Abstract: It is of great significance to carry out machine learning and early warning of precursor characteristics of various types of rock burst for ensuring the safety of rock burst mines. Based on I010203 fully mechanized caving face with various types of impact in Kuangou Coal Mine, the main types of rock burst occurred in this face were analyzed by means of field investigation, theoretical analysis and machine learning, and the distribution characteristics of microseismic energy level and frequency during mining under solid coal and in the transition area between solid coal and goaf were studied, and the monitoring and early warning indicators of rock burst time series were optimized, and the precursor characteristics of large energy events were analyzed. A time risk level early warning model (Bayes-LSTM) combining Bayesian optimization algorithm (Bayes) with long-term and short-term memory network (LSTM) is constructed, and the model is tested in engineering practice. The research results show that the large-energy induced impact events in this working face mainly occur in the area under the solid coal, especially in the area near the coal pillar, and the focal energy is greater than other areas, and the induced impact events are mostly concentrated in the range of 50–150 meters ahead of the working face. Microseismic events show a trend of high energy and high frequency in the middle position of working face under solid coal, and the coal pillar area is dominated by high energy, and the transition area between solid coal and goaf is a high frequency area. In contrast, the energy and frequency under goaf are obviously low. B value, A value, A(b) value, lack of earthquake, δ F, S and H(t) are determined as the multivariate monitoring and early warning indicators of rock burst in the time series of the working face, and the stage when the early warning indicator curve evolves to the extreme value, drops near the extreme value or suddenly increases to a higher value indicates the precursor of large energy events. The established model realizes the graded early warning of daily impact risk. The confusion matrix is used to analyze the test results of Bayesian optimization algorithm combined with long-term and short-term memory networks. The accuracy of the test set is 84.8%, which can accurately warn most “strong” level events. The model has good applicability and practicability in the time early warning of rock burst, and the research results provide technical support for rock burst mine monitoring and disaster prevention and control.