itself a trendable signal
Here is the twist that continuous monitoring adds: drift is a slow, monotonic trend — which makes it detectable the same way a real degradation is. A transmitter whose as-found error grows a little each calibration, or whose reading slowly departs from a redundant partner, is telling you it’s ageing. Trend the deviation and you can schedule the recalibration before it drifts far enough to raise a false alarm.
The discipline cuts both ways: a good predictive platform must separate “the machine is degrading” from “the sensor is drifting.” Cross-checks and the quality flag are what stop a drifting instrument from being mistaken for a failing machine — and sending a crew to a healthy asset.
Trusted data first, then predictions
Bluestream’s predictive platform carries a quality flag on every reading and validates signals before the rule engine ever sees them — range and rate limits, cross-signal checks, and derived features that flag a drifting or frozen sensor. Predictions are only as good as the data underneath them, so we guard the data first.
Talk to us about condition monitoring →Key takeaways
- Ask “process or instrument?” first — corroborate against related signals before dispatching a crew.
- Zero and span are the two calibratable errors and the usual cause of a slow AIH/AIL.
- Record as-found / as-left — the as-found error is your drift data and sets the interval.
- Automate validation — range, rate, flatline, voting, cross-checks — and put a
good/bad/uncertainquality flag on every reading. - Never act on non-
gooddata — a dead sensor must not trip a work order.