Machine learning methods as part of the digital twin of a gas turbine engine: surrogate modeling and technical condition diagnostics

Abstract

The article considers the application of machine learning methods as an element of the digital twin of a gas turbine engine (GTE) at the stages of design, testing and operation. A classification of machine learning methods for engine-building tasks is proposed (surrogate modeling, fault diagnostics and classification, condition prognostics, hybrid physics-data models, anomaly detection), as well as the architecture of a hybrid neural network module integrated into the unified information space of the digital twin. Using a methodological example of a surrogate neural network model of the throttle characteristic of a turbofan engine trained on the results of calculations in a simulation thermogasdynamic model, it is shown that a multilayer perceptron provides an average relative approximation error of thrust, specific fuel consumption and turbine exit gas temperature of 0.3–0.8% with a maximum error of no more than 2.1%, which is comparable to the scatter embedded in the training data. The possibilities of using neural network models for parametric diagnostics and assessment of the technical condition of gas turbine power plants (based on GTE) are discussed. doi 10.54708/19926502_2026_30311379