Localization and Quantification of Multiple Damages on One-Dimensional Beam with Sensitive Modal Data and Jensen-Shannon Divergence-Based Artificial Neural Network Regression

Abstract

Purpose: This paper introduces a new method to localize and quantify multiple damages in one-dimensional beams by simultaneously minimizing three loss functions in an Artificial Neural Network. It enhances modal amplitude sensitivity by elastically adding small mass, demonstrating that this approach yields more accurate results compared to traditional methods.

Design/methodology/approach: Free vibration dynamic Finite Element Analysis (FEA) is used to generate modal amplitude data for a damaged one-dimensional beam. Three loss functions—Jensen Shannon Divergence, Mean Squared Error, and Mean Absolute Error—are minimized simultaneously. Damaged Finite Element heights are converted into Probability Mass Functions for accurate loss reduction. The method is validated on beams with four boundary conditions: cantilever, simply supported, overhanging, and propped cantilever. Adding a small mass strip enhances input data, improving prediction accuracy over traditional methods.

Findings: This research shows that adding a small mass increases the sensitivity of modal amplitude data. It lowers total training and validation losses, reduces Jensen-Shannon divergences, and improves damage prediction accuracy in all four beam cases compared to no mass increment, demonstrating the effectiveness of this mass-enhanced approach.

Originality/value: The proposed Artificial Neural Network architecture uses three loss functions—Jensen Shannon Divergence, Mean Squared Error, and Mean Absolute Error—simultaneously, which is novel and unreported. Adding a small elastic mass generates extensive training data, enhancing accuracy. It uniquely predicts multiple damages in the 0.1 mm to 0.3 mm range with high precision, making the approach original and innovative.