Abstract: We relocate the 1017 earthquakes provided by the Sichuan Earthquake Agency and then use the 578 relocated earthquakes with a high signal-to-noise ratio as templates. The continuous waveform data recorded by 7 stations of Zipingpu reservoir seismic network from 2005 to 2008 are scanned by graphics processing unit-based match and locate （GPU-M&L） method. Then, based on the denoising results by applying the deep learning algorithm called DeepDenoiser, a convolutional neural network model is designed to further verify these newly detected events. Finally, the double-difference location method is used to accurately relocate them. A total of 16836 events are eventually identified, which is about 13 times as many as the events listed in the local catalog of Sichuan Earthquake Agency. The magnitude of completeness is reduced from M_L1.4 given by the local catalog to M_L−0.1. The relocated results show that the seismic events in the study area are linearly distributed in the northeast direction. The dominant focal depth indicates the location of the detachment layer in the crust of the region. Combined with the analysis results of b value and the focal depth, it is speculated that the earthquakes after impoundment in the study area are mainly natural seismic activities caused by the accumulation of tectonic stress, accompanied by the occurrence of reservoir-triggered seismicity.