An artificial neural network for locating possible damaged zones in beam structures

In the scientific literature, many vibration-based damage detection methodologies use artificial neural networks (ANNs). The majority of these methodologies are designed to determine the existence of damage in structures (or lack thereof). However, when it comes to locating and quantifying damage in large structures, ANNs are hindered by the large number of training cases needed to guarantee proper damage detection. This paper proposes a perceptron multi-layer neural network to locate damage zones in a beam structure, working from the assumption that such identification represents a step to be done prior to damage quantification. The input data vector was based on the modal flexibility matrix, and the output vector indicated whether a specific zone in the structure might be damaged. Only a few initial modes were measured at a specific quantity of points on the beam. A postprocessing computation was then employed to improve the ANN results. A confidence level was achieved in terms of expected damage zones. If 50% of the zones were considered to be "probably damaged", the methodology turned out to be approximately 90% successful. Yet, this reliability may have been affected by the damage extent of the damaged elements. The use of free-of-noise measurements led to an identification level close to 100%. In summary, the results point to the proposed methodology's ability to detect damage as highly prejudiced by incompleteness and noise (in the measurements).