Data-Driven Models Applied to Predictive and Prescriptive Maintenance of Wind Turbine: A Systematic Review of Approaches Based on Failure Detection, Diagnosis, and Prognosis

Abstract

Wind energy has achieved a leading position among renewable energies. The global installed capacity in 2022 was 906 GW of power, with a growth of 8.4% compared to the same period in the previous year. The forecast is that the barrier of 1,000,000 MW of installed wind capacity in the world will be exceeded in July 2023, according to data from the World Association of Wind Energy. In order to support the expected growth in the wind sector, maintenance strategies for wind turbines must provide the reliability and availability necessary to achieve these goals. The usual maintenance procedures may present difficulties in keeping up with the expansion of this energy source. The objective of this work was to carry out a systematic review of the literature focused on research on the predictive and prescriptive maintenance of wind turbines based on the implementation of data-oriented models with the use of artificial intelligence tools. Deep machine learning models involving the detection, diagnosis, and prognosis of failures in this equipment were addressed.

1. Introduction

The development of a society occurs with the use of energy, especially electrical energy. The supply of this form of energy is carried out through different sources that can be of a clean and renewable nature or of a non-renewable and polluting nature [1]. The choice of the type of energy source to be used depends on several factors, such as availability, technical, economic, and political issues. In the long term, the economic and environmental benefits of wind energy are superior to those of conventional energy, setting a new trend in the future development of electric power generation [2].

The planning of the energy supply must take into account all the aspects that may influence the load of the electrical system, since the energy supply must guarantee, at least, the supply of the existing demand, and it is important to know the behavior of each aspect and its externalities. This growth has been particularly noticeable in the wind energy sector, which occupies a leading position among renewable energies. Furthermore, not only has the sector undergone great development in the last two decades, but it should also continue its expansion in the coming years, duly reinforced by the energy plans of the great world powers. Wind power generation is an important measure for developing a circular economy and alleviating resource constraints [3].

The World Wind Energy Association—WWEA—has gathered statistical data on the increase in installed wind power capacity in the world. In the last 10 years, there has been an increase of approximately 623 Gigawatts of wind power. In 2022, around 76 Gigawatts were installed, which corresponds to an increase of 8.4%. As a result, the global installed capacity in the same year was 906 Gigawatts (Figure 1).

Given the scenario of strong growth in the sector, there is a forecast that the barrier of 1.0 million MW of wind capacity installed in the world will be surpassed in 2023. China continues to be the most representative country with 365.4 GW, followed by others markets, such as the USA and Germany, with 144.2 and 67.0 GW of installed wind power, respectively, according to data from December 2022.

In Brazil, wind energy ended 2022 with 904 plants and 25.63 GW of installed wind power, representing a growth of 18.85% in power compared to December 2021, when the installed capacity was 21.57 GW. Around 109 new wind farms were installed in 2022, resulting in a record 4.06 GW of new capacity. According to data from the Global Wind Energy Council (GWEC), Brazil was the third country that installed the most wind farms in the world, which resulted in the growth of its Global Ranking of Installed Capacity to the sixth position. Still, in 2022, an average of 9277.9 MWh of wind energy was generated. It can be seen in Figure 2 that the annual sum corresponds to 111.3 GWh.

With the increase in the installed capacity of wind energy in the world, the costs of operation and maintenance (O&M) of wind farms have grown significantly. Within this context, the asset management of these projects acquires great relevance in the sector, as it is crucial to seek optimal maintenance strategies that allow for improving the reliability of wind turbines and reducing their cost, resulting in an increase in their availability [6].

In general, existing maintenance strategies for wind turbines may include corrective maintenance (MC), preventive maintenance (PM), and condition-based maintenance (CBM), which can be identified by the operating condition of the equipment [7]. Corrective maintenance is used after the failure or breakdown of a wind turbine component or system. In the preventive maintenance strategy, a plan is executed based on the analysis of periodicity according to the operating rules, which are estimated based on the manufacturing specifications and engineering experience. Condition-based maintenance uses predictive maintenance to monitor certain components [8,9].

Several problems can cause the loss of production of a wind turbine. When the problem is easy to solve, such as by making some adjustments to the turbine parameters or even replacing certain electronic devices, the time associated with the repair is not too long, and, many times, some of these actions can be performed remotely by the operator from the control center. However, if the problem is bigger and requires the replacement of large equipment, the interruption time is great, which, in turn, will directly lead to prolonged unavailability, resulting in a reduction in profits in this type of scenario [10].

Wind turbines can be considered complex electromechanical equipment and typically have a useful life of between 20 and 30 years. The reliability of the systems that make up this equipment is a decisive factor in the success of an energy project. In recent studies, it was observed that the costs associated with the maintenance and repair of wind turbines could represent around 25 to 30% of the life-cycle cost [11], in addition to the fact that low reliability would directly reduce their availability [12].

With the advancement of information technology, wind farms today have a system known as SCADA, which stands for supervisory control and data acquisition and monitoring. These devices collect historical data from the sensors put in the equipment, which may then be used to facilitate fault detection and diagnosis using deep learning algorithms [13,14].

