Anomaly detection is not engineering inspection

Anomaly detection and inspection sound like the same thing, and the difference is the entire value of a monitoring platform. Anomaly detection answers "did this number move unexpectedly?" Inspection answers "is this asset behaving as it should, and if not, what should someone do about it?" The first is a statistical trigger. The second is an engineering judgement.

The naïve approach sets a threshold — alert if output drops below some line — and it fails immediately on a solar plant, because a solar plant’s output is supposed to move. It falls at dusk, under cloud, in dust, in heat. Set the threshold loose enough to avoid firing on every passing cloud and it misses a genuinely dead string; set it tight enough to catch the dead string and it fires all day on normal weather. Either way the owner learns to ignore it, and an ignored alert is worse than none, because it looks like coverage while providing none.

The deeper problem is that a threshold has no idea what "normal" is for this string, on this site, at this hour, in these conditions. Without an expectation, a deviation is just a number crossing a line.

Inspection replaces the fixed line with an expectation. Each string is scored against a weather-adjusted expectation for that string, so the question becomes "is this producing what it should, given today’s conditions?" rather than "is this above or below a hard number?" A drop that matches the weather is not a fault; a drop that the weather does not explain is worth looking at.

Then inspection demands persistence and context before it commits. A signal must survive across cycles and pass weather, soiling and time-of-day checks before it is allowed to count — a single odd reading is not an event. This filtering is not cosmetic: it is what cut alert noise by 76% on the systems we monitor, turning the same underlying faults into a fraction of the messages.

Finally, inspection classifies severity and puts a human in the loop. Complex events are surfaced for a solar engineer to investigate, not pushed to the owner as confirmed faults. The AI recommends; people decide. What reaches the owner is a verified, ranked action — not a shrug in the form of a notification.

An owner does not want to know that a number moved. They want to know whether their plant is honest, and if not, what to do first. Anomaly detection can only ever hand over the first; inspection is built to deliver the second. That is why we describe SPC as inspection, not a dashboard — the thing being protected is the performance of the owner’s investment, not the completeness of a chart.

It is also why the finding that 61% of monitored inverters had a dead string or went silent within 60 days is an inspection result, not an anomaly count. Anyone can raise 61% worth of alerts. The work is deciding which of them a person should act on — and being right often enough that the person keeps listening.