Listening to batteries - using feature extraction and machine learning to identify and predict gas evolution and particle cracking with acoustic emission analysis

Acoustic emission (AE) testing is a non-invasive method for detecting degradation in lithium-ion batteries. Despite numerous studies, analysis and evaluation of batteries using AE remains subject to inconsistency, partly due to variability in test conditions and method settings. This study therefore presents a systematic approach to identify key AE characteristics to classify degradation-related processes in battery cells. Controlled degradation experiments were conducted on lithium-ion battery test cells by inducing specific degradation. Compared to nominal cycling, these conditions led to acoustic activity several orders of magnitude higher, demonstrating the method's sensitivity to safety-critical events. A labeled dataset was generated from the induced degradation data, and a supervised machine learning model was trained to classify AE signals associated with gas evolution and particle degradation. The model approach reduces analytical errors and minimizing cognitive bias. Feature importance analysis confirmed established features and highlighted additional parameters. When applied to nominal cycling of a full NCM-graphite cell, the model accurately identified 50 % of acoustic events with a confidence exceeding 75 %, proving AEs capability to detect specific degradation events with a high confidence. The method thus allows for a detailed operando diagnosis that can be suitable for use in BMS. • Supervised numerical analysis can pinpoint degradation during use. • Approach shows potential for early detection of critical degradation. • Models are trained on induced degradation. • Key signal features can uncover specific degradation. • Framework enables development of smarter, safer battery monitoring systems.