In deep coal mine strata, characterized by high ground stress and extensive fracturing, predicting the strength of fractured rock masses is crucial for stability analysis of the surrounding rock in coal mine strata. In this study, rock samples were obtained from construction sites in deep coal mine strata and intact, as well as fissured, rock specimens were prepared and subjected to triaxial compression tests. A numerical model based on the discrete element method was then established and the micro-parameters were calibrated. A total of 288 triaxial compression tests on the rock specimens under different conditions of confining pressure, loading rate, fissure dip angle, and fissure length, were conducted to obtain the triaxial compressive strength of the fractured rock specimens under different conditions. To address the limitations of traditional back propagation (BP) neural networks in solving stochastic problems, a modified BP neural network model was developed using a random factor and an interlayer mean square error corrected network model evaluation function. The traditional and modified BP neural network models were then employed to predict the triaxial compressive strength of the fractured rock specimens. Through comparative analysis, it was found that the modified BP neural network prediction model exhibited smaller errors and significantly reduced overfitting, making it an effective tool for predicting the strength of fractured rocks in deep coal mine strata.