Table 1 Summary of algorithms, strengths, and challenges in wind turbines fault detection.

Rizk et al. (2024)¹¹

Hyperspectral imaging + 3D Convolutional Neural Networks (3D CNN)

Captures spatial and spectral features; high accuracy

High-dimensional data processing; requires hyperspectral imaging equipment

Roelofs et al. (2024)¹²

Transfer learning with autoencoders

leverages pre-trained models for generalization across turbine types

Dependence on quality of pre-trained models

Letzgus et al. (2024)¹³

Explainable AI framework for power curve modeling

Enhance transparency and robustness

Complexity in developing explainable models

Song et al. (2025)¹⁴

Self-Attention Mechanism and LSTM

High accuracy and generalization ability

Depends on the quality, volume, and labeling of SCADA data

Asy’Ari et al. (2025)¹⁵

Bispectrum image analysis and CNN, CNN-LSTM, CNN-BiLSTM

Effectively capturing the gearbox failures.

Hybrid models are complex to implement and tune.

Ma et al. (2025)¹⁶

Hybrid 1D CNN-BiLSTM-AdaBoost

Leverage data from multiple sources

Without careful tuning, overfitting may occur

Wang et al. (2025)¹⁷

FCNet-5

High diagnostic accuracy and enhances reliability by incorporating multiple sensor

Struggles to differentiate between fault types with minute variations.

Lee et al. (2025)¹⁸

PDCNN + AEPSO-XGBoost Framework

Incorporates physical knowledge and addresses the issue of class imbalance in SCADA datasets

Computationally intensive and could have a risk of overfitting with noisy data

Dai et al. (2024)¹⁹

DFD-kNN

Computational efficiency due to the reduced data redundancy

Parameter sensitivity and dependency on historical data

Guo et al. (2021)²⁰

Hierarchical framework: Haar-AdaBoost for region detection + CNN for classification

Efficient detection and classification of damage types

Complexity in multi-step processing

Zhu et al. (2022)²¹

LSTM + Fuzzy Synthesis + Transfer Learning for gearbox operational state prediction

Effective fault detection using SCADA data

Requires extensive SCADA data

Jia et al. (2024)²²

AQUADA-Seg

Enhanced segmentation by integrating thermal and optical data

requires synchronized multi-modal data

Davis et al. (2024)²³

YOLO and Mask R-CNN

Real-time object detection

Trade-off between accuracy and speed

Zhang et al. (2024)²⁴

LSTM-AVAGMM

Root cause analysis and incremental training

Complexity of hybrid model

Sun et al. (2023)²⁵

Spatial-temporal multi-learner neural network

Handles imbalanced SCADA data

requires careful learner selection

Liu et al. (2024)²⁶

STGNN

Models spatial and temporal dependencies

Graph construction complexity

Ran et al. (2022)²⁷

AFB-YOLO

Better localization accuracy

Requires careful tuning

Bielecki et al. (2021)²⁸

Unsupervised real-time monitoring with ART-2 neural network + Gaussian mixture models

No need for prior training data

Potential sensitivity to noise - method may have false alarms

Sheiati et al. (2024)²⁹

Siamese CNN for blade identification + deep learning segmentation

Supports blade tracking over time

Segmentation accuracy affects tracking

Lin et al. (2024)³⁰

Random forest for feature selection + LSTM for early fault prediction

Effective time series analysis

Needs extensive historical data

Silva et al. (2025)³¹

Image processing on audible noise spectrograms

Early fault detection via abnormal noise

Noise variability - requires good noise data quality

Manshadi et al. (2022)³²

Offshore hybrid system power prediction

Predicts net power effectively

data availability

Jamil et al. (2022)³³

Deep boosted transfer learning gearbox malfunction ID

Improves learning across conditions

Transfer learning challenges

Ahmed et al. (2023)³⁴

Deep autoencoder on vibration signals for anomaly detection

Early fault detection without manual engineering

Needs vibration data

Kang et al.³⁵

Adam-optimized CNN-LSTM

High diagnostic accuracy for gearbox faults through adaptive optimization.

Computational complexity and limited to controlled environments

Zhang et al.³⁶

CNN-LSTM cascade model

Early fault detection using SCADA temperature data

Limited by low-frequency data.

Qi et al.³⁷

CNN-LSTM vibration-based fault detection

High accuracy on vibration-signal-based bearing fault detection

Relies on clean, high-quality data and is computationally intensive.

Maldonado-Correa et al.³⁸

Anomaly Transformer and TranAD,

Use self-attention mechanisms to capturing long-range dependencies and complex temporal patterns

Computational limits; only up to 18 variables could be processed due to resource constraints

Raju et al.³⁹

HARO

High prediction accuracy and effective feature selection using Lasso regression

Computational complexity and dependency on high quality data

Dwivedi et al. (2024)⁴⁰

Attention-based Vision Transformers (ViT) on drone images

attention mechanism improves focus on relevant features

Large data requirements and computationally intensive

Qiao et al.⁴¹

1DCNN with Model-Agnostic Meta-Learning

Meta-learning enables rapid adaptation

Experimental validation is limited to one wind farm

Wang et al.⁴²

SL-GPN

Addresses real-world challenges like small-sample and label error problems

high computational complexity