Layer Normalization Enhanced Geometric Feature Fitting Neural Network for Flush Air Data Sensing

The Flush Air Data Sensing (FADS) method based on Artificial neural networks (ANNs) has been widely studied and applied in air data sensing for advanced aircraft. Most current methods focus more on numerical fitting but lack analysis and discussion of the physical meaning of the fitting process. That makes the fitting process highly susceptible to the quality of the training data, imposes high requirements for the training data and creates more uncertainties in practice. In this paper, the FADS method based on the Geometric Feature-fitting Neural Network (GFFNN) is proposed to optimise performance under conditions of limited training data and improve the interpretability of the fitting process. The GFFNN is constructed to fit the aerodynamic angle in the air data and the pressure paxameters in the air data are calculated in the analytical method. For that, a new dimensionless normalization method is proposed to enhance the geometric distribution feature of the neural network’s input and the fitting process is decoupled to reduce the complexity of the fitting objective. To validate the accuracy and stability of the proposed method, three distinct training-testing datasets combinations were constructed based on wind tunnel test data to simulate various real-world scenarios. Experiment replications are conducted to mitigate instability in training results, ensuring a more accurate assessment of the method’s effectiveness. The results show that the proposed method achieves superior stability in air data errors across diverse simulated scenarios. The stability is increased by 24.5% to 83.7% compared to the current methods with comparable or even higher accuracy, and the advantages would become more pronounced under stringent conditions.