Applying edge artificial intelligence to closed-loop real time control and monitoring of laser based battery pack welding

The manufacturing of battery packs plays an essential role in e-mobility. In this context, laser welding of batteries has taken a considerable momentum in recent years as it can quickly adapt to different form factors and battery arrangements which is crucial for short runs and on-demand manufacturing. In this paper, we present an assisted battery pack welding system applying a novel sensor fusion strategy. The presented work introduces a multi-camera solution combining a camera in the visual range and an infrared camera in a compact scanner head that applies graphics processing unit acceleration for image processing and real-time field programmable gate array processing for process control. In order to acquire accurate temperature measurements, the infrared camera has been characterized and calibrated for the temperature ranges of the welding process. In addition, we propose a method for image space coordinate transformation based on affine transformations used to transform the coordinates of the detected battery joints into different image spaces to adapt to the welding process in real time. A single shot multibox detector Mobilenet-v2 neural network has been retrained with custom images to detect and classify a set of battery pack types allowing the extraction of the battery cell joint coordinates using image processing, which enables the application of real-time corrections to the laser aiming system and a closed-loop laser power control. Finally, we present some battery welding results using different welding strategies on the described system.