By Todd Pree
Property maintenance has traditionally used two broad approaches. Reactive maintenance fixes equipment after failure. Preventive maintenance performs work on a schedule, whether or not the individual asset shows signs of a problem.
Predictive maintenance adds condition and performance data to the decision. The goal is to identify developing failure or degradation early enough to act, while avoiding unnecessary service on healthy equipment. It can improve uptime and labor planning, but it is not a universal replacement for inspections or scheduled maintenance.
Predictive maintenance begins with failure modes
A useful program asks how a particular asset fails and which signals appear beforehand. A bearing may produce vibration and temperature changes. A filter may create increasing pressure drop. A valve may show a mismatch between command and actual flow.
Collecting every available sensor reading without a failure hypothesis can produce noise. Maintenance, engineering, and data teams should define the expected failure modes, consequence, warning period, and available intervention.
The selected signal must provide enough time to act. Detecting a failure seconds before it occurs may help safety controls but not maintenance scheduling.
Data can come from several sources
Condition monitoring may use vibration, temperature, pressure, current, sound, oil analysis, run hours, starts, alarms, and control behavior. Work-order history, manufacturer guidance, and technician observations provide additional context.
Some equipment already exposes useful points through a building-automation system. Other assets may require added sensors or portable inspections.
The cost and reliability of the measurement should match the asset. Instrumenting a low-cost, noncritical fan with an expensive sensor system may not be justified.
Rules can be as valuable as machine learning
A simple threshold or engineering rule can detect many problems. If supply and return temperatures are inconsistent with a commanded valve position, the system can flag the condition.
Machine-learning models are useful when normal behavior changes with load, weather, occupancy, or operating mode and is difficult to capture with a fixed rule. They can identify unusual patterns or estimate remaining useful life.
Complexity should be earned. A transparent rule that operators trust may produce more value than an opaque model with slightly better laboratory performance.
Asset context prevents false alarms
A temperature spike may be normal during startup and abnormal during steady operation. A vibration level may differ by speed. Equipment can be intentionally off because a space is unoccupied.
The analytics need operating state, schedule, load, location, and relationships to other systems. This context reduces alerts that are technically unusual but operationally harmless.
Commissioning data and known-good periods can provide a baseline, though the baseline should be updated when the equipment or control sequence changes.
Alerts must enter the maintenance workflow
An alert creates value only when someone evaluates and acts on it. The system should identify the asset, evidence, likely issue, urgency, and recommended checks. It may create a work order or request a review.
Technician feedback should record whether the issue was confirmed, what was found, and what action was taken. That outcome helps improve thresholds and models.
A separate analytics dashboard that no one checks is not a predictive-maintenance process.
Prioritization should reflect consequence
Not every anomaly is urgent. Prioritize using failure probability, safety, tenant impact, energy loss, replacement lead time, redundancy, and repair cost.
A small efficiency loss across hundreds of similar units may have significant portfolio value. A rare warning on a critical chiller may require immediate inspection even if the probability is uncertain.
Risk-based prioritization helps maintenance teams direct limited labor to the most valuable work.
Preventive work still has a role
Regulatory inspections, safety tasks, lubrication, filters, and manufacturer-required service may remain schedule based. Some failures have no reliable precursor or cost less to handle reactively.
Predictive information can adjust frequency or target inspections, but changes should respect warranty, regulation, and engineering requirements.
The goal is an optimized maintenance strategy that combines reactive, preventive, condition-based, and predictive methods by asset class.
Data quality can undermine the model
Sensors drift, points are mislabeled, clocks differ, work orders use inconsistent asset names, and failure codes are incomplete. A model trained on these records may learn maintenance practices rather than physical failure.
Data-quality checks and asset master data are foundational. Technicians should be able to report bad sensors and incorrect mappings.
Historical work orders may be too sparse for supervised failure prediction. Anomaly detection or engineering rules can be a more realistic starting point.
Measure avoided cost carefully
A program may claim every detected anomaly as a prevented failure. That exaggerates value because some conditions would not have progressed or would have been caught during routine work.
Useful measures include unplanned downtime, emergency work, mean time to repair, maintenance labor, parts use, energy waste, tenant complaints, and failure recurrence. Compare similar assets or use a phased rollout where practical.
Document the counterfactual: what would likely have happened without the alert? Estimates should be conservative and transparent.
Start with a focused pilot
Choose an asset class with meaningful failure consequence, available data, repeatable behavior, and an operating team willing to participate. Define success and the feedback workflow before developing the model.
Validate alerts with technicians and inspect false positives. Expand only after the process proves useful, not merely because the model produces an anomaly score.
The operating team should help design the interface and thresholds. Adoption is a technical requirement.
Final perspective
Predictive maintenance can shift property operations from fixed schedules and emergency response toward evidence-based intervention. Its value comes from detecting a meaningful condition early and connecting that evidence to a well-managed action.
Sensors and models are only part of the system. Asset context, data quality, technician feedback, prioritization, work orders, and measurement determine whether predictive maintenance reduces risk or simply creates more alerts.
Related reading
- Smart Buildings Need Better Data
- Digital Twins in Real Estate
- Real-Time Analytics vs. Batch Processing