Predicting the end-of-life or remaining useful life of batteries in electric vehicles is a critical and challenging problem, predominantly approached in recent years using machine learning to predict the evolution of the state-of-health during repeated cycling. To improve the accuracy of predictive estimates, especially early in the battery lifetime, a number of algorithms have incorporated features that are available from data collected by battery management systems. Unless multiple battery data sets are used for a direct prediction of the end-of-life, which is useful for ball-park estimates, such an approach is infeasible since the features are not known for future cycles. In this paper, we develop a highly-accurate method that can overcome this limitation, by using a modified Gaussian process dynamical model (GPDM). We introduce a kernelised version of GPDM for a more expressive covariance structure between both the observable and latent coordinates. We combine the approach with transfer learning to track the future state-of-health up to end-of-life. The method can incorporate features as different physical observables, without requiring their values beyond the time up to which data is available. Transfer learning is used to improve learning of the hyperparameters using data from similar batteries. The accuracy and superiority of the approach over modern benchmarks algorithms including a Gaussian process model and deep convolutional and recurrent networks are demonstrated on three data sets, particularly at the early stages of the battery lifetime.

翻译：预测电动汽车电池的寿命终止或剩余使用寿命是一个关键且具有挑战性的问题。近年来，主流方法利用机器学习技术预测电池在反复充放电循环中健康状态的演化。为提高预测估计的精度（尤其是在电池寿命早期），多种算法已整合了从电池管理系统采集数据中提取的特征。除非使用多个电池数据集直接预测寿命终止（这有助于粗略估算），否则此类方法不可行，因为未来循环的特征无法预知。本文通过改进高斯过程动态模型（GPDM）开发了一种高精度方法以克服这一限制。我们引入核化版本的GPDM，使可观测量与隐变量坐标之间的协方差结构更具表达力。该方法结合迁移学习追踪直至寿命终止的未来健康状态，可将不同物理可观测量作为特征纳入模型，且无需获取数据截止时间之后的值。利用相似电池数据，迁移学习优化了超参数的学习过程。在三个数据集上的实验表明，该方法尤其在电池寿命早期阶段，其精度和优越性均超越现代基准算法（包括高斯过程模型、深度卷积网络与循环网络）。