Fault Diagnosis on Steel Structures Using Arti cial Neural Networks

Abstract

The goal of this effort is to diagnose fault on steel structures by using non destructive techniques. Ultrasonic techniques are usually applied in engineering for faults determination, thickness measuring, adhesive layers, and in metallurgy to establish the quality of welds in metallic pieces. But the ultrasonic techniques could be difficult or impossible to apply in structures with reduced space, i.e. car frameworks. Acoustic signals have been employed since ancient times for detecting faults. Striking an object produces a sound whose differences may be heard when the object is damaged, therefore the vibration signals can be applied to detect differences into a metallic structure. Moreover, the Frequency Response Function (FRF) is used in this work to detect damages in metallic structures. The FRFs are used as input in an artificial intelligent system such as neural nets to detect damage. In general, non destructive evaluation is applied to detect and localize structure faults by using a signal with wavelength smaller than the detecting fault. The method requires analyzing the object in numerous small sections just only to detect the fault. Damages in metallic structures cause small changes in resonance. This work considers global non destructive tests focused only on the estimation of the integrity of the system. Therefore, the whole structure is analyzed to detect damage with only one measure. Traditional fault structure detection practices usually require testing in numerous small sections. The technique used in this research decreases the fault detection costs drastically. The number of FRF spectral lines used to input the neural net is a small fraction of the total frequency range. The designing of a practical implementation requires the usage of a simple method for damage detection, i.e. neural networks. A supervised feed-forward network with Levenberg-Marquardt backpropagation algorithm is applied for testing goals. The net structure is a three-levels layer. The net has only one hidden layer and one output neuron that classifies the damage in the steel beams. The particular selection of the forty two spectral lines values results in forty two neurons as system inputs. After training on a small set of data, the neural network is able to identify the damaged beams with considerable accuracy. The network converges in average in less than twenty epochs. Focusing in the technological implementation, the artificial neural network obtains excellent results with few neurons.