Body odor is a stubborn problem. Not just for people, but also for sensors. Sensors and the computing attached to them struggle to perceive armpit odors in the way humans do, because B.O. is really a complex mix of dozens of gaseous chemicals.
The U.K.’s PlasticArmPit project is designing the first machine learning-enabled flexible plastic sensor chip. Its target audience: those who think they might stink. The prototype chip will be manufactured and tested in 2019.
The project is part of a broader effort Arm has been involved in to drive the cost of plastic IoT devices down below US $0.01 so that they can be embedded in all sorts of consumer goods, including disposable ones.
The plastic chip will consist of sensor arrays, a bespoke machine learning processor, and their interface. All of these will be built on a thin plastic film, although the sensor array is currently made on a separate piece of plastic. A battery and display may be integrated into the device later.
The team consists of: Arm, which designed the machine learning circuitry and developed tools that will make it easy for others to produce such designs; PragmatIC, which makes amorphous-oxide-based flexible electronics, NFC, and RFID chips and the systems to build them; the University of Manchester, which developed the plastic gas-sensing technology and a model of human smell perception; and Unilever, which lent its consumer-products expertise and its UK odor-testing lab.
The sensor arrays are a collection of field-effect transistors made from organic semiconductors that have been chemically modified.
“We tune the materials within the device to be responsive to different gaseous analytes,” explains Krishna Persaud, who developed them at the University of Manchester along with his colleague Michael Turner.
As gases bind to the transistor’s semiconductor channels, they alter the characteristics of the device’s performance.
Each device has eight different types of sensors. But they’re not specific to one chemical each. Instead, each responds to a spectrum of molecules within a class of chemicals, says Persaud. It’s their collective response, interpreted by the machine learning system, that indicates whether your armpit odor is so strong it’s wearing your shirt for you.
It took a different kind of machine learning system to make that interpretation. “We had to forget everything we knew about machine learning,” says James Myers, senior principal research engineer at Arm.
Deep learning gets most of the machine learning-world’s attention these days, but it’s not yet well suited to a flexible electronics system. Although plastic electronics have the potential to be much cheaper than silicon, they are limited in the number of transistors they can contain, because the transistors are much larger.
With fewer than 1,000 logic gates to work with; Arm needed to find a seriously cut-down classifier circuit. Arm researchers turned to a type of machine learning based on naïve-Bayes classifiers.
That type of classifier isn’t trendy these days because it’s not as good at interpreting images as deep learning, explains Myers. Despite being much simpler than deep learning, the naïve-Bayes classifier required “a lot of work” to make it fit within the limited number of gates, he adds.
Bringing B.O. (instead of something less awful) to the project was Unilever’s idea. But the technology could also be used for food freshness and perhaps help limit food waste. 1.3 billion tons of food was lost or wasted in 2016, according to the Food and Agriculture Organization of the United Nations, including 40 to 50 percent of all root crops, fruits, and vegetables, and 35 percent of fish.
Rather than a “best before” date, disposable plastic sensors and machine learning circuits in food packaging could take the guesswork out of shopping and stocking grocery stores, potentially leading to less waste.
Sensing freshness is likely a more difficult problem than telling the difference between a merely malodorous armpit and chemical-weapons-grade B.O.
“Bad fish smells different from bad burgers and from bad milk,” points out Emre Ozer, principal research engineer at Arm.
And with such limited circuitry, they’d need a different design for each nose-wrinkling application.
Those designs will need to be easy to devise, if flexible circuits and sensors are to be cheap enough to throw away. But that hasn’t been the case, so far. Myers likens the state of flexible chip design to silicon chip design in the 1970s. In other words, it’s slow and requires a lot of expertise.
So the Arm team has developed design automation tools that turn the machine learning algorithm, written as Python code, into a circuit design you can send straight to PragmatIC’s flexible-chip foundry. “It’s a hardware generator that will spit out an app-specific classifier,” Myers says.
Using the new design tools should make it easy for engineers without any particular machine learning experience to make new classifier circuits.
If you want to sense the freshness of apples instead of tomatoes, “you just push the button again,” says Myers.
(Such push-button design is one of the main goals of Darpa’s US $1.5-billion Electronics Resurgence Initiative. The idea there is to produce design tools that can design a whole system in 24 hours starting from just a high-level description of what the system has to do.)
PragmatIC, the Cambridge-based startup doing the flexible IC manufacturing, is trying to make the process as familiar to silicon IC engineers as possible.
“We don’t want to be novel where we don’t have to be novel,” says Catherine Ramsdale, the company’s vice president of device engineering.
They already have a production line in Sedgefield, County Durham, that uses mostly standard chip-making equipment to move glass wafers to which a flexible plastic film has been applied. The circuits, which are based on amorphous oxide thin-film transistors, are built on the plastic, diced into plastic “FlexICs,” and peeled off the glass, which then goes back into the system with a fresh plastic sheet. The Sedgefield facility is capable of making a billion FlexICs per year, according to the company.
The 60-person company has developed a “fab in a box,” says Ramsdale. It’s a self-contained cleanroom housing a complete FlexIC production line.
Ramsdale predicts that the cost of the system will be low enough that it can be located at the end user’s production line. If it all works out, you might one day buy a shirt with two plastic stench detectors that were designed, constructed, and attached at the apparel-maker’s site.
Source: IEEE Spectrum