Keywords: neural networks, damage, identification, structural wave propagation.

All industry branches like aerospace, mechanical and civil engineering are interested in less intrusion and more accurate failure assessment techniques. They are mostly interested in damage such as cracks, delaminations, disbanding, etc. A damage detection and assessment technique was developed in this paper. This method based on variations in structural wave propagation for undamaged and damaged structure. Soft computing methods (artificial neural networks) [1] were proposed to solve an inverse problem and predict damage parameters. Structural health monitoring (SHM) approach presented here is useful especially in large, complex and inaccessible structures.

The finite element models were created. Defects in the form of local changes of stiffness and material loss in an elastic rod were simulated in a finite element model. This simulation provided the possibility of extending the set of damage cases and improved the neural network generalization properties.

Based on earlier promising results with this approach [2,3] a set of laboratory tests were carried out on various elements. During investigations both steel, Plexiglas and composite materials were used. Several failure cases were introduced by cutting, delaminating or drilling the samples. Piezoceramics (PZT) elements served as transmitters and receivers of elastic waves. A scanning laser vibrometer was used for measuring the Lamb waves in the specimens. Different groups of excitation signals (continuous sine wave, one, four and six sine wave impulses) and frequency (from 2 to 150 kHz) were applied to introduce a wave into the structure.

During the laboratory experiments, advanced signal processing techniques were adopted. The time and frequency analysis of recorded signals were carried out to extract wave features and create damage parameter database (DPD). The measured signals were preprocessed by wavelet transforms in order to remove noise. Replication techniques were adopted for the experimental data. Neural networks (NNs) were trained to realize dependences between input (frequency spectrum) and output data (height and width of the damage).

To get the best possible neural network performance, several input combinations and network architectures were tested. A regularization technique (adding the normalized Gaussian noise) was adopted to improve network generalization. Results presented in this paper have proved the reliability and the usefulness of the proposed approach.

1

Z. Waszczyszyn, L. Ziemianski, "Parameter Identification of Materials and structures", Z. Mroz and G.E. Stavroulakis eds. CISM Courses and Lectures, 469, Springer, Wien-New York, 256-340, 2005.

2

L. Ziemianski, G. Piatkowski, "Use of neural networks for damage detection in structural elements using wave propagation", In: B.H.V. Topping, editor, Computational Engineering Using Metaphors from Nature, 25-30, Civil-Comp Press, Edinburgh, 2000. doi:10.4203/ccp.64.1.4

3

P. Nazarko, L. Ziemianski, "Damage identification in rods using Neural Networks", C.C. Baniotopoulos ed. Proceedings of the 2nd Intern. Conf. on Nonsmooth/Nonconvex Mechanics with Applications in Engineering, Thessaloniki, Greece, 107-114, 2006.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £62 +P&P)