so many alarms, none get answered

Problem statement

Commercial buildings consume ~40% of US energy, and HVAC systems account for roughly half of that. Automated fault detection and diagnostics (AFDD) systems promise to identify equipment malfunctions, control errors, and efficiency degradation before they waste energy or cause comfort failures. However, deployed AFDD systems in real buildings typically generate false positive rates of 50–90%, creating alert fatigue that causes building operators to disable or ignore the systems entirely. The result is a technology that works in laboratory demonstrations and controlled field trials but fails to deliver savings at scale because its interface with human operators is broken.

Why this matters

PNNL estimates that HVAC faults waste 15–30% of commercial building energy — $42B annually in the US alone. AFDD could recover much of this waste, but the false alarm problem has limited market penetration to <5% of commercial floor space despite decades of R&D. Building operators, already stretched thin managing multiple properties, rationally disengage from systems that cry wolf. This alert fatigue pattern extends beyond HVAC to building fire alarms, industrial process monitoring, and IT security — any domain where automated detection outpaces the human capacity to investigate alerts.

What’s been tried and why it hasn’t worked

Rule-based AFDD (ASHRAE Standard 36 sequences, expert system rules) generates false positives because fixed thresholds don't adapt to seasonal variation, building schedule changes, or occupancy patterns. Data-driven approaches (supervised learning on labeled fault data) require training data that is expensive to obtain — most buildings have never documented their fault history. Unsupervised anomaly detection flags anything unusual, which in a real building includes legitimate operational changes (setpoint adjustments, zone reconfigurations, temporary events) alongside actual faults. Even when detection is accurate, diagnosis is often wrong — the system correctly identifies "something is off" but misattributes the cause, leading operators to waste time investigating phantom problems. The fundamental challenge is that building HVAC systems operate in an open-world environment where the boundary between "normal variation" and "fault" is context-dependent and shifts continuously.

What would unlock progress

A hierarchical detection-diagnosis-verification architecture that separates pattern anomaly detection (automated, sensitive) from fault classification (model-based, specific) from operator verification (targeted, actionable) could dramatically reduce the false positive burden on humans. Alternatively, self-supervised learning approaches that continuously update "normal" baselines using building-specific data without requiring labeled fault examples could adapt to each building's behavior. Crucially, the HCI design — how alerts are presented, prioritized, and grouped — may matter more than algorithmic accuracy.

Entry points for student teams

A team could instrument a campus building's HVAC system with additional sensors, deploy an existing AFDD algorithm, and systematically categorize every alert as true positive, false positive, or ambiguous over a semester. This empirical false positive taxonomy would reveal which fault types are reliably detectable and which generate noise. Alternatively, a team could design an alert prioritization interface that uses energy-waste magnitude (not just detection confidence) to rank alerts and test it with real building operators. Skills: HVAC engineering, data science, human-computer interaction, building automation systems.