To improve the accuracy of dynamic disaster prediction and control in coal mines, uniaxial compression tests were conducted on different sandstone samples to obtain the stress-strain characteristics and variations in elastic modulus during the uniaxial compression process. By combining acoustic emission monitoring, energy-stress relationships, and Matlab modeling, the stress values and positioning coordinates of internal acoustic emission source points in different sandstone samples were calculated. A random forest algorithm was used to construct a mapping model between acoustic emission parameters and stress, yielding predicted stress-time curves and stress values for the internal units of the sandstone. By integrating machine learning and cubic interpolation methods, the evolution patterns of local stress fields in different sandstone samples were inverted and compared with surface stress monitoring using VIC-3D and post-peak failure morphology. The results indicate that the established mapping relationship between acoustic emission parameters and stress significantly improved the prediction accuracy of stress, especially in regions of local stress concentration, and confirmed the reliability of local stress field inversion.