Convolution Neural Network with Parallel Multi-Attention Framework for Structural Health Monitoring using Vibration-based Measurements

Structural health monitoring (SHM) ensures infrastructure safety by detecting anomalies early. The present study proposes a deep learning (DL)-based SHM technique using a 2D convolutional neural network (CNN) with a parallel multi-attention framework (2D CNN-PAMF) to improve anomaly detection. At first, vibration signals are transformed into images using the gramian angular summation field (GASF) technique. These images serve as input to the 2D CNN, where the parallel multi-attention module captures global and local features for better classification. The model is validated using the IASC-ASCE benchmark dataset and shows high performance in assessing structural health. Comparative analysis with 1D CNN, considering raw vibration signals, and 2D CNN-based methods, considering transformed matrices from vibration signals, high-lights its superior effectiveness. Future work aims to optimize this model for deployment on resource-constrained devices for real-time structural health assessment.