The reliable detection of cracks in engineering materials remains a fundamental challenge in nondestructive testing, especially in applications that require automated inspection, reduced instrumentation costs, and robustness under noisy operational conditions. Traditional nondestructive evaluation techniques often rely on complex sensing setups or expert-dependent interpretation, which can limit scalability and real-time applicability. In this context, this study addresses the scientific problem of achieving reliable and automated crack detection using simplified sensing architectures combined with intelligent data-driven analysis. This work proposes an intelligent signal classification framework for crack detection in polymeric materials based on machine learning and ensemble learning techniques. The scientific novelty of the proposed approach lies in the integration of vibration-based sensing with a structured decision pipeline that emphasizes robustness, stability, and statistical reliability rather than solely focusing on algorithmic complexity or physical crack modeling. The framework is specifically designed to operate effectively under reduced instrumentation scenarios, making it suitable for practical nondestructive testing and condition monitoring applications. Experimental vibration signals were acquired from a single piezoelectric sensor mounted on polymeric specimens under different structural conditions. The acquired signals were processed through a systematic pipeline including signal normalization, dimensionality reduction, and supervised machine learning-based classification. A comprehensive comparative analysis was conducted using Random Forest, Support Vector Machines, k-Nearest Neighbours, AdaBoost, and Boosted Trees. Model performance was evaluated under a consistent five-fold cross-validation protocol, ensuring reliable estimation of generalization capability and robustness. The results demonstrate that the Random Forest classifier achieved superior and more stable performance compared to the other evaluated methods, reaching a mean classification accuracy of 73.18% and exhibiting reduced performance variability across validation folds. Statistical analyses confirmed the consistent dominance of the ensemble-based approach, highlighting its robustness in handling noisy and nonlinear vibration signals. Beyond predictive accuracy, the Random Forest model also provides intrinsic interpretability through feature importance analysis, enabling insights into the signal components most relevant to crack detection. These findings demonstrate that machine learning-driven ensemble methods can provide an effective, scalable, and reliable solution for automated crack detection in nondestructive testing. The proposed framework supports the development of intelligent condition monitoring and decision-support systems, offering a cost-effective pathway toward practical implementation in industrial inspection and structural health monitoring applications.
Silva et al. (Thu,) studied this question.