Telemetry / Data integrity
Data Quality for AI-Driven Data Center Operations
An AI system can make a plausible recommendation from a bad sensor reading. Reliable operations start with knowing whether each signal is fresh, calibrated, comparable and fit for its intended decision.
Sections in this article
Before allowing a sensor value to influence an AI recommendation, verify its timestamp, unit, plausible range, calibration status and relationship to nearby measurements. Missing or uncertain data must be visible to the model and to operators.
A rack inlet sensor reporting 22°C can be useful data. It can also be a stale value last recorded before a workload increase. Those possibilities look identical unless the telemetry system checks freshness. AI adds value only if operators can trust—or at least qualify—the inputs behind its predictions.
Data quality for physical infrastructure is not a generic spreadsheet-cleaning task. It combines instrumentation, transport protocols, timestamps, engineering context and the consequences of a wrong decision. This article concentrates on the data layer, rather than the design of cooling safety interlocks.
Five ways telemetry becomes misleading
| Failure | What an operator sees | Check |
|---|---|---|
| Frozen or stale point | Temperature remains flat during changing IT demand. | Age-of-data check; heartbeat; independent comparison. |
| Calibration bias | A sensor is consistently warmer or cooler than a calibrated reference. | Scheduled comparison and drift trend. |
| Transient spike | One unrealistic jump followed by an immediate return. | Rate-of-change validation and nearby instrument checks. |
| Timing mismatch | Power and cooling signals describe different time intervals. | Synchronized timestamps and known sampling windows. |
| Unit or mapping error | Percent interpreted as a fraction, or one rack mapped to another. | Schema, unit and device-identity validation. |
No automated rule can simply discard every suspicious spike. A true rapid temperature increase or electrical event may look anomalous precisely because it is dangerous. The validation layer should preserve the raw event, flag uncertainty and use corroboration before suppressing or acting on it.
A sensor-quality gate before model inference
- 01Verify identity and unit
Confirm equipment ID, engineering units, expected operating range and source configuration.
- 02Measure freshness
Compare acquisition time with the decision interval. A lag acceptable for monthly analytics may be unsafe for active control.
- 03Check physical plausibility
Look for impossible values or implausible rates of change, without confusing extreme but real events with bad data.
- 04Compare independent evidence
Correlate related rack, return-air, flow and power signals where physically meaningful.
- 05Attach a quality state
Mark the value valid, suspect, stale or unavailable, and propagate that status to AI outputs and operator displays.
Example: two sensors disagree after a rack move
A newly installed 30 kW rack reports inlet air at 21°C, while adjacent instruments read 26°C and a load meter confirms higher demand. The difference does not by itself prove that the first sensor is defective: airflow, sensor placement and containment geometry can create real local variation.
The engineer verifies sensor placement, checks the point mapping and timestamp, compares the location with a calibrated handheld instrument and examines supply-air readings. Only after diagnosis should the team decide whether to recalibrate, relocate or replace the instrument.
Physical consistency checks may help. For a fluid stream, sensible thermal power can be estimated as Q = ṁ × cₚ × ΔT, where ṁ is mass flow rate in kg/s, cₚ is specific heat in kJ/(kg·K), and ΔT is temperature difference in K; Q is then in kW. The expression requires compatible units and an appropriately defined measurement boundary. It cannot, on its own, pinpoint a faulty rack sensor or prove the full facility energy balance.
What to do when data is missing
There is no universal threshold—for example, “15% of sensors missing”—at which every installation should abandon automation. The correct response depends on which measurements are unavailable, the model's role and whether independent protective controls remain active.
Use qualified estimates for analysis, while displaying uncertainty.
Require operator review when critical inputs cannot be validated.
Disable affected AI commands and follow site-tested control procedures.
Virtual sensors or interpolation may help an analytical model remain useful, but an imputed reading must not silently replace a safety-critical measurement. Engineering staff should decide whether any estimate is acceptable for that decision and document how its quality is shown.
Make instrument quality measurable over time
A useful data-quality dashboard distinguishes model performance from telemetry health. Track the percentage of stale samples by sensor group, calibration overdue status, time synchronization issues, missing intervals, disagreement with reference instruments and incidents in which the model used suspect data.
- Assign an owner and service history to each critical sensor or gateway.
- Maintain calibration and inspection intervals according to manufacturer guidance and operational risk.
- Record rack relocations, containment changes and device remapping as configuration changes.
- Retain raw events and quality flags for investigation under a defined retention policy.
- Revalidate relevant model assumptions after instrument or layout changes.
The best AI architecture cannot compensate for an unmaintained instrumentation estate. Good telemetry governance is practical maintenance work: know the point, trust the timestamp, compare reality and preserve the option to stop acting on uncertain information.
Sources and further reading
Primary frameworks and reference materials. Engineering decisions require site-specific validation; legal classifications require assessment of the actual use case.
- Guide to Operational Technology SecurityNIST
- AI Risk Management Framework CoreNIST AI Resource Center
- Thermal Guidelines, datacom facilitiesASHRAE