Prediction of helicopter rotor loads and fatigue damage evaluation with neural networks

Abstract

Machine learning algorithms have undergone rapid growth in recent years thanks to the ever-increasing amount of data and the parallel growth of computational power. Among the machine learning algorithms, one of the most famous and most effective classes for performance and flexibility are artificial neural networks, algorithms capable of learning relations between the data. In this work, neural networks are exploited to infer the relation between flight mechanics parameters and resulting loads of an articulated rotor configuration. The accuracy of these algorithms is closely related to the quality of the dataset used for their training. Since rotor loads are time-periodic signals with a precise harmonic content, a dedicated neural network is trained to predict each harmonic separately. The load time history is then reconstructed a-posteriori by combining all the predictions given by every single network. Different types of network architectures are tested, and a sensitivity analysis is conducted on hyper-parameters to determine the optimal configuration for the specific application. Furthermore, such predictions are then used to feed a fatigue damage calculation algorithm.