Desarrollo de sensores virtuales para la estimación de parámetros vibratorios en palas de turbinas eólicas.
Loading...
Date
2026
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
El daño estructural en las palas de turbinas eólicas representa una fracción significativa de los costos de mantenimiento del sector eólico, mientras que su monitoreo directo mediante instrumentación permanente resulta técnica y económicamente poco viable a gran escala. Como alternativa, la presente memoria de título desarrolla y valida un sensor virtual capaz de estimar la parámetros vibratorios de una pala de turbina eólica de escala reducida a partir de mediciones indirectas obtenidas en la góndola, evitando así la necesidad de instrumentar la pala en un escenario de operación real. Para ello, se diseñó e implementó un banco de ensayos a escala reducida compuesto por un aerogenerador de tres palas impresas en PLA, operado en un túnel de viento en configuración Open-Jet. El sistema de adquisición de datos, basado en microcontroladores ESP32-S3 con arquitectura de firmware multiprocesamiento (FreeRTOS) y sensores inerciales BMI160 bajo protocolo SPI, permitió registrar vibraciones sincronizadas en la góndola y en la pala a una frecuencia de muestreo de 1000 Hz, con una alineación temporal residual aproximada de 1 ms tras la corrección de la deriva del reloj mediante la detección de impactos mecánicos. A partir de una campaña de 17 ensayos que cubrió un rango de 47.0 a 232.4 RPM, se construyó un dataset tabular de estadísticos temporales (RMS, factor de cresta, curtosis) y componentes principales de la densidad espectral de potencia, empleado para entrenar una arquitectura de redes neuronales tipo Perceptrón Multicapa (MLP). Tras comparar un enfoque de modelo único multisalida frente a una configuración de submodelos independientes por dominio (frecuencia y tiempo), se adoptó esta última al demostrar una mejora sistemática en las 10 variables objetivo, elevando el coeficiente de determinación (R2) de test promedio de 0.428 a 0.711. Los resultados muestran que el sensor virtual reconstruye con alta fidelidad la dinámica global de la pala, alcanzando un R2 de test de 0.9929 para la componente principal del espectro y una correlación morfológica superior al 90% en la reconstrucción de la densidad espectral de potencia. Se identificó, además, un comportamiento diferenciado según el eje de medición: mientras los ejes transversales (edgewise y flapwise) se estiman con alta precisión, el eje longitudinal exhibe un desempeño sistemáticamente inferior. Se determinó que esta diferencia responde a la naturaleza distinta de la carga mecánica que domina cada eje: los ejes transversales flexionan ante la carga aerodinámica del flujo incidente, cuya amplitud responde con sensibilidad a la condición operativa, mientras que el eje longitudinal está gobernado por esfuerzos axiales de origen geométrico-gravitatorio, de amplitud casi constante y con escasa información predictiva disponible. Estos resultados validan la hipótesis planteada y confirman la viabilidad técnica de estimar el campo vibratorio de una pala mediante mediciones indirectas y aprendizaje automático, sentando una base metodológica para su futura extensión hacia sistemas de monitoreo de condición e integridad estructural.
