Occasionally, you might decide to delete a device from your Sense app. Maybe you no longer own that toaster, or you never even had one to begin with? Or maybe that kettle just keeps getting confused with your coffee maker? After you hit that “Delete Device” button, you’ll no longer see the device in the Sense app. But what’s actually happening behind the scenes?
Back to basics
First, let’s go back to basics. A ‘device’ is a pattern in your electrical signals that Sense has learned to recognize. We train machine learning models to detect these patterns using your whole-home electrical history. It’s important to note that Sense doesn’t simply look for patterns in your home and then match them up with some global “dictionary” of devices. All homes look different and Sense must build your local device detections from the ground up, albeit with help from known device signatures. Once the model is trained, it runs on the Sense monitor and looks for instances of this pattern in your whole-home data. When it’s found, device statistics get sent via the cloud to your app.
When you ‘delete’ a device, it is the device model that you delete, both from the cloud-based learning algorithms and from the monitor itself. Going forward, Sense will no longer recognize that specific device pattern in your home (though, it may re-detect it…more on that shortly). However, we do not permanently remove all of the historic data; we freeze the device stats and archive them. This data doesn’t get re-incorporated in the learning process, but we save it so you won’t lose your historic stats in the Timeline and Trends.
Why, then, might Sense re-detect a device after you’ve deleted it? Sometimes you want this, in the case of a model not perfectly detecting the on/off events of a device, but in other cases, you might not.
Let’s say a week after you deleted your toaster, Sense pops ups with “Sense found a new device and named it Toaster.”
When our automated device detection runs overnight, it attempts to improve existing models, then it searches for unexplained patterns in your signal history. If it finds a pattern, it makes a new model. In the toaster example, there is some distinct pattern in your signals that Sense keeps identifying. The plot above shows two prominent patterns, a repeated high frequency spike, and a large multi-stage load with an odd ramp at the end.
While you may not find that this detection accurately reflects reality in your home, Sense is continually finding something significant in the pattern. The best option here might not be to delete at all, but to let Sense continue to gather more data. If you don’t have a device at all like what Sense is suggesting, maybe Sense just got the name wrong? Try setting up some custom notifications and doing a little investigative work in your home.
We won’t always find the same device again after you delete it. With only a small amount of data, noise may look like a statistically significant pattern and Sense will create a model as soon as it thinks it found a device. But later, with the benefit of a larger dataset, Sense may be able to clearly identify it as noise and thus will not trigger an (inaccurate) detection. In this case, after you delete the device, the learning algorithm will ignore the noise. We leave it up to you to decide if a model is bad and should be deleted. The only time we will proactively delete a device is if the model is interfering with a different good device.
When good devices go bad
What about in cases where your device triggering correctly but has gotten a bit flaky lately? There’s a couple possibilities there.
Sense may initially discover two modes of an appliance as separate devices, say the toaster on “dark” vs “light” mode, or the different modes of the clothes washer motor plotted above. With only a few examples of an appliance, it’s hard to extrapolate the complete pattern. As we record more data and improve the detection model, the two clusters of patterns converge and so do the two models. They become duplicates or near duplicates. As a user, you may experience this in many ways: One device may no longer trigger at all. One device may trigger some of the time and the other the rest. One device may flicker on briefly, then the other device turn on. In cases of exact duplicates, if you delete one of these devices, the pattern remains explained by the other device (so feel free to do so!). If you’re experiencing “misfires” where Sense doesn’t consistently recognize the device as turning on, it should improve as it acquires more data.
Another possibility: As the learning algorithm incrementally improves a model, it may get stuck in a rut where small changes to the model no longer improve it (what we call a “local maxima” in computer science). The model might only match part of the pattern. Or it might trigger on another device too, but Sense can’t determine which to split out. In cases like this, it can be wise to delete the device. By deleting the device, you’ll start device detection with a different set of initial examples and may net a better model.
If we keep rediscovering a device that you are sure isn’t in your house, help us out. Something is making a distinct pattern in your electrical usage, even if it isn’t what we say it is. Hunt down the appliance responsible for the pattern (the Power Meter and aforementioned notifications can be a huge help here) and update the device name. That will feed into our data science research and into the Community Names neural network, helping all Sense users.