How Elite Energy Management detects and validates HVAC faults
Elite Energy Management reads the data your buildings already produce, learns each rooftop unit's normal behavior, screens every unit nightly with layered AI and physics-based checks, and scores it 0 to 100. When a score falls, the platform explains why. And every detection claim we publish is validated against technician-confirmed service events, not marketing estimates.
What data does the platform analyze?
The platform connects to the energy management system or connected thermostats a building already has and reads what they produce: zone temperatures, setpoints, cooling and heating stage activity, fan runtime, and outdoor air temperature, at minute-level resolution, every unit, every minute. This page is about method, and the method is the same whether the platform is watching one building or a thousand. Nothing is installed, so the data that powers detection is data that was already flowing and going unanalyzed.
How does the platform learn what normal looks like?
Every unit gets its own baseline. A rooftop unit over a kitchen behaves nothing like one over a dining room, and a unit in Phoenix behaves nothing like the same model in Minneapolis, so the platform compares each unit to itself: how it has historically performed at this outdoor temperature, at this time of day, under this schedule. Baselines begin forming as soon as data flows and sharpen over the first few weeks. Clear-cut failures do not wait for the baseline; they surface almost immediately.
How are developing faults detected?
Three layers screen every unit every night, because no single method catches everything:
- Physics-based engineering rules watch for the known signatures of specific failures: the falling cooling performance of refrigerant loss, stages that no longer respond, short cycling, and airflow problems.
- AI anomaly detection catches patterns too subtle or too unusual for fixed rules, the combinations of readings that do not match how healthy units behave. More than one model screens every unit, and the models check each other's work; when they disagree, the disagreement itself gets a closer look.
- Behavioral drift tracking compares each unit to its own baseline, catching the machine that still hits setpoint but works a little harder every week to do it, which is how most failures announce themselves before anyone in the building feels a thing.
More happens behind those layers than a summary can show. Every comparison is weather-normalized, so a brutal week in August is not mistaken for a failing compressor. Units with different duty profiles are modeled differently: a rooftop unit over a kitchen lives a different life than one over a dining room, and the platform knows it. Problems must persist to escalate, so one odd night does not alert anyone, while a signature that deepens across nights rises in severity. The platform also forecasts what each unit should do next and treats reality drifting from the forecast as another early warning. And new detection models have to earn their place: they run silently alongside the production system for weeks and are promoted only when they prove more accurate against real fleet outcomes.
The specific models, features, and thresholds are the product, so this page describes what the system does rather than how it is built.
What does the health score mean?
Every unit gets a 0 to 100 score every night. Above 70, the unit is running well. Between 40 and 70, something is worth watching, and the platform explains what changed. Below 40, action is needed. One honesty rule built into the scoring: units without enough sensor coverage to score reliably are labeled as data quality issues rather than being scored misleadingly, because a confident number built on thin data is worse than no number.
How are detection claims validated?
When a technician services a monitored unit, the outcome is logged against what the platform had already flagged: was the unit flagged before the visit, how many days of warning did the operator get, and did the score recover after the repair. In technician-confirmed cases, 87% of repairs were flagged before the technician arrived, with a median of 6 days advance warning. We track the misses with the same discipline, because a detection stat that ignores its misses is not a stat. Post-repair score recovery closes the loop: if the score does not recover, the problem was not fixed, and the platform says so.
Frequently asked questions
Does the platform need new sensors or hardware?
No. It reads the data your existing energy management system or connected thermostats already produce. If a site has no connected controls at all, basic hardware has to come first, and that is the only case where hardware enters the picture.
Can the platform be wrong?
Yes, and we say so. Some flags resolve on their own, and occasionally a fault arrives without measurable warning. Both are logged against the service record, which is what keeps the published detection numbers honest.
How long after connecting does detection start working?
Data flows within days of connecting. Obvious problems surface almost immediately, and detection sharpens over the first few weeks as each unit's baseline forms.
Every number on this page comes from the same production system that scores the fleet nightly, refreshed as the fleet grows. See what the platform found across the monitored fleet on the blog, or see it on your own buildings with a free 30-day pilot: read-only access to your existing systems, first health scores within days, no hardware, no commitment. Pricing details here.