Large-scale adoption of lithium-ion batteries in electrical transportation and energy storage necessitates efficient battery diagnosis for cell screening and aging monitoring. Currently, battery diagnosis usually relies on lumped performance quantities, including capacity, resistance, and voltage variations. This work proposes a data-driven battery diagnosis method which quantifies key physico-chemical parameters through combining electrochemical impedance spectroscopy and a machine-learning technique. A systematic parameter-sensitivity analysis is conducted to identify physico-chemical parameters that can be estimated with confidence from impedance data, based on a new definition of sensitivity and a physical impedance model. Afterwards, the multi-output relevance vector regression is used to determine the expectation values and confidence intervals of selected physico-chemical parameters from impedance data. The effectiveness of this method is demonstrated using Monte-Carlo simulations for specific cases of cell screening and aging monitoring. Important issues in practical application of this method are also discussed.

锂离子电池在电力运输和能量存储中的大规模采用，需要对电池筛选和老化监测进行有效的电池诊断。当前，电池诊断通常取决于集总的性能量，包括容量，电阻和电压变化。这项工作提出了一种数据驱动的电池诊断方法，该方法通过结合电化学阻抗谱和机器学习技术来量化关键的理化参数。基于灵敏度的新定义和物理阻抗模型，进行了系统的参数敏感性分析，以识别可以从阻抗数据中放心估计的理化参数。之后，多输出相关性向量回归用于从阻抗数据中确定所选理化参数的期望值和置信区间。使用蒙特卡洛模拟方法对特定的细胞筛选和衰老监测案例证明了该方法的有效性。还讨论了该方法在实际应用中的重要问题。