Battery fault diagnosis and failure prognosis for electric vehicles using spatio-temporal transformer networks

Battery-powered electric vehicles (EVs) are poised to accelerate decarbonization in nearly every aspect of transportation. However, safety issues of commercial lithium-ion batteries related to the faults and failures in real-world applications are still serious concern. Even a small increase in risk during the battery's operational lifetime may evolve into a safety hazard-fire and explosion, named as thermal runaway, after long-term incubation. Modelling and predicting the evolution of nonlinear multiscale electrochemical systems is challenging due to uncertainties in materials and manufacturing processes, dynamic environmental and operating conditions, as well as a lack of high-quality datasets. This challenge is further complicated when solving real-life physical problems with missing and noisy data and uncertain boundary conditions. In this study, we address these challenges by developing a specialized Transformer network architecture called BERTtery (Bidirectional Encoder Representations from Transformers for batteries) based on field data of EVs. By using charging voltage and temperature curves from early cycles before exhibiting symptoms of battery, the two-tower Transformer with temporal-wise encoder and channel-wise encoder is demonstrated as a powerful tool to capture early-warning signals across multiple spatio-temporal scales under a wide range of operating conditions. The method reliably predicts the evolution of faults in battery systems using only data provided by the onboard sensor measurements of battery performance.