Previous Abstract Return to Session C5 Next Abstract

Session C5: Multisensor Integrated Systems and Sensor Fusion Technologies

An Intelligent and Adaptive Autonomous Navigation System for Unmanned Aircraft Systems in Urban Environments
Suraj Bijjahalli, Alessandro Gardi, and Roberto Sabatini, School of Engineering - Aerospace Engineering and Aviation RMIT University, Australia
Location: Atrium Ballroom
Alternate Number 4

The requirement to safely accommodate Unmanned Aircraft Systems with varying capabilities into the existing airspace alongside manned aircraft has led to the concept of a UAS Traffic Management (UTM) system, in order to perform safety- and mission-critical functions such as authentication, dynamic geo-fencing, constraint management, trajectory optimization, and separation assurance. High-performance navigation systems are a prerequisite to fully realize the implementation of a UTM framework. In defining navigation performance, lessons can be learned from the civil aviation context wherein requirements are defined in terms of system accuracy, integrity, continuity and availability. Given that several UTM roadmaps are aimed at achieving Beyond Line of Sight (BLOS) flight in dense urban environments, a Performance-Based Navigation (PBN) framework that dictates airspace access on the basis of navigation system capability would provide efficient use of the airspace without compromising safety. Under PBN frameworks, UAS navigation systems are required to predict/detect off-nominal events that violate assigned limits on system accuracy/integrity/continuity/availability, and consequently impact the ability to execute safety-critical functions such as trajectory conformance and separation assurance, as well as contingency measures such as Return To Launch (RTL), emergency landing etc.
In multi-sensor navigation systems, off-nominal navigation events such as a loss of accuracy owing to sensor measurement biases are conventionally detected based on filter measurement-residual consistency checks. Consistency check thresholds are typically assigned to balance missed-detection and false-alarm rate requirements.
These types of filter consistency checks are inherently dependent on the number of available and valid redundant sensor measurements in order to detect a fault and assure system integrity. Redundant measurement availability is typically a limiting factor in dense urban environments where the Global Navigation Satellite System (GNSS) cannot be fully exploited due to low satellite availability and multipath, leading to hard trade-offs between missed detection rates, false alarm rates and system availability.
In order to overcome a loss of trusted autonomous navigation system capability, this paper proposes an Artificial Intelligence (AI) -based augmentation system to exploit secondary observables derived from sensor data to infer navigation system performance and accordingly drive a decision-logic to choose from a set of feasible actions given the operational scenario. Feasible actions include navigation-mode selection i.e. selection of optimal sub-sets of sensors for state-estimation, the annunciation of caution/warning integrity alerts and the execution of contingency measures such as RTL. Candidate sensors in this paper include GNSS, inertial sensors, and visual sensors integrated in an Extended Kalman Filter (EKF)-based architecture.
The proposed augmentation system employs a knowledge-based approach based on Artificial Neural Networks to account for uncertainty in the sensor-derived observables which include satellite elevation angle, and time-averaged Carrier-to-Noise ratio (for GNSS), and brightness, blur, and spatial entropy (for visual sensors). An inference engine is developed that predicts navigation system performance over a time horizon, and accordingly selects appropriate prevention/redressal actions to mitigate off-nominal events. The inference engine is aided by contextual information such as flight-phase, horizontal and vertical alert limits, and the required false alarm and missed detection rates.
Mathematical models of sensor error propagation are developed, and correlations between selected inference metrics and error-propagation in a reference multi-sensor data-fusion architecture are investigated in order to inform system design. The system design elements presented in the paper include data-clustering and tuning of network hyperparameters, selection of appropriate contextual information in a UTM context, and rule-base development to facilitate timely-decision making under uncertainty.
The modelling phase is followed by simulations to verify system design under conditions of multiple sensor faults and measurement noise uncertainty. Representative UTM scenarios in urban environments are simulated, under conditions of varying GNSS measurement availability and noise levels. The simulation case-studies will be used to refine system design aspects including the number of data clusters, and the number of rules. The developed system will be benchmarked against a conventional navigation system architecture which does not employ the proposed AI-based performance augmentation system. The results will verify enhanced performance in terms of navigation system accuracy, integrity, continuity and availability in an urban environment characterized by intermittent and error-prone sensor measurements.



Previous Abstract Return to Session C5 Next Abstract