Rock Discontinuity Parameter Generation Based on Differentiated WGAN-GP Models: Heterogeneous-Architecture Ensemble and Source-Wise Reconstruction

The generation of rock discontinuity parameters constitutes a fundamental prerequisite for equivalent rock mass construction and rock engineering stability analysis. Conventional Monte Carlo (MC) independent sampling fails to reproduce the correlation structure among parameters. Building upon heterogeneous Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) models, this study proposes a heterogeneous-architecture ensemble method and a source-wise reconstruction method. The heterogeneous-architecture ensemble method performs row-level weighted combination of samples generated by distinct individual models. The source-wise reconstruction method restores marginal distributions and the linear dependence structure through parameter-column-wise source selection, empirical-quantile calibration, and Iman–Conover rank reordering. Using 122 measured discontinuities from an open-pit mine slope, the proposed methods were evaluated using three metrics: the mean KS statistic across four parameters, the normalized Wasserstein distance, and the mean absolute error of six Pearson correlation coefficients. The results show that the source-wise reconstruction method achieved the lowest mean KS statistic (0.168) and normalized Wasserstein distance (0.046) among the six methods. The heterogeneous architecture ensemble reduced the correlation error from the individual-model mean of 0.264 to 0.248. Both methods outperformed the Monte Carlo method, which yielded a correlation error of 0.326. Ablation experiments indicated that the performance improvement mainly resulted from quantile calibration and rank reordering. The proposed methods preserve the linear correlations among parameters while maintaining accurate marginal distribution fits, providing a feasible approach for generating discontinuity parameters.