The basis for the implementation of these algorithms can be found in a methodology that helps in the maintenance processes of wind turbines. It corresponds to the application of techniques based on Prognostic Health Management—PHM—for the continuous monitoring of the “health” of the turbines. In this type of methodology, the processes of data acquisition, condition assessment, diagnosis, prognosis, and decision support are generally identified with regard to maintenance actions [15].

The usual maintenance procedures in wind turbines do not follow the expansion of the wind source, the need for complementing the load (demand), or the criterion for establishing assured energy, even due to variability in the wind speed. This becomes even more complex when these assemblies show component wear and tear that can evolve into failures that, depending on the type of wind turbine, present very different modes and mechanisms, such as those associated with wind turbines with and without a gearbox and with and without self-excitation.

Ref. [16] conducted a research study on wind turbine failures. The failure rates of fixed-speed (with multiplier) and variable-speed (direct drive) turbines were compared. The main discoveries were as follows:

• Failure rates in systems that make up the gearbox are higher than those of frequency inverters when analyzed in indirect-drive turbines;
• Systems composed of frequency inverters and electronic components present higher failure rates in direct-drive turbines when compared to the failure rates of gearbox systems used in indirect-drive turbines;
• The direct-drive wind turbine generator system has failure rates at least twice those of the indirect-drive system.

Ref. [17] analyzed four different configurations of gearboxes, generators, and energy converters for wind turbines. The findings revealed that turbines with permanent magnet generators and fully nominal power converters will have greater availability and cheaper operating and maintenance costs than turbines with doubly fed induction generators. It was also discovered that in turbines with permanent magnet generators, the direct-drive design offers the highest availability and the lowest operating and maintenance costs, followed by turbines with two-stage and three-stage gearboxes. The powertrain system, which includes the rotor, main bearings, gearbox, generator, and power converter, is responsible for 57% of all turbine failures and 65% of downtime [18].

Wind turbines, an integral part of this energy growth, are increasing in size and power, and unscheduled stoppages due to failures or component breakages can compromise this expected growth. These turbines operate under a wide range of complicated and dynamic loads caused by environmental circumstances such as wind shear, turbulence, gusts, and so on. Wind farms require condition-based maintenance, as well as prognostic and health measures, to maintain reliable and efficient performance. This is possible with the application of data-oriented models that, due to the reduction in sensing costs, can use an amount of data for the development of techniques aimed at machine learning and deep learning, enabling the diagnosis and prognosis of failures [19,20,21] (Figure 3).

Fault diagnosis has a high degree of importance, but as the operating system of wind turbines is complex, advance knowledge of the state of components and fault prognosis are state of the art in the operation and maintenance of wind farms.

Although some advances have been made in diagnosing wind turbine failures, little research has been carried out in the field of wind turbine failure prognosis. As a result, there is an urgent need to develop prognostic methodologies for such complex systems working in real-world situations that, due to the involvement of several components and the presentation of multiple failure modes, cannot be tackled using a simple modeling approach [22].

The phases involving a data acquisition process are integral parts of implementing PHM. In real systems, such data obtained by sensors installed in the equipment may contain contaminants, identified through noise, the absence of relevant points, and even redundancy. The importance of this analysis, which involves pre-processing tools, has been identified as a point to be improved in research used to develop ML and DL models.

Given this context, this work presents a systematic literature review (SLR) involving the analytical research of recently published works that deal with the predictive and prescriptive maintenance of wind turbines through the implementation of data-oriented models, based on the use of deep machine learning, that carry out the detection, diagnosis, and prognosis of faults in this equipment. This process aimed to have a broad view of the current technologies and methodologies employed [23,24]. The structure of this work is divided with the aim of providing a sequence of information that will allow a growing and contextualized understanding of the subjects addressed. Section 2 provides an overview of the theoretical foundations, establishing understandings regarding condition-based maintenance, the structure of PHM, machine learning and deep machine learning models, and wind turbine systems and arrays. Section 3 presents the state of the art, obtained through an extensive search of articles that deal with data-driven models using artificial intelligence and prescriptive maintenance approaches applied to wind segments. Section 4 identifies the results found and opportunities for research under development. And, finally, in Section 5, the final considerations of the research work are made.

2. Theoretical Foundations

2.1. Condition-Based Maintenance (CBM)

According to the European standard [25] that deals with terminologies associated with maintenance, CBM can be considered a form of “preventive maintenance that includes assessment of physical conditions, analyzes and possible resulting actions”. In its simplest form, it acts as a type of threshold-based maintenance, raising alarms when thresholds are reached. However, CBM has evolved far beyond this and is now used in the framework of predictive and prescriptive maintenance as well [26].

In general, the tools used in predictive or prescriptive CBM can be divided into the approaches described in the following subsections.

2.1.1. Data-Driven Techniques

This type of approach does not require prior knowledge of the system being analyzed. However, there is a need to obtain a large amount of data that can be used in artificial intelligence models for, for example, fault detection [27].

2.1.2. Physics-Based Techniques

In this type of approach, in-depth knowledge of the system or physical process is required, using scientifically established relationships to describe them. The residuals between the relative system measurements and the expected output of the physical model must be used to monitor its operational state [28].

