IoT AI specialist SensiML has partnered industrial automation developer onsemi to create a machine learning solution for autonomous sensor data processing and predictive modeling. The collaboration combines SensiML’s Analytics Toolkit development software with onsemi’s RSL10 Sensor Development Kit. The two companies are aiming to create a platform for edge sensing applications such as industrial process control and monitoring.

SensiML’s technology supports AI functions in a small memory footprint. Combined with onsemi’s RSL10 sensing and Bluetooth Low Energy connectivity, the joint solution allows smart sensing without the need for cloud analytics of highly dynamic raw sensor data.

Developers using the RSL10-based platform and the SensiML software together can add low latency local AI predictive algorithms to their industrial wearables, robotics, process control, or predictive maintenance applications regardless of their expertise in data science and AI. The resulting auto-generated code enables smart sensing embedded endpoints that transform raw sensor data into insight events right where they occur and can take appropriate action in real time. The smart endpoints reduce network traffic by communicating data only when it offers valuable insight.

“Cloud-based analytics add unwanted, non-deterministic latency, and are too slow, too remote and too unreliable for critical industrial processes,” said Dave Priscak, vice president of Applications Engineering at onsemi. “The difference between analysing a key event with local machine learning versus remote cloud learning can equate to production staying online, equipment not incurring expensive downtime, and personnel remaining safe and productive.”

“Other AutoML solutions for the edge rely only on neural network classification models with only rudimentary AutoML provisions, yielding suboptimal code for a given application,” said Chris Rogers, SensiML’s CEO. “Our comprehensive AutoML model search includes not only neural networks, but also an array of classic machine learning algorithms, as well as segmenters, feature selection, and digital signal conditioning transforms to provide the most compact model to meet an application’s performance need.”