How to verify liquid-cooling sensors and synchronized alarms

What proves that liquid-cooling telemetry is trustworthy?

Trace the sensing element, wiring, transmitter, controller, gateway, building management system (BMS), data center infrastructure management (DCIM) and historian as one measurement chain.

Calibration establishes the relationship to a reference; adjustment changes the instrument; verification checks the result against requirements. Retain as-found and as-left results. NIST's traceability policy, using JCGM 200:2012 §2.41, treats traceability as a property of a measurement result supported by a documented calibration chain and uncertainty [S2].

The owner and design authority approve limits; the OEM defines behavior; the calibration provider supplies uncertainty; the integrator proves mapping; the witness records disposition. Anvilfield's 2026 acceptance worksheet, called P1 below, supplies test selections for project adoption; its thresholds are not universal OCP or OEM requirements. Use CDU commissioning for the equipment boundary and loop commissioning for hydraulic readiness.

Which points support each cooling claim?

P1 verifies 100% of points supporting acceptance claims, including required alarms; the 12-point estimating package does not cap critical coverage. Record serial number, channel, register/object identity and mapping revision beside each tag.

OCP's liquid-to-liquid CDU methodology, Revision 1.0, distinguishes the coolant distribution unit (CDU), technology cooling system (TCS) and facility water system (FWS) boundaries [S1]. For liquid-to-air units, document the heat-rejection interface instead of an FWS pair. Separate probes, averages and calculations.

Table A uses worksheet IDs. Record each point's location, identity, timestamp origin, update rate and quality method. Calculated dew point also needs valid ambient-temperature and humidity inputs and the installed control mode.

What does a cross-vendor point check actually compare?

Read the vendor map before decoding. MODBUS Application Protocol V1.1b3 (2012), §§4.2 to 4.4, defines 16-bit registers and zero-based protocol addresses; the OEM supplies meaning, scaling and multi-register layout [S14]. Verify function code and address convention. Modbus sends the most significant byte first within a register; multi-register word order still needs the device map.

Vertiv CDU 121 Rev A Appendix G labels primary temperature register 7 as signed with a 0.1 °C/°F scale; with Celsius selected, raw 402 represents 40.2 °C [S6]. Retain raw, local and export values. Test low, middle, high and permitted negative inputs.

Use the DCIM integration guide for asset naming.

How accurate must CDU sensors be?

P1 selects ±0.50 °C for the installed temperature chain; OCP Revision 1.0 does not mandate this tolerance. Assign each purchased tolerance before accepting results.

Tolerance is allowed error, resolution is the reporting increment, repeatability is scatter, and control stability is process variation around a target. A stable display can be biased. Table B separates these specifications.

What changes with the sensor, fluid and signal path?

Temperature comparison requires representative mixing, adequate immersion, known wiring and settled readings. WIKA's July 2026 instructions exclude class A and AA with two-wire connections [S4].

Promag 50L TI00097D/06/EN/18.16 requires at least 5 μS/cm conductivity for general fluids and 20 μS/cm for demineralized water; remote cable length also affects applicability [S5]. Verify coolant suitability; include current-output conversion error when used.

The CDU 121 manual describes its secondary meter as calorimetric [S6]. Use that meter's coolant configuration and specification, not the Promag formula. Pressure subtraction and calculated dew point likewise need budgets for their own inputs.

How do you verify the whole measurement chain?

P1 selects 30, 40 and 50 °C plateaus for a 30 to 50 °C acceptance band, with ten paired reference/channel observations after settling at each level. Space readings to capture independent information; ten rapidly repeated readings do not automatically provide ten independent observations.

Use a stable comparison bath or suitable dry block for removable probes and an installed comparison for placement effects. Verify that plateau excursions fit the thermal-comparison uncertainty allocation before collecting readings. Record corrections, immersion, wiring and lag; define the reference correction as the quantity added to its indication.