2.1.3. Stochastic Techniques

Stochastic approaches are thought to have and employ intrinsic randomness in their modeling, and they often produce a probability distribution of the variable of interest as output. Data are often collected and evaluated to determine the required model parameters [29].

2.1.4. Hybrid Techniques

It is difficult to predict parameter trends using a single method in fault prediction or prognostic processes. The techniques can leverage each model’s strengths. Hybrids frequently combine knowledge of a physical failure process with data that cannot be effectively modeled, such as environmental or inspection data [30].

2.2. Prognostic and Health Management—PHM

In order to apply CBM to assets, it is necessary to develop a structure that involves the use of models, equipment, and people. The structure is known as Prognostic Health Management—PHM. In order to manage the health of the assets, this structure must perform some fundamental steps, such as collecting and storing data, detecting equipment defects, diagnosing defects, performing equipment prognosis, and, finally, ensuring that this information is properly used in decision support [31,32,33]. Figure 4 identifies the PHM cycle.

2.3. Machine Learning

According to [34], machine learning, in the area of data science, corresponds to the development of models aimed at data access so that, from these data, they can recognize and reproduce patterns with the objective of achieving high performance when performing a given task.

Machine learning models can be identified as supervised, unsupervised, semi-supervised, and reinforcement. In the supervised ones, there is previous labeling that differentiates the data considered normal from the discordant observations and, therefore, is used during the training phase of the model. In unsupervised models, there are no previously established labels, so learning occurs with a search for which points deviate from the majority distribution of the data. On the other hand, semi-supervised models are only trained with data that present a specific behavior, so, in the test set, we try to identify the points that differ from the original distribution [35]. Finally, the reinforcement learning model explores possible solutions based on a reward function through trial and error procedures, identifying the highest bonuses [36].

2.4. Artificial Neural Networks and Deep Learning

According to [37], artificial neural networks are computational techniques that, with the implementation of mathematical models, try to imitate how the brain performs some action or activity of interest by using electronic components or computer simulation. The learning process is called a learning algorithm, which has the task of modifying the network connection weights to achieve a certain goal. For [34], when a neural network has a relatively large number of layers, it is understood that it is part of the family of deep learning algorithms. In the following, briefly, some models are presented.

2.4.1. Perceptrons and Multilayer Perceptron—MLP

In 1943, the neuroscientist Warren MacCulloh was the one who first proposed the neural network algorithm, presenting a simplified mathematical modeling structure of data processing in biological neurons. As soon as 1957, the structure of the Perceptron was presented, which was composed of only a one-step-function operator connected to a weighted summation of the input data $x\in$ Rn. This operator transforms a linear combination of input data into a binary value. Perceptron training is represented by the search for weights, w, that minimize an output error function.

The limitation of using a Perceptron network is its ability to solve only linear problems that do not correspond to most real classification and regression problems [34]. The solution to this problem came with the use of neural networks called Multilayer Perceptrons (Multilayer Perceptron—MLP), composed of an input layer, an output layer, and at least one hidden layer. It is verified that the neuron outputs are linearly combined and processed from an activation function of the posterior layers. The nature of the output of the neural network depends on the range of the activation function of the last layer. The training of an MLP network consists of a structured search of weight values from the minimization of classification or regression error measures (Figure 5).

2.4.2. Recurrent Neural Networks—RNNs

RNNs have characteristics that provide them with the ability to retain information with the advancement of model training, providing subsidies for temporal dependencies to be absorbed [36,39]. However, due to the insertion of signal recurrence, there is a significant increase in the computational cost, which does not allow an adequate level of computational parallelism, making it difficult to use it in problems that present large volumes of data.

Long Short-Term Memory—LSTM—networks, developed by researchers Hochreiter and Schmidhuber in 1997, manage to solve a problem that affects RNNs, known as the disappearance of the gradient, which occurs with the evolution of learning. The internal architecture of the neural cell presents mechanisms (gates) that control the information coming from the temporal states of the cells, that is, functions of the sigmoid type and hyperbolic tangent that control the propagation of signals inside the cell. Consequently, the model is able to select information that will be forgotten or propagated [40,41,42].

2.4.3. Convolutional Neural Network—CNN

Convolutional neural networks are considered deep learning models and have great capabilities in resource extraction through mathematical convolutions, reducing their dependence on human intervention and specialized knowledge [37,43] (Figure 6).

In general, mathematical convolution is the first step and the main component of CNNs. The application of low-dimensional filters results in the extraction of features from the network input signal, and the number of features extracted from the input signal is directly associated with the volume of filters applied in the convolution step. Consequently, the increase in filters in the convolution step increases the computational cost. Secondarily, the pooling technique is applied. The aim of the technique is to accentuate the features extracted by the convolution, as well as to apply a reduction in the dimensionality of the resulting matrices [38,44].

Other applications presented in the literature that use CNNs for fault diagnosis include monitoring the integrity of bearings [45,46], monitoring gear wear [47], identifying and classifying faults in rotary switches using machine vibration data [48], and identifying disturbances in the voltage level of circuit breakers [49].

