Performance optimization of a diagnostic system based upon a simulated strain field for fatigue damage characterization

• Published in: Mechanical systems and signal processing 2013-11, Vol.40 (2), p.667-690
• Main Authors: Sbarufatti, C., Manes, A., Giglio, M.
• Format: Article
• Language: English
• Subjects: Algorithms; Artificial Neural Network; Computer simulation; Crack; Cracks; Damage; Diagnosis; Diagnostic systems; Fatigue failure; Finite Element Model; Position (location); Strain; Structural Health Monitoring
• ISSN: 0888-3270; 1096-1216
• DOI: 10.1016/j.ymssp.2013.06.003

The work presented hereafter is about the development of a diagnostic system for crack damage detection, localization and quantification on a typical metallic aeronautical structure (skin stiffened through riveted stringers). Crack detection and characterization are based upon strain field sensitivity to damage. The structural diagnosis is carried out by a dedicated smart algorithm (Artificial Neural Network) which is trained on a database of Finite Element simulations relative to damaged and undamaged conditions, providing the system with an accurate predictor at low overall cost. The algorithm, trained on numerical damage experience, is used in a simulated environment to provide reliable preliminary information concerning the algorithm performances for damage diagnosis, thus further reducing the experimental costs and efforts associated with the development and optimization of such systems. The same algorithm has been tested on real experimental strain patterns acquired during real fatigue crack propagation, thus verifying the capability of the numerically trained algorithm for anomaly detection, damage assessment and localization on a real complex structure. The load variability, the discrepancy between the Finite Element Model and the real structure, and the uncertainty in the algorithm training process have been addressed in order to enhance the robustness of the system inference process. Some further algorithm training strategies are discussed, aimed at minimizing the risk for false alarms while maintaining a high probability of damage detection. •A SHM methodology based upon strain field measures has been developed.•A FE model has been validated to estimate the damage effect over the strain field.•Simulated data have been used for the Artificial Neural Network training.•Experimental data have been used for the Artificial Neural Network testing.•Good performances have been obtained for damage detection, localization and quantification.