Autoencoders and Multidimensional Clustering Algorithms in the Identification of Engineering-Geological Elements

The article investigates the use of modern machine learning techniques in engineering geological analysis. Key methods include autoencoders for dimensionality reduction and uncovering hidden patterns, alongside clustering algorithms (k-means, fuzzy c-means, DBSCAN, GMM, and Self-Organizing Map (SOM)) to identify engineering geological elements (EGEs). Indicator kriging is utilized for spatial interpolation of categorical data. The study analyzes a dataset of 90 soil samples, focusing on parameters such as soil density, modulus of deformation, and internal friction angle, using PCA, Variance Threshold, and correlation analysis for feature selection. Among clustering algorithms, k-means demonstrated the best results (silhouette score = 0.635, Davies-Bouldin index = 0.454), making it ideal for soil transitional layers. Fuzzy c-means and DBSCAN showed advantages for transitional zones and irregular densities, respectively. Indicator kriging produced detailed spatial cross-sections using an exponential variogram, significantly reducing modeling uncertainty. Autoencoders without intermediate layers achieved high reconstruction accuracy (R2 = 0.72, MSE = 0.02), effectively supporting clustering in latent space. The study emphasizes the importance of proper normalization, particularly for heterogeneous datasets, and highlights the potential of machine learning to improve classification, interpolation, and spatial modeling of EGEs. Future research should address autoencoder sensitivity to normalization and explore feature selection techniques, such as random forests, to enhance robustness and scalability.