2.4.4. Autoencoders

Autoencoders are artificial neural networks that have the ability to learn dense representations of input data, identified as latent or encodings, without any type of supervision. These encodings acquire a much smaller dimensionality than the input data, helping in this reduction, mainly for visualization purposes [34].

The most basic possible autoencoder consists of a network of only three layers: an input data layer, a latent space layer, and an architecture output layer (Figure 7).

The optimization of this neural network consists of searching for the weights of the encoder (WT) and the decoder (VT) in order to reduce a loss function, as it is a regression model.

2.4.5. Variational Autoencoder—VAE

This neural network category was introduced by Diedrik Kingma and Max Welling in 2013 and has become one of the most popular types of autoencoders. In general, the VAE tries to encode the training data, not in a reduced-dimensional vector space but in a probability distribution whose density function is defined a priori, and during the optimization process, the best parameters are sought for that distribution. The architecture contains a process for sampling a multivariate normal distribution by simulation, and the network process converts the input data into parameters of this distribution [51].

2.5. Main Components of a Wind Turbine

2.5.1. Generators

There are four main types of generators used in wind generation systems.

Table 1. Classification and characteristics of wind generators [52].

Types of Generators	Technical Characteristics
Squirrel-Cage Induction Generator (SCIG)	An electronic converter controls the rotor; The machine has a greater number of poles; The gearbox can be removed; To create the magnetic field of the synchronous generator rotor, it is necessary to include another converter connected to the slip rings.
Doubly Fed Induction Generator (DFIG)	It features a multiplier box; It is connected directly to the mains; The rotor is connected through a directional electronic converter; The converter acts as an excitation source; It allows disturbances in the network; The machine is required to have slip rings on the rotor.
Wound Rotor Synchronous Generator (WRSG)	The multiplier box can be removed; The electronic converter has a larger number of poles; To create the magnetic field of the synchronous generator rotor, it is necessary to include another converter connected to the slip rings.
Permanent Magnet Synchronous Generator (PMSG)	There is no need for a multiplier box; There are no rotor windings; The need for maintenance is reduced; It eliminates the need for an excitation circuit; It has complete control of active and reactive power, acting in a wide range of wind speeds.

shows the main characteristics of each type [52].

In general, generators can be classified as synchronous or asynchronous; however, the permanent magnet synchronous generator (PMSG) is identified as one of the most widely adopted solutions due to its greater efficiency and higher weight/power ratio when compared to electrically excited machines. On the other hand, PMSG-type generators make use of electronic power converters that are properly sized for the nominal power of the machine, precisely because of restrictions imposed by the external network [53,54,55].

2.5.2. Converters—Power Electronics and Electrical Controls

The generators coupled to turbines have voltage generated with varying amplitudes and frequencies due to the seasonality of the wind regime. This interconnection of wind generation systems to the grid is only possible with the use of electronic converters that are responsible for directly controlling the injected power, as well as the generator torque. Indirectly, they are responsible for controlling the speed of the turbine and for isolating the generation system from the power grid [52].

Electronic controls account for about 1% of the cost of a wind turbine but cause 13% of failures. It is very important to increase the diagnostic effort for electronic controls. The costs for the power electronics of variable-speed, direct-drive turbines are much more significant than the costs associated with constant-speed turbines. When compared to other techniques, the diagnosis of power electronics is more complex, because there is a very short time between the appearance of the failure and the event that generates the unavailability of the asset [56,57].

2.5.3. Turbine Rotors, Blades, and Hydraulic Controls

Wind turbine rotors are subjected to different types of loading as well as creep fatigue, which can even be responsible for cracks in the blades. The imbalance between the rotor and its aerodynamic asymmetry may exist due to manufacturing defects, dirt, humidity, or processes that involve accumulated damage to the blades themselves. The diagnosis of blade failures has been studied based on strain measurement techniques. For the blades of small wind turbines, Ref. [58] used a piezoelectric impact sensor that monitored and monitored this wear.

Pitch movement is normally driven by hydraulic actuators or electric motors. For situations of great aerodynamic load, hydraulic systems are considered more resistant to failures. According to the study carried out by [59] regarding the operation of a Swedish wind farm, in the period from 2000 to 2004, 13.3% of failure events were due to hydraulic systems.

2.5.4. Wind Turbines with and without Gearbox

Wind turbines that have a fixed-speed drive system, known as a gearbox, use equipment known as a gearbox to increase the speed and, consequently, the generated power ratio. This type has a lower mass (generator). Variable-speed generators, also known as direct-drive generators, have larger and more complex generators and are capable of converting low blade rotation into alternating wind energy, meeting the desired frequency criteria.

The ability to detect and diagnose the premature wear of the wind turbine gearbox gears allows these assets to obtain more reliability and availability in their production process. Ref. [60] carried out a comparative study with a variety of methods that were used to monitor conditions in vibration, with an emphasis on the cepstrum, which identified problems in the teeth of the gears, while the envelope analysis can detect most of the failures that occur related to the bearings.

Even with the use of direct-drive wind turbines, due to the lower presence of mechanical components, the vibration analysis can be focused on detecting faults in the main bearings. In the study by [61], carried out on a test bench, models of the pre-processing of temporal data were used for the detection of faults using wavelet and Fourier transforms.

