Computational & Technology Resources
an online resource for computational,
engineering & technology publications
PROCEEDINGS OF THE NINTH INTERNATIONAL CONFERENCE ON COMPUTATIONAL STRUCTURES TECHNOLOGY
Edited by: B.H.V. Topping and M. Papadrakakis
Inference Models for Structural Systems Integrity Monitoring: Neural Networks and Bayesian Enhancements
Department of Structural and Geotechnical Engineering, University of Rome "La Sapienza", Italy
S. Arangio, "Inference Models for Structural Systems Integrity Monitoring: Neural Networks and Bayesian Enhancements", in B.H.V. Topping, M. Papadrakakis, (Editors), "Proceedings of the Ninth International Conference on Computational Structures Technology", Civil-Comp Press, Stirlingshire, UK, Paper 154, 2008. doi:10.4203/ccp.88.154
Keywords: complex structural systems, integrity monitoring, damage identification, neural networks, Bayesian methods, hierarchical strategy.
The realization of high-cost and safety-critical construction has resulted in the necessity of advanced approaches able to consider the intrinsic complexity of the structures. This complexity is related to different factors, such as nonlinearities, uncertainties or interactions between components. Only considering these aspects, a consistent evaluation of the overall performance can be obtained.
In this perspective, the quality of a complex structural system can be described in a comprehensive way by its dependability that is a global property that describes the overall quality performance and its influencing factors.
All these aspects are connected to the integrity of the structural system, considered as the completeness and consistency of the structural configuration. During the service life, the integrity and consequently the overall dependability could be lowered by deterioration and damage and an effective structural monitoring represents an essential tool for dependability assessment. Considering the large quantity of data coming from the monitoring process, suitable techniques, able to extract useful information from the measured data, are needed. In this work, an adaptive model is investigated, the neural networks, and the enhancements obtained formulating this model in a Bayesian viewpoint are presented.
The various concepts have been applied for the dependability assessment of a long suspension bridge and a multi-level strategy, including two steps, for the identification and quantification of damage has been proposed .
In the first step the structural behaviour has been monitored in different measurement points. For every point a neural network has been trained to approximate the response time-history in the undamaged situation. Then, new time-histories, obtained simulating different damage scenarios, have been proposed for the trained models. Analyzing the increment of the error in the approximation of these signals with respect to the error in the approximation of the response in the undamaged situation, the occurrence of damage has been detected and the damaged portion of the deck has been identified.
The second step aims at the identification of the specific damaged element and the quantification of such damage. In this step a pattern recognition approach has been used. A neural network has been trained using as inputs the errors in the approximation of the response in three points of the damaged section, and as output a vector including the location of the damage and its intensity. After training, the network is tested with some patterns not included in the training set: the location can be detected in 90% of the cases, whereas the intensity in 66% of the cases.
In both steps Bayesian techniques have been applied: in the first step the model optimization does not change the results in a substantial way; on the contrary, in the second step, the choice of the optimal model leads to an improvement of 20% in the location and 16% in the quantification.
purchase the full-text of this paper (price £20)