WIKA's July 2026 instructions, §8.1, require depressurization and electrical isolation before probe removal [S4]. Vertiv CDU 121 Rev A, §§1.1 and 1.5, requires electrical isolation before internal connections and stopped rotating parts for maintenance [S6]. Connect simulators with supplies isolated, then energize the approved test arrangement.

For pressure, record gauge versus absolute reference, port elevation, zero, ascending/descending points and overrange protection. Test both inputs when differential pressure is calculated. For flow, record fluid/concentration, direction, full-pipe condition, grounding, installation geometry and usable low-flow range. Compare against a calibrated meter in the same flow path, accounting for bypasses and reference installation effects. Pump command percentage is not a flow reference.

Record the injection boundary. Physical temperature or pressure exercises more of the chain than an electrical simulator; a forced BMS value proves only the downstream path. Coordinate these boundaries with BMS controls commissioning.

What should a 4 to 20 mA scaling check contain?

For the selected 0 to 100 °C range, subtract 4 mA, divide by the 16 mA span, then multiply by 100 °C. P1 tests both directions, comparing raw input, controller, gateway and historian.

How do uncertainty and guard bands change acceptance?

NIST TN 1297 (1994), §§3 to 6, combines independent standard uncertainty contributions by taking the square root of their squared sum, then applies a coverage factor for expanded uncertainty [S3]. Include covariance for correlated components; averaging does not remove shared bias.

Table D evaluates the mean indication error of the device under test (DUT) at a settled plateau using ten paired observations. Use certificate values and supported comparison estimates; retain scatter and response time separately.

Anvilfield estimates 0.13 °C expanded uncertainty from these allocations using NIST TN 1297. Complete-channel comparison includes sensor/transmitter behavior; do not add catalog limits again. Retain resolution unless independent dithering supports reduction; avoid duplicating quantization in repeatability.

Method: the unrounded combined standard uncertainty is 0.0653197 °C and expanded uncertainty is 0.1306395 °C. Use unrounded values for boundary decisions. This budget assumes negligible covariance and sufficient effective degrees of freedom; k = 2 represents approximately 95% coverage only with a suitable distribution and degrees-of-freedom model.

P1 requires absolute error plus expanded uncertainty to fit within tolerance for a pass. JCGM 106:2012 supports uncertainty-aware conformity decisions; this particular three-way rule is Anvilfield's selection [S11]. Anvilfield estimates 0.10 K transient pairing error: selected 0.50 K/s temperature change multiplied by 0.20 s channel separation.

How do you test timestamps, clock drift and alarm latency?

P1 selects a maximum absolute clock error of 100 ms per timestamping system, including comparison uncertainty; record the actual bound achieved for each system. Offset is error at an instant; drift is its change over elapsed time. Two timestamping systems each within 100 ms of a common reference can disagree by 200 ms, so those clock bounds alone cannot establish the order of events only 150 ms apart.

Verify offset directly against the reference where supported. A common electrical edge includes clock error plus input scan/detection delay, so characterize that contribution before calling an edge difference clock offset. Use isolated test inputs, not operating protection circuits, for the common-edge exercise. Retain reference uncertainty, synchronization state, selected source and delay/dispersion. A Network Time Protocol (NTP) connection or low stratum does not establish accuracy.

Separate sensor scan, controller evaluation, persistence, transport, historian write, notification and physical action. Faster logger polling does not increase the device's native update rate. Capture both UTC and monotonic elapsed time through normal traffic, time-source loss and restart; measure physical detector response separately from electrical injection.

Which timestamp belongs to the event?

OPC UA Part 4 v1.05.07 distinguishes a value's source timestamp from each server's receipt or known-valid timestamp, so an unchanged source timestamp alone does not prove stale data [S8] [S9]. Preserve originating UTC time through forwarding; label gateway-generated time when the device supplies none. Confirm freshness through acquisition semantics, heartbeat/sequence, status and server knowledge.

OPC UA Part 5 v1.05.06 §6.4.2 defines event Time as occurrence time, ReceiveTime as server receipt and EventId as identity [S15]. These event fields differ from DataValue timestamps. Retain source, receipt, historian ingestion and notification times separately, including during backfill and local-time display changes.

