Transforming intrusion alarm systems with power-intelligent analog computing architecture

Sponsored content

Most choices in life have tradeoffs. If you want to renovate your house, you might find a superb general contractor who is expensive or a meh general contractor who’s cheap, but how often can you find a general contractor who does both excellent work and who doesn’t charge a small fortune?

It’s much the same with today’s battery-powered intrusion-detection systems, including glassbreak sensors. Unfortunately, long battery life and excellent detection accuracy have always been inversely proportional. You can design a glassbreak sensor that has excellent battery life but is likely to trigger a false alarm on a sound that’s not glassbreak. But no one is a fan of false alarms: They diminish brand satisfaction and can prove expensive if law enforcement charges a penalty in response to false alarms. The alternative is to design a sensor with minimal false alarms but poor battery life—an option that’s equally unappealing as low-battery alerts always seem to happen when it’s most inconvenient.

Fortunately, a new more power-intelligent approach that leverages analog machine learning (analogML) is now available, freeing designers from having to compromise between these two important requirements.

How does analog computing deliver all the benefits of machine learning in ultra-low-power analog? And what do you need to know about this new approach to intrusion-detection before designing your next glassbreak sensor?

Traditional glassbreak sensors

Most of today’s wireless glassbreak sensors use a sensing methodology on which we’ve relied for decades: The microphone collects all ambient sound data, which are then digitized and analyzed by a microcontroller (MCU) or digital signal processor (DSP).

These MCUs and DSPs are programmed to interpret a loud low-frequency “thud” followed by a loud higher-frequency “shatter” as a glassbreak, and when these two sounds are heard in the proper order, this triggers an alarm. While this combination of amplitude and frequencies might represent a glassbreak, other common household sounds are similar enough to result in false alarms. A book dropping on the floor, for example, produces an initial low frequency “hit” sound followed by a high-frequency smack, which consistently false-triggers these sensors. But that’s just one of many similar household sounds that can fool the sensor. Dog barks and hands clapping are just as likely to set off the glassbreak sensor. Compound this type of indiscriminate data analysis with duty cycling—a technique that turns components on and off to save power, and which can lead to missing an event that occurs randomly and sporadically—and you can see why there are accuracy issues with the vast majority of today’s glassbreak sensors.

Does it really matter if these inaccurate glassbreak sensors can last up to five years without a battery change if the homeowner can’t rely on them?

Newer options with tiny machine learning (tinyML) chips

There’s a ton of industry buzz about tinyML chips—and that’s for good reason. These chips support the movement of sophisticated processing and data analysis from the cloud to the device edge, supporting an explosion of smart, always-on connected devices with embedded machine learning. Designers can use tinyML chips in countless always-on sensing applications in the home, including those that are always listening for voice or other audio sounds.

Newer intrusion-alarm systems, particularly those designed for the homeowner to install, rely on tinyML chips for more discriminating detection of events such as glassbreak. The good news is that the machine learning incorporated into these chips delivers improved detection-accuracy. The bad news is that these chips are still clocked processors that operate within the digital domain and require the digitization of all data prior to analysis. Because of their architecture, these systems continuously consume >3mW of power digitizing all incoming sound data, although in reality, a glassbreak hardly ever happens. And that’s a design problem because we estimate <<1% of the sound data digitized by these systems represent an actual glassbreak—which translates to an enormous waste of power resources.

While traditional tinyML chips offer substantially better accuracy, the homeowner has to trade this high accuracy for a reduced battery life of just one to three years.

AnalogML, the best of both worlds  

Extending battery life while maintaining performance requires us to intelligently reduce the volume of data that’s digitized, so the digital system only processes relevant data. The only way to achieve this is to analyze the content of the data while they are still in their natural analog state at the start of the signal chain. We call this approach “analog machine learning” or analogML.

AnalogML enables a transformative power-intelligent approach to always-on sensing that makes it ideal for detecting rarely occurring events such as glassbreak. By pulling the first level of sophisticated machine-learning analysis into the low-power analog domain, you can determine whether a glassbreak has occurred prior to digitization, using near-zero power. The analogML architecture further reduces always-on system power by keeping the digital system off unless a glassbreak event is actually detected. The result is a glassbreak system that has an always-on system power of a mere 125µW (analog microphone + analogML chip), a 97% reduction from a tinyML solution that operates within the standard digital paradigm. And because the software algorithms in the analogML solution are trained and programmed specifically to detect glassbreak, you don’t need to worry about false alarms from other household sounds.

Altogether, analogML provides the best of both words—minimal false alarms and a battery life of five years or longer. And because it’s a flexible, programmable analog technology, you can use the same analogML chip to improve the accuracy and battery life of other always-on applications in and beyond the home.

Tom Doyle is founder and CEO of Aspinity, the pioneer in the design and development of analog processing chips that are revolutionizing the power- and data-efficiency of always-on sensing architectures. Tom brings over 30 years of experience in operational excellence and executive leadership in analog and mixed-signal semiconductor technology to Aspinity. Prior to Aspinity, he was group director of Cadence Design Systems’ analog and mixed-signal IC business unit, where he managed the deployment of the company’s technology to the world’s foremost semiconductor companies. Previously, he was founder and president of the analog/mixed-signal software firm, Paragon IC solutions. Tom holds a B.S. in Electrical Engineering from West Virginia University and an MBA from California State University. Connect with Tom on LinkedIn

Editor's Note: For more information on the analogML approach to glassbreak detection, register for this year’s Sensors Converge (June 27-29, 2022 in San Jose, Calif.) and attend the technical session, Smarter Homes—ML-Based Intrusion Detection that Won't Compromise Battery Life, June 29 at 2:00 p.m. or visit Aspinity in booth #933