Soil liquefaction assessment by using hierarchical Gaussian Process model with integrated feature and instance based domain adaption for multiple data sources

For soil liquefaction prediction from multiple data sources, this study designs a hierarchical machine learning model based on deep feature extraction and Gaussian Process with integrated domain adaption techniques. The proposed model first combines deep fisher discriminant analysis (DDA) and Gaussian Process (GP) in a unified framework, so as to extract deep discriminant features and enhance the model performance for classification. To deliver fair evaluation, the classifier is validated in the approach of repeated stratified K-fold cross validation. Then, five different data resources are presented to further verify the model’s robustness and generality. To reuse the gained knowledge from the existing data sources and enhance the generality of the predictive model, a domain adaption approach is formulated by combing a deep Autoencoder with TrAdaboost, to achieve good performance over different data records from both the in-situ and laboratory observations. After comparing the proposed model with classical machine learning models, such as supported vector machine, as well as with the state-of-art ensemble learning models, it is found that, regarding seismic-induced liquefaction prediction, the predicted results of this model show high accuracy on all datasets both in the repeated cross validation and Wilcoxon signed rank test. Finally, a sensitivity analysis is made on the DDA-GP model to reveal the features that may significantly affect the liquefaction.