Identification of acoustic emission sources for structural health monitoring applications based on convolutional neural networks and deep transfer learning

Abstract

In the present work, different types of acoustic emission (AE) sources are identified by means of computational intelligence. The goal is to characterize the type of AE source and to successfully differentiate between sources that are related to an internal damage, such as a fracture initiation, or an external load represented by an elastic impact. A Hsu-Nielsen source (pencil-lead break) and two steel ball impacts of different diameters are selected for the excitation of an aluminum plate equipped with four piezoelectric transducers to record the acoustic emissions. Furthermore, 25 different areas for the AE sources are defined to collect a large database. Three different machine learning architectures are considered, which can predict the type of the AE source. Time domain signals of the acoustic emissions are used for the training of an artificial neural network and a 1D convolutional neural network. Additionally, the wavelet transformation is performed on the captured signals to generate RGB images of the sensor responses and to train a 2D convolutional neural network in combination with deep transfer learning. An error evaluation of each machine learning model is performed to discuss the classification results. The proposed methodology demonstrates that computational intelligence can be applied to accurately identify the type of AE source based on the captured acoustic emission signals.