Identification of features in low-quality CT images of prismatic Li-ion cells

Process variations during the production of LIBs can lead to significant differences in the physical properties, inhomogeneities, and defects among cells within the same production batch. Battery module manufacturers can employ CT imaging at the cell level to ensure cell quality. Due to the large number of cells, an automated inspection is necessary. ML-based algorithms can be used to automate identifying and measuring features in CT images. However, training accurate models typically requires a large amount of labelled data, but generating high-quality CT images for training is costly. A transfer learning approach using a foundation model trained on other cells’ CT images reduces the required data. An additional challenge for detecting cell features is posed by low-quality CT images. ML techniques, such as GAN or SRCNN, can enhance image quality to identify features more accurately. This research focuses on the question: ‘How can transfer learning in combination with ML techniques increase the accuracy of feature extraction from low-quality CT images of prismatic Li-Ion cells?’ The results indicate promising potential for accurately identifying features in prismatic battery cells, even when dealing with lower-quality images.