[steventhedev]

Temperature sensors are a great example for alerting because they fluctuate constantly, have multiple seasonalities, and failures can be subtle. In the end, you'll want:

1. If the sensor dies and there is no data at all

2. If the sensor gets stuck (giving same value)

3. If the sensor slowly drifts (adjusting for daily, weekly, and yearly seasons) - indicating a clogged filter or leaking refrigerant

4. Statistical spikes - this is the hardest to tune so you need to treat it as a model that detects anomalies and it takes a long time to label extremely rare events

5. Static thresholds, over varying windows to deal with sensor error and transient spikes.

It also raises questions like "if the sensor is reporting 400C then either the building is on fire or the sensor is broken", or "how do we get the alert if the building is indeed on fire" and the inevitable followup: do we even need to get an alert if the building is on literal fire?

[quickthrowman]

> It also raises questions like "if the sensor is reporting 400C then either the building is on fire or the sensor is broken", or "how do we get the alert if the building is indeed on fire" and the inevitable followup: do we even need to get an alert if the building is on literal fire?

Haha, I was looking at a BAS front end with a customer and we were trying to figure out why an air handler kept tripping out. The return air duct static pressure sensor was reading 65536 kpa, so either the sensor was faulty or the building was moments away from being vaporized in a massive explosion. Replacing the sensor corrected the issue and hundreds of lives were saved ;)