On one reference clock, the worked inputs are stimulus at 0.00 s, assertion at 120.40 s, remote record at 121.10 s and action feedback at 122.00 s. Remote transport is 0.70 s; stimulus-to-action is 122.00 s and includes persistence. Under P1, a measured 4.90 s transport delay plus a 0.20 s clock/comparison bound reaches 5.10 s and remains on hold against the selected 5 s limit.

RFC 5905 (June 2010), §8, calculates offset and round-trip delay from four exchange timestamps [S12]. T1 is client send, T2 server receive, T3 server send and T4 client receive. For T1 = 0, T2 = 15, T3 = 17 and T4 = 22 ms, the server-minus-client offset is: ((15 - 0) + (17 - 22))/2 = +5 ms. Subtract server processing time from total exchange time for delay: (22 - 0) - (17 - 15) = 20 ms. Network asymmetry still limits the offset inference.

Anvilfield estimates +110 ms final offset from a selected +20 ms initial offset and constant +50 parts per million (ppm) oscillator error over 1,800 seconds: the added error is 90 ms. To calculate drift from test results, divide the offset change by elapsed time in matching units, then multiply by one million. This exceeds P1's bound; use measured clock behavior on site.

Which alarm stimuli prove the cooling sequence?

Test the trigger, persistence, reset, latching, acknowledgement, identity, delivery and required physical action separately. P1 selects three complete cycles per scenario, retaining every cycle and the worst response. Record mode, priority, recipient, reset role and cause-and-effect revision.

Approach analog thresholds from both directions with calibrated stimuli clearly on either side after allowing for uncertainty, resolution and deadband. Apply a pulse shorter than persistence and a sustained pulse; check whether the timer clears or retains state as specified. A large step proves assertion but does not locate the threshold.

Use an isolated test arrangement or the approved operating procedure for intrusive challenges, documenting overrides and restoration. Acknowledgement records operator handling; it does not prove return to normal or restored protection.

What if the manual and controller settings disagree?

Vertiv's CoolChip CDU 121 Rev A manual lists a 10-second P906 temperature-alarm delay but describes two-minute delays for A24/A25, so acceptance must resolve the discrepancy against the installed configuration and witnessed behavior [S6]. Obtain OEM resolution for the model, firmware and mode before passing the contested sequence.

The same manual lists P702 Secondary T2 Period as 10 seconds while its disagreement narrative describes 30-second persistence [S6]. Record both: an acquisition period and an alarm timer are different quantities. Verify the timer independently and obtain the parameter definition.

P1's remote-delivery target measures local assertion to the specified remote event record; it excludes configured persistence and does not establish physical protection time. Record stimulus crossing, assertion, receipt, notification, commanded action and feedback to assess each purchased interval.

For dew-point mode, challenge ambient temperature, humidity and their faults, then verify setpoint override. CDU 121 Rev A specifies a 3 °C dew-point offset, but describes low-temperature assertion at or below dew point for three minutes and a 20 °C high-temperature default when override is active [S6]. Obtain the applicable firmware sequence from Vertiv and witness it before accepting this mode; use Table F's fixed-mode thresholds only in fixed mode.

Use the leak-detection guide for detector coverage. CDU 121 Rev A permits alarm-only or alarm-plus-shutdown leak configuration, so accept sensing, routing and action independently [S6]. Pump telemetry belongs here; redundancy and concurrent-maintenance testing proves hydraulic survival and spare capacity.

When must missing or stale data stop acceptance?

Under Anvilfield's 2026 missing-data rule, losing any required stimulus, alarm, action or reset transition invalidates that test scenario even when the overall trend meets its 99% completeness target. Repeat the full scenario; interpolation cannot recover the missing evidence.

A brief alarm can assert and clear between polls. Prove an event log, latch or purchased delivery mechanism that retains it, and verify historian deadband/compression preserves every required transition. Use an uncompressed acceptance export when the configured trend cannot do that.

