Identification of Structural Damage Location and Severity Using Deep Learning

(These authors contributed equally to this work) Structural members suffer damage due to various reasons over time, affecting the structure’s performance. Finding the location and severity of the damage could help in avoiding the possible catastrophe. There are a bunch of different NDTs to identify the damage, but these tests are time consuming and need a unique set-up and skilled human resources. So, to avoid these circumstances, modern techniques like deep learning could be promising in localising the damage and its severity. The stiffness of the material changes when damage occurs, which directly affects the structures’ natural frequency. The change in natural frequency can be used as a parameter to determine the damage and its severity. The objective of the work is to use the change in natural frequencies and corresponding mode shapes to find the location and the severity of the damage using deep learning algorithm. The natural frequencies are calculated using finite element software for both damaged and undamaged cases. In the damaged case, a crack has been introduced in an aluminium cantilever beam at twenty different locations with ten different sizes and the first five modes of natural frequencies and corresponding mode shapes are recorded. But the natural frequencies of the cantilever beam are functions of the geometry of the beam, so to make them independent of the geometry, the natural frequencies are multiplied with a factor, and new terms are calculated. The term is named dimensionless frequencies. Deep learning is an artificial intelligence tool in which an artificial neural network is trained with experimental data and created a model, which is then tested and validated with test data. In this work, some dimensionless frequencies and corresponding mode shapes are used as input, and the location and severity as the network’s output.