Conventional Deep Learning (DL) methods for bearing health indicator (HI) adopt supervised approaches, requiring expert knowledge of the component degradation trend. Since bearings experience various failure modes, assuming a particular degradation trend for HI is suboptimal. Unsupervised DL methods are scarce in this domain. They generally maximise the HI monotonicity built in the middle layer of an Autoencoder (AE) trained to reconstruct the run-to-failure signals. The backpropagation (BP) training algorithm is unable to perform this maximisation since the monotonicity of HI subsections corresponding to input sample batches does not guarantee the monotonicity of the whole HI. Therefore, existing methods achieve this by searching AE hyperparameters so that its BP training to minimise the reconstruction error also leads to a highly monotonic HI in its middle layer. This is done using expensive search algorithms where the AE is trained numerous times using various hyperparameter settings, rendering them impractical for large datasets. To address this limitation, a small Convolutional Autoencoder (CAE) architecture and a hybrid training algorithm combining Particle Swarm Optimisation and BP are proposed in this work to enable simultaneous maximisation of the HI monotonicity and minimisation of the reconstruction error. As a result, the HI is built by training the CAE only once. The results from three case studies demonstrate this method’s lower computational burden compared to other unsupervised DL methods. Furthermore, the CAE-based HIs outperform the indicators built by equivalent and significantly larger models trained with a BP-based supervised approach, leading to 85% lower remaining useful life prediction errors.

Wind turbine supervisory control and data acquisition (SCADA) datasets available for research usually contain a limited number of failure events. This limitation hinders the successful application of deep learning (DL) methods for fault detection and prognosis, as they require large datasets for robust training and generalisation. This work proposes a method using conditional generative adversarial networks (cGANs) to generate synthetic SCADA time series that replicate wind turbine behaviour under controllable operational, environmental, and degradation conditions. Given a set of SCADA time series representing these conditions, the cGAN generates temperature and pressure time series simulating gearbox operation. Results show that augmenting the training set of an artificial neural network (ANN) fault detection model with synthetic time series reduces false positives in the detected gearbox faults by 84 % on average, enabling the model to blindly detect a fault in a test wind turbine without prior knowledge of the event. Furthermore, training a convolutional autoencoder-based unsupervised health indicator (HI) model with both real and synthetic SCADA time series leads to an HI that more accurately captures the expected degradation trend. Using this HI, the gearbox's remaining useful life (RUL) can be predicted within the defined error bounds from around 4.5 months before the detection of the fault, while the HI obtained without the synthetic data fails to produce reliable RUL estimations.

State-of-the-art Deep Learning (DL) methods based on Supervisory Control and Data Acquisition (SCADA) system data for the detection and prognosis of wind turbine faults require large amounts of failure data for successful training and generalisation, which are generally not available. This limitation prevents benefiting from the superior performance of these methods, especially in SCADA-based failure prognosis. Data augmentation approaches have been proposed in the literature for generating failure data instances within a SCADA sequence to reduce the imbalance between healthy and faulty state data points, which is relevant to fault detection tasks. However, the successful implementation of DL-based failure prognosis methods requires the availability of multiple run-to-failure SCADA sequences. This paper proposes a data-driven method for generating synthetic run-to-failure SCADA sequences with custom operational and environmental conditions and progression of degradation. An Artificial Neural Network (ANN) is trained with signals that represent these factors to reconstruct the SCADA signals. Then, it is used to generate synthetic SCADA datasets based on data available from a wind turbine that experienced a gearbox failure. Synthetic data sets generated are evaluated on the basis of the similarity of their signal distributions, the temporal dynamics within each signal, and the temporal dynamics among different SCADA signals with those in similar field datasets. The results show that the generated synthetic datasets are consistent with their field counterparts, with a comparatively lower diversity in their dynamic behaviour in time.

In this paper, a set of best practice data sharing guidelines for wind turbine fault detection model evaluation is developed, which can help practitioners overcome the main challenges of digitalisation. Digitalisation is one of the key drivers for reducing costs and risks over the whole wind energy project life cycle. One of the largest challenges in successfully implementing digitalisation is the lack of data sharing and collaboration between organisations in the sector. In order to overcome this challenge, a new collaboration framework called WeDoWind was developed in recent work. The main innovation of this framework is the way it creates tangible incentives to motivate and empower different types of people from all over the world to share data and knowledge in practice. In this present paper, the challenges related to comparing and evaluating different SCADA-data-based wind turbine fault detection models are investigated by carrying out a new case study, the “WinJi Gearbox Fault Detection Challenge”, based on the WeDoWind framework. A total of six new solutions were submitted to the challenge, and a comparison and evaluation of the results show that, in general, some of the approaches (Particle Swarm Optimisation algorithm for constructing health indicators, performance monitoring using Deep Neural Networks, Combined Ward Hierarchical Clustering and Novelty Detection with Local Outlier Factor and Time-to-failure prediction using Random Forest Regression) appear to exhibit high potential to reach the goals of the Challenge. However, there are a number of concrete things that would have to have been done by the Challenge providers and the Challenge moderators in order to ensure success. This includes enabling access to more details of the different failure types, access to multiple data sets from more wind turbines experiencing gearbox failure, provision of a model or rule relating fault detection times or a remaining useful lifetime to the estimated costs for repairs, replacements and inspections, provision of a clear strategy for training and test periods in advance, as well as provision of a pre-defined template or requirements for the results. These learning outcomes are used directly to define a set of best practice data sharing guidelines for wind turbine fault detection model evaluation. The guidelines can be used by researchers in the sector in order to improve model evaluation and data sharing in the future.