Wind energy is one of the most sustainable and reusable clean energy; thus, more and more wind turbines are installed worldwide. Since wind turbines are operated in a very harsh environment, the key components, such as blades, generators, and bearings, are prone to failure. Therefore, early anomaly detection of wind turbines is critical to keep safe and reliable operation to improve the electricity generation efficiency and reduce the cost due to downtime.

Deep learning has been an effective approach with its ability to extract more informative representation from massive data. Deep learning-based anomaly detection methods have gained much attention from academia. Existing works are mainly composed of the following two ideas. One is that a deep neural network is first utilized to extract informative features from the raw healthy data, followed by an anomaly detection algorithm. The other is reconstruction error-based anomaly detection methods, such as autoencoder-based or generative adversarial network-based anomaly detection methods.

In our work, we proposed the Supervisory Control and Data Acquisition (SCADA)-based deep autoencoder Support Vector Data Description (SVDD). The SCADA dataset collected from wind turbines is the input of the network. It trains one autoencoder network by minimizing the reconstructed error and the volume of the enclosed hypersphere to the center in the latent space of the autoencoder simultaneously. These two loss function terms are balanced by the hyperparameter w, which adjusts the contribution of two loss functions. During the network testing phase, the reconstruction error and the volume of the hypersphere to the center are used to obtain the anomaly score of the test sample. A higher anomaly score will be obtained if the test sample is anomalous.

The proposed method is validated on one real wind farm dataset, the SCADA dataset. Experimental results approve the effectiveness of the proposed method and show more competitive anomaly detection performance compared with state-of-the-art methods.

One of the main challenges in the wind energy industry is reducing the costs related to operations and maintenance, which contribute to a significant portion of the levelised cost of wind energy, especially offshore, where operating conditions are harsher and access for maintenance is limited [1]. Therefore, it is necessary to implement predictive condition-based maintenance schemes which require the development of methods for predicting the remaining useful life (RUL) of wind turbine components [2]. 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 [3]. Several methods have been proposed in the literature for extracting the monotonic trend from sensor signals of a degrading component. They either adopt time-consuming trial-error [4] and computationally expensive search algorithms [5], or include ad-hoc assumptions, such as assuming a linear or cubic regime for the degradation process [6]. 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 performance of this method in constructing HIs and predicting the RUL based on them is verified by two case studies. 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 published solutions 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 and construct monotonic HIs quantified by Mann-Kendall monotonicity metric; and c) the proposed method is able to accurately predict the RUL of the test set components.

The transition from corrective and preventive maintenance to predictive maintenance is crucial for the successful widespread adoption of wind energy. Predictive maintenance involves using data and advanced analytics to predict when a component is likely to fail, allowing for proactive repairs to be made before a fatal failure occurs. Thus, predictive maintenance allows for the cost-effective operation of wind energy systems [1].

On the one hand, it is common for wind turbines that are approaching the end of their expected lifespan to lack specific condition monitoring systems. Condition monitoring systems, such as vibration sensors and ultrasonic testing equipment, are used to identify potential issues with wind turbine components before they become critical [2]. These systems can be expensive to install and maintain, and as a result, they are often not implemented on older wind turbines. On the other hand, the use of Supervisory Control and Data Acquisition (SCADA) data for condition monitoring of wind turbines is a cost-effective approach, specially for wind turbines approaching the end of their lifetime expectancy. SCADA systems are widely used in the wind energy industry to monitor and control equipment performance, and the data generated by these systems can provide valuable insights into the condition of wind turbines [3]. By leveraging SCADA data for condition monitoring, wind energy operators can identify potential issues before they become critical, allowing for timely repairs to be made.

In this work, a fault prognosis strategy is proposed for wind turbines using only SCADA data and transformer neural networks (TNNs). TNNs have recently demonstrated effectiveness in time series forecasting [4], despite their origin in natural language processing and image processing. TNNs are a type of deep learning model that utilizes self-attention mechanisms to process sequential data, such as time series data. These mechanisms allow TNNs to capture long-range dependencies in data and make accurate predictions about future values. TNNs have shown particularly strong performance in tasks involving multivariate time series [5], where multiple variables are changing over time and may be interdependent, as in this work. In particular, the proposed strategy uses SCADA data from a twelve-turbine wind farm, and the TNN model is trained and tested on data from each individual turbine. The results show that the failure of interest is successfully anticipated several months in advance, with no false alarms reported in the rest of the farm. This outcome is indicative of the effectiveness of the strategy.

Wind turbines are complex mechanical systems that convert kinetic energy from the wind into electrical power. To better exploit wind power, they are often located offshore or in remote environments, which makes their maintenance costly in terms of time and money thereby going to increase the Cost of Energy. Continuous vibrations during operation can cause mechanical degradation of components such as gearboxes, motors, sensors, and blades, leading to system failure. To date, the reduction in energy production due to system faults has been a significant limitation, leading to periodic maintenance interventions to inspect component status. However, periodic maintenance is inefficient overall, as failures can occur between interventions if intervals are too long, while unnecessary and costly inspections can be planned and executed if intervals are too short.

Practical and automatic diagnostic and prognostic algorithms are necessary to transition from a "time-based" to a "condition-based" maintenance scheduling approach. These algorithms should be able to track wind turbines' actual health and usage status continuously, identifying anomalous behavior related to incipient faults. Early detection of failures allows for prompt maintenance and prevents irreversible damage to the system. Given the recent successes of machine learning and signal processing in many industrial health and usage monitoring applications, we propose a diagnostic framework for wind turbines that combines these techniques with the initial aim of detecting blades pitch misalignment. On the one hand, our approach leverages signal processing techniques to extract effective time and frequency domain features from the continuously measured variables of the system, leveraging measures that are actually available in commercial machines.
On the other hand, it leverages machine learning capabilities to associate the values of these features with the health and usage status of the wind turbine components of interest. Combining time and frequency domain features is particularly promising in this scenario, due to the intrinsic periodicity of the system. When a component experiences a failure, its vibrational signature changes, and the deviation from the expected behavior varies with the severity of the damage.

Our health and usage monitoring system has the potential to extend the life cycle of wind turbines, schedule maintenance interventions more effectively, and maintain a constant level of energy production. By continuously tracking the system's status, in fact, one can avoid unnecessary operations and identify potential issues before they result in costly failures. The idea proposed here for pitch misalignment, will later be extended to other wind turbine sub-components.