Automation of image-based measurement of battery cell features by computed tomography and synthetic training data

Abstract

Due to process variations in the production of lithium-ion batteries (LIBs), cells of one production batch can show a variation in physical features, inhomogeneities, and defects. These can impact the performance and safety of the cells and should be detected, and if accepted in tolerances should be measured accurately. The cell features are often unknown to manufacturers of battery modules and packs. Hence, computed tomography (CT) imaging could provide insight into the cells’ quality, allowing the measurement of relevant battery cell features. However, the high number of cells requires an automation of cell inspection. This work focuses on the challenge of automated image processing and provides an image-based workflow measuring multiple cell features based on a single CT scan. Both classical computer vision (CV) and machine learning (ML)-based image algorithms are applied within the developed workflow. To train, test, and validate the convolutional neural network (CNN)-based algorithms, artificially generated training data is created and used due to the scarcity of training data, which can form a bottleneck in CNN-model development and evaluation. Hence, the generation of synthetic training data shown in this work can reduce the need for costly laboratory CT scans before adoption in serial production environments. The results show the promising potential of synthetic training data and the automated approaches to measure cell features, specifically the electrodes’ windings, the corresponding length and width, as well as the anode overhang.