Our facial skin presents subtle color change known as remote Photoplethysmography (rPPG) signal, from which we could extract the heart rate of the subject. Recently many deep learning methods and related datasets on rPPG signal extraction are proposed. However, because of the time consumption blood flowing through our body and other factors, label waves such as BVP signals have uncertain delays with real rPPG signals in some datasets, which results in the difficulty on training of networks which output predicted rPPG waves directly. In this paper, by analyzing the common characteristics on rhythm and periodicity of rPPG signals and label waves, we propose a whole set of training methodology which wraps these networks so that they could remain efficient when be trained at the presence of frequent uncertain delay in datasets and gain more precise and robust heart rate prediction results than other delay-free rPPG extraction methods.