The world demand for renewable energy resources has recently increased significantly due to the shortage of conventional energy sources and their adverse effects on the environment and climate. The current severe shortage of energy and degradation of the environment have compelled individuals and authorities around the globe to explore promising sources of alternative energy [1]. There are many sources for generating renewable energy, and one of them is wind energy, which is the most common due to its abundance and simplicity in generating power. Wind energy is an alternative source demonstrating the benefits of being sustainable, environmentally friendly, and clean [2–4]. Because of this, wind energy has attracted significant interest as a form of plentiful, environmentally friendly renewable energy resources and has recently seen unparalleled growth in development [1]. With just 2000 offshore wind megawatts (MW) constructed at the end of 2009, wind energy has been acknowledged as one of the primary renewable energy sources. Its progress has primarily involved the development of onshore wind farms [5]. Up till recently, onshore installations have accounted for the majority of the utilization of wind energy. Compared to other renewable energies, wind energy has a higher technological sophistication and higher availability of wind resources, which account for this selective progress [5]. Hence, we can witness a vast orientation of the developed countries to harness wind energy for power generation, as it all started with one wind turbine in 1887 to produce electricity in Scotland, and now it is more than 70,800 in the U.S. only. Even though wind turbines have shown numerous expansions, they are still significantly vulnerable to physical damage [6].

Fatigue occurs when a material under stress weakens with time, especially cyclic stresses [7]. Cycling loading and fatigue can cause cracks to spread even at nominal maximum stress levels substantially lower than the material’s ultimate stress limit, which can ultimately cause the structure to fracture [7].

During the operational phase of a wind turbine, most of its components experience fluctuating mechanical stresses due to the variation in wind speed. Consequently, this process results in gradual deterioration of individual components, ultimately culminating in a failure. The initiation of this process occurs at the micro-scale as a result of irreversible alterations in the microstructure. Over time, it progresses and eventually becomes evident as a defect, ultimately causing the loss of functioning in a specific component [8]. Understanding the process of component degradation over time is of significant importance, as it allows for the estimation of the remaining usable life of the component before the loss of its functionality. To enhance the operational longevity of a wind turbine, it is vital to understand the utilized lifespan, often referred to as cumulative damage. However, it should be noted that the cumulative damage in wind turbines directly correlates with the fatigue load they endure [9].

If fatigue loads are continuously monitored while wind turbines are in operation, an up-to-date evaluation of lifetime consumption concerning design criteria may be feasible [10]. To avoid fatigue as much as possible and improve the wind energy production process, many mechanisms could be used to predict fatigue, including the ANN (artificial neural network).

ANN is a set of interconnected units or nodes, as seen in Figs. 1 and 2. More accurately, it combines many simple processing units running in parallel, each capable of learning from its environment and storing that knowledge in its connections [11]. ANN provides obvious benefits over conventional data processing techniques when handling fuzzy, random, and nonlinear data [12]. Remembering that all artificial neural network methods exhibit optimal performance when dealing with nonlinear dependencies between the inputs and outputs is crucial. ANNs can model and identify linear relationships. However, it is worth noting that in some instances, the outcomes obtained through ANNs may be inferior to those achieved using more straightforward basic statistical procedures [13].

Within a neural network, individual nodes engage in basic computations, while connections facilitate the transmission of signals between nodes. These connections are characterized by numerical values referred to as “Connection Weight”, which signify the degree to which a signal is either amplified or attenuated by the connection [14]. The artificial neuron is designed to replicate the functionality of a real neuron, specifically by receiving multiple impulses from nearby neurons and processing them according to a predetermined and straightforward method [13].

There are many types of ANN methods. In this paper, we investigated the literature for its uses in the prediction of wind turbine fatigues, and the paper aims to make it a reference for researchers in this field or any related field.

The paper is structured as follows. Section 2 introduces an overview of wind turbine fatigue. The fundamentals of artificial neural networks are presented in Section 3, which provides a comprehensive review of the essential aspects of ANNs, including their basic principles, the learning process, and their specific application in the context of wind energy systems. A previous study on wind turbine fatigue is investigated in Section 4, which comprehensively examines the current research literature. The application of ANNs in wind turbine fatigue prediction is illustrated in Section 5, which includes an extensive review of the application of ANN in fatigue prediction. Section 6 presents a previous case study where ANNs had been applied in investigating wind turbine fatigue, in addition to a simulation that we conducted based on the literature. The current limitations and challenges in using ANNs for wind turbine fatigue prediction are presented in Section 7, and valuable perspectives for future research and development in this field are provided. A summary of the main findings and the significance of ANNs in advancing the understanding of wind turbine fatigue is presented in Section 8.

