Accurate estimation of the state of health (SOH) of lithium-ion batteries is
essential for ensuring the safety, reliability, and performance optimization of
electric vehicles. In practical operating environments, however, data quality is
often compromised by noise interference, frequent fluctuations in load
conditions, and the inherently non-stationary nature of battery degradation
features. These challenges reduce the effectiveness of conventional modeling
approaches, which often struggle to maintain both high prediction accuracy and
strong generalization capability. To address these issues, this study develops a
comprehensive SOH estimation approach encompassing data quality enhancement,
degradation feature extraction, and hybrid deep learning-based modeling. In the
first stage, multi-stage anomaly detection techniques are applied to remove
noisy or inconsistent measurements. A week-based indexing strategy is introduced
to generate temporally coherent labels, ensuring that time-series dependencies
are preserved. This procedure ensures a minimum per-vehicle valid-data ratio of
87.59%, thereby guaranteeing consistent data availability across all vehicles
and significantly improving both reliability and temporal alignment. In the
second stage, a set of degradation-sensitive health indicators is extracted from
raw sensor measurements, including voltage, current, and temperature profiles.
These features are then aggregated at a weekly resolution to enhance stability
and reduce short-term variability. In the third stage, a hybrid deep learning
model is constructed by combining a Temporal Convolutional Network (TCN) for
local pattern extraction, a Bidirectional Gated Recurrent Unit (BiGRU) for
long-term dependency modeling, and an attention mechanism for adaptive feature
weighting. Experimental results under 10-fold cross-validation show that the
proposed approach achieves a root mean square error of 1.30% and a mean absolute
error of 1.03% on the test set, outperforming single-model baselines in both
accuracy and robustness. The proposed framework offers a practical and scalable
solution for high-precision SOH estimation in real-world scenarios and provides
a strong basis for deployment under diverse and extreme operating
conditions.