Answers are grounded in this paper and cite their sources.
Accurate and cost-effective detection and localization of structural damage are important goals in structural health monitoring (SHM), particularly for lightweight metallic structures that are vulnerable to fatigue cracks. Such cracks can reduce the strength and reliability of a structure before they are identified through conventional inspection methods. In this study, an finite- element FEM and machine learning-based approach is proposed for detecting and locating cracks in an aluminum beam using a surface-mounted piezoelectric (PZT) actuator and sensor. A parametric finite element analysis was carried out using COMSOL Multiphysics, where a rectangular crack was progressively moved along the beam from 20 to 180 mm in 5 mm increments. The resulting floating electric potential was recorded at the four measurement points under transient conditions, producing more than 500 time-series responses. Two additional multi-crack configurations were also modeled to examine more complex damage scenarios, although these configurations were not included in the machine-learning dataset. From each transient response, six features were extracted at each measurement point: root-mean-square amplitude, peak amplitude, peak-to-peak amplitude, kurtosis, Hilbert-envelope arrival time, and dominant frequency. This resulted in a 24-dimensional feature vector for each response. A support vector machine (SVM) classifier was then trained to distinguish between healthy and cracked beam conditions. Grid search was used to select the most suitable kernel from linear and radial basis function (RBF) options. For crack localization, a Gaussian process regression (GPR) model with a Matèrn covariance kernel was trained using only the cracked-state data. The model achieved a mean absolute error of 25.09 mm and provided 95% prediction confidence intervals. Principal component analysis (PCA) was also used to visualize the distribution of the cracked-state features. In addition, the trained SVM classifier was applied sequentially along a simulated monitoring timeline to demonstrate how the probability of damage could evolve over time. However, the results should be interpreted as a proof of concept rather than as a validated damage-detection accuracy. In particular, the healthy-state data were synthetically generated by augmenting a single cracked-beam simulation instead of being obtained from an independent healthy-beam simulation. The study demonstrates the potential of combining piezoelectric sensing, FEM simulations, and machine learning for structural damage detection and crack localization.
No takes yet. Share an insight, caveat, or question.
Akash M A (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: