Data-driven frameworks for robust and interpretable damage detection in wind turbine blades
Wind energy has rapidly become one of the pillars of the energy transition generating more than 20% of the world’s renewable electricity in 2019. By the end of 2020, 743 GW of wind capacity had been installed globally. However, to limit global average temperature increases to 1.5°C, approx. 180 GW (including both onshore and offshore wind) must be installed annually until 2030. To ramp up wind energy’s contribution to the energy transition, the cost competitiveness of this technology must be addressed. The levelised cost of electricity (LCOE) from wind is heavily influenced by operational costs. Approx. 30% of the LCOE from onshore wind installations corresponds to operation and maintenance (O&M) costs. This share is higher for offshore wind installations due to the harsher marine environment, skilled labour and specialised vessels required, among other reasons. Maintenance, particularly unscheduled interventions, contribute to downtime. Blades represent a critical component for the reliable operation of wind farms with a failure rate of 0.7 occurrences per year and per turbine in the case of offshore wind. Economic and safety consequences are anticipated when such failures are not timely identified. Blade failures have been identified as the main cause of accidents in wind turbines between 2012 and 2020. Therefore, the research presented in this dissertation is motivated by the frequency and severity of failures in wind turbine blades (WTBs). Vibration-based structural health monitoring (VSHM) systems can contribute to reducing the number of failures and accidents, enhancing profitability, reliability and safety of wind farms. With the rapid development of sensing technologies, VSHM systems enable online and continuous remote integrity monitoring. VSHM systems, which are based on the measurement of vibration signals on structures, have been successfully implemented in offshore oil platforms, aerospace structures and bridges where the damage location is generally unknown and/or accessibility issues arise. Wind turbines share these challenges, yet VSHM systems have not been widely implemented. One of the prevailing challenges for damage detection in WTBs using VSHM systems is the influence of environmental and operational variabilities (EOVs) on the structure and, ultimately, on the sensor readings. EOVs such as temperature, wind conditions and moisture and humidity levels can either camouflage damage or falsely indicate damage. Despite significant efforts to mitigate the influence of EOVs on damage detection and the various approaches adopted by the scientific community to face this challenge, there is no consensus on one single approach that can be generalised and identified as the most accurate and reliable methodology for damage detection. This dissertation introduces a robust, interpretable, data-driven and semi-supervised damage detection framework based on machine learning techniques to address the influence of EOVs. Contrary to conventional damage detection frameworks, where the damage sensitive features (DSFs) are normalised, in this dissertation’s proposed framework, the normalisation of the EOVs is aimed at the novelty index. The effects of EOVs remain perceptible in the novelty indices and are later mitigated through a supervised regression model that is trained to learn the relationships between the DSFs extracted from vibration measurements from an in-operation wind turbine blade, EOVs and novelty indices. The damage detection framework proposed results from the combination of two approaches: one approach addresses the robustness of the damage detection process and the second approach addresses the issue of interpretability of data-driven novelty detection techniques. Different combinations of techniques including artificial neural networks (ANN), gradient boosted decision trees (XGBoost DT) and Shapley Additive Explanations (SHAP) were tested on data extracted from an operating Vestas V27 wind turbine blade and compared to the widely adopted Mahalanobis distance (MD)-based approach. The experiments and analyses conducted throughout this dissertation have shown that the ANN-based approach provides a mitigated novelty index that leads to high damage detection accuracy and the interpretable ML techniques, i.e., the XGBoost DT - SHAP sequence, facilitate the understanding of the effect of EOVs on the monitored structure without mitigating this effect. The XGBoost - SHAP sequence proposed results in a trade-off between accuracy and interpretability, however, this approach still provides higher accuracy relative to the MD-based approach. Ultimately, the choice of approach for damage detection will be application-specific and requires domain knowledge and expertise. The operator should decide whether the trade-off between accuracy and interpretability is acceptable for the monitored structure.