Sensor Hybridization Through Neural Networks for Rockets Guidance Applications

The guidance, navigation and control, known by its initials GNC, of ​​aircraft and autonomous space vehicles, and more concretely artillery rockets, has been one of the spearheads of research in the aerospace field in recent times. Improving accuracy is cornerstone for ballistic projectiles. Using inertial navigation systems and Global Navigation Satellite Systems (GNSS), accuracy becomes independent of range. However, during the terminal phase of flight, when  movement is governed by non-linear and highly changing forces and moments, guidance strategies based on these systems provoke enormous errors in attitude and position determination. Employing additional sensors, which are independent of cumulative errors and jamming, such as the quadrant photo-detector semi-active laser, can mitigate these effects. Despite of this fact, it is well known that the costs associated with this autonomous navigation, both in terms of sensors, and in costs associated with the development of the algorithm itself or in characterizing the different aircraft  such as mechanical tests, wind tunnels, etc. are very high. Trying to reduce these costs through the development of advanced algorithms for estimating data regarding the position and attitude of vehicles, and the elimination of sensors seems essential for the progress of aerospace engineering. In line with this last point, the development and training of neural networks with applications in guidance, navigation and control of aerospace vehicles seems to be currently very popular for this purpose. This research presents a new non-linear hybridization algorithm to feed navigation and control systems, which is based on neural networks. The use of a neural network for the estimation of parameters based on the dynamics of the projectile presents the advantage that once this network is trained, it is no longer necessary to know the physical-mathematical foundations that govern the dynamics of the vehicle, but it is the network that, based on input data, returns dynamic parameters that can later be used within the GNC algorithm. The objective is to accurately predict the line of sight vector from multiple sensors measurements, such as GNSS, Inertial Measurement Units (IMU) and semi-active lasers. It should be noted that the advantage of such a combined system over the  GNSS / IMU one is the ability to avoid distortion and modify final impact angles. Non-linear simulations based on real flight dynamics are used in a first stage to train the neural networks. Once training is completed, neural network hybridization is tested and simulated together with modified proportional navigation techniques and novel control methods. Monte Carlo analysis is conducted to determine closed-loop performance across a full spectrum of uncertainty at  initial  conditions,  sensor  data  acquisition,  atmospheric  conditions,  and  thrust  properties.  Simulation results demonstrate the  performance of the presented approach in a six degrees of freedom simulation environment showing high accuracy and robustness  against parameter uncertainty.