Resumen de Robust structural damage detection by using statistical hybrid algorithms
This thesis presents the results of applying a statistical hybrid approach for structural health monitoring using piezo actuating signals. Where, by combining statistical processing based on Principal Component Analysis (PCA), cross-correlation functions and pattern recognition methods it was possible to detect, classify and locate damages under varying environmental conditions and possible sensor faults. The proposed methodology consists of first transmiting/sensing guided waves along the monitored structure surface by using piezoelectric (PZT) devices. Then, cross-correlated piezoelectric signals are statistically represented by means of a PCA model. Later, damages are identified through error indexes computed from a statistical baseline model. Finally, clustering methods and scattered plots are used to verify the performance of the proposed algorithm. Improved or new techniques are presented in this thesis which were focused to achieve more reliable diagnosis with high robustness and good performance. Specifically, differential genetic algorithms are used for automatically tuning parameters in a PCA-SOM damage detection/classification approach. Additionally, Ensemble Learning is explored as approach for obtaining more efficient diagnosis with high separable boundaries between undamaged and damage conditions taking advantages of learner algorithms built from Non-Linear PCA and a Multiactuacting active scheme of piezodiagnostics. Also, a modified version of the Reconstruction Algorithm for Probabilistic Inspection of Damage – RAPID is implemented to solve location tasks in SHM. The proposed methodology was experimentally evaluated on different structures such a a carbon-steel pipe loop, a laminate plate, aircraft wings and a scale tower wind, among others; where different damage scenarios were studied, including leaks scenarios, mass adding and cuts. The effectiveness of the proposed methodology to detect, locate and classify damages under varying environmental and operational conditions is demonstrated. Likewise, the feasibility for continuous monitoring is validated by embedding the code of the proposed algorithm whose capacity to detect structural damages was demonstrated. As a result, the combination of piezodiagnostics approach, cross-correlation analysis, principal component analysis, clustering techniques and Ensemble Learning become as promising solution in the field of structural health monitoring and specifically to achieve a robust solution for damage detection and location.
