The most reliable finding in plant data is also the simplest: two things built to do the same job, under the same conditions, that have stopped behaving the same way.
It is reliable because it needs no baseline, no model of correct operation, and no agreement about what good looks like. The two units are each other's reference. If one has drifted from the other, something changed, and that something usually has a cost attached.
Why the control system will not tell you
A modern control system is very good at holding a value and raising an alarm when a limit is crossed. Divergence between two units is neither of those things.
Both units can sit inside every alarm limit while being nowhere near each other. A pump at 84% duty and its twin at 39% are both running. A basin at 5.35 mg/L and its twin at 0.49 mg/L are both aerated. Nothing is out of range in the sense the alarm system understands, so nothing fires, and the gap persists for years because it never presents as an event.
That is the gap worth watching, and it is visible in data a plant already records.
The three shapes it takes
Duty that never alternates
Lead and lag assets are supposed to swap. When the alternation is disabled, or was never configured, or was turned off during a commissioning problem and never turned back on, the lead unit accumulates all the wear.
In one plant's record, return-activated-sludge pumps showed 504 run hours against 232 in the same twenty-five days, and the lifetime counters showed the gap compounding: 44,299 hours against 30,953. The lead unit reaches overhaul years earlier than it needs to while the standby sits idle — and the standby is the one you will be relying on when the lead finally fails.
Lead-lag alternation costs nothing to enable. What makes this finding persuasive rather than theoretical is that the same plant's MBR blowers and permeate pumps were balanced, which marks it as an oversight on one asset rather than a design choice.
A shared supply, unequally split
Twin basins on a common air header should see similar dissolved oxygen at similar load. When one sits far above the other with its blower command pinned wide open, the control strategy is not the problem — the air is not arriving where it is supposed to.
This one is cheap to investigate and cheap to correct, which is what makes it worth finding. A valve position or a damper, not a capital project.
Instruments that no longer agree
Two turbidimeters on the same permeate stream, one reading a median of 0.11 NTU with a 90th percentile of 0.26, the other reading 0.14 and 2.12 — three and a half times the plant's posted upper limit.
One of them is drifting. Until you know which, compliance decisions are resting on an instrument that cannot be fully trusted. The encouraging part, in that case, was that the true value looked excellent.
Why frequent starts cost more than running
A related pattern worth naming on its own, because it is consistently underestimated.
Pumps are often judged on run hours. The more expensive variable is usually starts. Every start draws locked-rotor current — several times the running current — and the heat goes into the windings. Manufacturer guidance for mid-size pumps is frequently around six starts an hour.
In one record, reclaimed water pumps were logging 13 to 28 an hour. One asset logged 143 in a single hour: a start every twenty-five seconds. No alarm, because at no point was any pump outside its operating limits. It is a control deadband, and the bill arrives years later in motors, starters and contactors.
How to look for it in your own data
You do not need a platform to start. With a historian export and a spreadsheet:
- List your parallel assets. Anything with a twin: pumps, blowers, basins, trains, filters.
- Compare run hours over the same window, and compare lifetime counters if you have them. A persistent ratio far from 1:1 is the signal.
- Count starts per hour, not just runtime, and compare against the manufacturer guidance for that size of machine.
- For process variables, plot the pair on the same axis. Not each against its limit — against each other. Divergence that holds for weeks is not noise.
- Check duplicates against each other. Two instruments on one stream that disagree are telling you something regardless of which is right.
What you are looking for in every case is persistence. A transient divergence is operations. A divergence that holds for the whole record is a setting, a valve, or an instrument — and all three are fixable.
Why this is where we start
When we read a plant's history, parallel equipment is the first thing modelled, because it produces findings that are concrete, cheap to verify, and hard to argue with. Nobody has to accept a model's opinion about what dissolved oxygen should be. They only have to accept that two basins on the same air header ought to look alike.
More on what a historian export produces: what 25 days of data actually shows.