DATA-DRIVEN FAULT DETECTION IN COMPOSITE CYLINDRICAL SHELLS: DIRECTING THE PROPER ORTHOGONAL DECOMPOSITION PROSPECTIVE INTO AN ARTIFICIAL NEURAL NETWORK VISION

This paper presents a vibration data-driven study for damage detection in a physical cylindrical shell made of carbon fiber reinforced polymeric plies. The manufacturing method followed for the synthesis of this structure introduces numerous fault sites as potential initiators of damage propagation. Using three accelerometers, vibration datasets have been collected for damage detection by a classical vibration analysis method based on the change of the curvature of vibration shapes extracted by sectioning-at the value of a dominant natural frequency-the surface-formed landscape of the distribution of the dataset over the processed frequency-space domain. The existence and the position of damage sites in the cylinder are detected by computing the pointwise ratio of the section-based vibration shapes of the datasets. To verify and improve this technique, unleased is the unparalleled data reduction-computational analysis of the advanced proper orthogonal decomposition (POD) transform to detect damage sites as variations of POD modal shapes, being identified as invariants of the motion. The data-driven spirit of POD processing of datasets is conveyed further to support a potential artificial neural networks vision for fault detection by learning aspects of the physical system with training based on the collected datasets.