Sleep plays a crucial role in maintaining human health and improving quality of life. However, traditional manual sleep staging methods are not only time-consuming but also heavily reliant on expert experience, limiting their feasibility for largescale applications. Therefore, developing high-precision and fully automated sleep staging methods is essential for assisting clinical diagnosis. To address this research need, we propose an innovative automatic sleep staging network, DilatedSleepNet. This model introduces a novel multi-scale dilated convolution strategy to effectively capture the waveform characteristics of EEG signals, enabling accurate sleep stage classification using only single-channel EEG input. We systematically evaluated the performance of DilatedSleepNet on three publicly available datasets, achieving classification accuracies of 86.8%, 83.2%, and 85.4%, respectively. Experimental results demonstrate that DilatedSleepNet exhibits outstanding generalization ability and robustness across multiple datasets, providing a strong technical foundation for the diagnosis and research of sleep-related disorders.