P1's steady-trend denominator is 600 scheduled observations per required point in a half-open ten-minute window, at seconds 0 through 599. Count one valid observation per slot; duplicates and unverified retained values cannot replace gaps. Apply completeness, freshness and critical-event rules together.

OPC UA Part 4 v1.05.07 §7.11.5 requires clients to inspect status before using a value; Bad is unusable [S10]. P1 also excludes Uncertain values unless a pre-agreed treatment supports the claim. Good status does not prove physical correctness or calibration, and the specification permits Good severity when a server does not support status information.

Document what the implementation actually knows. A gateway heartbeat proves gateway availability; sensor freshness requires evidence from the originating acquisition chain. Preserve last-good value and acquisition time separately from current stale/invalid status, including native null values where appropriate.

Is the fault in the sensor, mapping, clock or cooling process?

Compare each point with an independent reference, locally and remotely. Local-to-reference error implicates measurement/installation; local-to-remote error implicates mapping/conversion/acquisition. Match load, fluid and flow before adjusting hardware.

Do not adjust probes merely to make them agree. During load changes, align sample times and account for response lag before interpreting a differential as cooling failure.

Worked example: accept, hold or reject a temperature point

Under Anvilfield's 2026 telemetry acceptance rule, a settled temperature channel with 0.20 °C mean indication error and 0.13 °C expanded uncertainty passes a ±0.50 °C tolerance because its guarded error is 0.33 °C. The selected reference input is 39.96 °C plus a +0.04 °C correction, giving 40.00 °C.

Calculate error by subtracting corrected reference from export. P1 passes when |e| + U ≤ L, fails when |e| - U > L, and otherwise holds, where e is error, U expanded uncertainty and L allowed absolute error. Agree this rule before testing; indicated error alone gives different near-limit decisions.

Pass closes this tested condition; hold requires improved measurement or adjustment and retesting under the agreed rule; fail requires correction and retesting. Record any authorized specification change as a separate deviation, preserving the original result. Repeat across the operating band.

How long does testing take, and what should it cost?

Anvilfield estimates two eight-hour field shifts with two technicians, eight analysis hours and $6,700 for P1's 12 analog points, ten alarm/fault scenario groups and one telemetry path, using November 2026 planning inputs. That package contains 36 plateau comparisons, 360 paired observations and at least 30 alarm cycles. Expand groups into repetitions for each probe or pump.

Allocate the 16 elapsed field hours to three hours of readiness/setup, six of analog comparisons, five of alarm/fault exercises and two of restoration/export. Schedule the 30-minute holdover and three startup-inhibit cycles explicitly; the published startup default consumes at least 60 minutes across those cycles before assertion/reset time. Allow longer for settling and unresolved OEM behavior.

The estimate assumes test ports, simultaneous logging and an isolated arrangement. Add piping work, load banks, travel, shutdowns, controls support, witnesses and retests as needed. Estimators developing this scope use PorchRocket Bid Control to find public HVAC and refrigeration bids with controls-testing requirements.

Method: Anvilfield estimates cost from selected hours multiplied by loaded planning rates, plus equipment and calibration allocations. An extra crew shift adds $2,000 before equipment/reporting. Access, point count, settling and alarm timers determine hours.

Field checklist and acceptance record

Failed readiness holds the affected test. Screenshots supplement machine-readable evidence.

