Detect and classify damage in bridges through MLP neural networks trained on SHM data

The paper presents an intelligent procedure to detect the presence, classify the type, and identify different levels of damage in existing bridges. Based on Multi-Layer Perceptron (MLP) neural networks, the procedure is tested referring to the Z24 bridge benchmark, where different damage scenarios were progressively induced. The proposed methodology trains MLP neural networks on data from both Forced Vibration Tests (FVT) and Ambient Vibration Tests (AVT), specifically trained for each sensor setup using frequency-domain selected features. A majority voting criterion is then applied to aggregate independent diagnoses into a robust final classification. Results demonstrate that this computationally efficient framework accurately distinguishes complex structural pathologies, offering a scalable solution for automated, real-time structural health monitoring. Future application of the proposed procedure involves training the MLP network on simulated damage scenarios of well-identified numerical models of real structures, which can lead to significant advancements in automated structural health monitoring.