An enhanced temperature field inversion model by POD-BPNN-GA method for a 3D wing with limited sensors

Abstract

Accurate inversion of global temperature is crucial to the state evaluation of high-speed aircraft since the real data can only be acquired by very limited local sensors. In this work, a temperature field inversion model combining real sensor data is developed for a 3D aircraft wing structure with heat transport paths. The model is trained by the Back Propagation Neural Network method with optimized critical hyperparameters, i.e., max epochs, width, depth and data dimensionality. The pre-generated sample temperature fields are fully decomposed into proper orthogonal modes, the principal features are extracted to form matched data dimensionality, and the model construction efficiency is significantly improved with very little accuracy compromise. A hybrid genetic algorithm is proposed to optimize the sensor locations and numbers simultaneously with integrated considerations of inversion error and cost, and the model performance is greatly enhanced by gathering sensors to high temperature gradient region. The test results indicate great performance with the mean relative inversion error, the mean absolute inversion error and the sensor number reduction of 0.063 %, 0.496 K and 60 %, respectively and the advantages of the TFI model are verified by the comparison with Random Forest, Radial Basis Function Neural Network and Convolutional Neural Network methods.