Damage localization and characterization using one-dimensional convolutional neural network and a sparse network of transducers

Early damage identification and continuous system monitoring save dramatically maintenance costs and increase the lifespan of priceless structures. Convolutional neural networks (CNNs) have attracted the attention of the structural health monitoring (SHM) community in recent years due to their great potential for identifying underlying data patterns. However, employing two-dimensional convolutional layers in a CNN necessitates the use of strong computing resources. Therefore, based on the present state-of-the-art technical solutions, a two-dimensional CNN is not suitable for real-time SHM applications with stand-alone processing units. Onedimensional convolutional networks (1D-CNN) have recently been employed in Ultrasonic Guided Wave-based (UGW-based) damage detection to address the aforementioned disadvantage. In this paper, a methodology for damage assessment at three levels - detection, localization, and characterization - based on 1D-CNN is put forward. Furthermore, the sequence length of the time-domain signals is significantly shortened by the application of a novel approach for processing them. Additionally, it is shown to what extend this method can improve the distinguishability between datapoints obtained from various damage scenarios. Consequently, by reducing the dimensionality of the problem, the proposed approach significantly reduces the memory usage of the classification algorithm. Experimental measurements as well as Numerical simulations, in which various damage scenarios such as corrosion, circular hole and cracks have been considered, are carried out to evaluate the efficacy of the proposed algorithm. It is shown that the suggested approach has benefits in terms of true classification rate of instances (above 93 percent for detection, localization, and characterization), computing time, in-situ monitoring, and noise resilience.