Enhanced Helicopter Vibration Prediction With Hybrid Sampling and Cost Mining Techniques

Helicopter vibrations increase pilot workload and accelerate fatigue and wear in structural and mechanical components, potentially resulting in higher maintenance costs and reduced operational safety. To address these challenges, this study develops a machine learning-based prediction model using vibration test data from the cockpit of a Korean utility helicopter. To mitigate the issue of class imbalance in the dataset, two hybrid sampling techniques are proposed and analyzed: first oversampling and last undersampling (FOLU) and first undersampling and last oversampling (FULO). In addition to conventional evaluation based on prediction accuracy, this study adopts a cost-aware perspective by applying both cost-insensitive and cost-sensitive learning frameworks. The models are compared in terms of misclassification-related cost losses under realistic operational conditions. Experimental results confirm that the proposed hybrid sampling methods outperform traditional oversampling and undersampling techniques in prediction performance. Among all configurations, the FULO-based models using deep neural network (DNN) and random forest (RF) achieved the highest prediction accuracy. Moreover, cost-sensitive learning generally reduced misclassification losses compared to cost-insensitive learning; however, in certain cases, the cost-insensitive model yielded lower total costs. These findings indicate that predictive model selection should not be based solely on accuracy metrics, but also on economic efficiency within operational contexts. This study contributes to the literature by demonstrating the practical effectiveness of hybrid sampling in helicopter vibration prediction as well as introducing a cost-aware model evaluation framework suitable for prognostics and health management (PHM) applications in military and civilian rotorcraft operations.