Abstract: Coal and gas outburst is one of the most severe coal-rock dynamic disasters in deep coal mining. The extreme geological conditions of “three highs, one low and one complexity” (high gas content, high ground stress, high structural destruction, low permeability, and complex geological structures) in Guizhou Province, formed by typical karst landforms, have made the outburst disaster mechanisms increasingly concealed and complex. Conventional contact-based static single-indicator prediction methods are no longer adequate for meeting the demands of continuous, dynamic and advanced early warning in deep mining operations. To overcome the technical bottleneck of precise early warning under complex geological conditions, this study, taking typical high-outburst coal mines in Guizhou as the engineering background and leveraging microseismic dynamic responses and artificial intelligence fusion technologies, proposes a comprehensive microseismic dynamic early warning technical system for coal seam outbursts, encompassing “fine data processing, geological anomaly identification, risk grading evaluation, and intelligent multi-source early warning.” At the signal processing level, a high-fidelity extraction and intelligent waveform recognition method for low signal-to-noise ratio microseismic signals is proposed, based on a cascaded architecture of frequency-domain fast singular value decomposition (FSVD) and hidden Markov model (HMM), achieving automated classification and elimination of operational interference signals (drilling, blasting, coal mining, etc.) from genuine coal-rock microseismic waveforms, with P-wave arrival picking errors reduced by over 70%. At the knowledge support level, a large language model (LLM) combined with a BiLSTM-CRF deep network is innovatively introduced to construct a gas outburst precursor knowledge graph encompassing four core dimensions (geological structures, gas parameters, microseismic responses, and manual observation phenomena), enabling high-precision extraction and structural transformation of outburst precursor entities from unstructured text. At the fusion decision-making level, a multi-source information deep fusion early warning algorithm based on the transferable belief model (TBM) is established, which quantitatively synthesizes microseismic dynamic physical indicators, gas monitoring time-series data, and knowledge graph prior knowledge through belief-level fusion and decision-level transformation mechanisms, effectively addressing the challenges of multi-source evidence conflicts and incomplete information conditions in reasoning and decision-making. A six-month field industrial trial was conducted at the No. 9 coal seam of Linhua Coal Mine in the northern Guizhou mining area. The results demonstrate that the FSVD denoising algorithm improves the signal-to-noise ratio of microseismic signals from below 5 dB to above 10 dB. The TBM-based multi-source fusion early warning model achieves an overall accuracy of 92% (23 out of 25), with 23 effective warnings issued and a zero missed alarm rate. In the scenario of crossing a concealed fault zone, the system achieved precise early warning 2.5 days in advance, successfully interrupting the disaster incubation chain. The proposed intelligent early warning method demonstrates significant advantages in early warning accuracy, advance response capability, and engineering practicality, providing reliable technical support and an engineering paradigm for continuous, dynamic and advanced prevention and control of coal and gas outburst disasters under deep complex geological conditions.