<p>Acoustic emission (AE) and electromagnetic radiation (EMR) can adequately respond to rock burst precursor information. Due to their different propagation modes and frequency–response ranges in coal and rock fracture, the response signals of precursor show diversity and chaotic characteristic, resulting in a single monitoring means or index cannot meet the requirements of accurate early warning. Moreover, the existing level of early warning based on real-time precursor response needs to be improved. This paper presents a multi-level fusion warning method considering the chaotic characteristics of AE and EMR. First, the AE and EMR monitoring signals in real time are pre-processed by morphological filtering. According to the phase space reconstruction, the corresponding parameters, such as delay time, embedding dimension, and Lyapunov index, are calculated to obtain the chaotic characteristics of different monitoring indicators, and the proposed chaotic-transformer model is input to predict their future change trend. Furtherly, the Symmetrized Dot Pattern (SDP) image feature mining method is introduced to perform fusion analysis of the four predicted index signals. Finally, the SDP fusion features are provided into the Swin-transformer model to realize multi-level fusion early warning of rock burst under the different precursor response. A case study in a coal mine in Liaoning Province shows that under different types of mine earthquakes, the average advance warning time and accuracy of the proposed method can reach 61.2&#xa0;h and 94.2%, respectively, which ensures the feasibility of the method and provides guarantee for the safe mining of coal.</p>