Hey, thanks for all of the info you are finding about your fridges!
On the topic of fault detection, we think this is something we will be able to do over time. We do see certain types of faults in the data, but we think that in order to do a decent job of this we are going to need to develop models of particular types of faults (failing compressor, refrigerant which has leaked etc). So, once we have a lot more data, we will be able to mine it for enough examples of certain types of failures and then figure out how to reliably detect them.
We don’t think the profiles coming out of the current application are good enough for this yet. For the examples here where compressors sometime stay on for longer than expected, this could be because the compressor really stayed on longer than expected (either because something was wrong with it, or perhaps it was just in some different mode temporarily), or could be because Sense missed the off transition! We of course try to minimize this sort of thing but aren’t doing a good enough job of this yet to use the lack of an off transition as a reliable indicator of a failure of the device itself.
On the topic of models changing over time, the models do adapt over time to new data coming in. We need to do this in case the operating mode of a device changes somewhat – changing of ambient temperature over time for example. Just having more data and adapting the models to the new data usually results in improved models over time. But, it does sometimes go wrong and something which was modeled correctly goes away or the performance becomes worse. This is yet another thing we are actively working on!