nRF52 RSSI to Distance Mapping and Machine Learning

[kmassa5]

Hi,

As has been explained numerous times in posts on this forum and many others, mapping RSSI to a distance measurement is notoriously difficult and inaccurate. I was hoping to implement a machine learning algorithm (such as a classifier) trained with a set of data collected in the specific environment the system will be used in. The system is 2+ nRF52832's picking up each other's advertisements and mapping the RSSI values to distances using the ML classifier.

My question is, will this actually work as implemented using the boards' MCU's alone (does the nRF52832 have enough processing power, will it draw too much energy, etc...)? Will I need to send information to a central processing unit to do the energy intensive calculations? Or do you recommend we ditch the ML altogether and go for a more classical algorithmic approach?

Thanks!

[Karl Ylvisaker]

Hello,

My question is, will this actually work as implemented using the boards' MCU's alone (does the nRF52832 have enough processing power, will it draw too much energy, etc...)? Will I need to send information to a central processing unit to do the energy intensive calculations? Or do you recommend we ditch the ML altogether and go for a more classical algorithmic approach?

This is a very broad question that is hard to answer, since it will depend heavily on which models and approaches you will be using - "ML classifier" is a very general term.
Which type of classifier are you thinking about, and how do you plan to do the training?
I think this is a very interesting topic, which definitely has the potential to yield some interesting results - however, I am not sure that the RSSI measurement alone will be enough for the classifier to learn from, it is hard to say.

What I however can say is that there has been an increased effort recently in the ML space to use microcontrollers, such as the TensorFlow Lite library. This development definitely holds potential for some very cool new applications of BLE and microcontrollers in general.
If you are looking into using TensorFlow Lite I highly recommend checking out the blog post and github repository linked by my colleague in this ticket. Please bear in mind that this has only been tested for the nRF52840, nRF9160 and nRF53, so it has not been tried on the nRF52832.

Best regards,
Karl