A novel unsupervised anomaly detection and traceability method for battery voltage inconsistency based on the time-segmented DBSCAN algorithm

Voltage inconsistencies in lithium-ion battery packs are a primary driver of performance degradation and thermal runaway. Accurate diagnosis from real-world operational data remains hindered by highly dynamic driving conditions, continuous measurement noise, and scarce labeled faults. To address these challenges, an unsupervised anomaly detection framework utilizing a time-segmented Density-Based Spatial Clustering of Applications with Noise algorithm is proposed. High-dimensional voltage features, capturing both transient fluctuations and long-term aging behaviors, are extracted via a multi-scale sliding window mechanism. Principal Component Analysis is subsequently applied to decouple absolute voltage levels from dynamic inconsistency patterns. Time-segmented clustering then facilitates highly efficient local anomaly detection, generating a continuous Consistency Health Index for the quantitative grading of individual cell health. The framework was rigorously validated using actual operational data from two distinct electric vehicles that experienced severe thermal runaway. The failed cells (Cell #126 in EV 1; Cells #52 and #53 in EV 2) were accurately flagged one full month prior to the catastrophic incidents, achieving significantly high Anomaly Signal-to-Noise Ratios of 8.98 and 10.18. Quantitative comparisons confirm the proposed methodology significantly outperforms benchmark algorithms in isolating true fault signals from background noise, fundamentally eliminating the ranking inversion issue across diverse battery systems.