During the excavation of deep coal roadways, affected by high in-situ stress, abnormal stress in the roadway surrounding rock can easily induce disasters such as coal bumps, roof falls, and rib spalling. Accurate positioning and monitoring are the prerequisite and foundation for preventing and controlling these disasters. Taking the deep excavated coal roadway in Zhaozhuang Coal Mine as the engineering background, a highly sensitive high-frequency microseismic positioning and monitoring system was established. To address the challenge of identifying complex and weak microseismic signals induced by coal roadway excavation, a denoising algorithm based on unsupervised deep learning and attention mechanism was introduced, achieving adaptive suppression of broadband noise. Based on the improved ELM intelligent algorithm, accurate and efficient picking of weak microseismic events was realized. The Hilbert-Huang Transform (HHT) time-frequency analysis method was used to deeply analyze the time-frequency characteristics of three typical microseismic waveforms, namely small-energy, large abnormal stress-energy, and coal bump waveforms, realizing accurate identification of complex microseismic waveform signals induced by coal roadway excavation. On this basis, the source location of coal bump events induced by abnormal stress was carried out, and the characteristics of the microseismic monitoring response law were analyzed, revealing the coupling law between the microseismic monitoring response characteristics of "precursory rise - peak mutation - attenuation and stabilization" and the dynamic evolution process of "stress accumulation - coal bump occurrence - stress release" during the excavation process. The research results provide effective technical support for the early identification and early warning of surrounding rock damage and catastrophes during the excavation of deep coal roadways.