I was investigating if these algorithms could be used in an embedded environment. Basically I implemented a load monitor on a raspberry pi using a pretty basic machine learning algorithm and checked performance. Performance wise it worked out pretty well but the accuracy was pretty bad on real world data. Like 50%. Of you’re interested in how it works I can give some more info.
Valid point about privacy issues. But actually this topic is also interesting in an industrial context when you want to know how much your machines are using to optimise your factory. But yeah, it always depends on how you use it.