Artificial Intelligence in Wind Turbine Fault Detection and Diagnosis: Advances and Perspectives

Abstract

The global shift toward renewable energy, particularly wind power, underscores the critical need for advanced fault diagnosis systems to optimize wind turbine reliability and efficiency. While traditional diagnostic methods remain foundational, their limitations in addressing the nonlinear dynamics and operational complexity of modern turbines have accelerated the adoption of Artificial Intelligence (AI)-driven approaches. This review systematically examines advancements in AI-based fault diagnosis techniques, including machine learning (ML) and deep learning (DL), from 2019 to 2024, analyzing their evolution, efficacy, and practical challenges. Drawing on a curated selection of 55 studies (identified via structured searches across IEEE Xplore, ScienceDirect, and Web of Science), the paper prioritizes research employing data-driven or model-based methodologies with explicit experimental validation and clearly documented data sources. The excluded works lacked English accessibility, validation, or data transparency. Focusing on high-impact faults in gearboxes, blades, and generators, these components are responsible for over 70% of turbine failures, the review maps prevalent ML and DL algorithms, such as CNNs, LSTMs, and SVMs, to specific fault types, revealing hybrid AI models and real-world data integration as key drivers of diagnostic accuracy. Critical gaps are identified, including overreliance on simulated datasets and inconsistent signal preprocessing, which hinder real-world applicability. This study concludes with actionable recommendations for future research, advocating adaptive noise-filtering techniques, scalable hybrid architectures, and standardized benchmarking using operational turbine data. By bridging theoretical AI advancements with practical deployment challenges, this work aims to inform next-generation fault diagnosis systems, enhancing turbine longevity and supporting global renewable energy goals.

1. Introduction

Renewable energy has emerged as a linchpin in addressing the dual challenges of climate change and sustainable development, playing a vital role in the global energy transition. According to the International Energy Agency (IEA), in 2023, renewable energy sources collectively accounted for 29.6% of global electricity generation, marking a significant increase from previous years [1]. This is an indication that there will be a growth in renewable energy and, as stated in the work of F. P. Garcia [2], the power generation from renewable sources will increase from 25% (2017) to 85% by 2050 and this increase will mostly be generated from wind and solar power sources [1,3]. By 2030, total greenhouse gas emissions are expected to reach 40 gigatons [4], proving that the world is facing major issues with reducing greenhouse gas emissions; hence the importance of renewable energy cannot be overstated and there is a need to carry out more research to enhance the operation of renewable energy systems. Taking European countries as an example, in 2017, there was a noticeable decrease in greenhouse emissions and the reduction even went lower than the 2020 target. This was a result of the increase in renewable power generation [1]. This transition not only mitigates the environmental impact of traditional fossil fuels but also addresses the increasing demand for energy in a sustainable manner.

After hydro power, the second fastest growing renewable energy sector is wind energy due to its simple infrastructure, cost-effectiveness, and advanced technology [5,6]. A wind farm is simply a collection of wind turbines set up either offshore or onshore [7] and is connected to the power transmission grid. Although most of the farms currently exist onshore, it is expected that there will be a growth in the construction of offshore windfarms [8,9]. IEA has identified a noticeable growth in the power being generated from offshore wind farms, specifically from the global leaders in offshore wind, such as China, Germany, and the UK [10]. In 2016, most countries had an average power generation of around 0.8 gigawatts, but in 2022, the average power generation from offshore windfarms went up to 2.2 gigawatts [11]. According to the study by C. Jung [12], there are 162 offshore wind farms around the world, and 26 are under construction. These offshore wind farms have proven to generate more power since there are higher wind speeds offshore than on land, and have less of an impact on people and landscapes [13]. Offshore wind farms are typically located in the oceans or the sea where the wind turbines will be exposed to more wind and hence generate more electricity [14].

For these reasons, many in the industry agree that it is better to prevent excessive maintenance and downtime caused by mechanical issues in order to reduce operating costs and improve financial gains [15,16]. It is important to explore what research has been conducted, specifically looking at fault diagnosis and detection. By understanding the methods used and what has already been achieved, we can encourage more research in this area. The process of detecting and classifying faults in rotating machinery can either be carried out using data-driven or model-based methods. Model-based approaches are normally based on an appropriate mathematical or physical model to detect and classify faults [17]. However, the primary difficulties with model-based methods stem from a lack of reliability due to the models not adequately replicating the true system’s complexity. Data-driven approaches use past datasets to predict the future condition of these rotating machines [18].

This paper utilizes a narrative literature review approach to systematically gather and synthesize pertinent research on fault diagnosis techniques in wind turbines. The methodology involved conducting a comprehensive search across academic databases, including IEEE Xplore, ScienceDirect, and Web of Science. The search strategy employed a combination of keywords and Boolean operators, specifically (“Fault Diagnosis” OR “Fault Detection”) AND (“wind turbines” OR “wind turbine”) AND (“artificial intelligence” OR “machine learning”) AND (“data-driven” OR “model-based”).

Focusing on papers published between 2019 and 2024, the resulting articles, based on inclusion and exclusion criteria, form this review paper. Moreover, to provide a comprehensive historical perspective, selected papers from before 2019 are included, highlighting their significance in the evolution of wind turbine condition monitoring.

• Inclusions

  -

  The searched terms mentioned above should be in the title or abstract.

  -

  The article must be published.

  -

  The paper uses either a data-driven approach or model-based approach for fault diagnosis.

• Exclusions

  -

  Articles that do not mention the source of data being used.

  -

  Articles that are not written in English.

  -

  Articles lacking experimental validation.

From 2019 to 2024, there has been a growing interest in research focused on wind turbine fault diagnosis and the role of ML in this field. As shown in Figure 1, the number of published research articles in this area has increased each year, highlighting the importance of conducting this review. This review has the following aims:

• Provide a comprehensive assessment of fault diagnosis techniques for wind turbines, focusing on both traditional and AI-driven approaches.
• Examine the role of machine learning (ML) and artificial intelligence (AI) in diagnosing faults across key wind turbine components, including bearings, blades, gearboxes, and generators.
• Identify and compare the strengths and limitations of data-driven and model-based fault diagnosis methods.
• Analyze recent advancements from 2019 to 2024, highlighting emerging trends and challenges in the field.
• Explore the potential of hybrid AI models and Digital Twin technology in improving real-world fault diagnosis, an area with limited existing research.

