Bond Graph Based Bayesian Network (BGBN) for Damage Diagnosis in Structural Health Monitoring

Abstract

On one hand, Bond Graph (BG) is a graphical language allowing the representation of energy transfers within complex dynamic systems and appreciated for its causal modeling properties in different dynamic systems for damage detection and isolation. However, it requires a prior deterministic knowledge of the characteristics of the system components. On the other hand, Artificial Intelligence (AI) can deal with uncertain knowledge and incomplete information by taking the advantage of probability reasoning. Such reasoning is performed with the aid of Bayesian Network (BN) in which the variables of the system form a causality network based on Conditional Probability Distribution (CPD). Nevertheless, the accuracy of the BN highly depends on the prior knowledge of network ordering. We present a hybrid model based on bond graph and Bayesian network—Bond Graph Bayesian Network (BGBN)—for damage diagnosis in the context of Structural Health Monitoring (SHM). Here, our focus is to embed a BG model in a BN model, that enables us to benefit from the causal properties of both models: (i) BG as a skeleton of the BN to overcome the variable ordering issue and (ii) BN as a tool dealing with variable uncertainties by virtue of historical data, expert knowledge, and other monitoring-induced data. To this end, one-story shear-building is considered and transformed into BGBN. By introducing corresponding CPDs based on expert opinion, the network is activated. The effectiveness of the network is investigated by applying different structural measurements as evidences in order to find and rank the most probable causes. It is seen that by utilizing BG, not only the task of complete determination of the system variables is carried out, but also the diagnosis of damaged components from BN is enhanced.