Abstract Summary
One of the main challenges in wind energy production is reducing the costs related to Operations and Maintenance, which contribute to a significant portion of the Levelised Cost of Energy produced by wind farms, especially offshore, where wind conditions are harsher and access for maintenance is far more limited. Therefore, it is necessary to replace the existing corrective and preventive maintenance interventions with predictive Condition-Based Maintenance schemes. This task necessitates developing methods for predicting the remaining useful life (RUL) of wind turbine components. These methods afford the wind farm operators the possibility of optimally planning the necessary maintenance. A typical step before predicting the RUL in many prognostic frameworks is constructing a health indicator (HI) correlated with the component degradation trend. One of the most promising approaches for constructing HIs focuses on their monotonicity (congruent with the irreversible nature of the component degradation) as a major performance metric. Several methods have been proposed in the literature for extracting the monotonic degradation factor from sensor signals of a degrading component. They either adopt time-consuming trial and error and computationally expensive search algorithms, or include ad-hoc assumptions, such as assuming a linear or cubic regime for the degradation process. In this work, a more efficient and rigorous method is proposed based on the development of a Convolutional Autoencoder (CAE) which is trained using the Particle Swarm Optimization (PSO) algorithm for simultaneous maximization of the monotonicity of the HI output from its inner layer and minimization of the mean squared error between the input and the reconstructed signals at the output layer. Subsequently, a Long Short-Term Memory (LSTM)-based Recurrent Neural Network (RNN) is trained on the constructed HIs to predict the RUL of the test set components. The feasibility of using the CAE-PSO approach to construct HIs describing the true component degradation and its performance compared to other solutions proposed in the literature is verified in case study 1, which is a synthetic dataset of a hypothetical degrading component’s sensor signals. Then, the generalization ability of CAE-PSO approach and the feasibility of using the constructed HIs for predicting the RUL with the LSTM-based RNN is evaluated in case study 2, which is the Aramis Challenge dataset presented in ESREL 2020 conference. This dataset includes simulated sensor signals of 800 identical components working in various operational conditions. No assumption is made about the type of the components. The results show that: a) the CAE trained with PSO is able to construct HIs which are congruent with the true degradation factor of the input signals and outperform other solutions published for case study 1 in terms of the monotonicity of the constructed HI; b) despite being trained on a limited training set, the model is able to generalize to an extensive test set quantified by comparing the monotonicity scores of the constructed HIs for the two sets; and c) the proposed method is able to predict the RUL of the test set components with a low error.
