Choose signals by working backward from a decision someone needs to make. Name that person, the evidence they need, and how old it can be. Then test whether the complete measurement path delivers that evidence. This gives you a useful starting point before buying sensors or adding automation.
Suppose a cooling loop keeps attracting attention. The pump runs, temperatures move, and someone wonders whether maintenance is needed. Collecting every available data point, often called a tag, can create a busy dashboard. A smaller set, tied to a clear question, can make the next investigation easier.
Start with a decision and an owner
Write the decision as an ordinary question. For the cooling loop, it might be: should the maintenance planner arrange an investigation during the next approved maintenance window?
That is different from diagnosing a failed pump or changing a chiller setting. Each needs different evidence and authority. Decide who receives the observation, what they may approve, and where they record the result.
- Question
- Does this loop need an investigation?
- Owner
- Maintenance planner, with an operator's context.
- Output
- A review request with linked records.
- Boundary
- No equipment commands from the observation tool.
Consider the cost of being wrong in both directions. An unnecessary request consumes attention. A missed condition can delay needed work. Agree how quickly a person needs the information and which existing procedure applies if evidence is unavailable. If no one can explain what changes after seeing a signal, reconsider collecting it.
Follow the whole measurement chain
A value on a screen has travelled through several steps. Check the physical location, the instrument, its conversion into engineering units, its asset label, and the stored record. A correct reading assigned to the wrong pipe still answers the wrong question.
- Locate: identify the asset and measuring point.
- Measure: document the instrument and calibration.
- Translate: verify scaling, units, and timestamps.
- Store: preserve identity and quality in the record.
- Check: compare the displayed result with its source.
NIST's metrological traceability guidance connects a measurement result to a reference through documented calibrations and their uncertainties. It also explains that traceability alone does not establish fitness for a particular use.
Ask whether measurement uncertainty, the doubt associated with the result, is small enough for your decision. Two displayed decimal places do not establish that. An instrumentation specialist can assess installation effects and the uncertainty of a comparison. Keep the supporting records available when instruments are replaced or mappings change.
Choose a small set with enough context
Start with the few measurements that could change the decision. Then add the context needed to interpret them. For this cooling-loop investigation, a candidate set might combine supply and return temperatures, flow, pump status, and production mode.
Here, supply water travels toward the equipment being cooled. Return water comes back from it. Flow describes the volume passing a measuring point per unit of time. Record those locations explicitly so a familiar label does not hide a different measuring point.
| Signal or record | Question it helps answer |
|---|---|
| Supply and return temperature | What temperature difference is present? |
| Flow and pump status | What circulation evidence accompanies it? |
| Production mode | Are the compared periods similar? |
| Maintenance notes | Did equipment or instrumentation change? |
For every reading, retain the asset identifier, measuring point, value, unit, source time, receipt time, and quality status. Maintain instrument and calibration details in linked asset records. Keep operating mode alongside the time series so later comparisons can use it.
The DOE's 2010 Operations & Maintenance Best Practices Guide includes a sample chiller log with equipment identity, operating conditions, and water temperatures. It is useful foundational context for recording related observations, rather than treating one temperature as a diagnosis.
Review what already exists before adding instruments. A manual observation or maintenance record may resolve the immediate uncertainty. Add a signal when you can explain which competing explanation it helps test.
Match timing to the question
Sampling is how often a measurement is taken. Freshness is its age when someone uses it. A frequently refreshed screen may display an old measurement. Record source and receipt times separately, with a documented time zone and clock checks.
- Sample time
- When was the process measured?
- Receipt time
- When did the record arrive?
- Decision time
- When is someone using that record?
For a hypothetical record sampled at 10:00 and reviewed at 10:06, its age is six minutes: 10:06 minus 10:00. Arrival at 10:05 means five minutes of delivery delay: 10:05 minus 10:00. These invented times demonstrate the distinction, not an acceptable delay for your process.
Choose a sampling interval with someone who understands the process and instrument response. Slow temperature drift and brief pressure events need different measurement plans. Faster logging cannot recover detail that the instrument or an upstream average has already removed.
