Enhancing inference speed in reparameterized convolutional neural network for vibration-based damage detection

Abstract

Structural health monitoring (SHM) technology has been widely used in civil engineering, and vibration-based damage detection (VBDD) technology is an important component of SHM research. With the advancement of deep learning, a plethora of deep learning-based algorithms have been applied to VBDD. The accuracy of VBDD is constantly improving with the assistance of various deep learning techniques. However, studies on the efficiency of VBDD tasks based on neural network are still relatively few, and lightweight network technology has been proven to be an effective way to improve efficiency of neural network. In this paper, a novel neural network based on reparameterization is presented, which can decouple the model training and deployment, and maintain high accuracy under the consideration of model inference speed. Specifically, a convolutional neural network with multiple 1 × 1 convolution is used in the training, and all layers of convolution are fused during testing and inference of the model to obtain a VGG-style network with a lighter structure and higher accuracy for deployment. Experiments on benchmark datasets from IASC-ASCE and the Z24 dataset show that the proposed method can make VBDD work better.