In pursuit of these objectives, this paper makes the following key contributions:

• It conducts a systematic review of fault diagnosis methodologies, integrating both model-based and data-driven approaches.
• It provides a comparative analysis of AI techniques used in wind turbine fault diagnosis.
• It identifies gaps and challenges in existing research, particularly in real-world validation and data availability.
• It discusses the potential role of Digital Twin technology for real-time fault monitoring, emphasizing its underexplored potential in wind turbine fault diagnosis.
• It presents recommendations for future advancements, including strategies for enhancing real-world implementation and the integration of hybrid AI models.

Figure 1. Bar chart showing the number of papers published over the last five years.

Figure 1. Bar chart showing the number of papers published over the last five years.

This paper is structured in the following sections, with Section 2 providing an in-depth overview of traditional and AI-driven fault diagnosis methods, including a comparison of their strengths and limitations. Section 3 discusses the various faults in wind turbines, focusing on bearing, blade, gearbox, and generator faults, and highlights AI techniques applied to address these faults. Section 4 presents a detailed analysis of the reviewed methodologies and provides recommendations for future research and advancements in the field. Finally, Section 5 concludes this paper.

2. Overview of Fault Diagnosis

Fault diagnosis is a critical process in maintaining the operational integrity of wind turbines. It involves identifying and classifying faults within the system after anomalies have been detected, allowing for timely corrective actions [19]. Unlike condition monitoring, which continuously assesses the health of components to detect potential issues before they escalate [20], fault diagnosis specifically addresses the identification of faults that have already manifested.

Advancements in Artificial Intelligence (AI) have significantly improved the accuracy and efficiency of fault diagnosis, enabling the detection of subtle and complex faults that traditional methods might miss. This paper systematically reviews fault diagnosis methods for wind turbines, focusing on advanced AI techniques (CNN and RNN) and their comparative performance in identifying faults across key turbine components such as gearbox, generator, blades, and bearings as they are the most addressed in the literature as shown in Figure 2, with a particular focus on how AI-driven techniques are enhancing the field. By understanding the strengths and limitations of these methods, we can better appreciate the role of modern technologies in fault diagnosis.

Fault diagnosis in wind turbines can be broadly divided into traditional methods and AI-driven methods [21]. While both approaches may use similar data sources, such as vibration signals or acoustic emissions, the difference lies in the analysis techniques employed. Traditional methods typically involve rule-based or statistical analyses, relying heavily on expert interpretation [22]. In contrast, AI-driven methods leverage advanced algorithms to automatically detect patterns and diagnose faults with greater accuracy and efficiency.

2.1. Traditional Fault Diagnosis Methods

Before the advent of AI, fault diagnosis relied heavily on traditional methods such as vibration analysis and Supervisory Control and Data Acquisition (SCADA) analysis. SCADA systems have been widely used to collect data such as wind speed, rotor speed, component temperature, and output power [23,24]. Despite the fact that the traditional method is used, whether it is vibration analysis or SCADA analysis, what is crucial in traditional fault diagnosis is the signal processing technique. In AI-driven approaches, a signal can be passed directly into the neural network and it will be able to extract and learn features continuously; that is not the case in traditional fault diagnosis. Some signal processing techniques that are used include statistical analysis, time-domain analysis, frequency-domain analysis, wavelet transformation, etc.

Vibration analysis is commonly used in fault diagnosis for rotating machinery, including wind turbines. It involves monitoring changes in vibration patterns to detect faults such as rotor imbalance and misalignment. The studies presented by [25,26,27,28,29,30,31] all used vibration analysis for fault diagnosis. While effective, vibration analysis can struggle as vibration signals are known to be nonlinear and nonstationary signals [32], so it becomes really difficult to extract deeper features and that is where an AI-driven approach can be useful.

SCADA analysis is also used widely and it involves monitoring key operational parameters like wind speed, rotor speed, and generator torque. Models and trends can be built from SCADA data to diagnose faults, and it has been implemented in [16,33,34,35,36,37]. However, this method has the tendency to lead to false diagnosis if the models built are not accurate. In order to build accurate models, they will be computationally intensive and not easy to implement [33].

Due to the low sampling rate of SCADA, it leads to inaccurate fault diagnosis when conventional condition monitoring is used [16,33]. Aside from that, models generated from SCADA data are normally poor due the noisy data from SCADA [38,39].

2.2. AI-Driven Fault Diagnosis

AI has significantly transformed fault diagnosis in wind turbines, allowing for more precise identification and analysis of faults after they have been detected. AI-driven approaches, such as Artificial Neural Networks (ANNs), Convolutional Neural Networks (CNNs), and Long Short-Term Memory (LSTM) networks, excel at processing large volumes of complex, nonlinear data that traditional diagnostic methods struggle to handle. ANNs, for instance, can learn from historical fault data to identify intricate patterns and relationships that might not be evident through conventional techniques [40]. CNNs are particularly adept at diagnosing faults through image-based data, such as detecting surface damage on turbine blades. Additionally, they can be effectively utilized to analyze various signal decompositions, including time–frequency representations, wavelet packet decompositions (WPD) [41], or intrinsic mode functions (IMFs) [42]. Recent advancements in deep learning-based fault diagnosis have demonstrated the effectiveness of 3D CNNs in capturing spatial relationships between sensor signals, leading to improved fault classification accuracy, particularly in asynchronous motors [43]. This approach can be extended to wind turbine fault diagnosis, where multi-dimensional data representations may enhance the detection of both mechanical and electrical faults.

Furthermore, these inputs allow the model to capture and learn from rich frequency and time-domain features derived from the original signal, enhancing its ability to detect subtle faults in wind turbine components. While CNNs excel in spatial feature extraction, Recurrent Neural Networks (RNNs), including variations like LSTM and GRU, have been optimized for wind turbine condition prognosis, offering improved predictive accuracy for long-term health monitoring [44]. LSTM networks are effective in diagnosing faults that evolve over time, making them ideal for monitoring the gradual degradation of turbine components [45].

The primary advantage of these AI-driven methods lies in their ability to improve diagnostic accuracy by continuously learning from new data, thus reducing the need for manual intervention and expert analysis [46]. This leads to more efficient and reliable fault diagnosis, which is critical for maintaining the operational efficiency and longevity of wind turbines. By integrating data from various sources—such as vibration signals, acoustic emissions, and operational metrics—AI models provide a comprehensive understanding of the fault’s nature and origin. All the methodologies used in [45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63] utilize AI-driven approaches to diagnose different parts of wind turbines. This paper mainly focuses on the AI-driven approaches and the paper mentioned above will be explored in detail in the next section.

