Scientometric Review of Artificial Intelligence for Operations & Maintenance of Wind Turbines: The Past, Present and Future

1. Introduction

• CBM supports fault detection, prediction, and diagnosis, helping identify degradation early and avoid costly failures or unnecessary maintenance outages.O&M also includes turbine and wind-farm design optimisation and forecasting vital parameters for performance assessment and energy-cost minimisation.Source: §1 · p. 1
• The field has shifted from signal-processing and physics-based approaches toward AI applied to turbine SCADA data, but limited perspective and trust may constrain uptake.The introduction notes that AI, especially deep learning, has benefited other domains more than wind energy.Source: §1 · p. 2
• Earlier reviews often focused on fault prediction or power forecasting, while broader trend analysis and recent AI applications remained underexamined.The authors identify a need to map changes over time, key data-driven O&M themes, and shifting applications of AI to CBM and performance assessment.Source: §1 · p. 2
• The paper presents the first wind-domain scientometric analysis of historical literature on data-driven O&M decision making, using statistical computing and scientific-mapping tools.Its stated scope centers on CBM while also including relevant performance assessment and analysis.Source: §1 · pp. 2–3
• The review analyzes 422 publications from the last decade, with publicly available data and an approach the authors say can be extended as new publications emerge.The authors describe the analysis as covering past and present research and drawing future insights from identified successes and failures.Source: §1 · p. 4
• The review traces conceptual and thematic evolution and identifies data quality and availability, black-box transparency, and real-time deployment as challenges for AI-based decision support.It also outlines possible strategies for addressing these challenges and future research opportunities.Source: §1 · p. 4

2. Data Collection

• The initial Web of Science search retrieved 818 records on wind-turbine CBM from 2010 onward; review articles, non-English or non-journal/conference records, and manually identified irrelevant papers were excluded.The retained publication types were journals and conference proceedings in English.Source: §2 · p. 4
• The filtered CBM corpus contained 734 records covering varied techniques, including signal processing and vibration analysis.These records were used to identify specific topics rather than rely on broad review perspectives.Source: §2 · p. 4
• A further logical criterion selected 422 AI-related records, mainly on CBM but also including some O&M performance-assessment and analysis tasks.The retrieved records were exported as plain text for subsequent scientific mapping.Source: §2 · p. 4

3.1. The Past: Prevalence of Signal Processing & Vibration Analysis

• Past signal-processing studies could diagnose faults without historical failure data, but often lacked validation metrics and robustness and could not estimate RUL or MTTF.Wavelet analysis could also be computationally intensive for fine-grained analysis and required careful parameter selection.Source: §3.1 · p. 6
• Among 734 total CBM records, 311 belonged to the analyzed period; past signal-processing and vibration-analysis publications included empirical mode decomposition, spectral kurtosis, demodulation, and wavelet methods.The review says EMD was used to predict faults, forecast power output, and design turbine control systems; it associates spectral kurtosis with vibration-signal modelling for turbine subcomponents, especially gearboxes, and demodulation with O&M signal processing.Source: §3.1 · pp. 5–6
• AI-based fault prediction was rare in the past, while a few conventional AI studies assessed turbine performance; the review notes a trade-off between added vibration sensors and SCADA data.Conventional AI examples included regression trees and hybrid probabilistic-optimisation models, which were simple to apply and interpret but generally outperformed by later approaches, especially deep learning for forecasting.Source: §3.1 · p. 7

