Electric heat on a PWM signal

This is the electric strip (4,8kw @ 240v) of my air handler during a defrost. It has been detected 3 days ago. During defrost a PWM (pulse width modulation) digital output of my hvac controller try to maintain 90F supply temp. You can see the last 2 minutes of defrost when demand is 100% because 4,8kw is not enough to handle the defrost load. At the beginning, the evaporator is not cold enough so the electric strip is pulsing on and off. However we can’t see that on the power meter view. Can someone tell me what is the precision of the power meter view?

I’m not sure what you are looking for. Your waveform shows your strips on for only 2 minutes. If your looking for it to cycle rapidly to maintain the 90 degree temperatures then it’s doing its job.
In the beginning when the strips first start to heat up, the waveform looks different because the strips themselves are not up to maximum temperature. The last half they are and then a constant draw from there. As far as accuracy. I don’t have the numbers or proof but feel what you see on the timeline is about as accurate as it gets
Your strips are running in defrost mode because while defrosting, you are actually running A/C and not heat. The strips come on be default in defrost to overcome the cold air you would feel.

The waveform at the beginning is not the exact waveform for sure. I can confirm the heat strip was pulsing on and off at that exact moment but get remodel by sense.

1 Like

Now that I’m paying attention to the person posting, I’m sure I’m wrong and you are correct.
Sorry about that @HVAC_Marc
I wasn’t aware this was something that happens with systems as I haven’t seen it before. Does sound like it would really save on energy and now you’ve peaked my interest about modulation and timeline accuracy.
I’m wanting the exact opposite as you as I ha e a washing machine that will show on/off 30 or more time for a single run cycle. It’s feom the drum switching direction I think. I would like sense to remodel so it combines everything or squeeezes them together where it’s not recognized as on and off but a continuous cycle.

Per the specs, Sense is sampling somewhere near a max of 4M 2 byte samples per second, though how it splits those samples between voltage and current, and mains vs. solar is not clear. But if you are looking at your snapshotted waveform in the Device Power Meter, you are looking at thrice or so abstracted data from the raw sample and sample resolution. Somewhere in the process, Sense:

  1. Does some local processing of the data, then compresses and transmits back to the mothership. Who knows what magic goes on locally though there seems to be enough going on in Sense monitor to allow it to do simple detections.

  2. The data is then converted to 2 readings/sec RMS power info for the main Power Meter

  3. The Device Power meter is synthesized into a 2 reading/sec RMS reading using the output of some kind of neural network which may or may not totally understand the PWM.

If the PWM pulses are less than 1/10 second, I would expect them get swallowed up in both the Power Meter and Device Power Meter RMS averaging. If the pulses / gaps are longer than 1 sec, then you should see them, at least in the main Power Meter. In the middle, you are likely to see a mishmash of results…


Exactly where in the data are detections originating? Is it from just the raw data sent to the servers and only happens there?
I’m curious because we are able to see very easily the undetected device as their cycle from the remodeled data shown on our timeline.
I’m thinking that sense should not only be looking at the original data but should also look at what we see.
I’m sure what the monitor sends to the servers is a mess to the human eye and cleaned up for our viewing. Do you know if this “second look” happens or not.
@RyanAtSense might also have an idea if this happens or not.

1 Like

Thanks @kevin1 for the answer.
This is a view from main power meter. You can see the heatpump compressor ramping up. Then i guess the littles spikes are on/off cycle of the heat strip. I guess there is a little confusion between compressor and heat strip at some point even if sense is able to catch the real power value of the heat strip when it’s full On. My controller pulse on a 1 second cycle (1 second on 9 seconds off = 10%).

We’ve actually done this to a certain extent with EV detections. Due to their long ramp-up/ramp-down times, ultra high-res data made finding them somewhat more problematic. Thus, we utilize a more macro-level timescale. But, this is not easily mappable to all devices and required a significant amount of labor to accomplish. Still, it’s certainly something we’ll continue to test.


I believe the strategy used for EV detections has worked pretty well. I don’t have one myself but judging by all the posts, there are many people satisfied with their detections.

Mon plaisir. I would have expected to be able to see the one second spikes more clearly in the main Power Meter. Perhaps you can zoom in further, maybe to a full screen resolution of 20-30 seconds ? Would be interesting.


Suspect that many of the detections are done at the monitor, at least the short ones

  1. Changes in waveforms are everything when looking for on-and off-events. There’s “no signal” in a waveform that is pretty much constant.

  2. Sense was right to go after the immediate (within 1/2 second) changes first because they are indicative of the fundamental device or device component types (resistive, inductive, capacitive) and magnitude. Much of the most distinctive data is hidden in that region of the waveform. To put it differently, longer on and off signatures (EV charging, washing machine cycles with DC motor) have been deliberately and artificially stretched out by electronics. For those device types, Sense can no longer use simple physics.

  3. When you get into those more complex device types, detecting and classifying changes is really about sweeping the spectrum of resolution looking for timing correlated deltas in current, including the half second / 10 minute window we humans use to identify “clear” waveforms. But there is a lot of “spectrum” territory to search. Betting most of that is still done back at the Sense mothership.