In the comments, Paul H pointed me to a great resource on digital filtering – i.e. this online book: The Scientist & Engineer’s Guide to Digital Signal Processing (1999), by Steven W. Smith. I’ve been reading in it for hours on end, it’s a fantastic source of information for an analog-signal newbie like me.

From chapter 21, it looks like the simplest filtering of all will work just fine: a moving average, i.e. take the average of the N previous data point, and repeat for each point in time.

Time for some simple programming and plotting. Here is the raw data I obtained before in gray, with a moving average of order 349 superimposed as black line (and time-shifted for proper comparison):

Note that the data points I’m using were sampled roughly 50,000 times per second, i.e. 1,000 samples for each 50 Hz sine wave. So averaging over 349 data points is bound to dampen the 50 Hz signal a bit as well.

As you can see, all the noise is gone. Perfect!

Why order 349? Because a 349-point average is like a 7-point average every 50 samples. Let me explain… I wanted to see what sort of results I would get when measuring only once every millisecond, i.e. 20 times per sine wave. That’s 50x less than the dataset I’m trying this with, so the 1 ms sampling can easily be simulated by only using every 50th datapoint. To get decent results, I found that an order 7 moving average still sort of works:

There’s some aliasing going on here, because the dataset samples aren’t synchronized to the 50 Hz mains frequency. So this is the signal I’d get when measuring every 1 ms, and averaging over the past 7 samples.

For comparison, here’s an order 31 filter sampled at 10 KHz, i.e. one ADC measurement every 100 µs:

(order 31 is nice, because 31 x 1023 fits in a 16-bit int without overflow – 1023 being the max ADC readout)

The amplitude is a bit more stable now, i.e. there are less aliasing effects.

Using this last setting as starting point, one idea is to take 500 samples (one every 100 µs), which captures at least two highs and two lows, and to use the difference between the maximum and minimum value as indication of the amount of current flowing. Such a process would take about 50 ms, to be repeated every 5 seconds or so.

A completely different path is to use a digital Chebyshev filter. There’s a nice online calculator for this over here (thx Matthieu W). I specified a 4th order, -3 dB ripple, 50 Hz low-pass setup with 1 KHz sampling, and got this:

Very smooth, though I didn’t get the startup right. Very similar amplitude variations, but it needs floating point.

Let me reiterate that my goal is to detect whether an appliance is on or off, not to measure its actual power draw. If this can be implemented, then all that remains to be done is to decide on a proper threshold value for signaling.

My conclusion at this point is: a simple moving average should be fine to extract the 50 Hz “sine-ish” wave.

Update – if you’re serious about diving into DSP techniques, then I suggest first reading The Scientist & Engineer’s Guide to Digital Signal Processing (1999), by Steven W. Smith, and then Understanding Digital Signal Processing (2010) by Richard G Lyons. This combination worked extremely well for me – the first puts everything in context, while the second provides a solid foundation to back it all up.