掩码表征迁移策略下的锂电池变工况健康状态预测

A masked feature transfer strategy for lithium battery state of health prediction under variable working conditions

中文关键词  锂离子电池  健康状态  掩码表征迁移策略  变工况迁移

英文关键词  Lithium battery  state of health  masked feature transfer strategy  variable working conditions transfer

中文摘要

      锂电池健康状态(SOH)预测可以对电池老化程度进行评估. 由于电池工作状况的差异, 锂电池训练数据(源 域)和在线应用数据(目标域)具有不同的分布, 而迁移学习是解决上述问题的有效方法. 然而, 一方面, 传统迁移学 习方法需要大量源域数据标签, 而SOH测量困难, 难以提供充足标签; 另一方面, 这些方法无法充分利用已有的专家 知识. 基于此, 本文创新性地提出了一种掩码表征迁移策略(MFTS), 实现了源域无标签场景下的变工况锂电池SOH 预测. 首先, 设计了一种掩码自监督框架, 其可以在无标签的情况下实现自动特征提取来表征源域数据. 其次, 提出 了一种专家知识模块, 引导所提特征逼近专家特征, 从而实现了专家知识的融入. 最后, 提出了一种双学习率的方 法对特征提取和SOH预测网络进行同步变速训练, 在迁移源域知识的同时实现了目标域SOH的准确预测. 基于NASA的公开数据集, 所提出的MFTS模型在6组实验上的预测误差均不大于4.08%.

英文摘要

      Lithium battery state of health (SOH) prediction can evaluate battery aging. Due to differences in battery working conditions, lithium battery training data (source domain) and online application data (target domain) have different distributions, and transfer learning is an effective method to solve the above problems. However, on the one hand, traditional transfer learning methods require a large number of source domain data labels, and the SOH measurement is difficult to provide sufficient labels. On the other hand, these methods cannot make full use of existing expert knowledge. To solve the above problems, this paper innovatively proposes a masked feature transfer strategy (MFTS), which realizes the SOH prediction of the lithium battery under variable working conditions with unlabeled source domain data. First, a masked self-supervised framework is designed, which can automatically extract robust representations in source domain data without labels. Secondly, an expert knowledge module is proposed to guide the extracted features to approach the expert features, thus realizing the integration of expert knowledge. Finally, a double learning rate method is proposed to perform synchronous variable speed training on the feature extraction and the SOH prediction network, and achieves the accurate prediction of the target domain SOH while transferring the knowledge of the source domain. Based on the NASA’s public data set, the prediction error of the proposed MFTS model in the six sets of experiments is all less than or equal to 4.08%.