3. Wind Turbine Faults

Wind turbines consist of several components that work together to generate electricity, and a fault in any component can ultimately affect the overall operation of the turbine. As illustrated in Figure 3, the key components include the gearbox, generator, blades, and several others [64]. This review specifically focuses on faults related to the gearbox, generator, blades, and bearings as they are the components most addressed in the literature. Figure 3 shows the percentage of papers that address each fault in this review paper. Each piece of literature typically covers a specific fault, but some studies address multiple faults.

Fault diagnosis in wind turbines employs various methods, combining traditional approaches with advanced mathematical models and smart technology models, as shown in Figure 4. This section delves into the utilization of mathematical models, including Fuzzy Logic [65], Gray Theory, and Fuzzy Petri Net (FPN) [66], as well as smart technology models such as ANNs, SVMs, and DL.

3.1. Bearing Faults

K-Nearest Neighbors (KNNs) [67] and Support Vector Machines (SVMs) [68] are two popular supervised ML algorithms that can be trained on labeled historical sensor data from wind turbines to detect and diagnose faults and failures accurately. Both techniques allow early detection of anomalies in turbine vibration, temperature, loads, etc., before secondary damage or costly downtime occurs. The trained KNN or SVM model basically serves as an automated fault diagnostic system for turbine operators by continually processing incoming live sensor data and alerting operators to any developing faults or need for preventative maintenance.

The domain of wind turbine KNNs and SVMs were compared by a study conducted by Vives J. [61]. A fault detection system was proposed to detect faults in the bearing brake failure by analyzing vibration signals that are obtained from accelerometers using ML. The methodology involves training the algorithm with feedback and using KNNs and SVMs for classification. Both SVMs and KNNs exhibit high processing speeds, robustness, and high accuracies of 95% and 94%, respectively, and they can replace the frequency spectral analysis, which is a traditional method.

ANNs mimic the human brain’s neural structure, enabling them to learn complex patterns from data. In wind turbine fault diagnosis, ANNs have demonstrated success in recognizing fault patterns from sensor data. Studies have applied ANNs to detect faults in various components, including gearbox and bearings, improving diagnostic accuracy.

Yang et al. [59] propose an automatic bearing fault classification system based on DL. The methodology involves extracting vibration signal features in the frequency domain and then configuring a dataset where each sample is initially assigned a random label. This dataset is used to train a Deep Neural Network (DNN) [69] for initial classification. After testing with sub signals, the sample labels are adjusted based on the evaluation results, and the modified dataset is used to train the DNN again. Through iterative training and testing, the network learns to cluster signals with similar fault characteristics into distinct classes.

This methodology was tested using the Case Western Reserve University (CWRU) bearing data [70] which is a standard dataset used for fault detection algorithms. In the study, the 12 k drive end, 48 k drive end, and 12 k fan end bearing data were classified into 7, 6, and 4 groups, respectively, demonstrating the method’s effectiveness in the automatic classification of vibration signals with different faults.

Zhang et al. [52] detected faults in the main shaft rear bearing using ANNs and SCADA data analysis by establishing a normal behavior model using ANNs. The ANN architecture that was finalized used a 5-10-1 structure with 1000 training cycles. The model calculates the theoretical value of related parameters and compares them to the real measurement of the same parameters, allowing for the early identification of component failure. Testing on a healthy turbine yielded an average difference between an actual and estimated rear bearing temperature of 0.026 °C and a root mean square error of 0.5 °C, which were within acceptable diagnostic limits. The percentage error was found to be less than 1.5%. The findings suggest that this approach can provide early warnings of developing faults of up to 10 days before failure for necessary maintenance actions, enhancing reliability and reducing operational costs.

3.2. Blade Faults

Fuzzy logic, introduced by L. A. Zadeh [71], is a mathematical framework that extends classical logic to handle uncertainty and imprecision in decision-making. Unlike traditional binary logic, fuzzy sets in fuzzy logic allow elements to have degrees of membership, representing the extent to which an element belongs to a set. This introduces the concept of “fuzziness”, enabling the modeling of real-world scenarios where conditions are not precisely defined. While powerful, fuzzy logic has some limitations, such as the need for extensive parameter tuning and potential subjectivity in rule interpretation. In wind turbine fault diagnosis, by representing linguistic terms and uncertainties in a rule-based system, fuzzy logic systems can effectively capture complex relationships within turbine data [72]. Fuzzy logic has been applied to fault detection in wind turbine components, considering variables like wind speed, temperature, and vibration patterns [31].

Qu et al. [47] proposed a non-singleton fuzzy logic and expanded linguistic rules. Supervisory Control and Data Acquisition (SCADA) data prediction errors are transformed into non-singleton fuzzy inputs to enable earlier detection of anomalies. This more sensitive fuzzy inference system detects four real wind turbine faults 3–5 days earlier than conventional techniques, including gearbox, cooling system, sensor, and mechanical wear faults.

Santos et al. [62] proposed an approach to diagnosing faults in wind turbines, focusing on detecting misalignment and imbalance in the transmission chain. It introduces a multi-sensory system that incorporates accelerometers for vibration evaluation, along with electrical, torque, and speed measurements. The vibration signals are processed using angular resampling techniques to extract relevant features for fault diagnosis. For classification, they tested SVMs with traditional and novel kernels, as well as Artificial Neural Networks (ANNs). Among the tested classifiers, linear SVMs showed the best performance, achieving an accuracy of 98.26%, outperforming the perceptron SVM (96.86%), Gaussian SVM (97.25%), stump SVM (97.85%), and ANNs (97.47%) in terms of accuracy, training, and tuning times.

A Multichannel Convolutional NN (MCNN) is an NN architecture designed to process multiple channels of input data simultaneously, allowing for the integration of diverse modalities or features. This parallel processing enhances the model’s ability to learn complex patterns [73].

