A Grad-CAM based visualizing method for important regions is proposed for the interpretability of RF fingerprint extraction and deep learning models of time-domain pulse signal samples. The impact of important regions on RF fingerprint recognition results is analyzed through multiple mask tests of important regions. Based on signal samples of 10 emitters, the test results of two ResNet models with different layers are compared. It is found that the proposed method can distinguish different types of signals and present individual differences. Analysis shows that this method can detect important regional localization differences when different emitters send the same signal, and can visually reflect the spatial distance of RF fingerprint characteristics, as well as the differences in feature representation and fingerprint localization accuracy of different models; At the same time, it is found that masks for important areas are more prone to false predictions, which proves the existence of RF fingerprints related to time-frequency characteristics in specific signals, and can assist in visualizing key points that affect the recognition of RF fingerprint samples.