With the rapid expansion of the electric vehicle market, the safety of
lithium-ion batteries, which serve as the main power source, has become a
critical concern. Current mainstream methods for battery fault detection
generally face a technical bottleneck of struggling to balance high accuracy
with a low false alarm rate. Furthermore, constrained by algorithmic complexity
and data processing efficiency, detection speeds often fail to meet the
practical demands of real-time monitoring. As a result, developing more
efficient and accurate fault detection technologies has emerged as a key
challenge urgently needing to be addressed in the industry. This paper proposes
a hierarchical fault detection framework for lithium-ion batteries that
integrates voltage change characteristics with a Local Outlier Factor (LOF)
scoring mechanism. The framework aims to achieve early identification and
accurate diagnosis of abnormal battery states through multi-dimensional feature
extraction and algorithmic fusion. In the first layer, decentralized voltage
data are standardized using the 3σ rule to identify potentially
anomalous batteries. In the secondary analysis phase, a sliding time-window
mechanism is adopted to dynamically capture voltage sequences. Within each
window, voltage variations are calculated along both vertical and horizontal
directions, and statistical metrics, including mean, standard deviation, range,
and increment are derived. Principal component analysis is then applied to
extract key features, and battery anomalies are evaluated and confirmed based on
the maximum LOF score. Experimental validation using datasets from vehicles that
experienced thermal runaway events, along with data from 1,000 normal vehicles,
demonstrates that the proposed method significantly improves the accuracy of
battery fault detection. It also provides early warnings up to 17 days prior to
the occurrence of thermal runaway.