Thrust Estimation for Aero-engine Based on Deep Convolution Neural Network

Aero-engines’ main task is to provide thrust for military or civil aircraft, so direct thrust control technology becomes one of the important research directions in the aero-engine control field. Thrust estimator design is an important part of direct thrust control. To overcome the shortcomings of traditional thrust estimation methods, such as low accuracy and poor robustness, a deep convolution neural network (DCNN) was introduced to design thrust estimator for aero-engine. Taking a turbofan engine component-level model as the objective, 1 data samples were generated through simulation for training and testing the estimator. Twelve measurable parameters, such as fuel flowrate ( W f ), combustion chamber exit temperature ( T 4 ), high pressure rotor speed ( HPS_N mech ), low pressure rotor speed ( LPS_N mech ), nozzle exit temperature ( T 8 ), etc. were selected as the input features of the estimator, and engine thrust was used as the output label. The effects of training data and test data partition ratio, convolution and pooling layer number, convolution kernel size, pooling window size and channel number of convolution/pooling layers on the performance of the thrust estimator were compared and analysed. Furthermore, the performance of the DCNN based thrust estimator was compared with that of traditional thrust estimators based on BP neural network and support vector machine (SVM). The validation results show that the DCNN based thrust estimator proposed in this paper has excellent thrust estimation performance, and its average accuracy can reach as high as 0.11%.