Crack Identification and Noise Sensitivity Analysis of Simplified Blades for Small Wind Turbines Based on Multi-Point Acceleration Responses

Abstract

This paper investigates crack-severity identification in a simplified isotropic steel blade using finite-element-generated multi-point acceleration responses. A finite element model of an equal-section S822 blade is established with four states: healthy, 1 mm crack, 3 mm crack, and 6 mm crack. Twenty transient response files are generated from five excitation amplitudes and five acceleration probes. Time-domain, frequency-domain, wavelet, and multi-point statistical features are evaluated using KNN, SVM-RBF, Random Forest, and ExtraTrees classifiers. To avoid optimistic results caused by correlated sliding-window samples, Leave-One-Force-Out validation and file-level majority voting are adopted. The results show that tree ensemble models can distinguish the four states under clean signals, whereas their performance rapidly approaches the random four-class level when the test signal-to-noise ratio decreases to 10 dB and below. This study therefore provides a compact numerical baseline for evaluating the noise sensitivity of conventional acceleration features and classifiers.