Performance prediction of gas turbine blade with multi-source random factors using active learning-based neural network

Rapid and accurate performance acquisition of high-temperature gas turbine blades is fundamental to energy system design, analysis, and evaluation. Data-driven surrogate models are a popular alternative to complex and expensive numerical methods, but they require numerous labeled data and are often inaccurate when influenced by multi-source random factors. In this paper, a data-driven framework is proposed based on neural networks and active learning. Gas turbine blade performance with multi-source random factors can be quickly and accurately predicted. Two active learning-based sampling strategies, the Bayesian Sampling strategy (BS) and the Combined Sampling strategy (CS), are proposed to improve the efficiency of dataset building. The established prediction model is called active learning-based neural network (ALNN), which significantly improves the prediction accuracy with small samples. Compared with the artificial neural network using traditional Latin hypercube sampling, the prediction accuracy for the maximum stress of the blade is improved by 31.9%, and the prediction accuracy for the maximum deformation is improved by 47.6%. In addition, based on the ALNN, the low cycle fatigue probabilistic life of the gas turbine blade with multi-source random factors is evaluated. When the material and processing parameters and operating condition parameters are varied randomly, the minimum value of the low cycle fatigue life is 7000 cycles.