重庆理工大学学报2025，Vol.39Issue(7)：1-8,8.DOI:10.3969/j.issn.1674-8425(z).2025.04.001
融合CNN与Transformer的锂离子电池健康状态估计
State of health estimation for lithium-ion batteries integrating CNN and Transformer
摘要
Abstract
Lithium-ion batteries are widely used on electric vehicles due to their high energy density and long cycle life.However,battery aging caused by electrochemical degradation during repeated charging and discharging poses significant challenges to their safety and reliability.Accurate state of health(SOH)estimation is essential for optimizing battery utilization,preventing failures,and extending their lifespans.Current SOH estimation methods are broadly categorized into model-based and data-driven methods.Among them,model-based methods,such as equivalent circuit models and electrochemical models,rely on precise parameter identification and simplified assumptions,leading to compromised accuracy in dynamic real-world conditions.Data-driven methods leverage machine learning to map measurable signals,such as voltage,current to SOH,but they suffer limitations including reliance on complete charging curves,complex feature engineering,and insufficient generalization across diverse operating scenarios.Existing studies often extract health features from full charging cycles,which are impractical in real-world applications where users rarely fully discharge batteries before recharging.Furthermore,conventional deep learning algorithms struggle to balance local feature extraction and global dependency modeling,while manual hyperparameter tuning causes computational inefficiency and suboptimal performance.Addressing these gaps requires a robust framework that integrates simplified feature extraction,adaptive feature learning,and automated optimization for practical SOH estimation. This paper proposes a hybrid deep learning framework combining convolutional neural networks(CNN)and Transformer,enhanced by Bayesian optimization,to achieve high-precision SOH estimation.Partial charging time sequences at equal voltage intervals are extracted as health features,eliminating the need for complete charging curves.These features exhibit a Pearson correlation coefficient exceeding 0.98 with battery capacity degradation,validating their effectiveness in representing aging characteristics.The framework employs a dual-branch architecture:a 1D-CNN with three convolutional layers captures local temporal patterns,while a Transformer encoder with multi-head self-attention mechanisms models long-term dependencies across the sequence.To optimize model efficiency,Bayesian optimization dynamically adjusts hyperparameters,including learning rate,kernel size,and the number of layers,using Gaussian processes.This approach reduces manual intervention and accelerates convergence compared to grid or random search. Validation experiments utilize aging data from eight Kokam lithium-ion batteries cycled under accelerated conditions(40 ℃,2C charge rate).A leave-one-out cross-validation strategy ensures unbiased evaluation,where each battery serves as the test set while others form the training data.Results demonstrate the method achieves a maximum error of 1.49％,root mean square error(RMSE)of 0.753％,and mean absolute error(MAE)of 0.632％,outperforming traditional models including standalone CNN(maximum error of 3.2％),LSTM(maximum error of 2.3％),GRU(maximum error of 2.1％),and CNN-GRU(maximum error of 1.9％).The integration of Bayesian optimization further reduces errors by 0.2％～0.5％compared to the non-optimized hybrid model,highlighting its role in enhancing precision.Robustness tests across different voltage ranges(3.1～3.4 V,3.5～3.8 V,3.9～4.2 V)further confirm the method's consistently superb performances,with optimal accuracy observed in the 3.5～3.8 V range(SOC 20％～80％),which aligns with typical user charging habits.Visualization of attention weights in the Transformer reveals its ability to prioritize critical voltage segments correlates with capacity loss,improving interpretability.关键词
锂离子电池/健康状态/贝叶斯优化/卷积神经网络/充电时间Key words
lithium-ion battery/SOH/Bayesian optimization/CNN-transformer/charging time分类
动力与电气工程引用本文复制引用
舒星,杨浩,刘西,陈飞,胡远志..融合CNN与Transformer的锂离子电池健康状态估计[J].重庆理工大学学报,2025,39(7):1-8,8.基金项目
重庆市自然科学基金面上项目(2024NSCQ-MSX0185) （2024NSCQ-MSX0185）
重庆市技术创新与应用发展专项项目(CSTB2023TIAD-STX0036,CSTC2021jscx-cylhx0006) （CSTB2023TIAD-STX0036,CSTC2021jscx-cylhx0006）
