Methods for Visualizing Deep Learning to Elucidate Contributions of Various Signal Features in Structural Health Monitoring

Abstract

Data-driven damage detection is gaining increasing attention. However, the operational principles of such methods often lack explainability. In this research, a 1D convolutional neural network (1DCNN) network is used to detect internal damage in aluminum plates, and the main innovation lies in proposing a visual explanation method to measure the contribution of different sampling points in the damage classification task. An aluminum plate damage data set was built to test the effectiveness of the developed contribution calculation method that contains six different sizes of damage and is detected by guided wave signals. The proposed 1DCNN achieved a 96% accuracy rate in identifying damage. The proposed visual explanation method highlights the specific features that played a role in the damage identification process. Moreover, we arrived at a novel finding: deep learning-based methods for damage identification typically depend on global features, contrasting with knowledge-driven approaches. This research helps researchers to understand how deep learning models work in structural health monitoring (SHM).