<p>Obtaining the precursor information of the acoustic emission (AE) and the quantitative identification of the damaged state of rocks during the spalling process can provide beneficial information for monitoring and accurate early warning of engineering geology disasters in deep underground engineering. In this study, a laboratory experiment with granodiorite specimens in biaxial compression was conducted to produce the rock spalling using a servo-controlled loading machine; during the testing process, the AE system was utilized to obtain the AE signals emitted by the rock fracture throughout the entire spalling process. The AE precursor of spalling failure is investigated by analyzing AE signals in the time-frequency domain; the influence of intermediate principal stress on AE precursor is studied. Subsequently, a damage identification model was proposed with four typical AE parameters exhibiting significant precursors during the spalling process (b-value, AE counts rate, main frequency, and amplitude) used as feature inputs and the damage degree as the desired output. Then, the proposed model was trained using the back-propagation (BP) neural network. The results indicated that the b-values were all characterized by a continuous decline immediately before the spalling, accompanied by a sudden increase in the AE count rate and the appearance of low-frequency signals with high amplitudes. These precursors provide beneficial information to predict an upcoming spalling. In addition, the evaluation of the damage identification model revealed that the evolution of predicted damage and the actual damage curve was consistent. The present study makes achieving a reliable damage prediction and early warning of spalling failure possible using the AE precursors and the proposed damage identification model.</p>