3. State of the Art

An extensive search was carried out in the Scopus databases, aiming to search for articles and journals that had a close relationship with the object of study. The keywords used for the search can be seen in

Table 2. Research to identify articles with the aid of Boolean operators.

Objectives	Keywords	Documents
Identify the most recent approaches and frameworks for applying fault detection and diagnostic models	(“PHM” OR “Prognostic and Health Management” OR “Condition Based Maintenance” OR “Predictive Maintenance”) AND (“Diagnosis” OR “Diagnostic” OR “Fault Classification” OR “Fault Detection”)	886
Identify applications of PHM models in similar contexts of direct-drive wind turbines	(“Permanent magnet” OR “PMSG” OR Direct-drive) AND (“WIND TURBINE” OR “EOLIC” OR “ENERGY”) AND (“Prognostic and Health Management “ OR “Condition Based Maintenance” OR “Predictive Maintenance” OR “Machinery Fault Diagnosis” OR “Machinery Prognosis” OR “PHM”)	27
Identify the latest prognostic modeling approaches	(“PHM” OR “Prognostic and Health Management” OR “Condition Based Maintenance” OR “Predictive Maintenance”) AND (“Prognosis” OR “Prognostic” OR “Remaining Useful Life”)	327
Research approaches related to the implementation of prescriptive maintenance of wind turbines	(“Prescriptive Maintenance” AND “Wind turbine” OR “EOLIC OR ENERGY) AND (“Prognostic and Health Management “ OR “Condition Based Maintenance” OR “Deep Learning” OR “Machinery Fault Diagnosis”)	16

.

It is observed that there was a considerable increase in publications after the year 2017, reaching the maximum in 2021 with 289 articles and 198 in the year 2022 (Figure 8). This can also be explained by the significant increase in scientific articles that reproduce constructive maintenance processes associated with the implementation of models based on artificial intelligence.

The level of research relevance can be analyzed through the impact factor based on the Journal Citation Reports—JCR—applied to a selection of the top 10 rankings of the journals in which the articles were published (Figure 9).

The highest percentage (40%) was identified for the range of journal impact factor values between 3.001 and 5.000, while the lowest (10%) was attributed to the range of impact factor values between 1.001 and 3.000.

For a conceptual analysis that establishes interactions between the vertices and nodes of topics and keywords related to the research, Figure 10 depicts a network that was assembled using links and criteria of importance of a quantitative nature.

3.1. Approach Based on Detection of Anomalies and Failures

Anomaly detection aims to identify instances that present deviations from the others, with patterns that can be generated by any other mechanisms. Fault detection can be understood in a simplified way as a binary classification task, such as checking whether the item of interest is performing well or not [62].

Ref. [63] proposed a methodology based on autoencoders (AEs) and LSTM recurrent layers, with the objective of detecting anomalies in bearings with a semi-supervised model. Initially, pre-processing was performed through wavelet decomposition and the Fourier transform to obtain information about the vibration. The extracted data formed inputs for five different autoencoders, whose objective is to extract more relevant information from the input vectors and, consequently, reduce the dimensionality of the data. The output of the five AEs was transferred to the LSTM recursive model, which detected anomalous situations. The results showed 99% accuracy in detecting anomalies in operating conditions.

Ref. [64] developed a methodology for anomaly detection and failure analysis for components of a wind turbine based on an autoencoder. In this context, the algorithm was trained for data reconstruction, and the anomaly was detected from the reconstruction error. Thus, anomalous data are those that the AE had difficulty replicating. A challenge of this approach was the determination of a threshold for the replication error. For this, the authors proposed the use of an adaptive threshold defined through extreme value theory. Such an approach has been tested and well evaluated on turbine vibration data.

In a context similar to the one presented, ref. [65] proposed an approach for the use of autoencoders, and in this case, the data originated from multiple sensors. The multimodal autoencoder was able to perform multiple and dependent time-series replication activities. Still in the context of multiple time series, ref. [66] applied a model based on a Denoising Autoencoder (DAE). The architecture developed by the authors was able to tolerate the intrinsic noise and fluctuations from the digital monitoring of a wind turbine (Figure 11).

As input, the Denoising Autoencoder (DAE) receives raw data from sensors monitoring wind generators located in Inner Mongolia. To mitigate noise problems, the model randomly and purposely corrupts the input data, changing the point data to zero through a corruption rate. This procedure prevents the AE from finding an identity function. In other words, the model will not mimic the inputs, thus obtaining the ability to remove noise and fluctuations.

Ref. [67] applied an approach similar to that of the aforementioned works and compared it with the classic approach of detection based on the threshold of an indicator. The results showed that the Variational Autoencoder (VAE) models were able to predict all the failure data of the vibration signals of rotating equipment, while the classical approach was not able to adequately attend to any of the points of the tested dataset.

Ref. [68] highlighted the non-stationary, non-linear, and noisy nature of the vibration data of a wind turbine gearbox as a problem for the construction of diagnostic methodologies based on these data. To resolve this issue, the authors proposed the use of a Stacked Multilevel-Denoising Autoencoder—SMDA. The results obtained were better than those produced by other types of autoencoders on the same vibration dataset.