3.2. The Present: Rise of AI in the Wind Industry

• Post-2015, AI use grew and shifted toward SCADA data and performance assessment; after 2017, deep learning rose, although interest dipped in 2017.Earlier approaches included conventional machine learning; later themes include feedforward networks for time-series prediction, fault diagnosis, operational optimisation, and system design.Source: §3.2 · pp. 9–10, Figure 6 · p. 9, Figure 7 · p. 9
• Present-day AI applications commonly use classification for fault diagnosis and regression for operational-parameter and time-series prediction, often alongside signal-decomposition methods.The review notes that signal-processing techniques such as wavelet transform and EMD continued to complement modern AI.Source: §3.2 · pp. 10–11
• Real-time deployment remains challenging: reconstruction-based fault models can be computationally expensive, while interfacing sophisticated deep learners with decision-support frameworks adds complexity.The review identifies model deployment for real-time decision support as a critical issue for the wind industry.Source: §3.2 · pp. 12, 14
• CBM research on anomaly prediction and incipient-fault identification remains embryonic, while an early SCADA temperature-anomaly approach lacked fault logs and still required technicians to classify faults.The study extended multilayer neural networks across a wind farm using a multi-agent system.Source: §3.2 · p. 15
• Deep CNN, LSTM, and GRU approaches show promise for fault and anomaly prediction, but black-box decisions lack rationales, and unsupervised methods without failure ground truth are difficult to validate.Denoising autoencoders use reconstruction error and sliding windows to model noisy SCADA features and temporal dependencies.Source: §3.2 · pp. 15–16
• Explainable AI applications remain limited, although LSTM-XGBoost transfer learning and attention-based models offer interpretable predictions and may support fault prediction in new domains without labelled failure histories.Attention mechanisms identify features influencing predictions; examples include wind-power prediction, blade imbalance detection, and causal analysis of SCADA faults.Source: §3.2 · p. 17
• Citation bursts peaked in 2016–2019: neural networks led in 2016, SVMs ranked second, and interest shifted from conventional methods to deep learning and, from 2019, CNNs and hybrid models.Support vector regression and recurrent neural networks were prominent among less popular techniques.Source: Figure 12 · p. 16, §3.2 · p. 17, Figure 13 · p. 17
• Natural-language generation could turn sequential SCADA and event data into human-readable diagnoses and maintenance suggestions, but wind-industry applications remain scarce and earlier systems handled only one or two parameters.Transformer-based models have demonstrated detailed diagnoses, action suggestions, and attention-based feature explanations; few-shot learning may help when training data are limited.Source: §3.2 · pp. 17–18
• Reinforcement learning has shown promise for turbine control, yaw control, and offshore maintenance planning, but mapping results indicate that the field receives little research attention.Deep reinforcement learning for yaw control outperformed traditional RL algorithms in the cited work.Source: §3.2 · pp. 18–19, Figure 14 · p. 19
• The publication network shows a stagnant rise of AI for CBM, with emphasis on prediction, classification, regression, optimisation, and SCADA feature engineering, but little focus on NLG or reinforcement learning.Neural networks, signal processing, and power curves feature across the mapped conceptual clusters.Source: §3.2 · p. 19, Figure 14 · p. 19

4.1. Data availability and quality ensurance

• Commercial sensitivity, scarce labelled failure histories, and unshared, dataset-specific SCADA data limit fault-prediction research, validation, and comparability.Only two of the openly available CBM data sources contain historical alarm and failure logs; unsupervised anomaly detection without ground truth cannot be validated or describe fault causes in detail.Source: §4.1 · pp. 20–21
• Imbalanced fault records, inconsistent alarm context, and noisy or complex datasets undermine AI decision support, while richer fault annotations, oversampling, and specialized language models offer possible remedies.SMOTE can balance minority fault samples; few-shot learning and pretrained language models could help generate fuller fault descriptions from limited, uneven alarm corpora.Source: §4.1 · pp. 21, 22, 24, 25
• The review recommends broader data sharing, transfer learning, common data standards, and investment in tools and training to improve data access and quality.Operators could anonymize sensitive information or share data from decommissioned turbines; transfer learning from related sources may help when labelled training data are scarce.Source: §4.1 · pp. 22–23

4.2. Challenges in deploying highly sophisticated O&M models for real-time decision support

• Real-time AI deployment is constrained less by continuous data logging than by complex, nonstandard high-resolution data, computing and memory demands, and security concerns; shallower models may sacrifice accuracy.One-second data intervals create added constraints for model inputs, continual updates, and retraining.Source: §4.2 · p. 27

