This paper employs four automatic detection methods for seismic signals: short-long time window ratio, multi-band filtering, deep learning, and template matching filtering. These methods are applied to process the continuous waveform data recorded by the YUS, ZOD, and YUT stations, which are located closest to the epicenter, for three significant seismic events with clear foreshock signals: the M7.1 earthquake in Yushu, Qinghai on April 14, 2010; the M5.9 earthquake at the junction of Shangri-La, Deqin, and Derong in Yunnan on August 31, 2013; and the M7.3 earthquake in Yutian, Xinjiang on February 12, 2014. The study explores the application effects of the four automatic phase-picking algorithms in detecting foreshock signals, supplementing foreshock activity information. By comparing the missed detection rates and false detection rates of different detection methods, it is proposed to use deep learning detection results to validate and complement the short-long time window ratio and multi-band filtering results. These results will serve as templates for matching and filtering, while also fully utilizing GPD for the automatic identification of P and S phases, which aids in improving foreshock phase information.