Ref. [69] used convolutional neural networks (CNNs) so that bearing defects could be detected using a database of acoustic signals. The ability to extract a greater amount of resources from signals was only possible with processes that involved transforming the signal into images and applying filters and mechanisms that managed to reduce noise signals. In the method used by [70], noise is eliminated through the use of a variation of autoencoders (CDAE), and Bi-LSTM is used to extract characteristics from the signals. The results were considered satisfactory for predicting the remaining useful life.

Ref. [71] compared the performance of six deep machine learning algorithms, C-AMDATS, Luminol Bitmap, SAX-REPEAT, KNN, Bootstrap, and Robust Random Cut Forest, which aimed to identify patterns of abnormalities in multivariate time-series data of a real floating production storage and offloading system. The results showed that unsupervised algorithms, such as C-AMDATS, were able to recognize and isolate abnormal events in rotating machinery with an accuracy of 99%.

In addition to the unsupervised approaches presented, there is a wide range of documents in the literature that propose the use of supervised algorithms, such as the work of [72], in which the authors proposed a CNN architecture for the classification of vibration data in an open database of images that are collected from the pre-processing of a time series. The CNN is capable of automatically extracting attributes from images. However, the use of this type of algorithm requires previously labeled data.

Ref. [73] used a model based on a BLSTM neural network together with an autoencoder, with the function of identifying faults in a speed multiplier box in a wind turbine. The performance of the suggested model was analyzed using a vibration vibration dataset.

Ref. [74] proposed a hybrid model composed of CNN and LSTM networks to detect abnormalities in the bearings of wind turbine generators based on deep learning. The results show that the CNN-LSTM model was able to detect abnormalities in the main bearing state earlier than the LSTM model. This condition was identified using mean squared error evaluation.

3.2. Approach Based on Fault Diagnosis

Different types of mechanical loads pose difficulties in extracting vibration data from bearings and therefore present problems for the implementation of models that can diagnose this type of failure. Reference [75] implemented multiscale models identified by MS-DSACNN that achieved, in a structured way and using transfer learning, excellent results when applied to different case studies.

Ref. [76] implemented an intelligent fault diagnosis method to automatically identify the operating conditions of a wind turbine gearbox. Unlike other techniques used, where feature extraction and classification are designed and performed separately, this work performed the automatic learning of fault features directly from raw signals while classifying the fault type in a single structure, establishing the learning of faults from end to end. In this study, a type of multiscale convolutional neural network architecture—MSCNN—was proposed, which performed feature extraction and classification simultaneously (Figure 12).

Ref. [77] proposed a model to diagnose faults in a direct-drive wind generator with a capacity of 3MW. The generator had the function of supplementing the energy consumption of an industrial plant located in Ireland. The study was supported by historical operating data collected at 10-minute intervals. The system provided records of anomalous conditions and operational failures, signaled by the Wind Energy Converter—WEC—and Remote Terminal Unit—RTU—columns. The data from the WEC corresponded directly to the information related to the turbine itself, while the data from the RTU corresponded to the energy controls.

Ref. [78] developed a model capable of performing fault diagnoses through monitored data, more precisely through vibration data in the time domain of a direct-drive wind generator. Four types of failures were studied, as well as the evaluation of the normal operating condition. The failures were described as mass imbalance, aerodynamic imbalance, failures in the yaw system (Yaw), and breakage of the generator blade. The authors pointed out that all vibration data were analyzed and studied for a speed of 270 rpm.

Ref. [79] implemented an Enhanced MixMatch (EMM) algorithm for diagnosing bearing failures using semi-supervised approaches and information flows, which used signals associated with acoustic and vibration emissions as input data. In this way, the model distinguished itself from other conventional fault diagnosis approaches that used only isolated signals. When compared with these types of models for diagnosing bearing faults, EMM was able to obtain better results.

Ref. [80] developed a model based on a Fusion Multiscale Convolutional Neural Network—FMSCNN—for the diagnosis of faults in bearings under conditions of speed variation using sound and vibration signals. The objective was to perform the classification of different types of failures when the speeds of the test samples were non-stationary and unknown to the trained model.

Ref. [81] proposed an intelligent fault diagnosis method for wind turbine gearboxes using a multiscale convolutional neural network—LMSCNN (Figure 13). The model was able to learn resources with the direct use of the vibration signal without any type of pre-processing, with a high capacity for versatility and excellent prospects for applications in industrial centers.

Ref. [82] proposed a multivariate analysis with the inclusion of vibration data. This was possible because they identified that a single method for the evaluation of electrical power signals for fault diagnosis in wind turbines was insufficient. For the development of the experiments, the authors used a test bench capable of gradually simulating the imbalance and misalignment of a wind turbine.

Ref. [83] performed a literature review study regarding anomaly detection and diagnosis for permanent magnet synchronous motors. In this study, failures were classified into electrical, mechanical, and demagnetization. For the case of electrical failures, the possibility of a failure in the internal stator (an open circuit) was observed, although this problem was associated with the transmission system. In this context, ref. [84] performed a study on the detection of this type of failure. The model separated failures into two categories: internal stator failure and inverter switch failure. It was, however, a mathematical model based on the physics of equipment failure, which did not address the use of machine learning algorithms.

