A Fatigue Reliability Analysis Method Based on Active-Learning IS-INGO-BPNN

To improve the accuracy and effectiveness of fatigue reliability evaluation, a method based on active-learning IS-INGO-BPNN is proposed for fatigue reliability assessment. This method employs an improved Northern Goshawk Optimization (INGO) algorithm to optimize the BPNN, thereby enhancing its fitting accuracy and accelerating convergence. To enhance the representativeness of the initial training samples, an active-learning function suitable for BP neural networks has been developed to identify optimal sample points. During the active-learning process, an important sampling method is introduced to concentrate more samples in the high-probability regions of the target distribution. Numerical examples were used to compare the predictive performance of active-learning IS-INGO-BPNN, active-learning INGO-BPNN, active-learning BPNN and the original BPNN, thereby validating the accuracy of the active-learning IS-INGO-BPNN model. When applied to the fatigue reliability analysis of the bogie frame, the results indicate that the IS-INGO-BPNN model achieves a predictive accuracy of 99.91% after 15 iterations of active learning, with the calculated fatigue reliability of the frame being 95.53%. The suggested approach guarantees not only the forecasting precision of the surrogate model but also enhances computational speed.