A recent study conducted by Wang et al. [58] used MCNN to detect faults in wind turbine blades, mainly blade angle, surface damage, and breakage through the processing of vibration signals. The methodology used in the study involves the development of models for normal and faulty wind turbine blade states. These models are based on the normal state and three common faulty states: blade angle anomaly, blade surface damage, and blade breakage. The vibration signals were captured using a triaxial vibration transducer and then processed through a human–machine interface system for wind turbine vibration analysis and detection in LabVIEW [74]. The MCNN algorithm was then proposed as an improvement over traditional CNNs, as it could simultaneously process vibration signals from the X-axis, Y-axis, and Z-axis for learning and identification. The paper reports that the MCNN achieved an accuracy of up to 87.8% in identifying the different blade states based on the vibration signals.

A digital twin is a virtual model or replica of a physical object, system, or process that mirrors its real-world counterpart. It is created using data from sensors and other sources that continuously update the digital twin in real-time. This allows the digital twin to simulate, monitor, and analyze the real object’s performance, behavior, and environment [75]. What distinguishes a digital twin from traditional methods of data collection (such as manual or experimental approaches) is that data can be fed directly into AI models in real time, allowing for more dynamic and immediate insights.

The concept of digital twin was used by Guo et al. [76] and it was combined with one-dimensional CNNs in order to carry out diagnosis for the wind turbine blade. Digital twin technology was used to create a virtual model of wind turbine blades, simulating their behavior under various conditions such as normal operation, structural damage, and icing. Vibration data collected through sensors are processed and fed into a CNN, which detects and classifies faults by analyzing patterns in the data. The model achieved a 95.4% accuracy in simulating blade behavior, closely matching real-world data. By analyzing vibration trends, the system predicts future failures, demonstrating its potential for real-time monitoring and proactive fault detection in wind turbines.

An image recognition approach was proposed by Yang et al. [56] in order to detect faults in wind turbine blades. A DL model was proposed for the automatic detection of wind turbine blade damage, focusing on common defects such as cracking, spalling, and sand holes. The model integrates transfer learning and a random forest-based ensemble learning classifier within a CNN. By utilizing the Otsu Threshold Segmentation method [77] to eliminate background interference, the model automatically extracts highly abstract features of the blade images, enabling the efficient and accurate detection of defects. The training and test datasets were obtained from a wind farm in Western China, containing defect-free and defective blade images. This study compares the proposed model with other methods, including SVM, Alexnet [78], Alexnet-tl [79], and Alexnet-rf, and evaluates their performance using accuracy, sensitivity, and specificity metrics. The results show that the proposed solution (Alexnet-tl-rf) has obtained an accuracy of 97% while other approaches achieved an accuracy between 92 and 96%.

Building upon the principles and capabilities of LSTM discussed earlier, a recent study by Dehghanabandaki et al. [55] integrated a framework for detecting incipient and subtle faults in wind turbine systems using Long Short-Term Memory (LSTM) and Mixture Density Network (MDN) [80] for fault diagnosis, along with real-time system identification based on Auto-Regressive Integrated Moving Average with exogenous input (ARIMAX) [81] to reveal changes in system dynamics. This study aims to reduce maintenance costs and increase system reliability by detecting faults of pitch angles. This paper compares the proposed approach with several benchmark approaches commonly applied in fault detection, including Dynamic Principal Component Analysis (DPCA) [82], LSTM, Stacked LSTM (SLSTM) [83], LSTM autoencoder (LSTMAE) [84], and Gated Recurrent Unit (GRU). The results show that the proposed approach outperforms the benchmark approaches in terms of F-score, a measure of a test’s accuracy that considers both the precision and recall of the test, with an F-score of 0.98 compared to the highest F-score of 0.94 achieved by the LSTM model.

Chen et al. [53] proposed an approach that focuses on detecting blade imbalance faults in wind turbines caused by ice accretion using a DL approach with an LSTM and NN model which is an advanced version of RNN. The dataset used was simulation data obtained from a simulation software called G. H bladed [85]. The study achieved an accuracy of over 98% in detecting the imbalance fault, which proves the effectiveness of the proposed approach.

Another study by Simani et al. [49] proposed a fault diagnosis technique using both fuzzy systems and neural networks to detect and isolate sensor, actuator, and system faults like pitch faults, torque converter faults, and drivetrain wear in an offshore wind turbine. The methods are tested on a high-fidelity simulator and hardware-in-the-loop rig (real-time hardware running) under various normal and faulty conditions. Both techniques show robust fault detection capabilities, with the fuzzy approach performing slightly better in hardware tests. The True Fault Detection Rate (TFR) is used as a measure of accuracy [86]. For the fuzzy system, a TFR of 0.995 was obtained whereas for the neural network, the TFR varied more significantly ranging from 0.798 to 0.989. This shows that while neural networks generally provided high accuracy in fault detection, their level of accuracy was slightly less consistent compared to the fuzzy systems. Aside from the True Fault Detection Rate (TFR), the paper also examines the Mean Fault Detection Delay (MFD), which for the fuzzy system ranged from 0.06 s to 0.76 s. In contrast, for the neural network, the MFD was as short as 0.014 s, but for complex fault scenarios, it extended up to 6.987 s.

3.3. Gearbox Faults

Another proposed method by Cheng et al. [63] utilizes rotor current signals and a deep classifier to identify characteristic frequencies associated with gearbox faults. The methodology involves obtaining the envelope of rotor current signals using Hilbert transform, which contains nonstationary frequencies proportional to the varying shaft rotating speed. An angular resampling algorithm is then applied to convert the nonstationary fault-related components in the envelope signal into constant frequency components. Subsequently, Power Spectrum Density (PSD) [87] analysis is performed on the resampled envelope signal to identify excitations at the characteristic frequencies of gearbox faults. SAE-based multiclass SVM classifier is employed for fault classification on the extracted fault features from the rotor current signals. The paper compares the accuracy of three classifiers for diagnosing gearbox faults in wind turbines. The proposed SAE-based multiclass SVM classifier achieved an overall accuracy of 89.3%, outperforming the traditional SVM and DBN-DNN deep classifiers, which had accuracies of 80% and 85.3%, respectively.

A Dilated Convolutional NN (DCNN) is a type of advanced computer algorithm that can analyze data from wind turbines to detect potential mechanical issues. It uses a special technique called dilated convolutions to efficiently capture important information from the data [88].

