To address the challenges of recognizing abnormal states, detecting subtle early
warning signs, and quantifying fault severity in scenarios involving
simultaneous multiple faults in lithium-ion batteries, this study proposes a
dual-layer fault diagnosis framework that integrates One-Class Support Vector
Machine (OCSVM) and Robust Local Mahalanobis Distance Quantile (RLMQD)
algorithm. First, a three-dimensional multi-scale feature space, incorporating
voltage, kurtosis, and voltage change rate, is constructed to detect abnormal
battery states via OCSVM and dynamically filter abnormal time periods with
improved adaptability. Second, a computationally efficient RLMQD-based
quantization algorithm is developed, which employs a small-scale sliding window
and adaptively selects healthy cells to construct reference distributions. By
incorporating low-quantile thresholds, the algorithm enhances early abnormality
detection and significantly reduces false positives. Subsequently, fault
severity is quantified through scale-weighted fusion and normalization, enabling
accurate evaluation across diverse abnormal modes. Finally, The diagnostic
performance of the proposed method is comprehensively validated through three
sets of simulation experiments and real-vehicle data collected under realistic
operating conditions. The results demonstrate that the proposed method
accurately identifies both single-point and clustered anomalies, corresponds
closely with actual fault conditions and exhibiting strong generalization
capability. In real vehicle validation, the method achieves 95.79% accuracy,
100% recall, and a 93.3% F1 score in abnormal detection tasks. Furthermore, It
demonstrates robustness and interpretability, enabling multi-type abnormal
detection and fault severity evaluation without reliance on extensive fault
datasets, thereby offering high suitablility for online monitoring and early
warning in actual Battery Management Systems.