Achieving the Elusive Glass Break Detector Performance Trifecta

Share

Wireless glass break sensors have only one job: to detect that a glass break has occurred and transmit an alarm.

The ideal glass break detector will:

  • Never miss a window glass break (have 100% True Positive Rate or TPR)
  • Never trigger an alarm unless a window glass break has occurred
    (have 0% False Alarm Rate or FAR)
  • Have an extended multi-year battery life so the homeowner does not need to change the batteries frequently.

Unfortunately, using the currently existing technology, these three parameters cannot be simultaneously achieved within the same sensor, so most glass break detectors perform well at only two of the three and are mediocre at the third.

The reasons for these performance compromises are relatively straightforward. Traditional glass break detection solutions utilize a specific set of rules to define the sound of glass break. One typical rule is to separate the glass break event into two phases which include a ‘thud’, indicative of an object hitting glass, and a ‘shatter’, indicative of the glass fracturing.  By analyzing the timing, frequency content, and amplitude of these two phases, a quick computation can be made to decide if an audio event has these characteristics.   When a sound meets all the predefined rules, it is assumed that a glass break has occurred.  Since there are only a finite number of rules which can be generated and checked, the power consumption required for these rule-based solutions is quite low and they can achieve a long battery life.

The downside to rule-based detection is that many common household sounds (disturbers) such as keys dropping, dogs barking or even a book dropping on the floor can have the same frequency, energy level, and timing attributes as glass breaking, resulting in an erroneous detection. According to Parks Associates, 62% of security system owners report that their system triggered a false alarm within the past 12 months and about two out of three security system owners have paid average fines of $150 for unnecessary emergency responses due to false alarms, all contributing to user dissatisfaction with their alarm systems that are supposed to bring them peace of mind.1

While It is possible to adjust the rules to help minimize the false alarms, reducing the false alarm sensitivity runs the risk of missing an actual glass break. So in general, rule-based glass break sensors trade off detection accuracy but achieve good battery life – not a compromise that homeowners want to make.

With the more recent introduction of tinyML − small, low power processors that can run sophisticated machine learning on local devices instead of in the cloud − these simple rules can be replaced by trainable neural networks which are capable of much better discrimination between target events and other events.  The obvious benefit of tinyML-based detectors is that they should have fewer false alarms while maintaining accuracy at detecting actual glass breaks. However, the improved accuracy comes at the expense of battery life because digital-based tinyML processors consume 5-6x more power than processors used in a rule-based detector. The result is a sensor with better accuracy and shorter battery life – still a losing compromise for the homeowner who is now stuck changing the batteries frequently.

Fortunately, the AML100 analog machine learning processor overcomes these challenges with an innovative new technology that enables next generation glass break detectors to achieve both high detection accuracy and a long battery life at the same time.

Aspinity’s glassbreak solution offers a system that can achieve the required TPR, FAR, and battery life simultaneously.  At the heart of Aspinity’s glassbreak solution is the AML100, an innovative analog signal processing and ML core which enables feature extraction and machine learning to be performed in the ultra-low power analog domain.  The AML100 determines when a glass break has occurred using near-zero power machine learning and accurately wakes the digital processor only when glass break has occurred.

The AML100 based glass break detection solution minimizes the amount of time that the digital processor is on, allowing for a higher performance digital processor to confirm the glass break without impacting battery life. The result is a glass break solution that unlike rule-based or tinyML based solutions, can finally achieve the high TPR, low FAR, and long battery life for homeowners – the glass break detector performance trifecta.

1 Parks Associate Whitepaper “Solving False Alarms: Bringing New Context for Monitoring”