French embedded AI startup Cartesiam has developed a tool to add machine learning classification to small microcontrollers which is being used in a sensor board by Bosch.
“Edge AI is huge, it’s like a tsunami,” said Joël Rubino, CEO and co-founder at Cartesiam talking to eeNews Europe . “We started in 2016 with this idea that there are billions of machines that are generating data and we need to listen to that data with a different approach, not to stream it to the cloud but to focus on the edge.”
The first version of the NanoEdge AI Studio tool, launched in February 2020, allows anomaly detection on ARM Cortex-M 32bit microcontrollers such as those from STMicroelectronics, Microchip and Renesas. The tool provides algorithm selection, optimisation and quantisation to create the library in around 4Kbytes that flags an anomaly in a signal. This is aimed at preventative maintenance.
“The technology giants in the cloud busines are all interested in the edge, not to run at the edge but to capture data from the edge,” said Rubino. “We started with the idea that our playground is the microcontroller, so we hired PhDs in machine learning and signal processing to rewrite the ML and signal processing algorithms to run in a microcontroller. It was a lot of hard work to test that in real environments, but V1 was mathematically correct and industrially proven,” he said.
Version two now allows a classification library to be built to identify the source of the anomaly from training data. This is built on a PC without using the cloud, and incorporated as a library in firmware in an edge AI design. This allows the classification library only to be called when an anomaly is detected for lower power operation. It can also be updated separately from the anomaly engine.
The company is to add the AI tool to an eight IoT