MIT's "smart power outlets" use machine learning to prevent circuits tripping unnecessarily

Arc-fault detectors can be invaluable protectors against electrical fires, but they can also be a nuisance if they start switching off devices and appliances when there's no safety risk. To cut down on the false alarms, an MIT team of engineers is working on new "smart outlets" that use neural networks to not only determine if a detected arc is dangerous, but also what type of devices are plugged into the outlet.

Electricity is like fire. It's an incredibly valuable servant, but it can also be an extremely dangerous enemy. From the earliest days of electrical engineering, preventing overloads, short circuits, and arcs was a top priority. One of the most basic protections was the fuse, which was a bit of wire that was designed to melt and break the circuit in the event of an overload.

Today, we have much more sophisticated versions in the form of circuit breakers and surge protectors designed to defend delicate digital devices from electrical spikes. Modern arc fault circuit interrupters (AFCI) perform a similar service. Even the best designed and built electrical systems undergo wear and tear or might be subjected to conditions that could generate arcs that could start fires or damage equipment. The job of the arc-fault detector is to, as the name implies, detect these arcs and shut off the power before they become dangerous.

The problem is that not all of these arcs are dangerous. Some are harmless or the device hooked up to an outlet might be something like an electric iron that wouldn't be bothered by a particular arc that would be disastrous for a computer. Unfortunately, modern detectors err on the side of caution and often indulge in what MIT calls "nuisance trips" that serve no real purpose.

The MIT approach is to make the AFCI a bit more intelligent. Current devices rely on microprocessors that use an algorithm to identify very simple, generic arc signatures in the electric current. What the MIT team wanted was something more intelligent that could differentiate dangerous signatures from benign ones.

They did this using a Raspberry Pi Model 3 microcomputer to monitor the incoming electrical data through an inductive current clamp installed around the outlet wire to sense the current. The clamp runs what it senses through a standard USB sound card with an integrated memory buffer, which is capable of handling very small signals at very high rates of 48 kHz, or 48,000 times a second.

As the data comes into the microcomputer, a machine-learning neural network analyzes it in real time. This network, at first, was trained to recognize the electrical patterns of an iMac computer, a stovetop burner, and an ozone generator – the latter electrically charges oxygen in the air, which can produce signals similar to those of a dangerous arc-fault.

Over time, the network is able to learn as more data is analyzed, so it is increasingly capable of accurately telling which signals to act on and which to ignore, as well as being able to tell which devices are plugged into it. The idea is that one day such smart outlets will operate in many people's homes to protect electronic devices, while a smartphone apps would allow consumers to analyze and share data about their electrical usage on a very detailed level. By anonymously sharing this data, the outlets will be able to further refine their algorithms.

"By making [the internet of things] capable of learning, you're able to constantly update the system, so that your vacuum cleaner may trigger the circuit breaker once or twice the first week, but it'll get smarter over time," says Joshua Siegel, a research scientist in MIT's Department of Mechanical Engineering. "By the time that you have 1,000 or 10,000 users contributing to the model, very few people will experience these nuisance trips because there's so much data aggregated from so many different houses."

So far, the smart outlet has shown a 95.61 percent accuracy in telling one device from another and 99.95 percent accuracy in telling good from bad signals, which is a little better than that of current AFCIs. In addition, the outlet reacted to a bad signal in less than 250 milliseconds.

Along with protection against arcs, the MIT team regards the smart outlet as an example of a future where appliances and other devices are intelligent, with the ability to diagnose themselves and anticipate the needs of their users.

"This is all shifting intelligence to the edge, as opposed to on a server or a data center or a desktop computer," says Siegel. "I think the larger goal is to have everything connected, all of the time, for a smarter, more interconnected world. That's the vision I want to see."

The research was published in Engineering Applications of Artificial Intelligence.

Source: MIT