Artificial neural networks (ANNs) are computational modeling methods recently developed and have gained widespread adoption in many fields for simulating challenging real-world issues [39]. ANNs could be described as thickly coupled adaptive basic processing units capable of executing enormously concurrently high-performance computations for data processing and knowledge representation [40,41]. Even though ANNs are blatant representations of their biological counterparts, the ultimate objective of ANNs is not to mimic the functioning of biological systems but rather to employ what is known about how biological networks work to address complicated issues. The extraordinary information processing traits of biological systems, such as nonlinear behavior, significant parallelism, durability, error and failure tolerance, learning, capability to cope with inaccurate and ambiguous input, and the ability for generalization, make ANNs attractive [42]. Such features make artificial models advantageous because (i) nonlinearity provides for an improved fit to the data, (ii) Noise-insensitivity allows for precise forecasting even in the presence of faulty measurements and questionable data, (iii) High parallelism ensures quick processing and hardware fault tolerance, (iv) As the environment changes, the system can adapt by updating (modifying) its internal structure, and (v) Generalization permits the model to be used with unlearned data. The primary goal of ANN-based computing, often known as “neurocomputing”, is to create mathematical formulas that allow ANNs to emulate the data processing and information acquisition processes seen in the human brain. Although ANN models possess an empirical nature, they can yield highly accurate answers for issues that are either precisely or imprecisely defined. Additionally, they are adept at handling phenomena that can only be comprehended using experimental data and observations conducted in real-world settings [43].

Artificial intelligence is distinctive in its predominant application for data analysis, notably advocated by social science and art scholars, and its utility in science and engineering [44]. ANNs have been extensively used in various domains, including business, education, economics, and everyday life. ANNs have demonstrated their applicability in multiple domains, including optimization methods [45], intrusion detection [46,47] and data analysis [48–50]. Numerous scholars have utilized machine learning (ML) methodologies to address classification challenges [51,52]. Artificial neural networks (ANNs) have demonstrated a high level of proficiency in identifying trends and patterns within datasets [53]. Consequently, they are well-suited for fulfilling forecasting and prediction requirements [53].

Within artificial systems, learning is commonly understood as the iterative process by which the system’s internal representation is modified in direct reaction to external inputs, enabling the system to execute a designated task effectively. Learning encompasses the alteration of the network architecture, encompassing modifications of link weights, the creation or removal of connecting links, and the adjustment of firing rules for individual neurons [41]. ANN learning occurs through iterative steps, wherein the network is exposed to training instances, resembling the process through which humans acquire knowledge through experiential learning. An ANN system is considered to have acquired knowledge when it can effectively process imprecise, fuzzy, turbulent, and probabilistic information without significantly compromising the quality of its responses. Additionally, a well-trained ANN system should be capable of using the knowledge gained from specific tasks to solve unfamiliar problems, thereby exhibiting the capacity to generalize its learning.

The forecasting process involves utilizing ANNs to analyze a time series dataset that represents a specific phenomenon inside a particular situation. This analysis entails training the ANN using samples from the dataset, enabling it to learn the patterns and relationships within the data. Once trained, the ANN can be applied to additional scenarios to anticipate or forecast the phenomenon’s behavior at future times [39].

One of the areas that the literature investigates is wind turbine fatigue reduction. A wind-induced stress transfer function changes by the nonlinearity of turbine reactions to various mean wind speeds, and numerous simulations handling thousands of environmental factors are needed to obtain the stress transfer function from time simulations in fatigue analysis [104]. Additionally, this requires a lot of time to complete, which could take a couple of days to complete the process. To minimize the required number of simulations, Kim et al. [104] used ANN models to account for the impact of mean wind speed on the fatigue analysis results. Whereas the stress transfer function, the stress spectrum, and the fatigue damage are potential outputs of the ANN models, the mean wind speed and frequency are supposed to represent the inputs. Significantly, using ANN models indicates that computational efficiency can be increased while retaining a maximum error range of 35% and reducing the number of simulations needed from 41 to 7 [104]. A thorough overview of ANN model development highlights the significance of sampling techniques in improving model performance. It suggests including crucial wind speeds, such as cut-in, cut-out, and rated wind speeds [104]. Furthermore, the superposition model concept, which focuses on correction factors close to stress spectrum peaks, can improve predicted outcomes without adding more sample data or neurons to the hidden layer. Although it is observed that the effect of lowly fluctuating resonant motion on fatigue damage in floating wind turbine systems is negligible, its inclusion in ANN models research is recommended, suggesting directions for further investigation [104].

