To reveal the tensile fracture and instability collapse mechanism of granite roofs in deep mining, three types of samples, namely fine-grained Luotian granite, coarse-grained Golden Ma granite, and medium-grained Yunfu granite, were selected for three-point bending tests. Combined with acoustic emission monitoring technology, the bending strength, dominant frequency distribution, damage evolution, and crack propagation patterns were systematically investigated. The results show that the bending strength of granite exhibits a significant particle size effect. The fine-grained Luotian granite has the highest strength of 17.79 MPa, with a sudden post-peak stress drop indicating brittle fracture. The coarse-grained Golden Ma granite has the lowest strength of 5.33 MPa, showing pronounced ductility characteristics. The strength of medium-grained Yunfu granite lies between the above two rock types, reaching 16.04 MPa, and exhibits brittle-ductile transition behavior after the peak. The dominant frequency signals of acoustic emission mainly appear near the peak load and in the post-peak stage, forming three main frequency bands of 0-40 kHz, 50-80 kHz, and 130-180 kHz. The high-frequency signals (130-180 kHz) of Luotian granite surge before the peak, while the dominant frequencies of Golden Ma granite and Yunfu granite are concentrated in the low-frequency band (0-40 kHz) after the peak. The damage evolution of granite undergoes three stages: initial damage, stable development, and accelerated development. Damage accumulation is highly concentrated in the post-peak stage. Golden Ma granite has the highest damage proportion in the post-peak stage, and Luotian granite releases the largest strain energy per unit damage. RA-AF crack mode analysis indicates that the failure of the samples is dominated by tensile cracks, and the proportion of tensile cracks is negatively correlated with mineral grain size. The proportion of tensile cracks in Luotian granite is 3.17 and 6.15 percentage points higher than that in Yunfu and Golden Ma granite, respectively. The quantitative early warning method for roof instability based on dominant frequency signals can provide a theoretical basis and engineering guidance for roof stability evaluation and dynamic disaster prevention and control in deep mining.