In He et al. [57], a method is proposed for detecting faults in wind turbine gearboxes using DCNN. The methodology involves preprocessing vibration signals and segmenting them into two-dimensional data, which is then used to train the DCNN model. The proposed method is compared with traditional ML methods such as SVM and Multilayer Perceptron (MLP) [89], as well as with traditional CNN. The dataset used in the study was obtained from an experimental wind turbine gearbox that experienced two Loss-of-Oil events resulting in internal damage. Vibration signals were collected from the gearbox using a sensor with a sampling frequency of 40 kHz. The results show that the DCNN-based method achieves the highest accuracy (100%) compared to MLP (99.15%), SVM (71.8%), and CNN (100%). In terms of computational cost, the paper reports that the DCNN-based method requires less computation than CNN, with a training time of 196 s compared to 207 s for CNN.

Huang et al. [90] propose an improved SVM model for detecting faults in wind turbine gearboxes. The study focuses on identifying potential issues such as bearing faults, fatigue spalling, cracks, and indentation on the surface of the bearing. The vibration signals caused by these faults are analyzed to extract the typical fault frequency of wind turbine gearboxes. The proposed approach involves optimizing the SVM model using an improved fruit-fly intelligent algorithm with decreasing steps, leading to enhanced accuracy in gearbox fault diagnosis. The study found that the proposed diminishing step fruit-fly algorithm optimized SVM (DSFOA-SVM) method for gearbox fault diagnosis demonstrated improved accuracy of 93% compared to 85% for traditional SVM and 90% for the Particle Swarm Optimization SVM (PSO-SVM) model [91].

Wang et al. [92] addressed the challenges in diagnosing faults in wind turbines planetary gears in the gearbox using digital twin technology. The key issues with traditional methods like data-driven approaches include poor data transmission timeliness, weak visualization, and delayed fault feedback [93]. The authors propose a novel method utilizing a digital twin model combined with empirical mode decomposition (EMD) [94] and atom search optimization-support vector machine (ASO-SVM) for real-time monitoring and fault diagnosis of planetary gears. The digital twin model was implemented using the Unity3D platform, allowing for real-time interaction between virtual and real data. The proposed system improves accuracy and timeliness in diagnosing planetary gear faults and enables predictive maintenance. Experimental results show that this method achieves a 94% accuracy in fault diagnosis, which is 6.67% higher than traditional SVM.

3.4. Generator Faults

Kabat et al. [48] developed two fuzzy inference systems to identify the faulty phase(s) and ground faults. The author used Fuzzy Logic Controller (FLC) [95] for fault detection and the inputs were the voltage and current. Through extensive simulations, it was found that FLC approach can rapidly and accurately detect different fault locations, resistances, and inception angles, with detection time ranging from 0.001 s to 0.014 s.

A DNN model was used by Teng et al. [60] to detect incipient faults in Direct Drive Wind Turbines (DDWTs). The specific fault being detected in this study is the fall-off of permanent magnets in a DDWT generator. The methodology involves several key steps: preprocessing of SCADA data to remove outliers, evaluation of variable importance using the random forest method, training the DNN model using historical healthy SCADA data, and using the Exponentially Weighted Moving Average (EWMA) [96] control chart to determine the fault threshold. The testing error from the trained DNN model is used as the metric for fault detection. The findings of the study indicate that the proposed DNN-based approach successfully detects the fall-off of permanent magnets in the DDWT generator. The DNN model exhibits excellent fitting ability between input variables and output variables, enabling the detection of the deviation degree of the online data from the normal state once the wind turbine is faulty.

Li et al. [50] presents an improved fuzzy synthetic condition assessment method for fault diagnosis in a wind turbine generator system (WTGS) by analyzing parameters such as temperature of components, wind speed, voltage, and current. This method detects changes in temperature and other physical parameters which helps to predict potential future failures or maintenance needs. For instance, a consistent upward trend in temperature readings might indicate a looming failure in the gearbox or other critical components, suggesting proactive maintenance might be required soon. The paper details a case study where the improved method’s effectiveness is illustrated by comparing it with traditional fuzzy assessments. For example, during a specific incident on 29 September 2009, the improved method accurately detected an adverse condition that led to the WTGS stopping due to high temperature, which the traditional method did not detect as a danger. Moreover, over the period from January to September 2009, the improved method more appropriately reflected the WTGS’s operational condition, detecting several ‘Danger’ conditions that the traditional method assessed as ‘Excellent’. However, the paper does not explicitly mention the percentage accuracy of both approaches.

3.5. Multi-Faults

Sun et al. [51] present a generalized model for identifying anomalies in wind turbines based on SCADA data, specifically targeting faults or abnormalities in parameters such as rotor speed, output power, and component temperature. The methods used for fault diagnosis include the development of prediction models using Back-Propagation Neural Network (BPNN) [97]. Through the prediction of component temperature, rotor speed and output power, it was predicted if there is a fault or not. The findings indicate that the prediction model trained by the current SCADA data of the wind turbine exhibits the highest accuracy and least mean absolute percentage error of less than 2%, and the proposed method for anomaly identification is more effective than traditional methods like threshold-based and single-model-based method.

A study by Dinh et al. [98] utilized digital twin technology to monitor the condition and diagnose faults in a 2MW double-fed induction generator (DFIG)-based wind turbine system. The authors implemented a digital twin using a 5D modeling approach [99], integrating machine learning algorithms, specifically DNN, RNN, and LSTM, to predict turbine output power. SCADA data collected over 18 months were used to train the models, and the DNN-based model demonstrated the best performance with a symmetric mean absolute percentage error (SMAPE) of 5.7%. The paper also introduced a fault diagnosis algorithm that compares real-time turbine data with digital twin predictions, using residuals to detect deviations and identify potential faults. Two specific fault types were detected: a generator-related fault and a gearbox-related fault. The system flagged these faults by identifying significant deviations between predicted and actual output power. Their future work aims to extend the digital twin’s capabilities to monitor subsystems and real-world implementation.

Yang et al. [54] also used LSTM for fault detection and diagnosis in wind turbines, utilizing data from a wind turbine simulator. Yang et al.’s study focuses on detecting critical faults such as blade pitch faults, generator faults, and gearbox faults, which can significantly impact the performance and reliability of wind turbines. The findings demonstrate that the LSTM-NN model effectively detects and diagnoses these faults with a high accuracy of 99%, showcasing the potential of advanced ML techniques in enhancing fault detection and diagnosis in wind turbine systems.

Table 1. Summary of fault detection and diagnosis methods for wind turbines.

