Electrocardiogram (ECG) signal generation plays a critical role in medical data augmentation (e.g., training arrhythmia classifiers) and cardiovascular disease diagnosis (e.g., virtual patient modeling). However, existing generative models face challenges in simultaneously capturing multi-scale morphological features (e.g., P-waves, QRS complexes) and maintaining strict temporal consistency in ECG signals.This paper proposes a diffusion model-based generative framework that integrates multi-scale feature extraction with temporal consistency enhancement techniques. First, a hybrid architecture combining wavelet decomposition and dilated convolutional UNet is designed to hierarchically extract features at different time scales. Secondly, a temporal consistency enhancement method based on Dynamic Time Warping (DTW) loss function and a lightweight Transformer module is introduced to improve the physiological consistency of waveform transitions. Experimental results demonstrate that, compared to the single-scale diffusion model (DDPM), our model significantly improves fidelity (MSE: 0.032 vs. 0.041, a reduction of 22%), diversity (Fréchet distance: 24.6 vs. 28.3, a reduction of 13%), and clinical effectiveness (expert rating: 4.2/5 vs. 3.5/5) on the MIT-BIH and PTB databases. Ablation studies further validate the key roles of multi-scale feature extraction and the DTW loss function.