Active learning multi-fidelity neural network framework for LCF reliability analysis of peak-shaving gas turbine components

Reliability assessment is of great importance for safeguarding the structural integrity of gas turbine components. However, traditional approaches frequently encounter accuracy and efficiency bottlenecks when confronted with the scarcity of high-fidelity data, alongside high-dimensional and strongly nonlinear scenarios. To address these challenges, this work proposes a novel active learning multi-fidelity neural network with an adaptive U learning function (AMFNN-AU) framework. Specifically, the MFNN comprising two fidelity sub-networks is designed to approximate the limit-state function. By fusing inexpensive low-fidelity data to capture global trends and exploiting complex cross-fidelity correlations, the proposed framework significantly alleviates the reliance on hard-to-obtain high-fidelity data. Furthermore, a multi-fidelity (MF) Monte Carlo Dropout scheme is developed, extending Monte Carlo Dropout into cascaded MF architectures to provide reliable uncertainty estimation. The derived predictive means and standard deviations of both fidelities then enable the proposed AU learning function to adaptively select the most informative candidates and their optimal fidelity levels. Numerical benchmarks encompassing both linear and high-dimensional nonlinear cases verify the framework's performance advantages in computational efficiency, accuracy, and robustness. Finally, an engineering application to a peak-shaving gas turbine disk demonstrates that the proposed method significantly reduces sampling costs while maintaining high assessment accuracy, providing an efficient paradigm for complex structural reliability analysis.