Abstract

The increasing demand for renewable energy technologies, particularly wind energy, has made it essential to develop an effective maintenance strategy that ensures the reliability and efficiency of wind turbines. Traditional corrective and preventive maintenance approaches are not reliable enough since unplanned downtime and high maintenance costs are often associated with them. Hence, this study developed a Predictive Maintenance (PdM) framework for wind turbines by estimating the Remaining Useful Life (RUL) of critical wind turbine components using a hybrid Deep Learning (DL) model. The study proposes a hybrid model that combines the strengths of Convolutional Neural Networks (CNN) in extracting spatial correlations and Bidirectional Long Short-Term Memory (Bi-LSTM) networks in capturing temporal dependencies. To novelize this approach, several features were added; an attention mechanism is utilized to ensure that the most informative features are given greater weight. Additionally, feature selection was applied using Binary Particle Swarm Optimization (BPSO), where only the most essential features were selected as input to the CNN-BiLSTM model. This helps reduce computational complexity and decreases the risk of overfitting, as the SCADA dataset may contain irrelevant features. The proposed solution also utilizes Bayesian optimization to select specific hyperparameters, including the number of filters in the convolutional layer and the activation functions used in the dense layer of the fully connected layer.

The proposed solution followed a structured pipeline, which starts with preprocessing and cleaning the high-dimensional SCADA dataset by removing outliers, normalizing the data, creating RUL labels, and generating sequential data segments. The next step involves PSO-optimized feature selection, followed by the development of the CNN-BiLSTM model architecture and the tuning, training, and evaluation of the model hyperparameters. However, to be able to evaluate the model performance fairly, it was compared to baseline models. Hence, three additional DL models were developed: a standalone CNN, a standalone BiLSTM, and a hybrid CNN-BiLSTM model without an attention mechanism. Each of these models will be tested twice: firstly, using all features provided in the original SCADA dataset, and secondly, using only the features selected by the PSO optimizer. The evaluation metrics used were the Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and the Coefficient of Determination (R²). The proposed method outperformed the other three models in both scenarios. The attention-based CNN-BiLSTM model with feature selection achieved an average RMSE of 56 timesteps (560 minutes or 9.3 hours). This means the model can estimate the RUL with a mean deviation of 9.3 hours. Compared to the standalone CNN model, which had a mean deviation of 13.9 hours, this is a 44% improvement. It also shows a 71% improvement over the BiLSTM model (30.47 hours) and a 33% improvement compared to the CNN-BiLSTM model (14.14 hours). In addition, MAE and MAPE were reduced by 36.9% and 11.4%, respectively, compared to the base CNN. The model also achieved the highest R² score of 0.9970, which indicates the best fit. These improvements done to the RUL estimation models will help in the decision-making process of when a wind turbine should be maintained.

Although the results of the proposed model are promising, it can be further improved by increasing the PSO population and iterations, as well as the number of calls in the Bayesian optimizer. However, the values were selected due to computational limitations, which did not allow a higher number of populations, iterations, and calls. This research contributes by integrating PSO-optimized feature selection with an attention-based CNN-BiLSTM, achieving up to a 71% improvement over baseline models and providing a computationally efficient PdM framework for wind turbines. This provides an efficient solution for maintaining wind turbines, making them a more reliable source of clean energy.

School

School of Sciences and Engineering

Department

Robotics, Control & Smart Systems Program

Degree Name

MS in Robotics, Control and Smart Systems

Graduation Date

Winter 1-31-2026

Submission Date

9-16-2025

First Advisor

Maki Habib

Committee Member 1

Amr El Mougy

Committee Member 2

Abdel-Latif Elshafie

Committee Member 3

Dina Mahmoud

Extent

143 p.

Document Type

Master's Thesis

Institutional Review Board (IRB) Approval

Not necessary for this item

Disclosure of AI Use

No use of AI

Download

Available for download on Thursday, September 16, 2027

Share

COinS