Data-driven approaches have demonstrated remarkable accuracy in battery’s state of health (SOH) estimation; however, they are susceptible to data quality and quantity. Therefore, an accurate data-based battery health estimation method is highly desirable in an unreliable industry environment when sensors’ random measurements unavailability is ubiquitous. Successful training under random data unavailability becomes a difficult task to undertake. Therefore, the main challenge is how an offline trained model can be reliable and accurate under random sensors’ measurements unavailability. This article develops an accurate SOH estimation model based on nonlinear autoregressive with exogenous inputs recurrent neural network for lithium-ion batteries whose features’ measurements are subjected to different random missing observations. To evoke the uncertainty of sensors’ measurements in online health diagnostic, missing observation occurrence is addressed by randomly eliminating sample data and then evaluating the model on the available measurements. Therefore, it does not require any imputation strategy for missing values. The accuracy of the estimator model is guaranteed when extracted underlying features are fused by adding their exponential moving average as the health features. The experimental results on two different datasets, Oxford and Toyota, under different battery chemistry and working operations demonstrate that the mean absolute errors (MAEs) and RMSs are well bounded below 2.70% and 3.10% for different random data missing rates of 1%–30%. It is a promising prediction model for numerous industrial applications with a high probability of random data unavailability.

中文翻译：

---

在随机传感器测量不可用的情况下开发锂离子电池的在线数据驱动健康状态估计

数据驱动的方法在电池的健康状态 (SOH) 估计方面表现出非凡的准确性；但是，它们容易受到数据质量和数量的影响。因此，当传感器的随机测量不可用性普遍存在时，在不可靠的工业环境中非常需要一种基于数据的准确电池健康估计方法。在随机数据不可用的情况下成功训练成为一项艰巨的任务。因此，主要的挑战是离线训练模型如何在随机传感器测量不可用的情况下可靠和准确。本文针对锂离子电池开发了一种基于非线性自回归和外生输入递归神经网络的准确 SOH 估计模型，其特征测量受到不同的随机缺失观测。为了唤起在线健康诊断中传感器测量的不确定性，通过随机消除样本数据然后根据可用测量评估模型来解决观察缺失的问题。因此，它不需要任何缺失值的插补策略。当提取的底层特征通过添加它们的指数移动平均值作为健康特征来融合时，保证了估计器模型的准确性。在不同的电池化学和工作操作下，牛津和丰田这两个不同数据集的实验结果表明，对于 1%–30 的不同随机数据缺失率，平均绝对误差 (MAE) 和 RMS 很好地限制在 2.70% 和 3.10% 以下%。对于随机数据不可用的可能性很高的众多工业应用，它是一个很有前途的预测模型。