anorlunda said:
...
A project I would like to do myself could be called a baseliner. It would be more fun to do it in a car than a house. It would simply log all possible measurements, and alert me to sudden changes or ramping trends. In a car for example, a sudden change in mpg signals something wrong. The cool thing is that the logging device has no need to know what is being measured, and various filters could be applied for expected daily/seasonal variations etc. ...
I actually had a great deal of success with something like this at my work.
(for the tl/dr version - picking your baseline is key!) We would have maybe a dozen production test/tuning systems running in parallel to handle the volume. We logged all the data for test results, and the electronic tuning settings ( for example, there might be 3 bits, values 0-7, dedicated to setting modulation levels). We had some PhD statisticians from corporate try to do some process monitoring, and while they were all very bright and hard working and had lots of resources available, they failed, I think because they just didn't grasp the 'system'.
They kept trying to apply traditional statistical analysis to detect sudden shifts in this data, and alarm on it. But we tune the product so that it
can handle variations in the supplied components. So a sudden shift in the mod setting from maybe a range of 5-6-7 to 2-3-4 is an alarm to that system, but we (and the product designers) know that might have just meant we got a batch of components from a different supplier, or a different batch from the same supplier. Ideally, you want them all the same, but that costs $. And the tuning algorithm is designed to accommodate that wider range, so it's just doing what it was designed to do. Now, if you start seeing a lot of 0's and 7's, yes, that can indicate you are running out of range, and might start seeing failures - but not necessarily, there may still be enough tuning range to accommodate the variation.
So anyhow, I took a different approach (with a narrower goal). I used a running average from that group of a dozen test benches for my 'baseline'. If one bench was running significantly different from the others, it was a warning that that bench might be out-of-calibration, or having some other sort of noise, or contact problem. So it didn't tell us so much about the product, but it was super-effective at alerting us to any test bench that was drifting from the others. It was just me and a summer intern working on it part-time, and we got it all going during her summer with us.
It was pretty funny, because our maintenance techs were skeptical, and the gruffest one in the bunch was complaining about having to sign off on this report at the start, middle, and end of the shift . A few days later, he came to me and said he was "getting a stupid alarm from that report, and he knew it was stupid, because they hadn't had any failures for that parameter on that bench all shift, and he didn't understand why he should have to 'fix' a bench that wasn't failing any product". But he said he checked it anyway, and the connector that is critical to that specific measurement was loose, ready to fall off. He fixed it before we saw any failures, saw the report come right back in line, and was a believer after that!
RE: "
various filters could be applied for expected daily/seasonal variations etc." ... Yes, I forget the details now, but I had an array or two for each parameter with some sort of 'ignore if' factors. I didn't want it to be too sensitive and create false alarms that would lead to the techs ignoring the alarms.
anorlunda said:
... More cool, and more sophisticated, just measure the total power consumption in your house and log it every second. Then you could learn to recognize things like water heater cycles, washer/dryer/fridge cycles, garage door open/close, and when people leave for work and come home. You could baseline it to watch for trends or changes, to optimize energy use, or simply have fun figuring out how the power company can spy on your privacy using nothing more than the smart meter data.
I've got a smart meter now, and for some odd reason, I rarely check it. I should take another look. Again, since the data is normally what I expect, there's no reason to look. But then I will miss an anomaly. I guess I could try 'scrubbing' their website and doing my own alarms?