To address the issue of motion artifact interference faced by wearable ECG monitoring devices in dynamic environments, this paper proposes an adaptive motion artifact removal framework based on improved Variational Mode Decomposition (VMD). By designing a parameter self-adjustment mechanism and a multi-feature fusion mode selection strategy, the algorithm's adaptability to non-stationary ECG signals and noise separation accuracy are enhanced. Experiments on the MIT-BIH Arrhythmia Database demonstrate that the improved VMD algorithm outperforms traditional wavelet transform, Recursive Least Squares (RLS), and conventional VMD methods in multiple performance metrics. Specifically, the signal-to-noise ratio (SNR) is improved by 5.17 dB, the Percentage Root Mean Squared Difference (PRD) is reduced to 49.13%, the correlation coefficient is increased to 0.88, and high real-time processing capability (Real-Time Processing Ratio, RTR = 22.5) is maintained, meeting the low-latency requirements of wearable devices. Moreover, case studies on pathological recordings (e.g., Wolff-Parkinson-White syndrome and third-degree atrioventricular block) reveal that the improved VMD better preserves clinically significant features such as delta waves and dissociated P waves. Furthermore, a downstream arrhythmia classification task using a CWT-CNN classifier achieves 91.67% accuracy on denoised heartbeats, which is 2.67 percentage points higher than that on raw noisy signals (89.00%), confirming the practical benefit of the proposed preprocessing for AI-based diagnosis. This study provides an effective processing solution for improving the signal quality of wearable ECG monitoring.