Session: 05-03: Vibration Based Methods

Paper Number: 90936

90936 - Random Vibration Based Robust Damage Detection on an Operating Wind Turbine Blade Under Variable Natural Excitation Conditions 

In-operation damage detection on wind turbine blades is important for safety, proper operation, and maintenance. In this study damage detection for a 13 m long blade of a full–scale Vestas V27 wind turbine under variable actual operating conditions is undertaken based on vibration signals collected over a 104-day-long campaign. Trailing edge opening damage scenarios of progressive sizes (15 cm, 30 cm, and 45 cm) are introduced on a blade, with the random vibration responses measured via an array of accelerometers. The problem is shown to be highly challenging as the effects of the environmental and loading conditions on the dynamics are significant and almost `mask’ those of the considered damages. Yet, in a recent study by the authors it was demonstrated that high detection performance is achievable when artificially enhanced excitation, applied via a custom actuator providing periodic impulse-like hits on the blade, and data-based Statistical Time Series type and unsupervised robust damage detection methods are employed. Evidently, the artificially enhanced excitation leads to significant enrichment and enhanced signal-to-noise ratios; facts facilitating damage detection.

The question addressed in the present study is whether similar damage detection performance may be obtained in the significantly more challenging case in which only naturally excited random vibration signals are available, that is without an actuator. To answer this question, a data-based and unsupervised Statistical Time Series type robust damage detection method is formulated based on m random vibration sensors, with the signals provided by m-1 of them being treated as `causes’ or pseudo-inputs and the m-th one as the `effect’ or pseudo-output. By properly selecting m via a multiple-coherence-based procedure, a multiple-input single-output transmittance data-based stochastic model is estimated based on available signals. This is of the stochastic MISO AutoRegressive with Exogenous excitation (ARX) form, and unlike conventional single-input single-out transmittance type models or the AutoRegressive (AR) models employed in our previous study, it remains invariant to varying excitation conditions, thus capable of providing previously unachievable robustness. The robust damage detection method is then built, with the model’s parameters constituting the feature vector, within an unsupervised framework based upon our recently introduced Hyper-Sphere-based approach which is in this study further enhanced via Principal Component Analysis based feature vector reduction.

The achievable damage detection performance is examined via thousands of test cases (experiments) to ensure high statistical reliability and is presented in terms of Receiver Operating Characteristic (ROC) curves that depict the True Positive Rate (correct detection rate) versus the False Positive Rate (false alarm rate) for varying detection threshold.  The obtained results are very exciting, indeed, resulting in outstanding damage detection performance with the use of 3 pseudo-inputs (m=4). In fact, the performance is simply ideal, achieving 100% correct detection for 0% false alarms for the two larger (30 and 45 cm) damages, while being superb and achieving 100% correct detection for about 3% false alarms for the smallest (15 cm) damage. It is impressive that this performance is achieved under largely varying and actual operating conditions, and it manages to exceed that of our recent study in which the actuator enhanced signals were employed.

Overall, the results of the study underscore the capabilities of advanced Statistical Time Series type random vibration methods for effective and robust detection of damage on the blades of operating wind turbines.

Presenting Author: Spilios Fassois University of Patras