Abstract: The characteristics of earth-rock dam compaction are crucial to construction quality. Previous predictions mainly focused on the single-output regression of the physical, mechanical and seepage compaction characteristics, lacking consideration of the correlation among the objectives of different compaction characteristics. To address this issue, we develop an improved multi-output Gaussian process regression (IMO-GPR) model that builds target-specific features using density-based spatial clustering of applications with noise and extends the MO-GPR model input space to improve its decoupling capability of complicated mapping in the high-dimensional feature space. This improved model considers effectively the correlation between multi-output compaction characteristic objectives through combining with the output covariance coefficient matrix used by MO-GPR, and can realize accurate predictions of multi-output dam material compaction characteristics. Compared with traditional GPR, MO-ELM, and MO-GPR models, its prediction accuracy is 24%, 20% and 17% higher respectively, and it has stronger robustness in the cases of noise interference, abnormal data, and insufficient data.

Key words: earth-rock dam material, compaction characteristics, improved multi-output Gaussian process regression model, target-specific feature, objective correlation