A good understanding of the precursor characteristics of rock failure is essential for geo-mechanical rock engineering. This paper proposes an inversion method for acoustic emission (AE) precursor signals based on a self-organizing map neural network. The feature of this method lies in a construction of cyclic segmentation iteration process. By segmenting and approximating the set of AE parameters, the AE precursor signals are extracted at 97% of the peak stress moment. The inversion results of the rock failure precursors in different lithology tests verified the rationality of this method. Compared with traditional AE precursor phenomena (including b-value decrease, fractal dimension decrease, and entropy sudden increase), the occurrence time of the precursor signals inverted in this study is closer to the time of rock failure. This indicates that these precursor signals are the approaching points from rock deformation to rock failure, proving the potential application value of these signals in short-term precursors and short-term warnings of rock failure. Considering the damage evolution characteristics of rock failure, the reasons for the generation of precursor signals were preliminarily explored, and the generation of precursor signals was attributed to the sudden increase in damage during the loading process. The obtained results will help develop a deeper understanding of the precursor phenomena of rock failure.