What is fault detection and diagnostics?

The mechanics

A fault is not a failure. A chiller with a stuck economizer damper still makes chilled water. An air handler in simultaneous heating and cooling still holds space temperature. Nothing alarms, nobody calls, and the building spends money continuously for years.

FDD closes that gap by running rules or models against operating data:

• Rule-based FDD encodes expert knowledge as logic: if outside air is below the economizer high limit and the damper is at minimum position and mechanical cooling is on, flag a stuck damper. Transparent, explainable, and the dominant approach in practice — ASHRAE Guideline 36 codifies a set of these for common HVAC sequences.
• Model-based FDD compares measured performance against a physical or statistical model of correct performance and flags the residual.
• Data-driven FDD learns normal behaviour from history and flags departures. It is powerful and it inherits every bad habit in the training period.

The distinction between detection and diagnosis matters for what a facility team can act on. "AHU-3 is using 22% more energy than expected" is a detection. "AHU-3's economizer is stuck at minimum position, causing mechanical cooling during 1,200 hours a year when free cooling was available" is a diagnosis, and only the second one produces a work order.

The evidence, which is unusually good for this field

Lawrence Berkeley National Laboratory's Smart Energy Analytics Campaign studied 6,500 buildings across 104 organizations and separated two populations:

	Median annual energy saving at year 2
Buildings running fault detection with a follow-up process	9% ($0.24/sq ft)
Buildings running an energy information system alone	3% ($0.03/sq ft)

Same sensors. Same data. Median simple payback across organizations with an EIS or FDD installation: two years. FDD savings across participants ranged from 1% to 28%. Recurring FDD software cost was measured at $0.02 per square foot per year against a $0.06 per square foot base cost.

The six percentage points between those two rows are the most important number in this field, because it is not attributable to the technology. LBNL named the success factors explicitly: a provider who analyses and prioritises faults, and a routine process for following up with operations. It named the barriers just as explicitly: participants were not reviewing the data, and there was no operations and maintenance staff time.

Confirmation of how much this depends on follow-through comes from New York State's Real Time Energy Management evaluation, published 30 September 2025. In commercial and industrial buildings, where follow-through existed, the programme realised 62% of promised electric savings — 6.08% of baseline. In multifamily, the same programme, same evaluators, same period, realised 9% — 0.76% of baseline, with 78% of sites reporting no savings at all and only 35% of recommended measures implemented where savings did occur.

How much is broken right now

Berkeley Lab analysed fault-detection records from more than 60,000 pieces of commercial HVAC equipment across roughly 90 fault types and found that 40% of air handling units and 30% of terminal units carry a reported fault on any given day — in buildings that all have a building management system. Twenty-one distinct AHU fault types appeared on a fifth or more of every unit monitored. (Crowe et al., Science and Technology for the Built Environment*, 2023, DOI 10.1080/23744731.2023.2263324; DOE-funded.)*

Worked example

A 250,000 sq ft office building.

• LBNL's measured FDD saving: $0.24 per square foot per year → $60,000 a year
• Recurring FDD software at LBNL's measured $0.02 per square foot per year → $5,000 a year
• In-house labour, measured by LBNL at 8 hours per month per building for FDD (against 1 hour for an information system) → 96 hours a year
• At a fully loaded $85 per hour: $8,160 a year of internal time

Net of software and labour, roughly $46,800 a year — and the labour line is the one that decides whether the programme survives. A facility with one engineer covering nine buildings does not have 96 hours a year per building to spare, which is why an FDD deployment that produces an unranked list of 400 faults produces no savings at all.

That is the reason to state the labour number rather than hide it. The relevant question at purchase is not "how many faults does it find" but "how many hours a month does acting on it cost, and which findings are worth those hours".

What FDD is not

It is not an alarm system. A BMS alarm fires when a value crosses a limit — a space is too warm, a pressure is too low. FDD identifies equipment operating incorrectly while every alarm limit is satisfied, which is where most of the wasted energy lives.

It is not automated correction. Detection and diagnosis are read-only. Acting on a fault means someone changes a schedule, replaces an actuator or re-commissions a sequence. Supervised write-back exists in some systems and belongs behind separate written authorisation.

A fault count is not a result. "Detected 1,247 faults" is a liability, not an achievement. The metric that matters is faults closed, verified as cleared, with the recovered dollars attached. A technician silencing forty alerts without reading them is the normal response to an unprioritised queue.

Electrical data alone cannot see every fault. MIT's field study identified seven HVAC faults detectable from electrical load alone; it could not see coil fouling, could not separate small loads at a centralised measurement point, and lost startup signatures behind variable-frequency drives. Code-recognised economizer fault detection — under California Title 24, for instance — requires outside, supply and return air temperatures accurate to ±2°F, which are not electrical points. The findings electrical data supports best are the unglamorous ones: equipment running outside schedule, loads that never turn off, standby power, staging that never sheds. Those are consistently where the money is.

It is not a substitute for a named owner. The single strongest predictor of whether a monitoring programme survives its first year is whether one person owns the queue. No software substitutes for that.