Define acceptable age for each intended use. Test whether timestamps remain trustworthy after reconnection. Preserve summaries as summaries, with their time windows and sample counts. Do not compare a momentary reading with a long-period average as if they describe the same interval.
Make weak evidence visible
Missing, stale, and suspicious readings need explicit states. Zero is a possible measurement. It is a poor substitute for missing data. Likewise, a repeated value might indicate a steady process, a frozen instrument, or a system replaying its last known value.
| State | Display and review response |
|---|---|
| Missing | Show unknown; identify the absent signal. |
| Stale | Show its age; withhold a current conclusion. |
| Rejected | Keep the reading and reason; request a data check. |
| Conflicting | Keep both records; request reconciliation. |
| Plausible | Eligible for review, with context still required. |
Check units, expected instrument range, duplicate records, gaps, and mismatched asset labels. Preserve any bad-quality status reported by the source. Flag disagreements rather than quietly selecting the more convenient reading. A value passing these checks is plausible, not proven correct.
Track where gaps occur. A nearly complete dataset can still omit the startup periods that matter most. Keep original readings, quality flags, and corrections distinguishable. If a chart fills gaps for readability, make those estimates visible and exclude them from decisions that require measured evidence.
Work through a cooling-loop decision
Hypothetical worked scenario
Consider a cooling loop serving production equipment. All values, identifiers, and records in this scenario are invented. They are not customer results, operating limits, or maintenance thresholds. The planner wants to decide whether to request an investigation.
| Recorded item | Snapshot A | Snapshot B |
|---|---|---|
| Production mode | Batch X | Batch X |
| Pump status | Running | Running |
| Supply temperature | 7 °C | 7 °C |
| Return temperature | 12 °C | 10 °C |
| Flow | 20 L/s | Unknown |
Each column represents readings aligned to a common source time, rather than a period average. Assume the available values meet the pilot's freshness and source-quality checks. That makes them eligible for review, not a diagnosis.
The return-minus-supply difference is 12 minus 7, or 5 °C, in A. It is 10 minus 7, or 3 °C, in B. That change alone does not establish a pump fault or quantify cooling performance. Flow in B is unknown. A matching production-mode label also does not prove equal demand.
The time-series store, a database of timestamped readings, retains both snapshots. Observation CL-01-OBS-01 links the readings, their quality flags, maintenance notes, and cooling-loop asset CL-01. The planner asks the operator about demand. An instrumentation lead investigates the missing flow record and records the finding.
The maintenance system holds the review request and its disposition as the system of record. The observation tool can prepare evidence, but cannot change settings or issue equipment commands. The planner approves an investigation through the existing workflow. If evidence remains incomplete, the request stays open with an owner and next review point. The intended business outcome is a documented next step with its uncertainty intact.
Run a pilot with a clear end
Start with observation on a defined asset and question. Agree the operating modes to cover, who reviews the evidence, and when the group will decide whether to continue. A calendar window alone cannot guarantee representative operation.
- Prepare: agree the decision, records, owner, and acceptance criteria.
- Observe: capture relevant modes and document gaps.
- Challenge: replay missing, stale, and conflicting test records offline.
- Decide: continue, improve the evidence, narrow the scope, or stop.
Define a usable review window as one with the required measurements, context, and agreed quality checks. Divide usable windows by all scheduled windows during relevant operation. Keep windows with data gaps in that denominator. Report coverage by operating mode so a strong overall result cannot hide weak startup evidence.
Also record review effort and whether each request reaches a documented disposition. Inspect misleading requests and missed known events. Set acceptance criteria before reviewing results. A pilot without enough relevant events cannot establish detection performance.
Simple software can apply timing checks and assemble records. Optional AI may summarize operator notes, provided reviewers can inspect the originals. It should preserve unknowns and have no control permissions in this pilot. People decide whether an investigation is warranted.
NIST SP 800-82 Revision 3, published as a final guide in 2023, addresses operational technology's reliability and safety constraints. For this pilot, keep existing safety functions independent of the observation tool. Have site specialists assess collection connections and any later control proposal separately.
Before choosing hardware, complete the decision brief above with an operator and a maintenance owner. Walk one existing record through the chain. That small exercise can reveal what to measure next.