Waveform signal-driven structural damage detection via a multiscale residual convolutional neural network with varifocal receptive field

Convolutional neural networks (CNNs) and residual networks have been widely adopted for structural damage detection (SDD). However, a key limitation of these models is their fixed receptive fields, which are constrained by kernel dimensions. While a common strategy to enlarge the receptive field is to increase kernel size, this approach leads to a significant growth in parameters and raises the risk of overfitting. To overcome this fundamental constraint, we propose a novel multiscale residual CNN architecture with adaptive varifocal receptive fields for damage detection. The core of our framework involves replacing standard residual blocks with varifocal modules that utilise bilinear interpolation, complemented by parallel multi-scale convolution branches to enrich feature diversity and discriminative power. Experimental cases are implemented on the flexible airship capsule and the building structure, and results demonstrate our method’s superiority over existing approaches, achieving high detection accuracies of 95.51% and 99.87%, respectively. Furthermore, noise robustness and uncertainty analyses confirm the effectiveness and reliability of the proposed method for practical SDD applications.