PCA–GPR-Assisted Sequential Bayesian Inversion of Slope Mechanical Parameters from Multi-Stage Deep Horizontal Displacement Monitoring

Reliable mechanical parameters are needed to predict deformation during staged slope excavation, yet deep inclinometer profiles are high-dimensional and repeated numerical inversion is costly. This study integrates principal component analysis, Gaussian process regression (GPR), and sequential Bayesian updating to infer six effective parameters from multi-stage horizontal displacement profiles of a highway slope in Shaoyang, China. FLAC3D simulations were performed for a 120-point Latin hypercube design. Three principal components explained over 99% of profile variance. Ten repetitions of five-fold cross-validation yielded mean PCA–GPR R2 values of 0.9769–0.9945. Empirical coverages of the 95% marginal prediction intervals ranged from 93.2% to 96.0%, and the GPR predictive covariance was included in the likelihood. Truncated multivariate Gaussian approximations were used to transfer posterior means and covariance structures between excavation stages. A five-strategy comparison showed that adding Stage 3 reduced parameter standard deviations by 10.0–22.0%, followed by a further 11.1–32.2% reduction after Stage 4. Sensitivity and posterior-contraction analyses indicated stronger constraints on the stiffness parameters, whereas ϕ2 remained weakly identifiable. The final posterior means reproduced the spatially held-out, within-site CX-2 profile with a mean absolute error of 0.18 mm. The framework quantifies the parameter-specific information gained from complete profiles across excavation stages; its findings remain conditional on the monitored site and the adopted modeling and error assumptions.