1. Match model, serial, firmware, mode and cause-and-effect revision; resolve manual/configuration conflicts and startup inhibits.
2. List every claim, physical point, unit, location and export field; close mapping omissions.
3. Agree tolerances, uncertainty rules, clock bounds, alarm timing and quality limits.
4. Attach reference certificates, corrections, uncertainty and coverage factors; verify intermediate checks.
5. Confirm probe placement, wiring, fluid, flowmeter setup and pressure ports.
6. Record overrides, isolation and restoration in the operating procedure.
7. Complete bidirectional signal checks and operating-band comparisons; retain as-found/as-left data.
8. Confirm update rates, protocol maps, timestamp origins and clock state; distinguish value/event timestamps.
9. Run common-transition and time-source-loss tests; identify steps and repeat invalidated timing claims.
10. Complete P1 alarm cycles, including startup, brief events and timer reset; distinguish trip, delay, reset, acknowledgement and latching.
11. Prove required physical actions with feedback, remote identity, routing and notification.
12. Test communications loss/recovery; retain stale state and original replay times.
13. Apply uncertainty/missing-data rules; assign pass, hold or fail with corrective action.
14. Remove overrides, restore protection/time sources, verify normal operation and deliver raw files and signed disposition.

What fields must the exported evidence contain?

Join point profiles, paired observations and event records by test ID. The point profile identifies claim, asset serial, firmware, mode, physical location, protocol/object/register, address convention, data type, scale, units and mapping revision.

Each observation retains raw export, reference indication, certificate correction, corrected reference, error, plateau ID, sample count, combined standard uncertainty, coverage factor, expanded uncertainty, tolerance, decision rule and outcome. Record native update, scan and poll intervals plus historian deadband/compression.

Each event retains origin and event ID, UTC occurrence, server receipt, ingestion and notification times, monotonic elapsed time, stimulus boundary, threshold crossing, assertion, command, action feedback, reset and acknowledgement. Include clock identity, offset, uncertainty, drift, synchronization state, quality, last-good acquisition, sequence/heartbeat, replay flag, priority, recipient, inhibits and overrides.

Attach certificates, configuration before/after, stimulus scripts, uncertainty calculations, excluded-row reasons, all repeats and signed dispositions. Preserve raw files separately from normalized analysis, with decision authority and corrective action for every open claim.

Use proof-packet assembly for turnover and integrated systems testing for larger failure scenarios.

How often should calibration and alarm verification be repeated?

Anvilfield's 2026 worksheet selects an initial 12-month calibration review and six-month intermediate comparison for acceptance reference instruments, with checks before and after each campaign. NIST's interval guidance, updated May 29, 2026, assigns intervals from accuracy needs, stability, environment and contractual requirements rather than prescribing annual calibration [S13]. ILAC-G24/OIML D 10:2022 supports risk-based selection and revision using history and intermediate checks [S13a].

Track permanent CDU sensors separately: compare at acceptance, then set intervals from OEM requirements, drift and criticality. Apply stricter contractual intervals. Shorten when drift consumes the uncertainty allowance; extend only with supporting history.

P1 selects a 12-month alarm-verification review and immediate affected-chain verification after sensor replacement, firmware, mapping, clock-source or sequence changes, or a failed intermediate check. Review earlier accepted results when as-found drift exceeds the budget. Keep calibration, alarm exercises and clock-health monitoring distinct.

Sources, review date and citation

Sources checked and page reviewed 2026-11-05. Update log: 2026-11-05: Initial worksheet; source and numerical recheck; clarified isolation, alarm modes, clock drift and acceptance decisions; repaired source links.

