A novel bayesian entropy network for probabilistic damage detection and classification

This paper introduced a novel Bayesian Entropy Network (BEN) classifier that can achieve fast learning by the ability to handle extra types of knowledge (e.g., not in the form of direct point observations). Such a classifier is based on the simple Bayes theorem and maximum entropy principle. The additional information such as mean, variance or range data about a certain feature can be coded together with the classical direct point observations. These knowledges were given in the form of constraints using the maximum relative entropy principle. The classifier is compared with a simple Naïve Bayes network. By encoding extra information into the network, the classifier behaves better when the training data size is small. In a simple example given in the presented work, the proposed BEN classifier can achieve fast learning comparing to the Naïve Bayes classifier. The proposed method is applied and demonstrated for composite structure damage detection and classification problem using numerical experiments (ongoing). The proposed method is shown to have advantages where there is not enough training data but empirical knowledge of certain feature is available. Some future work is listed in the conclusion part.