Cenwei Ma

This paper studies a structural reliability analysis method based on adaptive neural network ensembles combined with Monte Carlo simulation (ANNE-MCS) during the water landing of wing-in-ground (WIG) aircraft where artificial neural networks are employed as the surrogate model. To guarantee the accuracy of the failure boundary fitting, a weighted mean square error loss function is constructed, which enhances the weights of samples located in proximity to the failure boundary. Consequently, the neural network model with locally accurate surrogate capability can be obtained with a minimum number of samples. Additionally, the proposed method combines with an active learning strategy, employing the U learning function to perform adaptive sampling in regions with high predictive variances and proximity to the failure boundary, thus reducing the number of required samples. Finally, the proposed method is applied to evaluate the reliability of the WIG aircraft belly structure under multiple water landing conditions, providing references for water landing operations. Fluid-structure interaction models are adopted to calculate the structural stress, considering pitch angles of 0°, 3°, 5°, and 7° at the landing speed of 24m/s. The results demonstrate that the error of the failure probability obtained by ANNE-MCS is only 0.5%-4% compared to the true values. A comparative analysis of the four cases reveals that the maximum structural stress values occur at the step in all cases. Furthermore, the failure probability of the WIG aircraft belly structure is found to be lowest when landing at a pitch angle of 7° in the defined step angle.