Structural damage in wind turbine blades accounts for a significant share of maintenance costs in the wind energy sector, while direct monitoring through permanent instrumentation remains technically and economically impractical at scale. As an alternative, this thesis develops and validates a virtual sensor capable of estimating the vibratory parameters of a reduced-scale wind turbine blade from indirect measurements taken at the nacelle, thereby avoiding the need to instrument the blade under real operating conditions. To this end, a reduced-scale test bench was designed and built, consisting of a three-blade wind turbine with PLA-printed blades operated in an Open-Jet wind tunnel configuration. The data acquisition system, based on ESP32-S3 microcontrollers with a multiprocessing firmware architecture (FreeRTOS) and BMI160 inertial sensors under the SPI protocol, enabled synchronized vibration recording at the nacelle and blade at a sampling rate of 1000 Hz, achieving a residual temporal alignment of approximately 1 ms after correcting clock drift through mechanical impact detection. From a test campaign of 17 trials spanning a range of 47.0 to 232.4 RPM, a tabular dataset of time-domain statistics (RMS, crest factor, kurtosis) and principal components of the power spectral density was built and used to train a Multilayer Perceptron (MLP) neural network architecture. After comparing a single multi-output model against a configuration of independent submodels per domain (frequency and time), the latter was adopted, showing a systematic improvement across all 10 target variables and raising the average test coefficient of determination (R2) from 0.428 to 0.711. The results show that the virtual sensor reconstructs the blade’s global dynamics with high fidelity, achieving a test R2 of 0.9929 for the principal spectral component and a morphological correlation above 90% in the reconstruction of the power spectral density. A differentiated performance was also identified depending on the measurement axis: while the transverse axes (edgewise and flapwise) are estimated with high accuracy, the longitudinal axis exhibits systematically lower performance. This difference was found to stem from the distinct nature of the mechanical loading dominating each axis: the transverse axes bend under the cyclic aerodynamic load of the incoming flow, whose amplitude responds sensitively to the operating condition, whereas the longitudinal axis is governed by axial stresses of geometricgravitational origin, of nearly constant amplitude and with scarce predictive information available. These results validate the proposed hypothesis and confirm the technical feasibility of estimating a blade’s vibratory field through indirect measurements and machine learning, laying a methodological foundation for its future extension toward condition monitoring and structural health monitoring systems.
Structural damage in wind turbine blades accounts for a significant share of maintenance costs in the wind energy sector, while direct monitoring through permanent instrumentation remains technically and economically impractical at scale. As an alternative, this thesis develops and validates a virtual sensor capable of estimating the vibratory parameters of a reduced-scale wind turbine blade from indirect measurements taken at the nacelle, thereby avoiding the need to instrument the blade under real operating conditions. To this end, a reduced-scale test bench was designed and built, consisting of a three-blade wind turbine with PLA-printed blades operated in an Open-Jet wind tunnel configuration. The data acquisition system, based on ESP32-S3 microcontrollers with a multiprocessing firmware architecture (FreeRTOS) and BMI160 inertial sensors under the SPI protocol, enabled synchronized vibration recording at the nacelle and blade at a sampling rate of 1000 Hz, achieving a residual temporal alignment of approximately 1 ms after correcting clock drift through mechanical impact detection. From a test campaign of 17 trials spanning a range of 47.0 to 232.4 RPM, a tabular dataset of time-domain statistics (RMS, crest factor, kurtosis) and principal components of the power spectral density was built and used to train a Multilayer Perceptron (MLP) neural network architecture. After comparing a single multi-output model against a configuration of independent submodels per domain (frequency and time), the latter was adopted, showing a systematic improvement across all 10 target variables and raising the average test coefficient of determination (R2) from 0.428 to 0.711. The results show that the virtual sensor reconstructs the blade’s global dynamics with high fidelity, achieving a test R2 of 0.9929 for the principal spectral component and a morphological correlation above 90% in the reconstruction of the power spectral density. A differentiated performance was also identified depending on the measurement axis: while the transverse axes (edgewise and flapwise) are estimated with high accuracy, the longitudinal axis exhibits systematically lower performance. This difference was found to stem from the distinct nature of the mechanical loading dominating each axis: the transverse axes bend under the cyclic aerodynamic load of the incoming flow, whose amplitude responds sensitively to the operating condition, whereas the longitudinal axis is governed by axial stresses of geometricgravitational origin, of nearly constant amplitude and with scarce predictive information available. These results validate the proposed hypothesis and confirm the technical feasibility of estimating a blade’s vibratory field through indirect measurements and machine learning, laying a methodological foundation for its future extension toward condition monitoring and structural health monitoring systems.
Description
Tesis presentada para optar al título de Ingeniero/a Civil Aeroespacial.
Keywords
Turbinas eólicas, Sensores, Mantenimiento