基于IC曲线的锂离子电池内短路早期诊断

Early diagnosis of internal short circuit in lithium-ion batteries based on incremental capacity curves

摘要

为提升锂离子电池内短路故障的早期检测能力,提出一种基于卷积神经网络(convolutional neural network,CNN)与长短期记忆网络(long short-term memory,LSTM)融合的内短路诊断方法。首先将电池充电过程中的电压曲线转换为容量增量(incremental capacity,IC)曲线,并利用快速傅里叶变换(fast fourier transform,FFT)对IC曲线进行平滑处理,从中提取电池内部状态变化特征,进而采用CNN提取IC曲线中的局部特征,再通过LSTM捕获长期依赖信息,最终实现内短路故障的分类检测。通过并联电阻模拟不同程度的早期内短路故障,构建包含5 000组样本的数据集。结果表明,以IC曲线为特征时,CNN-LSTM融合模型对故障等级的分类准确率达88.62%,显著优于单一模型,并能有效区分故障严重程度。该方法为电池管理系统智能故障诊断提供了有效方案。

Abstract

To address the challenges of detecting the concealed early-stage characteristics of internal short circuits (ISC) in lithium-ion batteries, and the inability of existing methods to simultaneously capture local features and model temporal evolution, this study proposes a diagnostic method based on a hybrid convolutional neural network (CNN) and long short-term memory (LSTM) architecture. This approach aims to enhance the battery management system′s (BMS) capability to perceive and classify micro-short circuit faults of varying severities. First, leveraging the high sensitivity of incremental capacity (IC) curves to internal electrochemical phase transitions, charging voltage data is converted into IC curves to amplify subtle fault signals. A fast fourier transform (FFT) is then introduced to filter high-frequency noise generated during the differentiation process, effectively preserving key peak and valley features of the curves. Subsequently, a series CNN-LSTM network is designed, a one-dimensional CNN extracts local spatial morphological features from the IC curves, while the LSTM captures the long-term dependency rules of fault evolution over time, establishing an end-to-end fault diagnosis model. In the experiments, four fault states ranging from early-stage micro-short circuits to severe short circuits were simulated by connecting resistors of different resistance values in parallel to 18650 batteries, constructing a balanced dataset containing 5 000 samples. The results demonstrate that the proposed method achieves a test accuracy of 88.62% in the five-category classification task, with the recognition rate for severe short circuits reaching 93.5%. Comparative analysis confirms that the performance of this hybrid model is significantly superior to that of standalone CNN or LSTM networks. It effectively distinguishes the severity of internal short circuits, providing a high-precision technical solution for intelligent battery safety early warning systems.

关键词

锂离子电池 / 内短路检测 / 卷积神经网络 / 长短期记忆网络 / 容量增量曲线

Key words

lithium-ion battery / internal short-circuit detection / convolutional neural network / long short-term memory network / incremental capacity curve