Bayesian incremental learning paradigm for online monitoring of dam behavior considering global uncertainty

Abstract

Advanced data-driven methods for real-time monitoring of dam structural behavior, especially displacement, are obtaining increasing popularity. Most existing data-driven models are offline and static, however, which cannot be dynamically updated with the increase of monitoring data. Additionally, the capability of monitoring models to identify anomalies, which is the ultimate goal of dam behavior modeling, has received little attention. This study proposes a novel Extreme Learning Machine (ELM)-based Bayesian incremental learning methodology to address these challenges. The proposed model is termed Incremental Global Bayesian ELM (I-GBELM), which couples ELM with global Bayesian and incremental learning for online monitoring of dam displacements. Firstly, a global Bayesian approach to the ELM is presented to account for both input and target uncertainties and to infer the network parameters analytically. Secondly, an incremental learning paradigm is theoretically derived, which aims to directly learn new knowledge from additional data to update model parameters without retraining from scratch. The I-GBELM model integrates the strengths of both methods so that it allows accurately predicting displacements, identifying anomalies, and efficiently updating parameters. For online deployment purposes, finally, a non-iterative pruning algorithm is developed to adaptively discard the redundant hidden nodes, yielding a more compact model. The performance of the proposed model is fully validated through comprehensive verification cases using benchmark and real-world monitoring datasets. The empirical results indicate that the proposed I-GBELM outperforms all other models in most cases, and hence, is a competitive modeling tool for online structural health monitoring.