Yin et al. [105] studied a big data-driven multi-objective prediction system for predicting wind farm power production and structural fatigue. The average power production of the wind farm and the equivalent thrust of turbines were used as the response variables in the prediction system. In contrast, the wind conditions, control settings, and turbine characteristics were used as the predictor factors. Five distinct data mining methods, including the GRNN (General Regression Neural Network), the RF (Random Forest), the SVM (Support Vector Machine), and the GBR (Gradient Boosting), were then used to build the prediction models. A wind farm simulation platform, FLORIS (FLOw Redirection and Induction in Steady State) created by the National Renewable Energy Laboratory (NREL) and the Delft University of Technology, has been used to evaluate the aforementioned data-driven machine learning algorithms. Based on the most recent version of FLORIS, the five methods’ prediction performances were compared and assessed. According to their results, all of these approaches can achieve a relative accuracy of 99% or higher, which is sufficient for practical applications [105]. The GRNN and SVM show the highest levels of accuracy, with the GRNN having the highest accuracy and performance when making predictions about thrust. The GRNN is the most efficient in terms of computational performance [105].

Luna et al. [106] focused on predicting wind turbine tower fatigue, and they suggest using the NARX due its capability to predict nonlinear time series as it uses historical data from expected or actual time series. He et al. [107] proposed a data-driven technique for predicting fatigue load and power to evaluate the structural performance of wind turbines precisely and effectively under yaw control. The power yield and fatigue loads at crucial turbine elements are computed under realistic yaw angles while considering wake impacts. Then, input-output mappings are created using a support vector regression (SVR) model that has been trained. Based on comparison evaluation and effectiveness assessment, they discovered that the suggested prediction technique can return very high regression coefficients with a minor deviation, confirming its preciseness and resilience [107].

Yilmaz et al. [108] used the MLP (Multilayer Perceptron) and RBF (Radial Basis Function) in their study to control the wind turbine pitch angle and avoid any possible fatigue that could occur in the wind turbine blades. Their study focused on controlling the blade pitch angle at wind speeds above 14 m/s. One hidden layer has been used in both approaches (MLP & RBF). Two input neurons, five hidden neurons, and one output neuron used in MLP. Three input neurons, ten hidden neurons, and one output neuron used in RBF. Per their results, if there is no control for the pitch angle, the generator won’t be able to generate power because there will be no overspeed prevention [108]. Also, compared to the common PID (Proportional Integral Derivative) controller, the MLP and RBF showed better results in controlling and generating power [108].

Ziane et al. [109] used ANN to predict the fatigue life of the wind turbine blades, as shown in Fig. 3, by proposing the following hybrid neural network models: BPNN (Backpropagation Neural Network), PSO-ANN (Particle Swarm Optimization Artificial Neural Network), and CSNN (Cuckoo Search Neural Network). Based on the results, it was observed that the CSNN outperforms the BPNN and PSO-ANN in terms of prediction due to its ease of use when applying mathematical models, and due to its capability to change weights without determining a gradient [109]. Also, CSNN showed better fatigue life predictions, faster computation, higher precision, and most importantly a high convergence rate toward the actual global minimum [109]. The advantages of CSNN that make it preferable to use over BPNN and PSO-ANN are that it explores the search space more effectively, potentially leading to better convergence and avoiding local optima [109,110]. CSNN is robust to noisy data and can handle complex, nonlinear relationships between input features and fatigue life. CSNN efficiently adjusts the synaptic weights of the neural network, improving prediction accuracy while minimizing computational resources [109,110].

The limitations of CSNN are as follows: 1-It requires tuning of hyperparameters for optimal performance, because poorly chosen parameters may lead to suboptimal results [109,110]. 2-CSNN performance can be sensitive to the initial population of solutions, as different initializations may yield varying results. 3-CSNN may not perform as well as other algorithms regarding local exploitation around promising regions [109,110].

Li et al. [113] used the Back Propagation (BP) neural network and Radial Basis Function (RBF) neural network for icing conditions prediction of the wind turbine blades. For predicting the blade icing mass, the neural networks trained by the usage of data sets showing a positive correlation between blade icing mass and natural. Per their findings, the BP neural network is more accurate than RBF in detecting the blade icing mass [113].

Abdallah et al. [114] created a machine learning (ML) method to statistically combine the fatigue load on the wind turbine blade from numerous concurrent simulators, considering the varying fidelity, complexity, and underlying assumptions to create a “most likely” fatigue load and thereby decrease the model-form uncertainty. To derive a map of probability, they developed clusters of simulator outputs as a function of wind speed, and they utilized bootstrap aggregations to obtain an aggregated estimate of the damage equivalent fatigue of the blade root bending moment [114].

