Abstract: Random characteristic rock fractures are vital to stability analysis of the rock mass on slope in hydraulic engineering, but previous methods are based on a univariate or bivariate statistical models and neglect the dependencies between different fracture parameters. This paper presents a new artificial neural network method for joint simulations of the multi-dimensional parameters of such rock fractures, namely ensemble wasserstein generative adversarial network (E-WGAN) that is based on the generative adversarial networks and ensemble learning. Compared with the traditional methods, this method improves the description of the multi-dimensional distribution characteristics of the fractures. Through examining a set of fracture data for the rock mass cases of three parameters (trace length, strike, and aperture), we show it can accurately model the three variates and their relationships that the traditional methods fail to express. A comparison of the discrete fracture networks generated from simulation samples shows that the trace maps simulated using E-WGAN samples are closer to the real trace maps. Besides, this algorithm is also applicable to higher dimension cases in geological analysis and has a broad prospect of application in rock mass analysis.

Key words: slope rock mass, generative adversarial network, multi-dimensional parameter, discrete fracture network, trace map