Fault detection method for electric vehicle battery pack based on improved kurtosis and isolation forest

Early detection of abnormal battery characteristics is crucial for ensuring personnel safety and minimizing property damage. Conventional fault diagnosis methods often struggle to detect minor faults in their early stages. To address this challenge, this paper proposes a fault diagnosis method for lithium-ion batteries in electric vehicles that utilizes real-world operational data. Initially, a wavelet threshold denoising algorithm is used to effectively remove voltage data noise while retaining fault characteristics. Subsequently, an early fault warning method based on improved kurtosis index is proposed, which is capable of capturing the weak abnormal features of the battery and issuing warnings. Furthermore, a faulty battery cell localization method based on two-dimension feature extraction and Isolation Forest algorithm is proposed to capture minor fault features and accurately locate faulty cells. Finally, experiments conducted on four operational electric vehicles demonstrated that the early fault warning method could promptly detect faults and issue alarms, while the faulty battery cell positioning method accurately identified the faulty cell, achieving 100$\% $% Accuracy and 0$\% $% False Positive Rate. This confirms the effectiveness of the proposed method. Additionally, Bootstrap was used to evaluate the performance metrics, further verifying the robustness of the method.