To address the limitation of existing data-driven models for battery state-of-health (SOH) estimation that neglect electrochemical mechanism constraints, a deep learning framework integrated with electrochemical mechanism constraints is proposed. First, a hybrid neural network architecture based on convolutional neural network-gated recurrent unit-autoencoder (CNN-GRU-AE) is designed to collaboratively extract temporal features from battery data. The CNN unit captures local degradation features to obtain feature vectors, while the GRU-AE unit models temporal dependencies and computes data reconstruction loss. To ensure consistency with electrochemical mechanisms during SOH estimation, a coupled degradation model is embedded into the framework. This model integrates a linear capacity degradation component, a nonlinear active lithium decay model, and a solid electrolyte interphase (SEI) growth mechanism. The entire framework is optimized via differentiable programming, enabling simultaneous learning of neural network weights and mechanistic parameters. Coupled with a dual-task learning architecture, it simultaneously realizes data reconstruction and lithium-ion battery SOH estimation. Finally, experimental results demonstrate that, compared to other models, the proposed model enhances both the accuracy and robustness of lithium-ion battery SOH estimation, achieving deep coupling between electrochemical mechanisms and data-driven modelling.