Study	Fault Detected	Data	Input Data Type	Method
[47]	Blade Angle Sensor	SCADA	N/A	Non-Singleton Fuzzy Logic
[48]	Faulty Phases and Short Circuit Fault in Double Fed Induction Generator (DFIG)	SCADA	Voltage and Current	Fuzzy Logic Controller
[49]	Pitch Faults, Torque Converter Faults, and Drivetrain Wear	Simulated Data	Blade Pitch Angle and Generator Torque	Fuzzy Logic and NN
[50]	Overall Condition of the Wind Turbine Generator System	SCADA	Wind Speed, Temperature of Components, Reactive Power, Voltage and Current	Fuzzy Synthetic Condition Assessment
[51]	General Anomalies in Wind Turbines	SCADA	Rotor Speed, Temperature of Components, and Output Power	BPNN
[52]	Main Shaft Rear Bearing	SCADA	Active Power Output, Anemometer-Measured Wind and Turbine Speed, Turbine Rear, and Front Vibration Signals	ANN
[53]	Blades	Simulated Data	Wind Speed, Rotor Speed, Electrical Power, Turbine Current, and Generator Torque	LSTM
[54]	Blade Angle, Generator, and Gearbox	Simulated data	Rotor speed and Pitch Angle	LSTM
[55]	Sensor and Actuator Faults (for Blade and Pitch)	Simulated data	Rotor Effective Wind Speed, Pitch Angle, and Rotor Speed	LSTM
[56]	Blade Faults	Dataset	Images	CNN
[57]	Gearbox	Real-time Data	Vibration Signals	DCNN
[58]	Blades (Blade Angle Anomaly, Blade Surface Damage, and Blade Breakage)	N/A	Vibration Signals	MCNN
[59]	Bearings	N/A	Vibrations Signals	DNN
[60]	Permanent Magnet of Generator	SCADA	Temperature, Vibration, Pressure, Wind speed, Current and Torque	DNN
[61]	Bearing Brake Failure	Simulated Data	Vibration signals	SVM and KNN
[62]	Blade Misalignment	Simulated Data	Vibration Signals, Current and Torque of Electric Drive	SVM
[90]	Fatigue Spalling and Cracks in Gearbox	Real-time Data	Vibrations Signals	SVM
[63]	Gearbox	Simulated Data	Rotor Current	SAE-Based Multiclass SVM
[76]	Blades	Digital twin	Vibration Signals	CNN
[98]	Generator and Gearbox	SCADA and Digital twin	Vibration Signals and Output Power	DNN, RNN, and LSTM

highlights a summary of this literature in terms of the fault detected, source of data, selection criteria, and the method used.

4. Analysis and Recommendations

4.1. Analysis of Fault Diagnosis Methods

The works in [47,48,49,50] leverage fuzzy logic for fault diagnosis, but with varying focuses and methodologies. Reference [47] utilized non-singleton fuzzy logic to enhance the sensitivity of fault prediction in wind turbines, successfully identifying critical faults 3 to 5 days earlier than conventional methods. Reference [48] shows the rapid and precise capability of fuzzy logic in detecting electrical faults, achieving detection times in the range of 0.001 s to 0.014 s. Building on the work carried by [48], an improved fuzzy synthetic condition assessment was proposed in [50] to detect overheating and other critical conditions in wind turbine generators more effectively than traditional methods, thereby facilitating timely predictive maintenance. Reference [49] compared fuzzy systems and neural networks in diagnosing offshore wind turbine faults, where fuzzy systems exhibited a higher and more consistent accuracy. True Fault Detection Rate (TFR) of 0.995 and an MDF ranging between 0.06 s to 0.76 s, showcasing their robustness in real hardware tests [100]. Unlike the models presented in [48,49,50], the work of [47] presents a model that predicts faults before their occurrence, as opposed to just detecting faults based on historical data. The improved fuzzy synthetic model presented in [50] does not explicitly include any numerical results regarding accuracy and detection time. Therefore, it cannot be directly compared with the fuzzy logic presented in [48]. Although ref. [49] has demonstrated that fuzzy systems exhibit a higher TFR and Mean Detection Time (MDT) than neural networks, it is important to note that this may not always be the case because neural networks heavily rely on large datasets for training and learning patterns from data, which were not provided in this context. In contrast, fuzzy systems are designed to handle imprecise or uncertain information using linguistic variables and fuzzy logic rules. The proposed methodologies in [53,54,55] utilize LSTM-NN models, each with unique implementations and focuses. Both refs. [53,54] utilize the traditional LSTM-NN models to detect different types of faults, achieving impressive accuracies of 98% and 99%, respectively. In contrast, ref. [55] enhances the LSTM model by integrating it with an MDN, which aims to boost performance beyond that of the traditional LSTM-NN approach. Despite this enhancement, the performance of the MDN-integrated LSTM in [55], while high, does not significantly surpass the accuracies obtained by the simpler LSTM models used both in [53,54].

Different CNN models were explored in studies [56,57,58] for fault diagnosis, showcasing varied applications and effectiveness. Reference [56] enhances CNN with transfer learning to detect damage on wind turbine blades, achieving 97% accuracy. Reference [57] applied a DCNN to gearbox fault detection, which emphasized the advantages of dilated convolution for managing larger receptive fields, achieving 100% accuracy. Reference [58] utilizes MCNN with vibration signals to diagnose faults in wind turbine blades, obtaining an 87% accuracy. The implementation of transfer learning and dilated convolution represents methodological advancements that enhance model performance by addressing specific challenges such as the spatial complexity of gearbox faults. Yet, the variation in accuracy (ranging from 87% to 100%) across these studies raises questions about the consistency and reliability of these advanced models under different operational conditions.

Both refs. [59,60] investigate the application of DNNs for fault detection in different settings. Reference [59] employs an iterative DNN training method to classify bearing faults using vibration signals, demonstrating the system’s ability to refine its classifications across various bearing data. In contrast, ref. [60] utilizes a DNN to detect incipient faults in DDWT by analyzing SCADA data and establishing fault thresholds through the EWMA control chart. Despite the differing methodologies and data sources, both studies highlight the potential of DNNs in achieving high accuracy for automated fault detection in their respective domains. The main difference between the two studies is that ref. [85] uses iterative DNN training on vibration signals to cluster signals into distinct fault categories, while ref. [86] trains its DNN on preprocessed SCADA data and uses the EWMA control chart to identify deviations from normal operational data. The DNN model then measures deviations in online data for fault detection, leveraging a broader range of parameters such as temperature, vibrations, pressure, wind speed, current, and torque, which can provide better insights than just using vibration signals alone, depending on how these multiple parameters are analyzed or preprocessed.