4.3. Lack of transparency in the black-box natured AI models:

• Black-box AI predictions lack rationales, making operators reluctant to adopt them despite their accuracy and the generally stronger performance of deep learners over interpretable conventional models.The review argues that explainability is essential for building trust in AI-supported decisions.Source: Figure 16 · p. 26, §4.3 · p. 28
• XAI has seen few wind-industry applications, focused mostly on turbine power prediction, with limited attention to fault and maintenance-action prediction.The paper presents XAI as promising for responsible, trustworthy decision-making and for helping identify dataset and feature-behavior issues while building engineers’ and technicians’ trust in predictions.Source: §4.3 · p. 30
• XAI adoption remains slow and static: XGBoost/tree methods, SHAP packages, and CNNs with attention receive more attention than LSTM-attention, Seq2Seq, and Transformer approaches.The authors attribute the latter approaches’ limited use partly to a lack of available insights and reluctance to apply AI in the wind industry.Source: §4.3 · pp. 30–31, Figure 17 · p. 31
• Natural-language generation could predict turbine alarm messages while explaining failure causes and describing appropriate maintenance actions.The authors call for broader exploration of XAI to support human-intelligible diagnosis and prognosis of turbine operational inconsistencies.Source: §4.3 · p. 31

5. Discussion

• The roadmap forecasts likely focus areas for wind-industry AI decision support over the upcoming five years, but the authors caution that the future is not certain.The roadmap is based on current trends, successes, and challenges.Source: §5 · p. 31
• The roadmap spans three groups—AI, Big Data, and Deployment—and a topic’s first appearance signals likely high priority, while its later disappearance signals reduced but continuing importance.For example, data quality appears in 2022 with a timeframe through 2024.Source: §5 · p. 31, Figure 18 · p. 32
• The roadmap anticipates early adoption of deep learning and natural-language-generation models, growing interest in explainable AI, and slower progress on data quality, HPC, and cloud computing.Transfer learning is also expected to grow, while IoT-based real-time decision support is not expected soon because of data-security and privacy concerns.Source: §5 · p. 32
• The authors suggest real-time decision support across most wind farms globally by 2026 is possible if current AI adoption trends continue, while acknowledging the possibility of another AI winter.They expect following the roadmap could bring O&M cost savings and wider wind-energy adoption.Source: §5 · p. 32

• The scientometric review traces past, present, and future data-driven decision making for wind-industry O&M through thematic and conceptual analysis.Source: §6 · p. 32
• AI applications for wind-industry decision support have grown significantly, with particular interest in deep learning.Source: §6 · p. 32
• Signal-processing techniques are expected to continue complementing AI models during the wind industry's transition.Source: §6 · p. 32
• AI in wind energy remains embryonic compared with advances in computer vision and NLP.Source: §6 · pp. 32–33
• The review identifies challenges in quality-data access, real-time AI deployment, and black-box transparency, and proposes tailored AI, deep learning, and natural-language generation for explainable decision support.Source: §6 · p. 33

• The research received no specific grant from funding agencies in the public, commercial, or not-for-profit sectors.Source: Funding source · p. 33

