Use of autoregressive model for feature extraction in the detection of faults in scaled and real-size structures

Modern infrastructure such as bridges, buildings, and aircraft deteriorates over time, making structural health monitoring (SHM) essential for extending service life and ensuring safety. Vibration-based monitoring is widely used due to its sensitivity to structural changes and the low cost of sensors. However, the large volume of data generated poses challenges for manual analysis. This study explores machine learning techniques, specifically outlier detection, to automate pattern recognition without requiring labeled data. Two methods were evaluated: One-Class Support Vector Machine (OCSVM) and Elliptic Envelope. Feature engineering focused on identifying indicators sensitive to damage but robust to operational variations, using autoregressive model coefficients and residual errors. Experiments were conducted on two datasets: one from a test structure at UFSC's Laboratory of Vibrations and Acoustics, and another from the Z-24 Bridge in Switzerland (SIMCES project). Performance was measured using the area under the ROC (AUC) and compared to a baseline of statistical features. Results showed significant improvements: for the LVA structure, OCSVM and Elliptic Envelope increased AUC by 13% and 5%, respectively; for the Z-24 Bridge, the improvements were 28% and 33%, confirming the effectiveness of the proposed approach.