Convolutional neural network and impedance-based SHM applied to damage detection

The impedance-based structural health monitoring technique uses measured signatures changes to identify incipient damages in structures. The purpose is to perform a correlation of these changes with the physical phenomena. However, since electromechanical coupling exists, some environmental influences such as temperature changes may lead to false decision regarding the condition of the structure. As a result, innovative machine learning tools have been extensively investigated to avoid errors in structural prognosis and, in this sense, recent applications of convolutional neural networks (CNN) have emerged within the scope of SHM research, focusing mainly on vibration analysis. However, studies that aim to combine neural architectures with intelligent materials for structural monitoring purposes have been poorly evaluated. Consequently, its integration with the electromechanical impedance method is still considered as being a new application of CNN. Thus, in order to contribute to the SHM area, this work presents a combination of the CNN architecture and the EMI methodology. In the present contribution, three aluminum beams subjected to three different steady temperature levels (0 degrees C, 10 degrees C and 20 degrees C) were studied. For this aim, a test chamber was used for humidity and temperature control. Artificial damages such as mass addition were taken into account so that impedance signatures related to both pristine and damaged conditions can be analyzed. Thus, a one-dimensional Convolutional Neural Network (1D CNN) was designed, trained and used for damage prediction purposes. In this context, a temperature robust model that is able to identify damage independently of environmental condition was developed.