Ref. [85] used a Kalman filter model and a time-series smoothing algorithm to construct a generator stator “health” indicator, as well as a threshold that would indicate failure. A time-series prediction model based on LSTM and RNN neural networks was used to calculate the remaining useful life (RUL). In a more recent study carried out by the authors [86], a Kalman filter model was applied to create a “health” indicator and perform electrical failure detection on a test bench with a PMSG-type generator.

Ref. [87] developed an intelligent fault diagnosis methodology based on models using a modified version of Transformer Neural Networks called T4PdM. (Figure 14).

Vibration data from the MaFaulDa and CWRU bases were used with the following results: 99.98% and 98%, respectively. The performance of the T4PdM model was compared with other published works and proved to be superior in the detection and classification of faults in industrial rotating machines.

3.3. Approach Based on Failure Prognosis and RUL Estimation

In the prognosis, the remaining useful life—RUL—of the products or systems of interest is calculated by future estimates generated using models that take into consideration the deterioration trajectory and the operational usage plans [88,89]. Research dealing with fault diagnosis makes use of data from tests run to failure, from which RUL labels can be derived. However, the easiest method to determine the RUL is by calculating time to failure [63].

From a machine learning perspective, the prognosis can be considered a regression problem because the goal value is in the actual domain. Prognosis aims to learn a function that identifies the condition of an item with its RUL estimates [90].

Ref. [91] developed a model for estimating the RUL of bearings using deep learning. An LSTM network and a CNN were used. Compared to traditional statistical analysis, features extracted from the CNN had better representation ability to withstand the degradation process and thus improved the RUL prediction performance. Compared with traditional regression methods, the LSTM network made use of the temporal information of the degradation process and obtained the lowest prediction error and better numerical stability than other methods.

Ref. [92] applied a data-driven model for the prediction of faults in permanent magnet motors using temporal voltage and current data for RUL prediction. A Kalman filter and bilateral attention-based LSTM were implemented and presented error rates substantially lower than those of the other tested models.

Ref. [83] developed a two-step approach to predict the residual service life of mechanical bearings. In the first step, the authors applied a model based on convolutional neural networks, with the objective of extracting the general characteristics of the raw vibration signal in the time domain. In the second stage, a sequential hybrid neural network was used, aggregating layers of one-dimensional convolutional neural networks with an LSTM recurrent neural network, to capture the temporal characteristics of the degradation phenomenon of the tested bearings.

Ref. [93] highlighted the need to obtain expressive volumes of data for models based on artificial neural networks that make use of recurrence and/or mathematical convolutions. If this does not happen, there may be a loss of performance, signaled by large adjustments in the model’s input data. Finally, the authors proposed the application of an artificial mechanical vibration database so that it could serve as inputs in the survey of estimates of the residual useful life of rotating equipment.

Ref. [94] developed a model capable of estimating the useful life of rotating equipment with a direct data supply. In this way, there would be no need to estimate degradation states or a failure threshold. The prediction model used was based on an SVM that, through the use of hyperplanes, created classifiers or regressors. The authors identified the need to define the most important attributes for model training, since the extraction of attributes that presented redundancy in their characteristics would imply an increase in the computational cost and in the accuracy of the estimates.

Ref. [95] estimated the residual service life of a wind turbine bearing under non-stationary operating conditions by applying a Gaussian learning model. The authors used wavelets to attenuate the intrinsic noise of the equipment operation dynamics. Vibration data were obtained from two wind turbines at the Tong Fa wind farm in Jilin Province, China.

3.4. Approach Based on Prescriptive Maintenance Applied to Wind Turbines

The operation and maintenance of wind turbines is considered a complex activity, as it requires the specific knowledge of a production method that is directly associated with the process of energy generation and distribution. This makes maintenance strategies that, globally, are identified as reactive or proactive decisive in obtaining the reliability and availability of the systems that make up this equipment.

A reactive strategy is one that conditions the equipment operation process until the occurrence of a certain failure. The next action will occur with the planning of a maintenance team that can carry out the necessary repairs, allowing the operational return of the equipment [96,97].

The choice of a corrective method as a maintenance strategy in a wind turbine can lead to failures that are considered serious, which generate prolonged unavailability and an increase in the costs involved in the repairs. The option for preventive maintenance, in turn, can generate extra costs associated with the need for periodic equipment downtime, as well as waste involved in replacing parts and components before necessary [98].

Proactive maintenance corresponds to actions that are validated through periodic and condition-based maintenance [99,100] (Figure 15).

When the measured parameters exceed the limits of the values established in the commissioning status by the manufacturers, interventions can occur through the planning of maintenance teams, or operators can make use of predictive and prescriptive processes that leverage the data collected to try to prevent a failure in advance [8,101].

Prescriptive maintenance uses technological tools, such as the use of artificial intelligence through machine learning, to prevent imminent failures. These technologies are described in the literature [102] and include classification models such as decision trees, random forest, neural networks, and dynamic Bayesian networks that predict failures and propose solution planning [103,104].