Furthermore, traditional ML algorithms highlighted in [61,62,90] emphasize the robustness of SVM and KNN in fault detection. This study underscored the efficiency of these models in handling high-speed, robust fault detection through vibration analysis. Reference [61] reported classification accuracies of 95% for SVM and 94% for KNN, effectively replacing traditional frequency spectral analysis. In addition, the work in [62] advanced SVM’s capabilities incorporating a multi-sensory system and angular resampling techniques, enhancing feature extraction and fault classification, outperforming ANN in terms of accuracy and efficiency by almost 1%. The work in [90] demonstrated significant improvements in gearbox fault detection with an optimized SVM model using a novel fruit-fly optimization algorithm, suggesting that while newer neural network models are advancing, traditional methods like SVM and KNN continue to provide effective, less resource-intensive alternatives. Lastly, ref. [63] presents a rotor-current-based fault detection method using a deep classifier, which claims an impressive 89.3% accuracy for classifying gearbox faults in wind turbines. However, this accuracy raises questions about its performance across more diverse conditions and emphasizes the need for further validation before considering it a definitive alternative to traditional spectral analysis methods.

One of the primary limitations and challenges in applying ML to wind turbine fault diagnosis is the availability and quality of data. Using simulation datasets for developing fault diagnosis methods, while beneficial for preliminary testing, often fails to capture the complexity and variability of real-life conditions. These datasets typically simplify operational environments, leading to models that do not perform accurately when exposed to the unpredictable dynamics of actual wind turbine operations. The studies conducted by [49,53,55,61,62,63,101] are limited by their reliance on simulated data rather than real-life data. This can result in overfitting, where models perform well on simulated data but poorly on real data, potentially leading to unreliable diagnostics and an increased risk of undetected failures in critical scenarios. Hence, while simulations are a useful tool, dependence on them without subsequent validation and adaptation to real-world data can severely limit the practical effectiveness of diagnostic techniques. Combining MDN with LSTM should theoretically improve accuracy since it can represent complex multi-modal distributions that standard LSTMs cannot handle. However, this was not the case in [55] and that is why study [55] had lower accuracy compared to [53,54]. Study [55] used a real-life dataset, while both refs. [53,54] used simulated datasets. LSTMs tend to perform better with simulated data because they are clean, controlled, and lack the noise and unpredictability present in real-life data, making it easier for the model to learn the intended patterns. Due to the mentioned reasons, both refs. [53,54] had higher accuracies. Moreover, the scalability of simulated data enables the training of LSTMs on larger datasets, thereby enhancing the robustness of complex dependencies.

The literature shows that there is a high dependency on vibration signals to detect faults in wind turbines. Vibration signals have been widely used in [57,58,59,60,61,62,90] to detect and classify different faults in wind turbines. Due to the complex fault patterns they present and their nonlinear and nonstationary nature, it becomes difficult to extract deeper and more useful features from them. This leads to the use of signal processing to extract and select features from the vibration signals. This is mostly required when using traditional ML such as SVM or KNN. Unlike CNN and other neural networks, these typically implement feature extraction and selection within their layers. Even though neural networks carry out feature extraction within their layers, ref. [58] tends to overlook signal processing which affects the overall accuracy. This can be noted since the lowest accuracy was by [58], which is 87.86%. The same issue is faced by [57]; the vibration signals are fed directly into the DCNN after being segmented into two-dimensional data. The overall accuracy for the DCNN shows 100% accuracy in both training and validation datasets. While this indicates excellent performance, it is also unusual to achieve such high accuracy, which could indicate potential overfitting. However, additional testing on unseen data would provide a better understanding of its practical effectiveness.

Moreover, the works in [61,62] both applied traditional machine learning algorithms and both extracted statistical features that are derived from the raw vibration signals. Normal vibration signals are transformed to time, frequency, or time-frequency domain and statistical features are then extracted. For example, ref. [61] extracts statistical features in the time-domain only while ref. [62] extracts statistical features of the vibration signal in the time domain, followed by spectral and time-frequency analysis. The domain from which the statistical features are extracted influences the overall accuracy because depending on time-domain only is not sufficient since some features can be captured from the frequency domain only [102]. This theory can be proven since [62] obtained a higher accuracy by almost 1.86% than [61]. There are a lot of other signal processing techniques that can be more powerful and enhance the accuracy of detection and classification.

Table 2. Advantages and disadvantages of current fault detection and diagnosis methods.

Model Type	Method	Advantages	Limitations
Mathematical Model	Fuzzy Logic	Simple to implement. Effective in systems with uncertainty.	May struggle with highly dynamic systems. Relies heavily on expert-defined rules.
	Non-Singleton Fuzzy Logic	Handles uncertainty and imprecision well. Flexible and adaptable to complex systems.	Computationally intensive. Requires expert knowledge to define fuzzy rules.
	Grey Theory	Effective in handling incomplete or uncertain information. Simple implementation.	Limited in modeling complex relationships. May lack precision in detailed analysis.
	Fuzzy Petri Net	Good for modeling dynamic and concurrent systems. Handles uncertainty and temporal aspects.	Complex to design and implement. Can be computationally expensive.
Smart Technology Models	Back-Propagation Neural Network	High accuracy in detecting complex patterns. Adaptable to various types of input data.	Requires large amounts of data. Prone to overfitting. Computationally expensive.
	Artificial Neural Network	Good at capturing nonlinear relationships. Versatile and adaptable to different problems.	Requires extensive training data. Computationally expensive. Black-box nature makes interpretation difficult.
	Long Short-Term Memory	Effective for time-series and sequential data. Captures long-term dependencies.	High computational cost. Requires large datasets. Sensitive to hyperparameter tuning.
	Convolutional Neural Network	Excellent at image-based tasks and spatial data. Automatic feature extraction.	Requires large datasets. Risk of overfitting. Computationally demanding.
	Deep Convolutional Neural Network	High accuracy in detecting complex patterns. Deep feature learning.	Very high computational cost. Prone to overfitting if not properly regularized.
	Multiscale Convolutional Neural Network	Captures multi-scale features effectively. Robust to variations in input data.	Computationally intensive. May require more data to achieve high accuracy.
	Deep Neural Network	Capable of modeling complex patterns. Highly flexible and accurate.	Requires significant computational power and large datasets. Risk of overfitting.
	K-Nearest Neighbors	Simple and intuitive. Good for small datasets. No training phase required.	Computationally expensive at runtime. Performance degrades with large datasets. Sensitive to noise and irrelevant features.
	Support Vector Machines	Effective in high-dimensional spaces. Robust to overfitting, especially with a clear margin of separation.	Computationally intensive. Sensitive to the choice of kernel function. Less effective with noisy data.

highlights the advantages and limitations of both mathematical models and smart technology models based on the literature that has been covered.

