Accurately estimating the state of charge (SOC) of lithium batteries is essential for an efficient battery energy management system.Traditional SOC estimation algorithms often struggle with inaccurate feature extraction from sampled signals and insufficient accuracy over long time series.This paper proposes a lithium battery SOC estimation method that combines improved wavelet threshold denoising (WTD) with deep learning to enhance both accuracy and stability.First, we apply an enhanced WTD algorithm to the sampled signals of lithium batteries to extract original features.This improved algorithm optimizes threshold selection based on traditional WTD, ensuring accurate signal feature extraction.Next, we input the processed signals into a multilayer long short-term memory (LSTM) neural network that incorporates an attention mechanism.By focusing on key signal features at critical time points, this approach enhances the accuracy of SOC predictions.Finally, experiments validate the effectiveness of our proposed method.Comparison results indicate that our method can control SOC estimation error within 1%, outperforming traditional methods.This demonstrates significant potential and practical value for our fusion-based WTD and deep learning approach in real-world applications.