A Hybrid Signal-Based Fault Diagnosis Method for Lithium-Ion Batteries in Electric Vehicles

Jiuchun Jiang,Caiping Zhang,Xinwei Cong,Shuowei Li,Yan Jiang,Weige Zhang

doi:10.1109/access.2021.3052866

Abstract

A large proportion of electric vehicle accidents are attributed to lithium-ion battery failure recently, which demands the time-efficient diagnosis and safety warning in advance of severe fault occurrence to ensure reliable operation of electric vehicles. However, serious battery system faults are often not caused by easily-observed cell state inconsistency, but derived from a certain cell failure with precursory signals untended, or occasional abuse, thus eventually thermal runaway. In this paper, a signal-based fault diagnosis method is presented, including signal analysis to eliminate the impact of state inconsistency on time-series feature extraction, feature fusion, and dimensionality reduction by manifold learning, with clustering-based outlier detection to identify abnormal signal features. The challenges in threshold determination of fused features can be effectively resolved by supplementary correction to largely reduce the amount of false alarms. Compared with the judgments from actual battery management systems, and other signal-based methods with single features, earlier detections can be achieved with robustness, verified by real-world pre-fault operation data of electric vehicles that suffered thermal runaway.

Highlights

Lithium-ion batteries are widely used as power supplies for electric vehicles (EVs), due to high energy density, no memory effect, long life span, low self-discharge rate, and other excellent performances [1]
SIGNAL PROCESSING WITH VMD AND Empirical mode decomposition (EMD) In order to extract significant components as the basis of analysis from raw voltage signals, we introduce relevant signal analysis methods
Based on several distances between voltage time series, quantification of state inconsistency is obtained from the static parts concisely, and introduction is mainly made on how the features of the IMFs are extracted and selected from dynamic parts

Summary

INTRODUCTION

Lithium-ion batteries are widely used as power supplies for electric vehicles (EVs), due to high energy density, no memory effect, long life span, low self-discharge rate, and other excellent performances [1]. Based on thresholds [30], machine learning [31], local outliers [32], correlation coefficient [33], information entropy [34], and other derivative features [35], [36], data-driven methods are usually employed online without battery modeling efforts, but the result may be obtained with poor robustness, accuracy, and interpretability. This is caused by inevitable cell inconsistency, measurement noise, and fragmented sparse data records collected and uploaded in the process of actual battery system operation.

SIGNAL PROCESSING WITH VMD AND EMD

DISTANCE-BASED STATE INCONSISTENCY DERIVED

GDI EXTRACTION BASED ON DYNAMIC PARTS

CLUSTERING AND MANIFOLD LEARNING IN INTEGRATED FEATURE ANALYSIS

RESULTS AND DISCUSSION

CONCLUSION AND OUTLOOK