Outlets with basic AI can “learn” to identify appliances and dangerous electrical spikes.
We often think of it as “blowing a fuse” when we turn on an appliance and the power to a whole room goes off, but much of the time interruptions like this are “nuisance trips.” This is when the detector installed behind the wall senses something that could potentially be a dangerous arc-fault in the electrical line, and trips an outlet’s electrical circuit.
According to a team of engineers at MIT, many of today’s arc-fault detectors often err on the side of caution and shut off an outlet’s power in response to electrical signals that are harmless.
This belief led the team to develop a solution that they are calling a “smart power outlet.” Their device analyzes electrical current usage from a single outlet, or multiple outlets, and is capable of distinguishing between benign arcs — harmless electrical spikes such as those caused by common household appliances — and dangerous arcs, such as the sparking of faulty wiring that could lead to a fire. What really makes these outlets smart is that the device can be trained to identify what’s plugged into an outlet, such as a fan versus a computer.
Custom hardware processes electrical current data in real-time, and software analyzes the data via a neural network of machine learning algorithms that mimics a human brain.
The smart outlet’s machine-learning algorithm determines whether a signal is harmful or benign by comparing a captured signal to signals that the system was trained on. The larger the amount of data used to train the network, the more accurately it can learn to identify the “fingerprints” used to differentiate good signals from bad ones—and even to distinguish appliances from each other.
The smart power outlet can operate as part of the Internet of Things (IoT) as it is able to wirelessly connect to other devices, said Joshua Siegel, a research scientist in MIT’s Department of Mechanical Engineering. Siegel pictures a network where users install smart power outlets in their homes, and an app on their phone, which they can use to analyze and share data on their electrical usage, such as where and what appliances are plugged in, when an outlet has tripped, and the reason why a trip occurred. This data would be shared securely and anonymously with the research team to help them refine their machine-learning algorithm, making it easier to identify appliances and distinguish dangerous events from benign ones.
“By making IoT 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,” Siegel said. “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.”
Electrical Fingerprints
Modern homes use arc fault circuit interrupters (AFCI) to help reduce the risk of an electrical fire. AFCIs interrupt a faulty circuit when it senses certain potentially dangerous electrical patterns.
“All the AFCI models we took apart had little microprocessors in them, and they were running a regular algorithm that looked for fairly primitive, simple signatures of an arc,” Pratt says.
A more discerning detector could discriminate between many signals to tell a benign electrical pattern from a potentially harmful pattern, so Pratt and Siegel set out to design one.
A Raspberry Pi Model 3 microcomputer records incoming electrical current data, and an inductive current clamp that fixes around an outlet’s wire without actually touching it senses the passing current as a changing magnetic field.
The team connected a USB sound card between the current clamp and the microcomputer which they used to read the incoming current data. The team found these sound cards, which are commodity hardware like that found in conventional computers, are ideally suited to capturing the type of data that is produced by electronic circuits, because they are designed to pick up very small signals at high data rates, which is similar to what an electrical wire gives off.
The sound card offers other advantages, including a built-in analog-to-digital converter which samples signals at 48 kilohertz, which means that it takes measurements 48,000 times a second. The card also has an integrated memory buffer, which enables the device to continuously monitor electrical activity in real-time.
Much of the microcomputer’s processing power is devoted to running the neural network, in addition to recording the incoming data. The team trained their network to establish “definitions”—that is, recognizing the associated electrical patterns—produced by four devices: a fan, an iMac computer, a stovetop burner and an ozone generator type air purifier that produces ozone by electrically charging oxygen in the air, which can produce a reaction similar to a dangerous arc-fault.
Each device ran multiple times under a range of conditions, and the gathered data was then fed into the neural network.
“We create fingerprints of current data, and we’re labeling them as good or bad, or what individual device they are,” Siegel said. “There are the good fingerprints, and then the fingerprints of the things that burn your house down. Our job in the near-term is to figure out what’s going to burn down your house and what won’t, and in the long-term, figure out exactly what’s plugged in where.”
Shifting Intelligence
Once the neural network was trained, the team ran the whole system — hardware and software — on new data from the same four devices. They found that the network was 95.61 percent accurate in discerning between the four types of devices (for example, the fan versus the computer). The system achieved 99.95 percent accuracy when identifying good from bad signals — slightly higher than existing AFCIs. The system also matched the performance of current certified arc detectors when reacting to trip a circuit, taking less than 250 milliseconds.
The smart power outlet will get more intelligent with increasing data, and Siegel envisions running a version of the neural network over the Internet with other users able to connect and report on their electrical usage. This will provide an increasing amount of data to the network and help it learn new definitions and associate new electrical patterns with new appliances and devices. These new definitions would then be shared wirelessly to users’ outlets, improving their performance, and reducing the risk of nuisance trips without compromising safety.
“The challenge is, if we’re trying to detect a million different devices that get plugged in, you have to incentivize people to share that information with you,” Siegel says. “But there are enough people like us who will see this device and install it in their house and will want to train it.”
Beyond electrical outlets, Siegel sees the team’s results as a proof of concept for “pervasive intelligence,” and a world made up of everyday devices and appliances that are intelligent, self-diagnostic, and responsive to people’s needs.
“This is all shifting intelligence to the edge, as opposed to on a server or a data center or a desktop computer,” Siegel says. “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.”
Siegel and his colleagues have published their results in the journal Engineering Applications of Artificial Intelligence. His co-authors are Shane Pratt, Yongbin Sun, and Sanjay Sarma, the Fred Fort Flowers and Daniel Fort Flowers Professor of Mechanical Engineering and vice president of open learning at MIT.
To learn more, check out Three Innovations in Smart Home Technology.