Metodologia para o monitoramento de integridade estrutural por meio de vibrações mecânicas

Abstract

Structural integrity of bridges is essential to ensure the safety and durability of civil engineering works, especially in regions subject to adverse conditions. However, traditional inspection methods are often time-consuming, costly, and invasive, which has driven the development of non-destructive techniques based on dynamic structural analysis. In this context, Structural Health Monitoring (SHM), which combines vibration measurements with artificial intelligence, emerges as a promising alternative for rapid, accurate, and automated damage detection and classification. Based on this approach, the main objective of this study was to develop an efficient methodology for damage detection in bridges through the analysis of mechanical vibrations. To achieve this goal, laboratory measurement protocols were established, a physical scaled-down bridge model was constructed, dynamic response analyses were carried out, and artificial intelligence techniques were applied to identify and classify the severity of structural damage. The proposed methodology was initially validated through tests on wooden beams with simulated defects, which enabled the creation of a robust and representative dataset for training Artificial Neural Networks (ANNs). Three hammer tips (plastic, metal, and rubber) were tested, with damage simulated by notches and different numbers of hidden layer neurons (ranging from 5 to 200). The best results were obtained using the plastic hammer tip, with accuracy ranging from 0.85 to 0.90 and kappa coefficient between 0.77 and 0.87, particularly with configurations of 25 to 35 neurons. The rubber and metal tips also showed solid performance, with accuracy between 0.82 and 0.88, and kappa values between 0.70 and 0.87, depending on the configuration. Subsequently, the approach was applied to a reinforced concrete bridge prototype, also at a reduced scale, equipped with 12 strategically positioned accelerometers and subjected to controlled excitations using an impact hammer. This practical application allowed the methodology’s performance to be assessed in a more complex and realistic structural system. Modal testing revealed that the peak response values of the accelerometers varied significantly with the progression of the notches. For example, accelerometer Ac2 showed an increase in response values from 0.92 (undamaged) to 1.17 (section 4), while Ac7 ranged from 0.54 (undamaged) to 0.65 (section 4). In contrast, some sensors such as Ac5 exhibited a sharp reduction in response, with values dropping from 1.45 (undamaged) to 0.03 (section 4), indicating localized loss of structural stiffness. Although the methodology proved effective in identifying dynamic changes in the prototype, the artificial neural network (ANN) showed an initially satisfactory performance (accuracy of 57.14% and Kappa of 0.36), with emphasis on the correct identification of the most severe damages (Section 4), which had a sensitivity of 66.7% and a positive predictive value of 100%. The limitation in the results of the other classes is related to the small variety of applied damages. It is expected that, with the introduction of new damage scenarios, the model will improve its learning and significantly increase its accuracy.