Predictive maintenance for HVAC: use the data you have

Predictive maintenance is sold as a product these days: a cloud platform, wireless sensors, a dashboard and a monthly subscription. The assumption that comes with it is stubborn — that you need a capital project and a building management system retrofit before you get anything out of it. That assumption is backwards for most commercial buildings in Greater Montreal. Predictive maintenance for HVAC starts with data your building already produces, for free, on every service visit and every utility bill. Hardware comes second, and only on the assets that justify it.

Predictive maintenance for HVAC: what it actually means

Predictive maintenance for HVAC means triggering work from an asset’s measured condition rather than from a date on a calendar. You trend indicators — combustion efficiency, vibration, bearing temperature, motor current, run hours — and you intervene when the trend leaves its normal band, before the failure.

The distinction is not cosmetic. The U.S. Department of Energy’s reference guide defines predictive maintenance as measurements that detect the onset of degradation, allowing the underlying stressor to be removed before the component physically deteriorates. Preventive maintenance stays time-based: change the oil at the scheduled mileage, clean the heat exchanger every summer, regardless of what the machine actually went through.

Reactive, preventive, predictive: three logics, one budget

No building runs purely on one of these. The real question is the mix — and what each slice costs.

The same Department of Energy guide, widely cited in facilities management, estimates that a properly functioning predictive program returns savings of 8% to 12% compared with a preventive-only program. That is not a magic number: it is the order of magnitude you get from allocating the same maintenance dollars better. The savings do not come from a technology, they come from no longer servicing what does not need it so you can service what does.

The data your building already generates

Before buying anything, inventory what exists. In a typical commercial boiler room, four data series are sitting in binders or in someone’s inbox:

• Combustion analysis reports. Every visit leaves an O₂, CO and flue gas temperature reading. On its own it is a snapshot. Lined up over three years it becomes a slope: efficiency slipping a point a year, CO creeping up, stack temperature rising — a signature of fireside fouling or waterside scale.
• Run hours and cycle counts. Most modern burner controls track them. A burner whose cycle count doubles from one season to the next at comparable degree days is not running better, it is short-cycling — and electrode, valve and blower wear follow the same curve.
• Supply and return temperatures. The gap between them, logged under comparable conditions, is one of the most honest indicators on a hydronic loop. A flattening ΔT points to rising flow, a fouling exchanger or a valve stuck open.
• Utility bills. Normalized to heating degree days, they expose an efficiency drift long before a tenant complains about comfort.

The obstacle is rarely technical: it is that this data is scattered. Properly kept cloud maintenance records turn a stack of work orders into a usable series. No consolidated history means no trend, and no trend means no predictive anything.

Which assets actually deserve condition monitoring?

This is where most programs go off the rails: buildings instrument what is easy to instrument instead of what matters. Three questions sort it out.

What does downtime cost? A single boiler serving a rental building in January is not the same asset as a rooftop unit over a storage room. Across a Greater Montreal portfolio, where losing heat becomes a habitability emergency within hours in deep cold, equipment without redundancy goes straight to the top of the list.

Is the degradation gradual? Predictive only works on failures that announce themselves. A pump bearing, a fouling exchanger, a slackening belt, a motor whose current is drifting: yes. A control board that dies after a thunderstorm: no. No sensor anticipates a random event.

Is the measurement repeatable? A vibration reading is worth something only if it is taken at the same point under the same load conditions. That is the principle behind the ISO 20816 series, which governs how machine vibration measured on site is evaluated. A reading taken “roughly in the same spot” does not produce a trend, it produces noise.

In practice, in a commercial building, the short list usually comes down to three or four machines: the lead boiler, the primary circulation pumps, and the hardest-working compressors.

Field case: the pump that had been announcing its failure for six weeks

In an office building in the Montreal area, the primary circulation pump on a hot water loop failed on a Friday in February. Emergency replacement, temporary rental pump, two floors underheated all weekend.

Reviewing the history afterwards, the signal had been there. For six weeks the technician had noted a bearing noise “to monitor” on every work order. Motor current, logged on each visit, had climbed several percent at constant flow. And the loop ΔT had pinched, consistent with an impeller losing performance. Three clues, three separate documents, never placed on the same page.

None of that required a sensor: the measurements already existed. What was missing was the habit of putting them side by side every quarter. That building now logs motor current and loop ΔT in a single table reviewed at each visit — and the second pump, which showed the same drift a year later, was rebuilt on a Tuesday in September, on schedule, at a fraction of the cost.

What predictive maintenance does not replace in Quebec

A predictive program does not exempt anyone from a legal obligation. Pressure equipment, boilers included, remains governed by the Régie du bâtiment du Québec: periodic inspections, an up-to-date register, defined operator duties. Work on gas-fired appliances falls under the CSA B149.1 code as adopted in Quebec. Condition monitoring sits on top of that floor, never in place of it.

That is also the clearest way to present it to an owner: the regulatory schedule guarantees compliance, the data guarantees availability. A well-run preventive maintenance program supplies the raw material — every mandatory visit generates readings that feed the trend. The two approaches reinforce each other rather than compete.

When does it start paying off?

The trade-off is asset by asset. Permanently instrumenting a machine whose failure costs one weekday service call makes no sense. Instrumenting the one whose failure triggers a weekend emergency, a rental unit and a round of tenant complaints pays for itself on the first avoided incident.

Between the two sits an often-overlooked middle path: the temporary measurement campaign. Leaving a data logger on a suspect asset for two or three weeks costs a fraction of a permanent installation and settles the question — either the drift is confirmed and justifies continuous monitoring, or it disappears and the budget goes elsewhere.

On the funding side, Quebec programs think in measured results. Hydro-Québec’s Solutions efficaces program supports efficiency measures in commercial and institutional buildings, and a documented energy optimization project — with before-and-after data — defends itself far better than a hunch. The same measurement series therefore does double duty: predicting the failure, and proving the gain.

Your first move this week

Do not start with a sensor tender. Pull the last three service reports for your most critical asset, put the combustion readings, run hours and loop temperatures on a single sheet, and look at the slope. Within an hour you will know whether an indicator is already drifting — and you will have the first row of the table you update at every visit from now on.

Reading those trends is what the Montréal Combustion team records on every intervention instead of letting it die inside a work order: a boiler room whose slope you know stops surprising anyone at the first deep freeze.