Scaled boundary finite element based two-level learning approach for structural flaw identification

Abstract

A comprehensive framework that combines classification and regression techniques through machine learning (ML) algorithms to address the inverse problem of flaw identification and quantification is proposed. The framework uses the scaled boundary finite element method (SBFEM) and quadtree (octree mesh in 3D) to compute displacements for various flaw shapes and specified boundary conditions. The flaw data, such as their numbers, location, and size, along with the displacement data, serve as the input for ML models. A random forest model is used to predict the number of flaws, while an artificial neural network (ANN) employing a fully connected multi-layer perceptron is utilized to predict flaw locations and sizes. The framework underwent rigorous testing to identify the number and location of multiple flaws (circular and elliptical) in 2D structures. The accuracy of the classification and regression models for flaw detection, ranging from one to four circular flaws, is observed to be 98%. As an extension of its capabilities, the proposed model successfully identifies multiple flaws and their features in 3D structures as well with 98% accuracy, demonstrating the robustness of the proposed approach.