Long-term seismicity is an effective tool to infer fault properties at depth, but the catalog construction is challenging because of the large data volume. We propose a new deep learning-based workflow that follows a “Train-Detect-Pick” procedure, which solves the generalization problem in AI pickers. We apply the new workflow on the preseismic phase (2008-2019) of Ridgecrest-Coso region. Results show that the new workflow realizes efficient and stable detection, and well substitutes matched filter. Our new catalog helps characterize the preseismic fault behavior: (1) the Ridgecrest area has a distributed deformation, and the 2019-ruptured segment has a persistent asperity; (2) the central Garlock fault is unfavorable for rupture propagation, because of its discontinuous geometry and low coupling ratio; (3) the Coso geothermal field generates intense and shallow seismicity, which has a high b-value that does not correlate with seismicity rate and industrial production, thus suggest a low stress level.