I hear what you’re saying about real-time detection, but I think that has some conceptual problems. A modified real-time approach might solve this. Imagine saying that you’ll route packages by the width and depth you see as they pass the scanner. Tag a signature with a device name, but if the owner has indicated that device is conflated with others, be able to take a second look.
Heating elements all have similarities; the total power draw, time to reach peak heat or to cool again are some of the details that might tell one from another. The space being heated is also a key that it seems Sense does not look at. The mass being heated will determine the duty cycling parameters. An oven is bigger than the toaster oven, which is bigger than the toaster. A soldering iron, a curling iron, an individual cup heater, a water kettle, a coffee maker. All of these want to heat up quickly, and are limited to the amount of power a socket can provide. Some of them will try to hold a temperature, some just reach a temperature and shut off.
Motors may have chaotic loads (clothes dryer with tumbling clothing), or may have constant loads (air blower). Some may have decreasing load (garage door). You use some of these details now. I think you are wasting another measurement (time and duty cycle) that could be very helpful.
I don’t know where the logic is that tags the power graph and puts entries in the timeline. It may be inside the Sense unit in my breaker panel, or it may be in ‘the cloud’ in Sense’s servers. Wherever it is, it must have a list of devices in my home that have been identified, and the timeline to display when I look at it. If the device list had an attribute that says ‘this device needs two looks to distinguish from another device’, you could queue up the second look to see the duty cycle, or total runtime, single heat, etc. When tagging the signature as an coffee maker, you would see that I’ve indicated it was confused by the toaster oven. (Both use maximum allowable power but one cycles and the other doesn’t). Queue it up for another look after three minutes and you’ll know which was right.
I think Sense does great identifying what it does now, but could be even better.