Application of neural networks in structural damage diagnosis and condition monitoring

The objective of this dissertation was to explore the potential of artificial neural networks in structural damage diagnosis and automatic monitoring.; First, an experimental study of a five story model steel frame was conducted. The goal of the experimental program was to investigate the sensitivity of the changes in vibrational signatures to the level of damages induced to the model structure. The vibrational signatures of the undamaged model structure were recorded during a shaking table test. Damage cases were experimentally simulated through gradual reduction of bracing areas in the first and second floors. Vibrational signatures were recorded for each damage case. Then, they were compared with those of the undamaged condition. The output showed that there was a clear and systematic increase in the changes of vibrational signatures as the level of damages were elevated. The effect of variations of input excitation on vibrational signatures was also investigated.; Next, the applicability of neural networks in analyzing the changes of vibrational signatures was investigated. Neural network based classifiers were used for diagnosing the damages associated with the changes in vibrational signatures. The classifiers were trained with a set of the experimental output. Then, their ability to diagnose damages correctly was tested against another set of the experimental output.; Finally, an analytical study of the model structure was conducted. The purpose of the analytical study was to investigate the possibility of generating the training samples, used to train the neural network classifier, through mathematical modeling of the structure.; Based on the results, a general framework for an integrated system for structure monitoring based on neural network classifiers is proposed. It is concluded that this framework provides a very strong base to advance the idea of automatic structure monitoring to practical reality.; Further studies are suggested in the area of quantifying the information needed to design and train the neural network classifiers, investigating the nature of the ambient vibration, and extension of the work to realistic structures.