The remaining useful life (RUL) of solid-state lithium-ion battery (SSLIB) is a crucial challenge for their future marketability due to the fact that it guarantees the safety and reliability for electric vehicles (EV) under complex degradation mechanisms. To address this issue, a novel RUL prediction approach based on improved convolutional neural network (CNN) and derived health indicators (HIs) from the limited HIs is proposed according to “No Free Lunch” theorem in this paper. To alleviate the limitation of real-time obtaining impedance in SSLIB and improve the expression ability of limited HIs, some HIs related to impedance are derived based on coefficient of variation and symbolic regression from charging and discharging curve. Then, mutual information based on k-nearest neighbor algorithm is used for scoring measured HIs and derived HIs to select valuable information. In many cases, the RUL prediction methodologies would need to be deployed on vehicle terminals with limited computational capacity and memory so as to support real-time decision and reduce the data communication cost. Notably, excessive attention has been paid to predicted performance and existing neural networks often have high complexity. Therefore, a novel knowledge distillation model based on CNN, which is assisted adaptively by energy circle sparrow searching algorithm (ECSSA-KDCN) is developed intricately to alleviate this issue. The ECSSA is inspired from the energy-consuming in sparrows. The results in real-world data of SSLIB show that the of student model in KDCN is above 0.92 and total parameters is only 0.076% of the original teacher model. Other indexes, such as RMSE, SMAPE, MLSE, MSPE also are competitive compared with the other state-of-the-art methods, which indicate that the ECSSA-KDCN has competent generalization properties and high precision regardless of the aging patterns to SSLIB data.

中文翻译：

---

基于改进CNN和健康指标推导的固态锂离子电池剩余使用寿命新方法


固态锂离子电池（SSLIB）的剩余使用寿命（RUL）是其未来市场化的关键挑战，因为它要保证电动汽车（EV）在复杂的退化机制下的安全性和可靠性。为了解决这个问题，本文根据“没有免费的午餐”定理，提出了一种基于改进的卷积神经网络（CNN）和从有限的HI导出健​​康指标（HI）的新型RUL预测方法。为了缓解SSLIB中实时获取阻抗的局限性，提高有限HI的表达能力，基于变异系数和充放电曲线的符号回归推导了一些与阻抗相关的HI。然后，使用基于k近邻算法的互信息对测量的HI和导出的HI进行评分，以选择有价值的信息。在许多情况下，RUL预测方法需要部署在计算能力和内存有限的车辆终端上，以支持实时决策并降低数据通信成本。值得注意的是，人们过度关注预测性能，并且现有的神经网络通常具有很高的复杂性。因此，为了缓解这一问题，开发了一种基于CNN、能量圈麻雀搜索算法自适应辅助的新型知识蒸馏模型（ECSSA-KDCN）。 ECSSA 的灵感来自麻雀的能量消耗。 SSLIB的真实数据结果表明，KDCN中的学生模型的参数在0.92以上，总参数仅为原始教师模型的0.076%。 其他指标，如 RMSE、SMAPE、MLSE、MSPE 与其他最先进的方法相比也具有竞争力，这表明 ECSSA-KDCN 具有良好的泛化性能和高精度，无论 SSLIB 数据的老化模式如何。