Analysis of a Convolutional Neural Network Model in Terms of Neuron Activation for Structural Damage Classification Approaches


Teixidó Marquès, Eduard


Detecting when a structure is damaged is key to prevent the structure from collapsing, as it have been in several cases such as bridges, buildings or telecommunication towers. Over the years, structural health techniques have been developed to be able to detect damages generated in structures before they fail. New non destructive testing methods have recently been implemented as they allow the structure to maintain its functions once testing is completed. Recently, a non- destructive method has been implemented based on the frequency response of the structure when it is subjected to vibrations. When a structure is damaged, this response is different from the pristine structure. Comparing these two frequency responses, a covariance matrix is obtained which, according to the values obtained, it can be determined whether the structure under study is damaged.
At the same time, the exponential increase of Artificial Intelligence field has been able to allow its application in damage detection discipline. Artificial Intelligence, and more specifically, Machine Learning, has allowed the possibility of training neural networks so that, automatically, through the matrix of covariance obtained, the structures can be diagnosed. Recent preliminary research has applied the use of networks along with frequency responses to detect damages on aluminium laminar structures, both experimental and simulated by finite elements.
However, the application of neural networks for the diagnosis of the state of these laminar structures has not proven to be a sufficiently reliable technique as it has evidenced to be inconsistent in its predictions. Consequently, to improve the accuracy and precision of this method, the aim of this work is to understand and identify the behaviour of the neural networks used for the detection of structural variations through the visualization of the Class Activation Maps of each model.



Pérez, Marco Antonio


IQS SE - Undergraduate Program in Industrial Engineering