针对当前模型对于复杂工况下的锂电池健康状态（State of Health, SOH）预测精度低的问题，本文提出一种结合Transformer和扩散模型(Diffusion)的锂电池SOH预测方法。通过融合非线性自回归外生输入（Nonlinear Auto-Regressive With Exogenous Inputs, NARX）机制与结构化状态空间模型，构建由卷积神经网络、Mamba模型和Transformer解码器和条件扩散模型组成的动态模型，旨在提高锂电池SOH的预测精度与泛化性能。首先，将多周期锂电池数据堆叠为三维张量并进行标准化处理，利用双路径特征输入模块分别提取多维时序特征和历史SOH关联信息。主路径通过卷积层和Mamba状态空间模型捕捉容量衰退的长程依赖关系，辅助路径使用历史SOH值实现时序特征增强。然后，Transformer解码器将根据NARX方法读取电池的动态行为，预测周期t时电池的SOH，以此完成NARX反馈环路，使模型能够合并其过去的预测。在模型输出端设计了条件扩散模型，通过条件特征引导扩散过程，有效抑制预测误差的累积并提升对噪声的鲁棒性。实验结果表明，所提出的模型能够实现准确的电池SOH预测，并且能够有效融合跨周期退化特征，实现了复杂工况下的锂电池SOH高精度预测。
英文摘要:
This paper proposes a lithium battery State of Health (SOH) prediction method that combines a diffusion model and a Transformer to address the issue of low accuracy in predicting the State of Health (SOH) of lithium batteries under complex operating conditions. By integrating the Nonlinear Auto Regression With Exogenous Input (NARX) mechanism with a structured state space model, a dynamic model consisting of convolutional neural network, Mamba model, Transformer decoder, and conditional spread model is constructed to improve the prediction accuracy and generalization performance of lithium battery State of Health (SOH). Firstly, stack the multi cycle lithium battery data into a three-dimensional tensor and standardize it. Use a dual path feature input module to extract multidimensional temporal features and historical SOH correlation information separately. The main path captures the long-range dependencies of capacity decay through convolutional layers and Mamba state space models, while the auxiliary path uses historical SOH values to enhance temporal features. Then, the Transformer decoder will read the dynamic behavior of the battery based on the NARX method, predict the SOH of the battery at cycle t, and complete the NARX feedback loop to enable the model to merge its past predictions. A conditional diffusion model was designed at the output of the model, which guides the diffusion process through conditional features, effectively suppressing the accumulation of prediction errors and improving robustness to noise. The experimental results show that the proposed model can achieve accurate prediction of battery State of Health (SOH) and effectively integrate cross cycle degradation features, achieving high-precision prediction of lithium battery SOH under complex operating conditions.