Acoustic emission (AE) is a low-cost, non-invasive, and accessible diagnostic technique that uses a piezoelectric sensor to detect ultrasonic elastic waves generated by the rapid release of energy from a localised source. Despite the ubiquity of the cylindrical cell format, AE techniques applied to this cell type are rare in literature due to the complexity of acoustic wave propagation in cylindrical architectures alongside the challenges associated with sensor coupling. Here, we correlate the electrochemical performance of cells with their AE response, examining the differences during pristine and aged cell cycling. AE data was obtained and used to train various supervised binary classifiers in a supervised setting, differentiating pristine from aged cells. The highest accuracy was achieved by a deep neural network model. Unsupervised machine learning (ML) models, combining dimensionality reduction techniques with clustering, were also developed to group AE signals according to their form. The groups were then related to battery degradation phenomena such as electrode cracking, gas formation, and electrode expansion. There is the potential to integrate this novel ML-driven approach for widespread cylindrical cell testing in both academic and commercial settings to help improve the safety and performance of lithium-ion batteries.
Read full abstract