On the Camera That Drifted on the Hot Aisle [/ɔn ðə ˈkæmərə ðət ˈdrɪftɪd ɔn ðə ˈhɑt ˈaɪəl/] n - The symptom. A fixed thermal camera watching a hot aisle (part of the Energy-Saver rig I built to catch hotspots early) slowly began disagreeing with the spot-check infrared thermometer. Over weeks its readings crept low, so genuine hotspots went under-reported. A monitoring system you cannot trust is worse than none, for people cease to double-check.

The dead end. [/ðə ˈdɛd ˈɛnd/] n - The textbook answer is “send the camera out for calibration.” That means pulling the sensor, losing coverage for weeks, and paying a lab; and it would return accurate only until it drifted anew. Uncooled thermal sensors drift with their own temperature; on a hot aisle the camera bakes, and that is precisely where you most need it honest.

The actual cause. [/ðə ˈæktʃuəl ˈkɔːz/] n - The drift tracked the camera body’s own temperature, which rose as the aisle’s load grew. The sensor was not failing; it was being asked to report absolute temperatures in an environment that was quietly biasing it.

The fix - limited and clever. [/ðə ˈfɪks/] n - Rather than chase absolute calibration, I mounted a small reference target of known, stable emissivity within the camera’s field of view, a fixed patch whose true temperature I could measure independently. Each frame, the software reads that patch, applies the offset needed to true it, then applies the same offset to the whole image. The camera corrects itself every frame against ground truth, so drift cancels without the sensor ever being removed. This pairs naturally with the stable thermal environment that Direct-to-chip liquid cooling is bringing to newer rooms.

The lesson. [/ðə ˈlɛsən/] n - If you cannot keep an instrument from drifting, give it something true to look upon and let it correct itself. A known reference in-frame beats a perfect calibration that begins decaying at once.

Part of Field Notes. Built on the tooling in Crafts.