Deep Acoustic Emission Detection Trained on Seismic Signals

In the structural health monitoring field, the acoustic emission technique (AE) is one of the most important and extensively applied methods. AE is a non-destructive testing (NDT) method which investigates acoustic ultrasonic waves caused by a sudden energy release due to cracks and micro-cracks opening in the investigated material that can be analyzed for structural health monitoring (SHM) purposes. Two essential qualities of the AE method are the capability to detect and locate the damage/crack, through a reliable and accurate onset time detection, and the ability to identify the cracking mode from recorded parameters. The aim of the present contribution is the automatic detection of the Onset time of AE signals with the help of artificial intelligence algorithms, more specifically artificial neural networks (ANN). Two different approaches have been used for automatic onset time detection. The foremost is related to adopting a convolutional neural network (faster R-CNN) with a pre-training on a very large dataset, it was possible to employ the transfer learning (TL) technique. The main benefits of TL include: speed up training considerably, saving of resources, improving the efficiency and removing the need for a large set of labeled training data. The latter approach involves the use of a convolutional recurrent neural network (CRNN), developed in the field of the sound event detection (SED). The SED’s objective is to identify sound events in a recording and their related starting and ending time instances. Considering the obvious parallelism between AE signals and seismic signals, the training of the two networks has been carried out with the latter, because of their larger availability. The dataset composed of seismic signals has been collected thanks to ITACA, Italian ACcelerometric Archive.