Experimenting with Vibration Sensors Challenge Summary
It has been another fun challenge. I was not able to achieve what I had proposed (TinyML model to identify people by their footstep vibrations), but as always I learned a lot working on the project. And I've developed some hardware and software tools that will be valuable for future projects.
The Kemet/Tokin VS-BV203 Piezo Vibration Sensor is ultra compact (8.4 x 11.4 x 2.9 mm) and the 3 meter miniature flexible cable makes it easy to attach to instrumentation at a reasonable measurement distance. It has a sensitivity of 10 m/s^2 and a buffered voltage output that centers the output at 1/2 the power supply voltage which makes it easy to interface to an ADC.
The sensor was probably not a good choice for my application (or at least my implementation). I think a more traditional seismic sensor with mechanical amplification might have worked better. From papers that I've been able to find on footstep detection/classification, it seems that the problem is best solved with an array of sensors rather than a single sensor. And in some cases audio sensors are utilized rather than seismic sensors to detect the footsteps. In all cases, a lot of data is required (a couple orders of magnitude greater than what I had proposed). I think a viable solution would require automatic data capture and automatic data labeling to deal with the large volume of data. A possible implementation could use a camera and image recognition for triggering and labeling the data capture.
I think a better TinyML application for the Kemet sensor that I'd like to try is monitoring my HVAC system. I've always wanted to install a smart, network connected, thermostat. I could use information from the thermostat to trigger and label data captured from the sensor mounted on the HVAC unit. Then, I could try to use TinyML for Predictive Maintenance. Something to think about.....
Thanks to Element14 for the challenge opportunity.