Monitoramento da integridade estrutural a partir de sensoreamento de baixo-custo: Investigação de técnicas de inteligência artificial para classificação de danos estruturais

Abstract

In engineering, a structure is designed considering its possible failure due to damage that occurs over time. In this context, in order to extend the useful life of components, avoid failures and operational problems, periodic maintenance is carried out. Several studies are conducted on maintenance based on real-time monitoring of the state of a structure, anticipating the detection of damage. The high precision and constant dependence on human resources to obtain conclusive parameters makes the process costly. Therefore, the biggest obstacle for monitoring is differentiating between damaged and undamaged structures using low-cost equipment and statistical pattern recognition algorithms. In the present work, supervised learning techniques were used to analyze simulated vibration signals and vibration signals of a structure for correlation with different types of damage. Using MATLAB, ADXL335 accelerometer and ESP-32 microcontroller to monitor the state of a simple two-story structure excited by external impact-type forces, it was possible to extract the natural frequencies and damping coefficients as main attributes for classifying the dimension of the associated damage. . At the end of the study, it was found that the different classifiers Medium Gaussian SVM, Medium KNN, Medium Tree and Linear Discriminant achieved accuracy greater than 85% to highlight the differences between damage conditions. Different performance metrics were analyzed to prove their accuracy, such as: true positive rate, true negative rate, positive predictive values, false discovery rate, false positive rate, ROC curve and AUC. By comparing the results obtained with the structure and simulated signals, the effectiveness of artificial intelligence and the feasibility of using economically accessible instrumentation to monitor structural integrity in small structures were evaluated. Despite the need for a large amount of data to obtain good results, it is believed that the model obtained can contribute to the generation of new algorithms capable of classifying different types of damage to structures.