Reinforcement Learning applied to flight control has shown to have several benefits over classical, linear flight controllers, as it eliminates the need for gain scheduling and it could provide fault-tolerance. The application to civil aviation in practice, however, is non-existent as there are multiple safety concerns. This research demonstrates the evaluation of longitudinal Handling Qualities of the Soft Actor-Critic Deep Reinforcement Learning framework with the aim to translate the unpredictable black box of Reinforcement Learning into classical flight control terminology. The framework is applied to a pitch rate command system of a jet aircraft and shows robustness to off-nominal flight conditions, center of gravity shifts and biased sensor noise. Accurate tracking performance is achieved, while adhering to Level 1 longitudinal Handling Qualities for all conditions.

The near-term artificial intelligence, commonly referred as ‘weak AI’ in the last couple years was achieved thanks to the advances in machine learning (ML), particularly deep learning, which has currently the best in-class performance outperforming other machine learning algorithms. In the deep learning framework, many natural tasks such as object, image, and speech recognition that were impossible to be performed by classical ML algorithms in the previous decades can now be be done by typical home personal computer. Deep learning requires large amount of data that has to be rapidly collected (also known as ‘big data’) in order to create robust model parameters that are able to predict future occurrences of certain event. In some domains, a large dataset such as CIFAR-10, MNIST, or Kaggle exist already. However, in many other domains such as aircraft visual inspection, such a large dataset is not easily available and this clearly restricts deep learning to perform well to recognize material damage in aircraft structures. As many computer science researchers believe, we also think that in order to achieve a performance similar to human-level intelligence, AI could and should not start from scratch. Introducing an inductive bias into deep learning might be one solution to achieve that humanlevel intelligence. In this paper, we give an example how to incorporate aerospace domain knowledge into the development of deep learning algorithms. We performed a relatively simple procedure: we conducted fatigue testing of an aluminum plate that is typically used in aircraft fuselage and build a deep convolutional neural network that classifies crack length according to crack propagation curve obtained from fatigue test. The results of this network are then compared to the results of the same network that was not injected by domain knowledge