Predictive maintenance for HVAC uses IoT sensors and machine learning to watch each fan, pump and chiller and predict failure before it happens, instead of servicing on a fixed calendar. A typical node reads 3-axis vibration to 5,000 Hz, bearing temperature to 150 °C and motor current, and a trained model flags drift 1 to 4 weeks early; sites report 30 to 50 percent less wasted maintenance labour and 20 to 40 percent fewer unplanned stops after rollout.
From Calendar to Condition
Preventive maintenance changes a bearing every 8,000 hours whether it needs it or not; predictive maintenance changes it when vibration, temperature and current say the film is gone. That shift moves work from a guessed date to a data-proven one, raising the share of repairs done before failure from roughly 50 percent to 80 to 90 percent. On a chilled-water plant with 30 machines, that typically frees one technician-day per week and cuts spare-parts holding 20 to 30 percent.
The Sensor Layer
The data starts at the edge. A vibration node samples at 10 to 25 kHz to capture bearing tones; a temperature probe on the bearing housing tracks the 60 to 90 °C band; a current transducer sees the 5 to 10 percent amp rise of a failing compressor. The table shows the core set per machine class.
| Machine | Sensors | Key range | Sample rate |
|---|---|---|---|
| Fan | Vibration, temp | 0 to 5,000 Hz | 10 kHz |
| Pump | Vib, temp, current | 0 to 3,000 Hz | 8 kHz |
| Chiller | Temp, pressure, current | 0 to 20 bar | 1 Hz |
| AHU coil | Temp, humidity | 0 to 100 percent RH | 0.1 Hz |
How Machine Learning Adds the Warning
Raw data is not a prediction. The platform collects 3 to 12 months of normal operation to build a baseline signature per asset, then applies anomaly detection: a 2x vibration rise, a 10 °C bearing climb, or a falling chilled-water delta-T each score against the model. After enough labelled events, a supervised model reaches 80 to 90 percent precision on bearing and motor faults and gives a remaining-useful-life estimate in days. The operator gets a ranked work list instead of a flood of raw charts.
Deployment Steps
Start with the top 10 to 20 critical machines, not the whole site. Install wireless or wired nodes, stream to a dashboard, and let the model learn the baseline for one season so it sees summer and winter load. Set alert thresholds: warning at a 2x rise over baseline, action at the 4.5 mm/s ISO 10816 zone C limit. Close the loop by feeding the work order result back so the model improves, typically reaching useful accuracy after 3 to 6 confirmed events.
When to Adopt Predictive vs Stay Preventive
- Adopt predictive on critical fans and chillers whose stoppage costs more than the sensor kit.
- Adopt it when you already have a BMS and can feed its points into the ML platform.
- Stay on calendar preventive for cheap, redundant, easily swapped small motors.
- Use route-based handheld checks below 10 machines where wireless nodes cost more than the risk.
Common Pitfalls
The biggest failure is poor data: a sensor on a painted surface reads the wrong temperature, or a node sampling at 1 Hz misses a 2,000 Hz bearing tone. Match sample rate to fault frequency and mount on clean metal. The second pitfall is alert fatigue, 50 false alarms a week that operators ignore; tune thresholds so only 1 to 5 real warnings reach the screen. Done right, predictive maintenance typically pays back in 1 to 3 years on critical HVAC plant.
Keruite manufactures this equipment at an 80,000 m² facility in Dezhou, China, with an in-house AMCA-210-style test laboratory; every unit is balanced to G2.5 and performance-tested before shipment, and baseline vibration and temperature data ship with each fan to seed your predictive model.
Data Architecture That Scales
The platform matters as much as the sensor: edge nodes should sample at the right rate, buffer through a Wi-Fi or LoRa gateway, and push to a dashboard that stores at least a year of 1-minute history for trend proof. A fan needs 10 kHz vibration capture, not 1 Hz, or the bearing tone is missed; a chiller needs 1 Hz pressure and current, not 10 kHz, or you drown in noise. Match the sample rate to the fault frequency, tag every point with asset and location, and feed the BMS points in so you are not installing two parallel sensor networks on the same machine.
Closing the Loop With Work Orders
A warning that creates no work order is just a number, so link the ML alert to the CMMS so a 2x vibration flag opens a task with the recommended action attached. When the technician closes it, feed the result, fixed or false, back to the model so precision climbs from 60 percent to the 80 to 90 percent range after 3 to 6 confirmed events. Sites that skip this loop report 50 false alarms a week and stop trusting the system, which wastes the 1 to 3 year payback the sensors were bought to deliver.
Case Study: A 30-Machine Chilled-Water Plant
A chilled-water plant with 30 machines fitted wireless vibration and current nodes and trained the model on one season of data. Within four months it flagged a 22 kW pump with a rising 2x peak and a 10 °C bearing climb, warning nine days before the seal failed. The planned swap cost a USD 200 bearing and two hours; the unplanned version modelled at a day of lost cooling and a USD 3,000 rewind. Across the year the programme cut unplanned stops 35 percent and freed one technician-day a week.
Calculating the Payback Before You Buy
Justify the rollout with a simple sum: take the cost of one unplanned stop on a critical machine, say a day of lost cooling at USD 2,000 to 6,000, multiply by the stops per year the programme is expected to cut, typically 30 to 40 percent, and compare that to the sensor and platform cost of USD 80 to 400 per node. On a 30-machine plant the avoided failures usually clear the capital in 1 to 3 years, and the spare-parts and labour saving on top is upside. Do this maths per machine class so you fit predictive where the stop is costly and stay preventive where a cheap spare already covers the risk, which is how the budget stretches to the assets that matter.
FAQ
What sensors are needed for HVAC predictive maintenance?
Fit vibration, temperature, current and humidity sensors; a typical fan node reads 3-axis vibration at 0 to 5,000 Hz, bearing temperature to 150 °C, and motor amps. A basic IoT kit per machine costs USD 80 to 400, and a site of 50 machines needs 150 to 300 tagged points.
How does machine learning predict failures?
ML trains on 3 to 12 months of normal data to learn each machine's signature, then flags drift such as a 2x vibration rise or a 10 °C temperature climb as anomaly. A model typically gives 1 to 4 weeks warning of a bearing failure with 80 to 90 percent precision after enough labelled events.
Is predictive maintenance better than preventive?
Preventive work on a fixed calendar wastes 30 to 50 percent of tasks because healthy parts are serviced early. Predictive does the task only when data shows degradation, raising wrench time value and cutting spare holding 20 to 30 percent. The trade-off is the sensor and platform cost, paid back in 1 to 3 years on critical plant.
Can small sites use predictive maintenance?
Yes via low-cost wireless nodes and a cloud dashboard; a 10-machine site can start at USD 1,000 to 3,000. Below that, route-based handheld checks every 30 to 90 days are cheaper. Predictive pays when one unplanned stop costs more than the sensor rollout.