A sparse SOH estimation framework based on clustered voltage segments and multi-head CNN under arbitrary charging condition

Rapid battery advances have accelerated electric vehicle adoption, but accurate state of health (SOH) estimation is still crucial and most methods depend on specific or complete charging cycles, limiting use under irregular real-world charging. To address this issue, this paper proposes a sparse SOH estimation framework based on partial charging segments. The proposed method first extracts incremental capacity features from randomly sampled short voltage segments, and then applies K-means clustering to group segments with similar aging patterns, which reduces model complexity and improves feature robustness. Next, a cluster aware Multi-head convolutional neural network is constructed. Each head models features within its assigned cluster, and the learned aging representations are fused through an attention mechanism to suppress low quality branches and strengthen informative branches. The fused representation is fed into a fully connected layer to obtain the final SOH estimate. Experimental results show that the framework maintains high predictive accuracy and robustness across datasets spanning diverse chemistries, C-rates, and charging protocols. When the input voltage window is constrained to 0.11 V, the RMSE is approximately 1.7 %. Meanwhile, the model is sparse and efficient, with 137.9K parameters and an inference latency of about 0.9 ms, demonstrating its potential for deployment in automotive battery management systems.