Development of Sensors for Structural Health Monitoring in Aerospace Applications by using a Machine Learning Approach

The inverse problems related to the characterization of piezoelectric composites using measurable properties pose strong requirements in terms of computationally feasible calculations, especially if partial experimental information is given. This study presents a comprehensive reverse engineering procedure that combines computationally efficient homogenization calculations by use of the Mechanics Of Structure Genome (MSG) concept through SwiftComp, machine learning modeling, and multi-objective optimization, making fast-paced characterization highly feasible. The proposed model combines MSG via SwiftComp for Representative Volume Element homogenization, utilizing machine learning models like Gaussian Process Regression (GPR), Artificial Neural Network (ANN), Random Forest (RF), Gradient Boosting Machine (GBM), and Support Vector Regression (SVR), achieving a significantly improved computational speedup (400×) while maintaining prediction errors below 5%. Validation is carried out using benchmark test case of an Active Fiber Composite (AFC) extracted from the literature, ensuring accurate results for every single combination of the elastic, piezoelectric, and dielectric constants. The study also compares the performance characteristics of different optimization algorithms for scenarios involving both comprehensive and partially (incomplete) experimental information from references, setting critical paths for a better interpretation, such as the use of single-objective optimization for scenarios requiring comprehensive experimental information, and multi-objective optimization when problems become theoretically unsolvable. Additionally, the study performs critical analyses by demonstrating that it is theoretically unfeasible for single-objective optimization using partially comprehensive experimental information, accepting perfect model fitting, and will catastrophically fail regardless, and how dual-objective optimization with exact regularization can guarantee physically viable numerical values.