For early damage identification utilizing vibration analysis measures Dervilis et al. [115] employed ML (Machine Learning) nonlinear neural network techniques such as auto-associative neural network (AANN) and radial basis function (RBF). The primary purpose was to improve the structural health monitoring of the wind turbine blade and assist with preventing catastrophic failure.

According to Malik et al. [116] the primary causes of a blade imbalance, as shown in Table 2, include mistakes made during manufacturing or construction, icing conditions, deterioration brought on by aging, or wear and fatigue in the wind turbine. Equipment shifts and wears to varied degrees over time due to imbalance on the blades and rotating shaft. Various factors, including control system errors and extreme wind shear, contribute to aerodynamic asymmetry. One blade pitch angle differs slightly from the other two due to a control system malfunction, which will cause aerodynamic asymmetry in the wind turbine. Malik et al. [116] used PNN (probabilistic neural network) to analyze the wind turbine defects caused by imbalance conditions. The simulated results demonstrate that the PNN has a greater diagnosis accuracy than other ANN approaches and requires less training and testing time, and it performs better than conventional approaches [116].

Due to their capability to transmit a relatively high load from blades while offering a high gear ratio (necessary for the generator to produce electrical energy efficiently), considerable power planetary gearboxes became widespread components of wind turbine drivetrains [117]. In this category of rotating machinery, gearboxes are at the top of the drivetrain components that are most prone to damage [117]. A single malfunction typically causes total gear damage, necessitating the replacement of the entire gearbox because a planetary gearbox distributes the load amongst multiple meshes while in operation [117]. Vibration analysis is one of the methods for evaluating the extent of damage to planetary gearboxes. Still, it is more difficult for such planetary gearboxes because wind turbines inherently operate in nonstationary circumstances with varying speeds and loads [117]. Strączkiewicz et al. [117] used methodology that considers the fluctuation of vibration signal characteristics under the effects of rapidly changing speed and load. The ANN algorithms trained at the machine typical operation. The provided value of a feature is compared to the value estimated using the taught net for each following unclassified observation. A linear regression is used, and the b parameter, also known as the error term, is observed to reduce the impact of changing conditions [117].

Chen et al. [118] used the CNN (Convolutional Neural Network) for fatigue prediction in the wind turbine gearbox, and the CNN was trained by using experimental vibration data that was processed by using standard deviation, skewness, and kurtosis in the time domain, and FFT (Fast Fourier Transform) and RMS (Root Mean Squared) in the frequency domain. Based on the results, they found that for over 20 test cases, the CNN model classification rate is better than that of SVM as it is more accurate. The CNN model had an accuracy range of 91.4% to 98.9%, whereas the SVM (Support Vector Machine) model ranged from 55.9% to 70.2% [118]. In contrast, Janssens et al. [119] used a CNN model to classify faults, as the CNN model is applied to raw frequency spectrum vibration data from bearings experiment rigs. The results showed that CNN has a higher classification accuracy than the conventional feature extraction methods, with an accuracy of 94% [119]. Based on that, we can see that Convolutional Neural Networks (CNNs) have proven to be highly efficient in predicting fatigue in spinning machinery, such as wind turbines [120–122]. The imported original vibration signals can be automatically and effectively analyzed to capture fatigue features and determine their condition using several CNN models [122]. This can be done by acquiring features directly from the one-dimensional raw vibration signals without manual feature extraction [123]. Using CNNs in this domain has yielded exceptional accuracy in recognizing and classifying fatigues [120,124].

Simani et al. [125] used NARX for fault diagnosis. NARX before comparing it with other methods like the GSKV (Gaussian kernel support vector machine), UDC (up-down counters), and GFM (general failure model), it was trained using data from routine and defective wind turbine simulators. When evaluated on nine failures, neural network estimates outperformed other methods, with just a 1.12-s delay between the incidence of the failure and its detection [125]. Additionally, it was shown that this method had significantly lower false alarm and missed fault rates than other methods. Based on that, we can tell that NARX is a type of recurrent neural network that has proven effective in predicting future values in time-series data. Within the domain of wind turbine monitoring, these models can predict future conditions of the turbine by analyzing its previous performance. This capability can detect and prevent fatigue early [125].

A wind turbine gearbox test rig experimental vibration data was gathered under fault-induced conditions by Jiang et al. [126]. To diagnose the faults in these data, the MSCNN (Multi-Scale Convolutional Neural Network) approach was used, and the results of this analysis were compared to those of conventional CNN and other techniques. Based on the results, the proposed MSCNN outperformed the different approaches, averaging 98.53%; the most significant average was 84.31% in fault classification and feature learning [126].