Predictive Maintenance for Pumps, HVAC and Elevators: Data to Start With

Start from the failure mode, not the catalogue

Predictive maintenance (also called condition-based maintenance) relies on measurements that detect the onset of degradation so you can intervene before a component is significantly damaged. The U.S. DOE's FEMP best-practice material, hosted by Pacific Northwest National Laboratory, contrasts this with preventive maintenance driven by time or run-hours and with reactive maintenance that repairs after failure. Because a predictive program schedules work from the machine's actual condition, it can cut unplanned downtime; the same guidance cites planning-level estimates of roughly 8-12% savings over a preventive-only program, with larger gains where reactive maintenance dominates. Treat these as planning figures, not guarantees.

The common mistake is to instrument everything. A more effective approach is to pick the asset most likely to fail, identify its dominant failure mode, then choose the cheapest signal that reveals it plus a few context readings. The condition-monitoring framework in ISO 13374-1 describes the pipeline you are really building: collecting, processing and presenting data so you can assess current health, predict failure and generate recommended actions. Keep that pipeline simple at the start — it is easier to add useful data later than to clean up noise.

Pumps: vibration, temperature and differential pressure

For a rotodynamic (centrifugal) pump, the measurements that pay off first are overall vibration velocity taken on the bearing housing, bearing and motor temperature, and a hydraulic signal such as differential pressure or flow. ISO 10816-7 gives guidance for evaluating vibration on rotodynamic pumps above 1 kW measured on the bearing housing, and uses two complementary criteria: the absolute magnitude of vibration and its change over time. ISO 10816-3 supplies advised severity zones — for example new machinery, unrestricted long-term operation, restricted operation and the damage range — that facilities commonly use as starting thresholds for medium machines.

Link the signal to the failure. Rising broadband vibration often points to imbalance, misalignment or looseness; creeping temperature suggests failing bearings or motor overload; a falling head or differential pressure at constant speed and flow hints at impeller wear, internal recirculation or erosion; cavitation shows up as noise and unstable pressure and flow. Whatever you measure, log the operating point — speed, load and whether a variable-speed drive is active — because vibration and temperature only make sense at a known condition.

• Rising vibration: imbalance, misalignment, looseness
• Creeping temperature: bearing wear, motor overload
• Falling head at constant speed: impeller wear, erosion, recirculation
• Noise plus unstable pressure/flow: cavitation

HVAC: mine the building management system first

The cheapest predictive maintenance for HVAC is often data you already own. Air-handling units, fan-coils and chillers are usually controlled by a BMS that records supply, return and mixed-air temperatures, duct static pressure, valve and damper positions, and fan or pump current. Filter fouling shows as a rising differential pressure across the filter; cooling-coil fouling as a smaller temperature difference across the coil or a drifting valve position; fan degradation as rising current for the same airflow. Baseline these at stable conditions and watch the drift.

Chillers add a refrigerant-side view: evaporator and condenser approach temperatures, superheat and subcooling, and refrigerant pressures drift as the charge changes or tubes foul. Published control and fault-detection logic — for example the rules for air-handling units in ASHRAE Guideline 36 — gives a practical template for which comparisons to compute. Note that these are performance and control data, not a replacement for periodic physical checks such as belt tension or refrigerant leak detection; combine both sources.

Elevators: controller fault logs and cycle counts before extra hardware

Much of the data for elevator predictive maintenance is already digital. Modern controllers keep fault logs with timestamps and often count door operations, starts and running hours, and record motor-drive temperature and current. In many fleets, doors produce a large share of service calls, so door-cycle counts and door-related fault codes are a sensible first signal. Ride-quality vibration captured on scheduled walkthroughs is useful and does not require an always-on sensor.

Adding sensors or IoT devices to a lift is not like doing so on a pump, because lifts are safety-regulated and connected to the safety chain. Third-party monitoring must not compromise safety functions; in practice you need the manufacturer or authorized maintainer involved, plus compliance with the applicable local code and standards such as the ISO 8100 series — its part 2, published in 2026, covers verification of door locks, safety gears, overspeed governors and buffers. Start by extracting the data the controller already records before attaching anything.

Turning signals into decisions: baselines, trends and pilots

A single reading means little on its own. You need a baseline captured at defined operating conditions and enough history to separate a real drift from noise or seasonal load change. Give the pilot three to six months to build that history before trusting alarms. Use both absolute thresholds — drawn from standards or OEM advice — and relative change from baseline; ISO 10816-7, for example, evaluates vibration by magnitude and by change over time.

Scope the pilot to the most critical or most failure-prone asset in each category: one pump, one air-handling unit or chiller, one high-traffic elevator. Define in advance what a technician will do when an alarm fires, because a predictive program fails when alerts have no owner or assigned response. Document each decision so the next season's thresholds become more accurate.