关键词: 岩体结构面, 多参数, 分步聚类, t-SNE, K-means算法

Abstract: Clustering of rock discontinuities is crucial for evaluating rock mass stability. The conventional clustering methods often rely on the orientations of rock discontinuities, without considering the influence of physical characteristics on rock mass stability. To address the limitations of single-factor grouping, a stepwise clustering method of rock discontinuities dominated by multivariate parameters based on student-distributed stochastic neighbor embedding (t-SNE) is proposed. This method takes into account the effects of dip direction, dip angle, trace length, opening, filling state and roughness of rock discontinuities. Firstly, the t-SNE algorithm is used to reduce the dimensionality of discontinuity characteristics except for the orientations. Subsequently, the simulated annealing algorithm is employed to search for the global optimal initial values of the K-means algorithm, and the stepwise clustering idea is utilized to accomplish the clustering. The research shows that the proposed method addresses the sparsity issue of high-dimensional data while preserving the local and global structures of the data. Compared to the conventional methods, the proposed method achieves more accurate partitioning of physical characteristics within the spatial distribution similarity zone, resulting in higher grouping accuracy. Furthermore, the proposed method effectively distinguishes the differences between orientations and physical characteristic parameters on rock mass stability without the need for complex weight value calculations. Finally, the proposed method is applied to the measured data of rock discontinuities in an open-pit slope in Xinjiang, China. The grouping results are found to be reasonable and reliable, which further validates the effectiveness of the proposed method in practical engineering. This research provides a reference for stepwise clustering of multi-parameter rock discontinuities.

Key words: rock discontinuities, multivariate parameters, stepwise clustering, student-distributed stochastic neighbor embedding (t-SNE), K-means algorithm