Convolution Neural Network Development for Identifying Damage in Vibrating Pylons with Mass Attachments

A convolution neural network (CNN) is developed in this work to detect damage in pylons by measuring their vibratory response. More specifically, damage detection through testing relies on the development of damage-sensitive indicators, which are then used to reach a decision regarding the existence/absence of damage, provided they have been retrieved from at least two distinct structural states. Damage indicators, however, exhibit a relatively low sensitivity regarding the onset of structural damage, further exacerbated by the low amplitude response to a variety of environmentally induced loads. To this end, a mathematical model is developed to interpret the experimental data recovered from a fixed-base pylon with a top mass attachment to transverse motion. Damage is introduced in the mathematical model in the form of springs corresponding to the cracking of the beam's lower end. Families of numerically generated acceleration records are produced at select stations along the beam's height, which are then used for training a CNN. Once trained, it is used to identify damage from acceleration records produced from a series of experiments. Difficulties faced by CNN in correctly identifying the presence/absence of damage in the pylon are discussed, and steps taken to improve the quality of the results are proposed.