Abstract: Accurate state-of-health（SOH） estimation of lithium-ion batteries is crucial for reducing risks and losses. However, the complex electrochemical mechanisms within the battery and the user-specific operating conditions pose significant challenges to this task. Current prediction methods based on modal decomposition often fall short in accuracy because they cannot simultaneously handle the multi-scale information related to both short-term capacity growth and long-term performance degradation specific to batteries. To address these limitations, a novel method for predicting battery SOH based on global multi-head self-attention mechanism and local multi-channel feature fusion is proposed. First, complete ensemble empirical mode decomposition with adaptive noise（CEEMDAN） is utilized to obtain multi-scale information on battery performance degradation. Then, a global self-attention mechanism based on the Softmax function is employed to capture long-term degradation features of the battery, while using various depthwise separable convolutional neural networks to extract short-term multi-channel and multi-scale information. By adaptively fusing these long-term and short-term features through residual connections with weight parameters, precise battery SOH prediction is achieved. Experimental results based on NASA and CALCE battery datasets demonstrate that the proposed method not only achieves higher prediction accuracy compared to mainstream model architectures, but also exhibits stronger generalization capability.