NiOA-tuned multilayer perceptron for wind turbine fault prediction in predictive-maintenance applications

As the demand for sustainable energy accelerates, ensuring the reliability and operational efficiency of wind turbines has become essential to minimizing unplanned downtime and reducing high maintenance costs. This study proposes a deep learning-based fault prediction framework designed to support predictive-maintenance applications in wind turbines and enhanced through the Ninja Optimization Algorithm (NiOA), a novel metaheuristic inspired by the stealth, adaptability, and precision of historical ninjas. The developed framework was evaluated on an encrypted wind turbine dataset comprising 20,000 training observations and 5000 test observations, where each set contains 40 predictor variables and one binary target variable indicating normal and failure states. Before model development, the dataset was processed through a structured preprocessing pipeline that included data type verification, missing-value inspection, median imputation of V1 and V2, duplicate checking, and feature standardization using StandardScaler . We developed and evaluated a multilayer perceptron (MLP) model trained on encrypted sensor data to predict turbine failures before they escalate. Although the baseline MLP demonstrated strong initial performance, NiOA-driven hyperparameter optimization yielded substantial gains, achieving a classification accuracy of 98.94% and marked improvements in sensitivity and F-score, metrics directly related to minimizing costly false negatives. In addition to the optimized MLP benchmark, the study further broadens the comparative analysis by examining NiOA-based optimization for RF, SVM, and KNN models to strengthen fairness in cross-model evaluation. These performance enhancements were further validated using statistical tests, including ANOVA and the Wilcoxon Signed-Rank Test, confirming the significance and consistency of NiOA across repeated trials. A convergence analysis was also conducted to examine the search behavior of NiOA relative to benchmark optimizers, and the impact of NiOA parameter settings on optimization behavior was discussed to clarify the role of population size, iteration budget, and exploration–exploitation balance. The findings establish NiOA-tuned MLPs as a statistically robust, scalable, and economically valuable solution for fault prediction and predictive-maintenance support in renewable energy systems. Broader validation across multiple datasets and additional deep learning architectures remains an important direction for future work.