近年来，锂离子电池因其卓越性能在电动汽车、储能电站和消费电子等领域得到广泛应用。同时，快充技术的创新、能量密度的提高和使用规模的扩大，进一步推动了锂电池发展。然而，锂电池安全性问题也日益凸显，相关事故不断发生，带来严重经济损失和社会危害。因此，迫切需要电池管理系统（BMS）采取更加有效的故障检测方法，及时预警并采取保护措施，以确保锂电池使用安全。目前，相关技术方案在检测精度、适用性和稳定性等方面仍存在诸多挑战，已成为产业界和学术界的研究热点和重点。在此背景下，本文使用三星18650-33G三元锂离子电池为研究对象，开展故障早期检测技术相关研究，旨在进一步优化对电池系统的安全管理。课题的主要研究内容如下：（1）对电池工作原理、故障演化机理和影响进行分析。本文关注电池故障产生原因、内部机理和表征方式间的因果关系，并结合不同电池模型，从极化角度理解电池故障产生影响。基于分析结果，选用二阶等效电路模型耦合热模型，作为故障检测的基础物理模型。（2）搭建电池测试实验平台，设计并开展一系列电池实验。一方面进行电池容量和一致性特性测试，并在多温度梯度下对电热耦合模型进行离线参数辨识；另一方面，基于故障仿真模型和实验平台进行一系列电池内短路故障测试实验，补充电池故障数据，从而为相关研究提供支持。（3）提出一种基于融合模型的电池故障早期检测方案。本文阐释了粒子滤波（PF）和扩展卡尔曼滤波（EKF）算法的原理与不足，并基于早期故障快速检测需求，提出一种EKF电池观测器与双向长短期记忆神经网络（BiLSTMNN）学习模型融合的方法。学习模型能够对观测误差进行有效补偿，从而消除故障检测过程中不确定性因素的干扰。（4）利用测试数据对所提出方法进行验证评价。首先，基于改进粒子群算法（PSO）对比不同学习模型融合后的补偿效果。然后，在不同工况下对融合模型进行交叉验证，突出其鲁棒性和准确性。最后，对不同演化阶段的电池内短路故障进行检测，结果表明本文所提出方法能够达到预期效果。
In recent years, Lithium-ion batteries have been widely used in Electric Vehicles (EVs), energy storage stations, consumer electronics and other aspects owing to their outstanding performance. Concurrently, the innovation of fast charging technology, improvement in energy density, and expansion of application scale have further propelled the development of the batteries. However, safety issues regarding the batteries are increasingly prominent, with related accidents occurring continuously, resulting in substantial economic losses and societal hazards. Therefore, there is an urgent need for battery management systems (BMS) to adopt more effective fault detection methods, timely warning and implement protective measures to ensure the safety of the batteries. Currently, the related technical solutions still face various challenges in terms of detection accuracy, applicability and stability, thus becoming a research hotspot and focus for both industry and academia.Based on the aforementioned background, this paper takes the Samsung INR 18650-33G lithium-ion batteries as the research object and carries out the investigation of fault early detection technology, aiming at further optimizing the safety management of the battery system. The main research contents are as follows:(1) The basic working principle, internal fault mechanism and effects analysis are performed for Lithium-ion batteries. The causal relationship between failure causes, mechanisms and symptoms are focused, integrating various battery models to comprehend the implications of battery failures from the perspective of polarization. Based on the analysis results, the second-order equivalent circuit model coupled with the thermal model is selected as the basic physical model for fault detection.(2) The experimental battery testing platforms are built, and a series of experiments are designed and conducted. Firstly, experiments regarding the battery‘s capacity and consistency are performed, alongside offline parameter identification for the electro-thermal coupling model under various temperature gradients. Then, a series of battery internal short circuit fault testing experiments, based on the fault simulation model and experimental platform, are carried out, supplementing the data to provide support for related research.(3) A method for early detection of battery faults based on the fusion model is proposed. This paper offers a comprehensive explanation for commonly used Particle Filter (PF) and Extended Kalman Filter (EKF) algorithms, and considered with the need for rapid detection of early faults, proposes a fusion solution incorporating an EKF battery observer and a Bidirectional Long Short-term Memory Neural Network (BiLSTMNN) learning model. The latter is capable of assessing the state of battery, then effectively compensating for observation errors and mitigating the interference of uncertainties during the fault detection process.(4) The proposed method is verified and evaluated. Firstly, the compensatory effects of integrating different learning models were compared, utilizing an optimized Particle Swarm Optimization (PSO) algorithm. Then, the fusion model is cross-validated under different electric vehicle driving conditions to underscored the its stability and robustness. Finally, battery short circuit faults at different evolution stages are detected respectively, and the results show that the proposed method can achieve the expected results.
