Since 2006, North Korea conducted 6 underground nuclear explosions at its test site near the China-North Korea border. Due to its large magnitude of mb(Lg)=5.56, the explosion detonated on 3, September 2017 caused subsequent cavity collapses, i.e., the 3 September 2017 event [mb(Lg)=3.95] 8 minutes after the test, the 23 September and 12 October 2017 events [mb(Lg)=3.3 and 2.9] a few weeks later. These collapses had been suspected to be nuclear tests. Both explosions and their aftershocks generated abundant broad band digital seismograms within regional distances.  Seismograms were collected to investigate event attributes, including the discrimination, yield estimation, and relocations of epicenters. The spectral ratios Pg/Lg, Pn/Lg, and Pn/Sn were calculated for all 6 explosions, 3 aftershocks caused by cavity collapses, 4 nearby nature earthquakes, and a mining earthquake. These sources show apparently different features in their spectral ratios. Taking the logarithmic spectral ratio log(Pg/Lg) as an example, for explosions their values raise quickly from -0.8 to 0.25 between 0.5 and 4.0 Hz, and keep above 0.25 at frequencies higher than that. In contrast, the log(Pg/Lg) values for earthquakes raise slowly from -0.8 to near 0 between 0.5 and 6.0 Hz and keep below 0 above 6.0 Hz. As a comparison, the log(Pg/Lg) values for cavity collapses raise quickly at low frequencies, similar to those from explosions, but start to decrease at approximately 4-5 Hz. At high frequencies, their log(Pg/Lg) values are low, even below earthquake populations. The spectral ratio from mining earthquake is similar to that from the cavity collapse. The results demonstrated that the network averaged P/S type ratios can unambiguously separate explosions, cavity collapses, and natural earthquakes. This research was supported by the National Natural Science Foundation of China (grants 41674060, 41630210 and 41374065).