AI as a co-diagnostician.
Industrial troubleshooting is a hypothesis game — ten things could cause a compressor trip, nine of them are wrong, and the cost of checking is measured in hours and refrigerant kilos. AI doesn't replace the technician's judgement; it prunes the tree fast enough that the right cause surfaces while the system is still warm.
Ten candidates. One right answer. Fast.
A field alarm is a question posed in bad lighting: what changed? The discipline is Bayesian — you hold a prior ("condenser fouling accounts for 70% of high-side pressure trips in my installed base"), you gather evidence cheaply (the fan is running, the ambient is 14 °C, the sight glass is clear), and you update. The technician who wins is the one who eliminates three causes in the time it takes the competition to check one.
AI's role isn't to replace that Bayesian loop — it's to expand the prior. A well-indexed language model has read a thousand more service bulletins than you have, seen failure modes on adjacent product families, and remembers what a desaturation fault on a parallel drive looks like at 3 AM. The technician still owns the call. The AI just makes sure no branch of the tree gets skipped.
Model, manual, and a good multimeter.
The AI-assisted diagnostic stack is layered — a reasoning model at the top for hypothesis work, a curated knowledge base in the middle for manuals and bulletins, instrumentation at the bottom for ground-truth. None of the layers matter without the others. The model hallucinates without grounding; the manuals collect dust without a reader; the instruments measure the wrong thing without a hypothesis.
The elimination loop.
A diagnostic session is a loop, not a line. Each step produces evidence that either confirms a branch or prunes it — and the loop continues until one branch is the last one standing. Speed comes from cheap eliminations early; accuracy comes from refusing to commit until the last branch is the only one that explains every piece of evidence.
One compressor, one drive, one motor.
The diagnostic tool below covers three invented units designed to exercise realistic fault modes. The compressor is a semi-hermetic CO₂ reciprocating unit; the drive is a mid-range AC VFD with braking and STO; the motor is a Super Premium efficiency induction machine. Between them: 66 alarm codes, all prefixed RGH-.
Feed it a code or a symptom.
Two doors into the same reasoning engine. If you have the alarm in hand, type the code or a fragment of its name — RGH-C05, discharge, igbt. If you only have an operator description — "it's burned", "it won't start", "it hums but doesn't spin" — pick from the symptom chips and get a ranked list across all three units.
"The chiller that blamed everybody."
Late-night call from an operator: a CO₂ chiller tripping repeatedly with three alarms stacked — RGH-C01 discharge pressure high, RGH-D04 drive output overcurrent, and RGH-M02 bearing DE temperature high. Classic cascade. Each alarm, taken alone, points somewhere different. Together, they tell one story — but only if you read them in the right order.
Parameter dump from the PLC shows the three trips landed inside a 90-second window. Ambient 32 °C. Load at 88% nominal. Last successful run 14 minutes prior.
Time-ordering matters enormously here. The PLC log puts RGH-M02 (bearing temp) first, then RGH-D04 four seconds later, then RGH-C01. That ordering changes the whole tree.
If the bearing was the root, the drive tripped because the motor's friction load spiked, and the compressor tripped because the drive yanked the shaft. That's one story. The competing story: the compressor was already fighting a blocked condenser, the drive was working overtime, the motor overheated its bearing from heat soak. Cost-ordered test: check the bearing first.
DE bearing PT100 reading 118 °C with the machine stopped for eight minutes. Condenser inlet 34 °C — perfectly normal for 32 °C ambient. Sight glass clear. Discharge pressure nominal during the cooldown. The condenser story is dead.
Turn the shaft by hand — noticeable resistance at one angular position. Vibration log from the previous 48 hours shows a growing 1× rotational signature. Every piece of evidence now fits: bearing failure caused the friction spike, the friction caused the overcurrent, the overcurrent triggered the pressure-side trip downstream.
Bearing replacement ordered. Customer informed the fault is mechanical, not refrigeration — their condenser is fine, they don't need a refrigerant top-up, the drive doesn't need replacement. Three alarms, one cause, one procurement line.
Why this is the AI leverage. A tired technician at 23:00 reads "discharge pressure high" and heads for the condenser — it's the 70% prior. The AI co-diagnostician flags the alarm ordering in the PLC log as the single most informative data point on the page, re-ranks the tree accordingly, and suggests the bearing test first. Ninety minutes saved before the first wrench is picked up.