The prescriptive maintenance process can be even more relevant in the management of wind turbine maintenance if there is a combination of modeling involving remote inspections and repairs with meteorological forecast improvement processes. A decision framework for prescriptive maintenance can still follow an opportunistic approach using planned and unplanned downtime to perform specifically scheduled maintenance activities [105,106].

Ref. [107] implemented a prescriptive model for optimizing a spare parts inventory using CBM data to predict the RUL and determine the quantity of stock more economically. The presented model was validated with monitoring data from an onshore wind farm.

The prescriptive tool proposed by [108] identified an integrated framework for maintaining a wind farm that incorporated the use of real-time sensor data with a specific optimization model. This tool allowed decision-making at the entire maintenance level, describing an ideal schedule that related economic interdependencies and maintenance between turbines.

Considering failure prognoses and O&M models, the work proposed by [109] used data from the Condition Monitoring System (CMS) and the use of prescriptive tools, resulting in the ability to decide on maintenance management processes’ prevention in wind turbines, minimizing their costs. The Markov decision process was used to represent the degradation of the system state, with three options available: no intervention, preventive maintenance, and observation.

4. Results Found and Gaps in Development

It is necessary to consider some limitations of the use of data-driven models in the implementation of predictive and prescriptive maintenance of wind turbines associated with the detection, diagnosis, and prognosis of failures. The absence of failure data is important because it is a basic input for carrying out model training, and the costs of implementing monitoring, associated with specific maintenance strategies, continue to be an obstacle to the implementation of this type of control. Even so, the search for improvements in the reliability and availability of these assets is growing. It can be observed, in recent years, that there has been an increase in the number of publications and, consequently, in the scientific interest in the wind energy segment, with a focus on the O&M of its equipment.

The models resulting from articles presented in the previous sections make up the frontier threshold of knowledge for the maintenance contexts described so far. Some articles do not necessarily point to the wind turbine segment. However, these are studies that provide techniques and/or models for operating on data from rotating equipment, of which wind turbines are a part.

Some works mentioned in this article present roller bearings and power converters as key components in the analysis of the failures of this equipment. In general, the vulnerable items in mechanical equipment are bearings, associated with speed multipliers, and electronic converters, associated with failures identified in PMSG-type generators. Some AI models stood out in the volume of jobs, as well as in the results achieved, such as models based on the AE and its variations (AE, VAE, and DAE); convolutional neural networks; Densely Connected Neural Networks (MLP); SVM-based models (regressive and classification models); recurrent neural networks (LSTM); and hybrid networks (CNN + LSTM and Transformers).

In general, the articles do not clearly present the structural category of the wind turbines used in their research. On the other hand, when there is an indication of the type of architecture, geared models are the most abundant. For the direct-drive-type structural category, the studies by [77,78] stand out, where both used models based on SVM and MLP, respectively. Despite finding deep learning studies applied to permanent magnet synchronous generators, especially for the purpose of creating health indicators via the Kalman filter, a range of studies were not found that use this application in the context of wind turbines with generators of the PMSG type.

Development gaps include the application of deep learning models to the context of direct-drive turbines and permanent magnet synchronous generators. These types of equipment, as they are mostly made up of electronic components, mainly in the frequency and energy conversion system, associated with the operation of thyristors, do not present a well-defined degradation process, such as that which occurs in mechanical components identified by multiplier boxes, which are associated with noise, temperature rises, and a specific decrease in performance. Most of the time, identifying that something is not compatible with normal operation is difficult to verify, and untimely failures end up occurring. All of these analyses aim to evaluate trends in changes in behavior of the main characteristics contained in databases originating from supervisory and control, in a timely manner, so that decisions can be made or actions can be suggested to mitigate the effects of failures.

Models that show the ability to extract temporal dynamics and that provide more accurate resource extraction are a possibility for innovative experimentation. Those that can perform failure predictions more broadly in assemblies and systems and not just in isolated components, applied in environments that comprise different types of similar operations, are also characterized as future challenges. Therefore, the architectures of convolutional, recurrent, and hybrid neural networks, as well as the application of autoencoders and their derivations, can be better evaluated and explored in complex systems, such as that which occurs in the maintenance of wind turbines, especially in types in which there is a predominance of electronic components, such as direct-drive ones and PMSG-type generators.

5. Conclusions

In recent years, wind energy has achieved a leading position among renewable energies. To support the expected growth of this sector, maintenance strategies for wind turbines must provide a means to increase the reliability and availability of these assets. The usual maintenance procedures present difficulties accompanying the expansion of this energy source. The use of computational models that can, through operational data and studies of failure modes and mechanisms, implement algorithms that can detect, diagnose, and predict failures in advance in complex systems, such as those present in the operation and maintenance of wind turbines, can optimize these production processes.

In this study, a systematic review of the literature focused on research on the predictive and prescriptive maintenance of wind turbines based on the implementation of data-oriented models with the use of artificial intelligence tools was carried out. Approaches based on deep machine learning models focused on the detection, diagnosis, and prognosis of faults in sets and systems related to wind turbines. The research results demonstrate that the studied models can, within specific scenarios and conditions, help predictive and prescriptive maintenance, with the latter being the type to be explored in the future with more tools based on deep machine learning.