4.2. Recommendations

Based on the review of current fault diagnosis techniques for wind turbines, several recommendations can be made to improve the accuracy, reliability, and efficiency of these methods. As shown in Figure 5, these recommendations are interconnected in one way or another to give us an enhanced Fault Diagnosis.

• Integration of Hybrid AI Models: While various AI techniques such as ANNs, LSTM networks, and CNN have demonstrated their strengths, integrating hybrid models that combine the benefits of multiple techniques could improve fault detection accuracy. For example, merging LSTM with CNN could address both temporal and spatial complexities in fault diagnosis, especially for components such as the gearbox and blades.
• Use of Real-World Data for Model Training: Many of the reviewed studies rely heavily on simulated datasets. While these are useful for controlled testing, they often do not fully capture the complexities of real-world turbine operation. To ensure the generalization of fault diagnosis models beyond simulation environments, real-world validation is essential. One approach is integrating SCADA data, which provides real-time operational parameters such as temperature, vibration, and power output for model assessment. Additionally, field data collection through collaborations with industry partners can offer real-world sensor readings for comparison with simulated results. Transfer learning can further enhance model adaptability by fine-tuning pretrained models with real-world datasets. Moreover, utilizing publicly available wind turbine datasets enables cross-validation under diverse operating conditions. Digital twin can also be a very powerful tool for extracting real-world and real-time data, yet there are limited studies focusing on its application in wind turbine fault diagnosis. Incorporating these strategies can significantly improve the reliability and practical applicability of AI-based fault diagnosis methods.
• Development of Advanced Signal Processing Techniques: Current diagnostic techniques that rely on vibration signals often struggle with nonlinear and non-stationary patterns. It is recommended that more advanced signal processing methods, such as Empirical Mode Decomposition (EMD) or Wavelet Transform, be combined with AI techniques to enhance feature extraction and classification.
• Direct Data Input vs. Signal Processing with AI: An important area for further investigation is whether AI models perform better when fed directly with raw data or when used in combination with signal processing techniques. While AI has the capacity to learn features directly from raw data [103], applying signal processing first may help extract critical features that improve model accuracy [104]. Comparing these two approaches across different wind turbine components and fault types could provide valuable insights into optimizing fault diagnosis techniques.
• Noisy Signal Management: SCADA data are often noisy, which can affect the performance of diagnostic models. Implementing advanced noise filtering techniques, such as adaptive noise cancelation, will improve the quality of the input data and subsequently enhance the accuracy of the fault diagnosis.
• Adaptation to Diverse Operating Conditions: The variability in wind turbine operating conditions adds complexity to fault diagnosis. Future research should focus on developing models that can adapt to a wide range of operating conditions, normalizing the effects of seasonality and fluctuating wind speeds, so that diagnostic systems are more robust.
• Focus on Early Fault Detection: As demonstrated by the success of non-singleton fuzzy logic and ML models like LSTM, early detection is critical to preventing major failures. Future development should focus on improving the sensitivity of these models, particularly in detecting subtle, evolving faults before they escalate into critical issues.
• Improved Computational Efficiency: Some AI-driven models, while highly accurate, are computationally intensive [105]. Research should focus on optimizing these models to balance between accuracy and computational efficiency, making them suitable for real-time fault diagnosis applications in operational wind farms.
• Computational Challenges in Real-Time Deployment: An important area for future exploration is the computational challenges associated with deploying AI-based fault diagnosis models in real-time industrial settings. High computational costs can be a limiting factor, as deep learning models often require significant processing power, making on-site deployment challenging. Additionally, real-time processing constraints must be addressed to ensure low-latency fault detection while handling large streaming datasets efficiently. Another critical aspect is model interpretability, as black-box AI models may not be easily trusted in industrial environments. Furthermore, seamless integration with existing industrial systems remains a challenge, particularly when dealing with legacy infrastructure. Investigating optimization techniques, lightweight model architectures, and hybrid edge-cloud solutions can enhance the feasibility of real-time fault detection and should be considered in future research.
• Implementation of Hybrid Models: Future research can explore the integration of hybrid models, such as CNN-RNN architectures, to leverage both spatial and temporal dependencies for improved fault detection in wind turbines. Additionally, transformer-based models with self-attention mechanisms present a promising approach for analyzing time-series sensor data, offering enhanced feature extraction and long-range dependency modeling compared to traditional RNNs. Applying transformers to fault diagnosis could improve predictive accuracy and robustness, particularly in handling noisy or incomplete data. Further studies should also focus on optimizing computational efficiency through techniques such as pruning, quantization, and edge computing, enabling real-time fault detection with reduced computational costs for industrial applications.

By following these recommendations, researchers and practitioners can further advance the field of wind turbine fault diagnosis, ensuring improved system reliability, reduced maintenance costs, and enhanced operational efficiency in the renewable energy sector.

5. Conclusions

This review has comprehensively examined the advancements in fault diagnosis techniques for wind turbines, with a particular focus on the role of AI technologies such as ML and DL. Traditional diagnostic methods, while foundational, have proven limited in handling the complexities of modern wind turbines, especially in nonlinear, non-stationary signal environments. AI-driven approaches have emerged as superior solutions, offering greater accuracy, efficiency, and the ability to diagnose a wide range of faults such as gearbox, blade, generator, and bearing issues.

By analyzing research from 2019 to 2024, it is evident that AI methodologies, especially those integrating hybrid models, hold promise for improving fault detection and operational reliability. Key techniques like CNN, LSTM networks, and SVM have shown high accuracy in various fault diagnosis scenarios. However, challenges remain, particularly in the reliance on simulated datasets and the difficulty in translating these models to real-world applications. Future research should focus on refining these AI models using real operational data, improving signal processing techniques, and optimizing computational efficiency to facilitate their deployment in real-time wind turbine monitoring systems.