Selecting the most effective location for a single accelerometer is a critical challenge in the structural health monitoring (SHM) of wind turbine blades (WTBs), where high installation costs, restricted access, and strong environmental and operational variability (EOV) often limit instrumentation. This study proposes a fully unsupervised, temperature-aware data-driven framework for determining the optimal accelerometer position directly from healthy-state vibration data, enabling reliable monitoring under ambient excitation. This study addresses this challenge using a full-scale dataset from a Vestas V27 wind turbine blade instrumented with 12 accelerometers. The complete multi-sensor configuration is employed to form the basis of a systematic, data-driven investigation into optimal sensor placement under realistic operating conditions. The framework begins with a frequency-domain fusion (FDF) stage, where a discrete wavelet transform (DWT) identifies a frequency split, separating low- and high-frequency regimes. Continuous wavelet transform (CWT) coefficients are then computed within each regime to capture both global blade dynamics and localized transient responses. Features from all sensors are merged and ranked using two complementary unsupervised feature selection strategies: the Laplacian score, which preserves local geometric structure, and a temperature-aware statistical filter that penalizes environmentally sensitive features. A sensor-level attribution analysis traces high-ranking features back to their source sensors across multiple compression thresholds, identifying the most consistently informative location at each operational speed (32 RPM and 43 RPM). An autoencoder trained exclusively on healthy-state data from the selected sensor is then deployed for anomaly detection, with reconstruction error serving as the damage index. The results reveal that, in both regimes, the accelerometer located at the leading edge near the mid-span provides the most stable and sensitive response to early-stage damage. Notably, the single optimally placed sensor achieves anomaly-detection performance comparable to that of the full twelve-sensor network, demonstrating a scalable and cost-efficient pathway toward interpretable, field-deployable SHM of operational wind turbines.
Ashkarkalaei et al. (2026) studied this question.