Washing machine faliture

How High Resolution Energy Monitoring Informs Fault Detection

Researchers think that it should be possible to use data science to detect when a device is failing to work properly. We can already use data science to detect a few device failures. Read on to learn how we do it.

The data science team at Sense is focused on making appliance detection as accurate and speedy as possible, but it is also important that Sense be able to detect when a device isn’t working the way it should . Today I want to discuss anomaly detection: the ability to diagnose that something in the operation of a device is not as it should be. Such a capability can be useful in various ways–it can guide the user to change the way an appliance is used, or to have it repaired or even replaced. These interventions may, in turn, lead to energy or cost savings, and could even avert a house fire or similar disaster.

Theoretical Examples of Anomaly Detection

Anomaly detection is an important aspect of many engineering disciplines, and various signatures–including acoustic, electromagnetic, chemical and other types–are widely utilized to assess the health of a wide variety of devices. Within the field that Sense is a part of (energy disaggregation, or “non-intrusive load monitoring”) a few recent papers show an increasing awareness of this potential application.

In the paper “Exploring The Value of Energy Disaggregation Through Actionable Feedback” (Batra, Singh and Whitehouse, presented at the NILM 2016 workshop,) the authors focus on two significant causes of energy inefficiency. First if a crucial sensor in a refrigerator malfunctions, the appliance will defrost far too frequently–thus wasting energy both to perform the defrosting function and to cool everything down again. Second, an inappropriate schedule for an HVAC can lead to large energy wastage with no benefit to user comfort. Batra, Singh and Whitehouse show that both of those anomalies in principle can be detected using typical disaggregation approaches. However, when they applied state-of-the-art methods to a standard database of appliance signatures, the researchers were  not able to demonstrate that this could  be done in practice. Yet, as I’ll discuss below, we believe that Sense is much more likely to succeed in this task.

Sense Discovers an Anomaly

Here I want to show a third type of malfunction that we have noticed during the development of the Sense algorithms. Firstly, consider the trace of a cycle of a gas furnace shown below. (As usual, this trace shows the wattage consumed–vertical axis–as a function of time elapsed on the horizontal axis; the total time span is about 8 minutes.)

We clearly see the three main elements of the furnace in operation: firstly, the inducer motor turns on, followed after a short delay by the spark igniter (which ignites the gas that is used as heat source.) The spark igniter cycles rapidly, so on this resolution it appears as continuous “band” of energy. After the spark igniter turns off, there is a slightly longer pause before the blower motor turns on, circulating the heat through the system. Finally, the inducer motor turns off, and then the blower motor.

How a furnace works

In houses with typical gas furnaces, this cycle repeats continuously throughout the winter months. However, we sometimes observe something like the trace below:

Broken Furnace

Instead of “firing” once during the furnace cycle, the igniter runs three times before the blower motor eventually starts up. It turns out that this a common, and potentially dangerous, failure mode for gas furnaces: the igniter is not fully functional, and therefore it sometimes fails to light the gas stream. If the igniter does not light the gas stream, then hopefully a safety function in the furnace will cut off the gas, but in a worse case scenario, this malfunction may result in a gas leak.

In this case, ignition eventually succeeds, and the cycle completes, but frequent occurrences of such “misfires” are an important indication that the igniter requires attention.

Sense’s High Frequency Measurements Make Anomaly Detection Possible

As an aside, these traces are another example of a point we have made previously, namely that high-frequency measurements are crucial for the type of real-time, high accuracy disaggregation that Sense performs. At lower time resolutions it becomes much harder, or even impossible, to understand the appliance activity in the  detail needed to reach this type of conclusion. With our measurements, we are confident that we will be able to detect energy-wasting issues with refrigerator defrost cycles and HVAC temperature schedules, which is hard to do with lower-frequency data sets (as Batra, Singh and Whitehouse demonstrated).

Finally, if you own a Sense monitor and experience a failure of one of your electrical appliances, please drop us an email at feedback@sense.com! We are always trying to understand the details contained in our data, and it could be that your failing device created an electrical signature that we can use to warn you, or other Sense customers, of likely future failures.