Détection, diagnostic et pronostic pour les éoliennes : développement de modèles basés sur l’autoencodeur variationnel utilisant des données SCADA

Abstract

Wind turbines are among the key technological solutions for energy transition. A significant increase in the wind energy installed capacity worldwide is advocated to limit climate change and its consequences. For instance, the "Energy Transition Outlook to 2050" (DNV, 2023) estimates that the global wind energy installed capacity should increase eightfold between 2022 and 2050. Increasing the penetration rate of wind energy globally relies on the continuation of the trend of decreasing the Levelized Cost of Energy (LCOE) of this source. This can be achieved by increasing availability rates and reducing operation and maintenance (O&M) expenditures of wind turbines through optimal O&M strategies. Operators currently analyze the health condition of wind turbines using data from the built-in Supervisory Control and Data Acquisition (SCADA) system. The quest for optimal O&M strategies motivates research in the field of Prognostics and Health Management (PHM). The PHM analysis comprises the detection, the diagnosis, and the prognosis of abnormal conditions. The models resulting from these analyses provide information to guide decision-making regarding O&M, especially in situ interventions such as inspections and repairs. Models for detecting and diagnosing operational anomalies are well developed and widely used in wind farm operation. In contrast, prognosis models still pose scientific and technical challenges. Additionally, the lack of interpretability hinders the use of AI models in the analysis of wind turbine health conditions despite their proven potential. In this context, this research investigates the implementation of the PHM analysis using SCADA data and interpretable AI models. Specifically, it sought to detect anomalies earlier than current methods, diagnose degraded conditions among multiple health conditions, and establish a prognostic model for wind turbines. The research is focused on degradation modes on critical wind turbine components, namely the generator, gearbox, main bearing, and blades. A SCADA database comprising more than two years of operation and nearly 120 wind turbines was used to develop and validate the proposed approaches. This research introducedPHManalysis using Latent Variable Models (LVM), which are variations of the Variational Autoencoder (VAE) model. A supervised LVM, the Variational Autoencoder Embedded with Classifier (VAEC), projects the high-dimensional physical space of SCADA measurements into a low-dimensional latent space. The geographical position of the projected point in the latent space informs the operational condition of the wind turbine. The latent space of the VAEC is representative of the physical condition of wind turbines, allowing the definition of relevant health indicators. Furthermore, the space can be defined in 2D to provide a convenient visualization tool that adds interpretability to the proposed approaches. The first paper presents the VAEC model and demonstrates its use for the analysis and visual representation of SCADA data. It introduces a new approach aiming at the health condition monitoring of wind turbines using the VAEC latent space. The Health Index (HI) uses the Mahalanobis distance and the exponentially weighted moving average (EWMA) control chart. Real case studies show that this approach can detect anomalies earlier than the SCADA control system and regression-based approaches. This first paper also demonstrates the use of the VAEC classifier for diagnosing abnormal operating conditions in wind turbines. Case studies illustrate the potential of the integrated visualization tool to improve interpretability and confidence in the model results. The second paper introduces a new approach for monitoring the health condition of complex systems based on a standardized representation of theVAEC latent space. An original contribution of this paper is the use of the Nataf isoprobabilistic transform to map the latent space into a standardized space. In the Nataf standard space, the healthy condition follows the standard normal distribution, allowing the definition of two complementary HIs and appropriate thresholds based on a predefined confidence level. Moreover, the use of the Nataf transform reduces the approach’s sensitivity to hyperparameters during the learning process. The implementation in two case studies demonstrates the potential of the proposed approach and its ability to detect anomalies earlier than competing approaches. First, the approach was applied to NASA’s Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) database. The second case study involves a database with SCADA measurements from modern wind turbines, including multiple abnormal operation classes. The second paper also demonstrates the use of the latent space and the Nataf standard space as visualization tools, tracking the evolution of probability distributions within the latent and the Nataf spaces. The third paper proposes an approach to estimate the remaining useful life (RUL) from the VAEC latent space projecting the SCADA data. The projection of time-oriented datasets defines trajectories in the latent space. Such trajectories inform about the evolution of the system’s health condition. The latent space projection is described as a smooth trajectory by using kinematic modelling and an appropriate regularization method. The approach uses Taylor expansion and the Monte Carlo method to estimate the RUL probability density, with a criterion defining the end of useful life (EOL) in the latent space. Application to post-mortem degradation case studies of wind turbine components shows consistency between the estimated and ground-truth RUL. In conclusion, this thesis presents innovative AI approaches for the detection, diagnosis, and prognosis of wind turbines using SCADA data. The proposed AI model is interpretable via a built-in visualization tool. The three proposed papers demonstrate the use of the VAEC model for analyzing the complex behaviour described by SCADA measurements. The VAEC latent space enabled the implementation of the PHM analysis process for wind turbines while emphasizing model interpretability to facilitate understanding and adoption by operators, analysts, and technicians. The results demonstrate the effectiveness of health indicators defined from the latent space, in particular for early anomaly detection and accurate diagnosis. The proposed prognosis model, based on the trajectory projected in the latent space, provides a reliable estimation of the wind turbines’ remaining useful life. These technical contributions, in addition to the scientific outcomes, lay the foundation for developing software solutions that can be integrated into wind turbine performance monitoring tools. Thereby, it can facilitate technological transfer and the industrial application of the proposed approaches. The outcomes of the research can support decision-making related to the O&M of wind turbines, potentially contributing to the reduction of costs and the increase of reliability and availability. The final chapter of the thesis presents recommendations for future research, for the technological transfer and deployment of the proposed methods.

Date	15 Jul 2024
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Antoine Tahan (Supervisor) & Francis Pelletier (Co-supervisor)