Online anomaly monitoring and graded diagnosis method for power batteries based on multi-dimensional time-frequency feature fusion

Artikel i vetenskaplig tidskrift, 2026

Effective battery fault diagnosis and safety warning are essential for EV safety. However, early-stage voltage anomalies in real-vehicle data are usually weak and easily obscured by operating-condition fluctuations and measurement noise, which may lead to false alarms or delayed warnings. This paper proposes an online anomaly monitoring and graded diagnosis method based on multi-dimensional time-frequency feature fusion. A slidingwindow mechanism is adopted to update voltage sequences in real time and extract multiple effective statistical features from both the time and frequency domains. Principal Component Analysis is then used for feature dimensionality reduction, and Local Outlier Factor is employed to quantify the abnormality degree of each cell. Based on this, a dual-threshold strategy is designed, where persistence-based assessment is adopted for minor anomalies to suppress false alarms, while instantaneous triggering is used for severe anomalies to ensure timely identification. Experiments on real-vehicle data show that the proposed method can detect early-stage minor abnormal cells during progressive fault evolution, and identify and localize severe abnormal cells more than 70 frames before thermal runaway, while maintaining a low false-alarm level under normal operating conditions. Overall, the proposed method shows promising application potential for real-vehicle battery anomaly monitoring and graded warning.