Reducing Experimental Data Requirements in CNN-based damage detection through Transfer Learning

Abstract

While neural networks represent a promising approach for evaluating sensor data to assess damage presence, location and severity, large amounts of data are required for training. However, the generation of experimental data is both labor-intensive and costly. Transfer learning is a well-established method in deep learning that enables the reuse of knowledge from pre-trained models to improve performance and reduce data requirements across various domains such as computer vision and natural language processing. Its application within the field of strain-based structural health monitoring (SHM) has received little attention from currently published literature. This work develops a resource-efficient transfer learning approach for strain-based SHM. While the Convolutional Neural Network (CNN) model learns relationships between strain distribution and crack geometry in an aluminum beam from Finite Element (FE) data, fine-tuning adapts it to experimental conditions, accounting for factors such as measurement noise, increasing overall accuracy and robustness. An encoder-decoder CNN (UNet) is initially trained with synthetic data from FE simulations. The model is then fine-tuned based on a smaller experimental Digital Image Correlation (DIC) dataset obtained from an aluminum beam subjected to a four-point bending fatigue test. For this purpose, the encoder part of the CNN is frozen, while parameters of the final layers of the decoder are updated. The approach is validated with respect to its accuracy, robustness and applicability for SHM systems. The presented approach significantly reduces the experimental data requirements while improving damage detection performance for an aluminum beam under four-point bending. This study demonstrates that transfer learning enables efficient adaptation to real-world conditions, offering a cost-effective and scalable solution for data-driven SHM.