Sensor Selection and Defect Classification via Machine Learning During the Laser Welding of Busbar Connections for High-Performance Battery Pack Production

The transition towards electric mobility requires the development of manufacturing systems capable of realising products with elevated electrical and mechanical performance and in-line qualification. Laser welding of thin sheets is an enabling technology for the production of battery packs. Given the numerosity of the joints and the stringent requirements, in-situ monitoring of the process and advanced data analysis with Machine Learning (ML) algorithms are fundamental tools which need to be explored. The current study presents a methodological approach for the process development and integration of a monitoring architecture for the realisation of dissimilar material busbar connections (0.2 mm Ni-plated steel over 0.6 mm Cu in lap joint configuration) for the production of a high-performance battery pack for an electric racing motorbike. A single mode fiber laser welding system was equipped with different sensors to retrieve data during the laser-material interaction. The monitoring system was composed of three photodiodes positioned off-axis respectively observing the visible, thermal near-infrared and laser back-reflection region. A spectroscope also sampled process emission from an off-axis perspective whilst another photodiode was positioned within the laser source to observe the process coaxially. Following a preliminary phase required to characterise the process and data provided by the sensors, experiments were designed to identify defects and variations with respect to the reference condition. On a single sensor basis, supervised classification machine learning algorithms were trained to discern joints performed on an out of focus workpiece or in the presence of gap between the sheets. Results indicate that photodiodes observing the laser back-reflected light are capable of providing process relevant information which can be exploited to identify drifts from the reference processing condition. ML algorithms exhibited high accuracy classification even with a reduced amount of data