Cross-domain transfer learning for vibration-based damage classification via convolutional neural networks

This work explores the application of computer-vision (CV) oriented convolutional neural networks (CNNs) to the problem of structural damage classification using vibrational-based features. It does so by taking general-purpose CV oriented CNNs and re-training them following a transfer learning approach. This is made possible by the use of a visually distinctive damage-sensitive feature: the complex frequency domain assurance criterion matrix, which exhibits distinctive degradation on its diagonal patterns when calculated using vibrational data acquired from different structural conditions. The use of this feature is compared to the use of frequency response function as commonly used in the literature. The method is applied to two different datasets: one where training, validation and testing datasets are generated using finite element models of a beam; and another where training and validation sets are generated using finite element models of a plate, but testing datasets were experimentally obtained. The impact of several factors relating to characteristics of the input features on the accuracy and sensitivity of the re-trained models are evaluated using Taguchi experimental designs to ensure statistical significance. Over all, this work demonstrates the viability of the proposed methodology and shows an improvement over commonly used methods found in the literature.