A. Summary of reviewed publications

• Signal-processing approaches use vibration data to characterize bearing and gear faults, while wavelet–EMD analysis addresses aliasing but can be computationally intensive.SK-based analysis identified damage severity and transients in real-world bearing signals; DWT characterized time-varying gear-signal components and energy distribution. Wavelet analysis requires careful parameter selection, and identified faults may lack performance-metric validation or estimates of RUL and MTTF.Source: A. Summary of reviewed publications · p. 35
• Regression, Gaussian-process, and hybrid models support turbine power assessment and forecasting, including forecasts accounting for wind conditions and approaches that provide prediction confidence intervals.Gaussian processes correct wind-speed data for up to one-day-ahead forecasts, while SVM-enhanced Markov models improved short-term forecasts over SVM-only models; modified firefly methods provide probabilistic predictions. Some approaches are generally outperformed by deep learning.Source: A. Summary of reviewed publications · p. 35
• Conventional ML supports SCADA-based fault detection and analysis, but reported limitations include binary-only classification, false alarms, reliance on power curves, and absent online model updates.Random forests with PCA can identify dominant SCADA signals and provide feature importance; ELMs identify component faults but may mistake deviations from ideal signals for faults. CNN image methods identify blade faults but do not predict anomalies in internal components.Source: A. Summary of reviewed publications · pp. 36–37
• Some systems offer real-time monitoring and decision support, but others lack real-world validation, transparent reasoning, or online updates, and may still require expert intervention.A cloud-based framework is conceptual rather than experimentally demonstrated; another system enables real-time analytics with an interpretable decision tree. Spatiotemporal SCADA fusion uses residuals from normal-operation training to flag anomalies but does not explain its decisions.Source: A. Summary of reviewed publications · pp. 37, 39
• Turbine-layout methods use optimization to identify positions for performance and deployment planning, accounting for wake effects, wind-speed time series, and cable parameters.Genetic algorithms and ant-colony optimization are reported for locating turbines, while layout searches can cover many possible cases and distributions.Source: A. Summary of reviewed publications · p. 38
• CNN–LSTM fusion learns multiscale spatial and temporal SCADA patterns and outperforms conventional ML baselines, while other deep-learning studies report component-specific anomaly detection and fault identification.The studies include ELM-based fault identification with shorter training and inference time than feedforward neural networks, CNN detection of main-bearing failures across turbines, and SCADA-based temperature anomaly detection in gearbox and generator components.Source: A. Summary of reviewed publications · pp. 37–39
• SCADA-based methods can detect faults from multiple features without historical fault ground truth and model noisy, fluctuating inputs and temporal dependencies, but may be hard to validate or explain.LSTM-XGBoost is also reported to provide transparent anomaly predictions across turbine sub-components and transfer predictions to domains lacking labelled training data, although source-domain training requires substantial data.Source: A. Summary of reviewed publications · p. 39
• IoT applications support turbine control, cloud logging of O&M parameters, and remote access, but integration with AI for real-time deployment and security and privacy remain constraints.Source: A. Summary of reviewed publications · p. 39
• Attention-based LSTM and CNN approaches support interpretable wind-power prediction and diverse applications, including wind-speed prediction, blade fault detection, and causal inference in SCADA features.The reviewed approaches generally outperform conventional ML techniques, but are more complex to model than simpler neural architectures such as ANNs.Source: A. Summary of reviewed publications · p. 39
• Natural-language generation ranges from rule-based summaries and case-based reasoning to transformers that generate detailed fault messages and maintenance strategies with feature-importance explanations.Combining case-based reasoning with rule-based generation shows promising results, while rule-based development is time-consuming and challenging, case-based output may be limited, and transformer development is more complex.Source: A. Summary of reviewed publications · pp. 39–40
• Reinforcement-learning methods show promise for multi-objective MIMO turbine control, while deep RL performs promisingly in yaw control and maintenance planning for offshore vessel transfers.Deep RL can outperform conventional RL but is more complex and slower to train; planning also requires historically labelled failure data and accounts for SCADA, weather, and fault type or severity.Source: A. Summary of reviewed publications · p. 40
• Transfer learning is applied to power and wind-speed prediction, blade-ice assessment, sub-component fault prediction, and monitoring O&M parameters during anomalies.Initial source-domain training generally needs high-quality labelled data, and transfer models can require substantial computational power.Source: A. Summary of reviewed publications · p. 40
• Data-filtering approaches can improve power-curve data quality and filter wind-speed or direction data, while SMOTE generates synthetic samples to address imbalanced datasets and scarce labelled anomaly records.Reported limitations include difficulty handling imbalance and weaker performance than more sophisticated ML filtering; RAMOBoost and ADASYN are noted as alternatives used in other domains but not yet applied in wind energy.Source: A. Summary of reviewed publications · p. 40
• Explainability methods can expose feature importance and transparent outputs for black-box predictions, supporting interpretable fault prediction across turbine sub-components and human-intelligible decisions.Transformer approaches outperform conventional Seq2Seq baselines in reported applications, but are more complex and require longer training; some explainability methods do not directly support complex RNN or CNN architectures.Source: A. Summary of reviewed publications · p. 41

• More sophisticated architectures, especially deep learners, often outperform the methods discussed in the passage in performance measures such as accuracy.Source: CRediT authorship contribution statement · p. 41
• Joyjit Chatterjee contributed conceptualization, methodology, software, formal analysis, investigation, and the original draft.Source: CRediT authorship contribution statement · p. 41
• Nina Dethlefs contributed supervision, methodology, and review and editing of the manuscript.Source: CRediT authorship contribution statement · p. 41