Edge Impulse Imagine 2022 Day 1: ML on the Edge — An Industrial Perspective with Hubertus Breier

Thank you, Cindy. Also a warm welcome from my side. I have the honor today to talk about an industrial perspective in machine learning on the edge. I'm head of technology at Lapp. We are 3600 colleagues working in industrial automation, sensors, network devices, and communication devices. We can make use of machine learning in many applications. For example, automotive robots with vibration sensors measuring joint angles, supporting battery assembly. We also measure variable pitch angle of wind turbine blades, and in semiconductor industry measuring wafer edges. Three challenges: solve real-world problems, decide on-premise vs cloud vs edge, and create value propositions. For the milling example, we detect cutting wear of tools using vibration sensors. Raw data shows little difference to the naked eye, but machine learning can classify with 99% accuracy on the sensor. On edge or cloud, accuracy reduces due to data transport. For an automotive elevator, we detect anomalies in vibration patterns to predict failures, preventing costly production line downtime. On-premise vs cloud: On the sensor, full raw data available, high accuracy but expensive sensor. On edge, less data, lower accuracy. On cloud, full computing power but latency. Value propositions: Start with business model canvas, understand customer's problem. Wind energy example: Preventing unplanned downtime is extremely valuable, use anomaly detection on sensor data in cloud. For automotive elevator, edge gateway for early warnings. For milling, real-time decision making on sensor is wiser. Summary: Not one size fits all. Need to solve real-world problems with domain know-how, decide where to run ML, and create value propositions that count. Know your problem, have great partners, build your industrial grade solutions. Thank you.