【作者】 刘欣；

【导师】 高德欣；

【作者基本信息】 青岛科技大学 ， 控制工程（专业学位）， 2022， 硕士

【摘要】 随着新型能源技术的快速发展,动力电池的市场需求量逐渐提高。由于近几年动力电池的安全问题频发,人们对动力电池的关注逐渐由容量需求转向安全需求。动力电池在循环充放电的过程里会出现老化现象,其性能逐渐衰退。持续使用临近退役的动力电池易发生故障,从而影响供电系统的正常运行。因此,及时检测动力电池的健康状态并估计其剩余寿命对于维护动力电池的使用安全性具有重要意义。本课题基于深度学习方法对动力电池进行检测与分析的内容如下:第一,介绍动力电池检测的研究背景意义及其国内外的研究现状,分析模型构建法与数据驱动法的特点,选择数据驱动中的深度学习算法对动力电池进行检测与分析。接着描述动力电池的结构组成和基本工作原理,设定电池的检测标准。采用NASA动力锂电池的公开数据源,阐述NASA电池在充放电实验中参数数据的变化过程,通过灰色关联度分析法提取与电池健康状态关联度高的参数作为健康因子。第二,考虑动力电池的参数数据具有时序性,深度学习算法中的循环神经网络(Recurrent Neural Network,RNN)常用于分析时间序列数据。通过查阅文献和公式推导发现RNN在计算梯度和调整权重矩阵时,存在梯度消失或爆炸现象。长短期记忆神经(Long Short-Term Memory,LSTM)网络在RNN基础上添加记忆单元的门控操作,处理RNN中的梯度问题。然后,构建基于LSTM深度学习的动力电池检测模型,并介绍模型中使用的Dropout技术和Adam优化器,对NASA的动力锂电池数据进行检测实验与分析。第三,为了建立动力电池数据的过去与未来之间的联系,将LSTM神经网络改进成双向长短期记忆神经(Bi-directional Long Short Term Memory,Bi LSTM)网络,即由前向LSTM隐藏层和反向LSTM隐藏层组成,优化了LSTM相对单一的时序分析能力。然后,设计基于Bi LSTM深度学习的动力电池检测模型,对NASA的动力锂电池数据进行检测实验分析。第四,根据一维卷积神经网络(One Dimensional Convolutional Neural Network,1D CNN)适用于一维数据特征提取的特点,将1D CNN和Bi LSTM相融合。通过卷积层和池化层提取动力电池数据中的深层特征,再由Bi LSTM层双向分析数据之间的关联性。CNN-Bi LSTM在Bi LSTM的基础上提高了检测模型的泛化能力。然后,建立基于CNN-Bi LSTM深度学习的动力电池检测模型,对NASA的动力锂电池数据进行检测实验分析。最后,综合多组实验结果和评价指标进行分析,基于CNN-Bi LSTM深度学习的动力电池检测模型对电池健康状态的检测能力更强,电池健康状态检测曲线与实际电池状态曲线之间的拟合度R~2均在0.9以上,RUL预测误差稳定在2以内,与RNN、LSTM和Bi LSTM检测模型的实验结果相比,CNN-Bi LSTM检测模型具有更高的预测精度和稳定性。更多还原

【Abstract】 With the rapid development of new energy technologies,the market demand for power batteries has gradually increased.Due to the frequent safety problems of power batteries in recent years,and people’s attention to power batteries has gradually shifted from capacity requirements to safety requirements.The power battery will age during the cycle of charge and discharge,and its performance will gradually decline.Continuous use of power batteries that are nearing retirement is prone to failure,thus affecting the normal operation of the power supply system.Therefore,timely detection of the state of health of power batteries and estimation of their remaining life are important to maintain the safety of power batteries.The detection and analysis of the power battery based on the deep learning method in this topic are as follows:Firstly,the background significance of power battery detection and its research status at home and abroad are introduced,the characteristics of model construction method and data-driven method are analyzed,and the deep learning algorithm in data-driven method is selected to detect and analyze power battery.Next,it describes the structure and basic working principle of the power battery,sets the battery detection standards.Using the public data source of NASA power lithium battery,the change process of the parameter data of the NASA battery in the charging and discharging experiment is described,and the parameters with high correlation with the battery health state are extracted by the gray correlation analysis method as the health factor.Secondly,considering that the parameter data of the power battery is time series,the Recurrent Neural Network(RNN)in the deep learning algorithm is often used to analyze the time series data.By consulting the literature and formula derivation,it is found that the gradient disappears or explodes when the RNN calculates the gradient and adjusts the weight matrix..The Long Short-Term Memory(LSTM)network adds the gating operation of the memory unit on the basis of the RNN to handle the gradient problem in the RNN.Then,a power battery detection model based on LSTM deep learning is constructed,and the Dropout technology and Adam optimizer used in the model are introduced,and the detection experiments and analysis of NASA’s power lithium battery data are carried out.Thirdly,in order to establish the connection between the past and the future of power battery data,the LSTM neural network is improved into a Bi-directional Long Short Term Memory(Bi LSTM)network,that is,the forward LSTM hidden layer and the reverse It is composed of hidden layers to LSTM,which optimizes the relatively single timing analysis capability of LSTM.Then,a power battery detection model based on Bi LSTM deep learning is designed,and the detection experiment analysis of NASA’s power lithium battery data is carried out.Fourthly,according to the characteristics of One Dimensional Convolutional Neural Network(1D CNN)suitable for one-dimensional data feature extraction,1D CNN and Bi LSTM are integrated.The deep features in the power battery data are extracted through the convolution layer and the pooling layer,and the Bi LSTM layer is used to analyze the correlation between the data bidirectionally.CNN-Bi LSTM improves the generalization ability of the detection model on the basis of Bi LSTM.Then,a power battery detection model based on CNN-Bi LSTM deep learning is established,and the detection experimental analysis of NASA’s power lithium battery data is carried out.Finally,after analyzing multiple sets of experimental results and evaluation indicators,the power battery detection model based on CNN-Bi LSTM deep learning has a stronger ability to detect the battery state of health,and the fitting degree between the battery state of health detection curve and the actual battery state curve is R~2Compared with the experimental results of RNN,LSTM and Bi LSTM detection models,the CNN-Bi LSTM detection model has higher prediction accuracy and stability.更多还原