基于改进残差网络的锂离子电池故障诊断

Fault diagnosis of Li-ion battery based on improved residual network

针对残差网络(ResNet)对特征提取准确率低和拟合度不够的问题,提出一种基于改进残差网络的锂离子电池故障诊断方法。首先,利用Simulink对电池容量变小、内阻变大、充电不足和自放电大等4种故障进行故障模拟,得到故障电压数据,作为输入,将首层提取的特征因式分解,分别加到后面的每一层;然后,引入注意力模块(SELayer)分支轻量化;最后,采用反卷积上采样,使远距离残差特征融合,加深特征提取能力,并降低计算量。改进残差网络故障模拟实验表明,与传统的ResNet50、ResNext、DensNet121和DensNet169等4种模型相比,所提模型的诊断准确率从88.63%提高到99.00%以上,参数量从2 500万减小到了2 470万,收敛速度上也具有一定的优势。

Aiming at the problems of low accuracy of feature extraction and insufficient fitting of residual network(ResNet),a fault diagnosis method for Li-ion battery based on improved residual network was proposed.Firstly,Simulink was used to simulate four kinds of faults,such as battery capacity becoming smaller,internal resistance becoming larger,under-charge and self-discharge becoming larger,to obtain fault voltage data,which was used as input.The feature factor extracted from the first layer was decomposed and added to each subsequent layer.Then the attention module(SELayer)branch was introduced for lightweight.Finally,deconvolution up-sampling was used to fuse the remote residual features,deepen the feature extraction ability and reduce the calculation amount.The improved residual network fault simulation experiment showed that compared with the traditional ResNet50,ResNext,DensNet121 and DensNet169 models,the diagnosis accuracy of proposed model was increased from 88.63%to above 99.00%,the parameter quantity was reduced from 25.0 million to 24.7 million,the convergence speed also had some advantages.