A novel damage identification algorithm by combining the boundary element method and a series connection neural network

Abstract

A novel damage identification approach based on a model-driven and a data-driven combined algorithm is developed. By using this approach with only boundary strains, the existence, location, classification as well as the extent of the damage in a plate can be predicted at once with high accuracy and efficiency. To accumulate the data, the boundary element method (BEM) is applied as a model-driven in modeling plates with one or multiple damages in the forms of circular or elliptical holes or cracks and for solving the boundary strains of the defective plates. The dimensionality reduction and semi-analytical characteristics of BEM not only can compress the feature data also can improve the accuracy of the database for the data-driven algorithm. The boundary strains are obtained directly from BEM models, which are also easily collected through the use of strain gauges mounted on the surfaces of structures being monitored in real applications. A series connection neural network algorithm is established to accomplish the novel damage identification assignments with deep learning. The number of the existing damages is firstly detected by a classification neural network model, then the extracted features are transmitted to the corresponding regression neural network model to prognosis the location, classification as well as the extent of each flaw. A high accuracy of about 99.86 % is achieved by the present combined neural network algorithm, which is promising in applications of actual structural health monitoring.