СИНТЕТИЧНА ГЕНЕРАЦІЯ ДАНИХ ДЛЯ НЕЙРОМЕРЕЖЕВОГО МОНІТОРИНГУ ТЕПЛОВОГО СТАНУ ЛОПАТОК ГТД НА ОСНОВІ ФІЗИЧНООБҐРУНТОВАНИХ МОДЕЛЕЙ

One of the key challenges in developing proactive monitoring systems for gas turbine engine (GTE) components is the limited availability of representative experimental data covering the full range of operating conditions. This is particularly relevant for cooled high-pressure turbine (HPT) blades, which operate under extreme thermal loads and whose thermal state is determined by the complex interaction of multiple parameters of the cooling system. The aim of this study is to develop and substantiate a methodology for synthetic training data generation based on physically justified mathematical models of heat transfer for training a hybrid neural network intended for thermal state monitoring of cooled HPT working blades. The synthetic dataset was formed with uniform coverage of five input parameters: inlet gas temperature, cooling air pressure, cooling air temperature, cooling air flow rate, and gas flow pressure ratio. The mathematical model used for data generation is based on heat transfer criterion equations and was calibrated against existing experimental measurements. The present paper proposes a virtual instrumentation scheme comprising thirty thermocouples distributed across three characteristic blade cross-sections and one optical pyrometer was developed. The scientific novelty of the proposed approach lies in the development of a comprehensive and reproducible methodology that simultaneously accounts for spatially distributed thermocouple measurements across multiple characteristic blade cross-sections and the integral signal of an optical pyrometer – a combination that is typical for real thermo-hydraulic testing conditions but has not previously been systematically reproduced in synthetic datasets for neural network training. The practical significance of the results consists in the fact that the proposed methodology enables uniform coverage of the entire operating range, creation of a balanced dataset for all blade condition classes, and investigation of the individual influence of each cooling system parameter without conducting costly full-scale experiments. The results confirm that a hybrid neural network trained on the generated synthetic data achieves a coefficient of determination R2 > 0,99 and a mean absolute prediction error not exceeding 2 % of the operating temperature range, which confirms the adequacy of the methodology and its suitability for training neural network models for GTE blade thermal state monitoring.