Abstract: In this paper,glass fiber composites with layered defects were selected as test materials and subjected to room temperature monotonic tensile test. The acoustic loading technology was used to dynamically monitor the loaded process. The K-means clustering analysis method was used to identify the damage patterns of the amplitude and peak frequencies of the acoustic emission signals,and the relationship between the damage evolution degree and the acoustic emission signal characteristics was discussed. A method for selecting precursor feature selection by calculating the timing signal density of the damage mode is proposed. Finally,with BP neural network,the precursor signal of instability and damage is detected and the instability state is predicted. The results show that the clustering of acoustic emission signal parameters can obtain the characteristic frequency of each damage mode. Through the feature comparison of different damage modes,the high frequency low amplitude is stronger than the intermediate frequency high amplitude signal. Using the signal timing intensity as the early warning threshold of the instability failure state can accurately and effectively predict the instability failure state. This study explores the damage evolution process and the precursory feature identification method of glass fiber composites with defects,and provides new ideas for large-scale instability failure,fracture and health monitoring methods in composite structures.