Validate sensors, points, alarms and asset data before trusting AI predictive maintenance decisions.

Infographic showing a Singapore commercial facility BMS data flow from sensors and equipment to validated analytics, with checklist items for calibration, naming, alarms, trends and asset tagging.

Predictive maintenance software can identify unusual equipment behaviour, prioritise work orders and support more proactive facility operations. However, the quality of those outcomes depends on the data entering the system.

For many Singapore buildings, warehouses and SMEs, the first challenge is not selecting an AI platform. It is determining whether the existing building management system (BMS), ACMV controls, electrical monitoring and critical-equipment records are accurate enough to support reliable analysis.

BCA’s Smart Facilities Management guidance describes Smart FM as an integration of systems, processes, technologies and people. Its built-environment AI guidance also highlights facilities management, data management and workflow automation as relevant areas for AI adoption. In practical terms, this means AI readiness starts with engineering discipline and operational verification.

Use the following checklist before connecting BMS data to a predictive-maintenance or energy-optimisation solution.

1. Define the maintenance questions first

Do not begin with a generic request to “connect all points”. Start by identifying the decisions the system must support. Examples include:

Which air-handling units show abnormal temperature, pressure or runtime behaviour?
Which pumps, fans or compressors may be operating outside their normal pattern?
Which electrical assets require closer monitoring for loading, temperature or trip events?
Which alarms should create an inspection task rather than only appear on a BMS screen?
Which warehouse or critical-area conditions require timely human response?

Each question should be linked to an asset, measurable variables, an acceptable operating range and a response process. This prevents unnecessary data collection and makes later model validation more meaningful.

2. Confirm sensor selection and physical suitability

AI cannot compensate for a sensor that is poorly selected or installed in the wrong location. Review whether the installed devices measure the variables needed for the intended use case.

Check temperature and humidity sensors against the actual environmental-control objective.
Confirm differential-pressure sensors are connected to suitable measurement points and tubing.
Review flow, energy, current, power-quality or vibration measurements where they are relevant to the asset.
Identify equipment that is monitored only through a binary run or trip signal when richer data may be required.
Record sensor location, measuring range, unit of measurement and associated equipment.

For warehouses, pay particular attention to sensor placement, zone coverage and conditions that may differ between loading areas, storage aisles, plant rooms and occupied offices. A technically functional sensor may still provide misleading data if its location does not represent the area or asset being assessed.

3. Standardise point naming and asset hierarchy

Inconsistent names are a common barrier to integrating BMS, CMMS, energy and work-order data. A point labelled “Temp 1” is difficult to use reliably across buildings, floors and equipment types.

Create a naming structure that identifies the site, building or zone, floor, system, asset, point type and unit where appropriate. The exact convention can be adapted to the organisation, but it should be documented and applied consistently.

Build an asset hierarchy that connects:

Site and building;
Floor, area or zone;
System, such as ACMV, electrical or water services;
Parent equipment, such as an air-handling unit or pump set;
Component and sensor point;
Maintenance strategy, responsible party and work-order reference.

Keep a controlled point list with descriptions, units, normal state, alarm limits, source system and last verification date. This list becomes a useful bridge between engineering teams, FM operators and any future AI platform.

4. Verify calibration and data plausibility

A calibration label or commissioning record is useful, but it should not replace operational checks. Compare sensor readings with a suitable reference instrument or an independent engineering measurement where practical. Record the date, method, result, tolerance applied by the organisation and any corrective action.

Also test for data that is technically present but operationally implausible. Examples include:

A room temperature that remains unchanged for an unusually long period;
A differential-pressure value that is permanently zero or fixed at one number;
An energy meter that resets unexpectedly or shows an impossible jump;
A run-status point that disagrees with the equipment’s actual condition;
A humidity or outdoor-air value that does not respond to known operating changes.

Outliers should be investigated rather than automatically deleted. They may indicate a failed sensor, incorrect scaling, a communication issue, a changed setpoint or a real equipment event.

5. Test connectivity, timestamps and data continuity

AI models need dependable time-series data. Confirm how information moves from field devices to controllers, the BMS, historian, cloud platform or maintenance system. Document the relevant interfaces and data ownership.

Check that:

Point values have consistent units and timestamps;
Time zones and daylight-saving assumptions do not create confusion when data is exported;
Communication failures are visible rather than silently recorded as normal values;
Trend logs have an interval appropriate to the equipment and use case;
Historical data can be retained, exported and matched to the correct asset;
System changes do not overwrite the audit trail without an explanation.

Do not assume that a point visible on a BMS graphic is automatically available at the required resolution or quality for analytics. Test an actual export or integration sample before approving a wider deployment.

6. Rationalise alarms and map them to action

A large alarm count does not necessarily indicate good control. Repeated nuisance alarms can cause operators to ignore important events, while poorly configured limits may hide developing faults.

Review each important alarm for its trigger, delay, priority, deadband, escalation path and expected response. Separate alarms that require immediate attention from advisories, maintenance reminders and information-only events.

For predictive maintenance, define how an emerging anomaly should be handled. The response may be an inspection, a trend review, a planned work order or a safety escalation. The appropriate action depends on the asset and risk; it should not be left entirely to an algorithm.

7. Tag critical assets and maintenance history

Sensor data becomes more useful when it can be linked to equipment identity and maintenance history. Confirm that critical assets have stable identifiers and that those identifiers remain consistent across BMS graphics, asset registers, service reports and work-order records.

Where available, associate operating data with installation details, equipment capacity, service dates, fault history, component changes and known operating restrictions. Avoid copying incomplete records into a new platform without marking uncertainty. A clearly identified data gap is more useful than an apparently complete but unreliable record.

8. Include human verification before production use

Facility engineers, technicians and operators should be involved in validating the data and the recommendations. They understand equipment behaviour, temporary operating arrangements and site conditions that may not appear in the database.

Run a controlled verification period. Select representative assets, compare analytics with site observations, review false positives and false negatives, and document what operators consider actionable. Include abnormal operating periods such as shutdowns, fit-out works, seasonal changes or occupancy variations where relevant.

Human verification should continue after deployment. Asset modifications, sensor replacement, controller changes and revised operating schedules can affect data quality. Assign ownership for point-list updates, calibration records, alarm reviews and model feedback.

A practical go/no-go review

Before investing in a larger AI predictive-maintenance deployment, ask:

Can each priority data point be linked to a known asset and location?
Are units, timestamps, names and status values consistent?
Have critical sensors and meters been physically checked?
Can missing, frozen or implausible values be detected?
Are alarms connected to clear operational actions?
Can BMS data be matched with maintenance and asset records?
Have FM personnel reviewed the data and tested the proposed workflow?

If several answers are “no”, the next step may be BMS rectification, sensor commissioning, data cleanup or workflow design rather than immediate AI deployment.

Conclusion

Singapore’s move towards integrated, data-driven FM creates opportunities for predictive maintenance, energy management and automated workflows. But reliable outcomes begin with trustworthy field data and disciplined commissioning.

A well-structured sensor and BMS readiness review helps facility managers, warehouse operators, building owners and SMEs understand what is usable today, what requires rectification and where AI can add practical value. Contact ISS to discuss engineering, facility management or AI automation requirements.

Singapore FM Sensor Data Quality: An AI-Ready BMS Commissioning Checklist