• [S1] OCP, Liquid to Liquid CDU Test Methodology and Performance Rating. Revision 1.0; §§3.2.4 and 3.3.2. Boundaries and approach.
• [S2] NIST, Metrological Traceability. Public policy; JCGM 200:2012 §2.41. Result traceability.
• [S3] NIST TN 1297. 1994 edition, §§3 to 6. Standard and expanded uncertainty.
• [S4] WIKA, TR10-L/TR10-0 instructions. 3370964.13, 07/2026, §8.1 and pp.29 to 30. Removal, element classes and wiring.
• [S5] Endress+Hauser, Promag 50L Technical Information. TI00097D/06/EN/18.16, pp.12 and 18. Error and conductivity.
• [S6] Vertiv, CoolChip CDU 121 Operation and Maintenance Guide. SL-80278 Rev A, 10/2025, §§1.1, 1.5, 4.3.3, Tables 4.27, 4.29, 4.36 and Appendix G.
• [S7] Fluke/Hart, 1502A User’s Guide. Rev 973101 ENG, §3.1. Readout specifications using a 100 Ω PRT; probe uncertainty excluded.
• [S8] OPC Foundation, SourceTimestamp. Part 4 v1.05.07, §7.11.3. SourceTimestamp.
• [S9] OPC Foundation, ServerTimestamp. Part 4 v1.05.07, §7.11.4. ServerTimestamp.
• [S10] OPC Foundation, StatusCode. Part 4 v1.05.07, §7.11.5. Value status.
• [S11] BIPM/JCGM, Measurement uncertainty in conformity assessment. JCGM 106:2012, §§8 to 9. Conformity decisions.
• [S12] IETF, RFC 5905. June 2010, §8. Four-timestamp calculation.
• [S13] NIST, Recommended Calibration Interval. Updated 2026-05-29. Calibration interval selection.
• [S13a] ILAC/OIML, Recalibration intervals. ILAC-G24/OIML D 10:2022. Recalibration and intermediate checks.
• [S14] Modbus Organization, Application Protocol. V1.1b3, 2012, §§4.2 to 4.4. Register encoding/addressing.
• [S15] OPC Foundation, BaseEventType. Part 5 v1.05.06, §6.4.2. Event identity and times.

Cite the point matrix, uncertainty budget, timestamp test, alarm stimuli or missing-data rules directly by section link.

How to cite this page: Anvilfield, How to verify liquid-cooling sensors and synchronized alarms, 2026, https://anvilfield.com/field-guides/datacenter/liquid-cooling-telemetry-calibration-acceptance/.

Related tools

FAQ

People also ask

• What is a CDU in a data center?
• What is direct-to-chip liquid cooling?
• What does DCIM stand for and what does it manage?
• What is BMS commissioning?
• Why do data centers need liquid-cooling leak detection?

Sources checked

• OCP, Liquid to Liquid CDU Test Methodology and Performance Rating Revision 1.0; §§3.2.4 and 3.3.2. Boundaries and approach.
• NIST, Metrological Traceability Public policy; JCGM 200:2012 §2.41. Result traceability.
• NIST TN 1297 1994 edition, §§3 to 6. Standard and expanded uncertainty.
• WIKA, TR10-L/TR10-0 instructions 3370964.13, 07/2026, §8.1 and pp.29 to 30. Removal, element classes and wiring.
• Endress+Hauser, Promag 50L Technical Information TI00097D/06/EN/18.16, pp.12 and 18. Error and conductivity.
• Vertiv, CoolChip CDU 121 Operation and Maintenance Guide SL-80278 Rev A, 10/2025, §§1.1, 1.5, 4.3.3, Tables 4.27, 4.29, 4.36 and Appendix G.
• Fluke/Hart, 1502A User’s Guide Rev 973101 ENG, §3.1. Readout specifications using a 100 Ω PRT; probe uncertainty excluded.
• OPC Foundation, SourceTimestamp Part 4 v1.05.07, §7.11.3. SourceTimestamp.
• OPC Foundation, ServerTimestamp Part 4 v1.05.07, §7.11.4. ServerTimestamp.
• OPC Foundation, StatusCode Part 4 v1.05.07, §7.11.5. Value status.
• BIPM/JCGM, Measurement uncertainty in conformity assessment JCGM 106:2012, §§8 to 9. Conformity decisions.
• IETF, RFC 5905 June 2010, §8. Four-timestamp calculation.
• NIST, Recommended Calibration Interval Updated 2026-05-29. Calibration interval selection.
• ILAC/OIML, Recalibration intervals ILAC-G24/OIML D 10:2022. Recalibration and intermediate checks.
• Modbus Organization, Application Protocol V1.1b3, 2012, §§4.2 to 4.4. Register encoding/addressing.
• OPC Foundation, BaseEventType Part 5 v1.05.06, §6.4.2. Event identity and times.