Previous Abstract Return to Session A5 Next Abstract

Session A5: Alternative Sensors for Aiding INSs and Precision Timing

Enhanced UAV Navigation in GNSS Denied Environment Using Repeated Dynamics Pattern Recognition
S. Zahran, A.Moussa, N. El-Sheimy, University of Calgary, Canada
Location: Big Sur
Alternate Number 1

Recently, small or micro unmanned aerial vehicle (UAV) plays an important role in different fields of our daily life, as their applications save effort, time, and does not expose human life to danger. Regardless of the assigned task, the UAV must be able to operate autonomously which require accurate, precise, and robust navigation system. Most of the commercial UAVs depends on the fusion between the Inertial Navigation System (INS) and Global Navigation Satellite Systems (GNSS) to get an appropriate estimation of its states (position, velocity, and attitudes). The small size of small and micro UAVs makes the operation areas more challenging like caves, tunnels, urban and natural canyons, and downtown areas. The problem arises when the GNSS signal is blocked, as the navigation solution will be massively deteriorated as a result of the exhibited drift in the IMU. So using another aiding sensors to aid the INS and limit the drift during GNSS signal blockage is necessary. However, it’s a must to take into consideration the small size of these UAVs, which imposes some limitations on the utilized sensors like the limited space, size, weight, power, and cost. From this point of view and taking advantage of the useful information which can be gained from UAV model to enhance the navigation is the appropriate solution for these small UAVs. During the flight there are repeated dynamic patterns, that can be detected from actuators according to the model, to aid in navigation solution and limit the INS drift during GNSS signal blockage. Benefiting from the previous flights available and with the aid of supervised Machine Learning (ML), these repeated patterns can be related to actuators and act as constraints on the UAV during GNSS signal outage. This ML model aided navigation approach will not add any weight, require space, or require additional power in order to be utilized.
This paper presents a novel approach to enhance the small and micro UAVs navigation solution during GNSS signal outage, without the requirement to add any extra sensor to the UAV taking into consideration the limitation imposed over these small UAVs. By benefiting from the useful information that can be obtained from recognizing the UAV Vehicle Model (VM), as it’s possible to relate the actuators to the UAV dynamic motion pattern. There are also other approaches to utilize the Vehicle Dynamic Modeling (VDM) to enhance the navigation solution. Although to use this approach a special equipment, and a massive effort is required to be able to model the VDM parameters, moreover, due to the existence of integrations process in the VDM state equations, perturbations in the modeling parameter may lead to massive drift in the solution. The proposed technique benefits from the previously available information during past flights when the GNSS signal is available and recognize the repeated patterns occurred during flights and relates it to the actuators. Through a hybrid machine learning approach (classifier) these patterns are detected and the proposed technique will act as constraints on the UAV during GNSS signal blockage.
The proposed approach was tested on a quadcopter platform through simulation environment (MATLAB) and the adopted configuration is the plus configuration (as it massively decreases the decoupling between the X and Y axis, in order to deal with each axis separately). One of the main challenges in the proposed algorithm is the selection of appropriate inputs during the learning process that will aid in accurately detecting and identifying the repeated patterns during the flight, as it is noticed that during mission either in simulation or in real environment, the quadcopter movement can be classified according to its motor behaviour into a set of repeated patterns, these patterns were identified and detected by two machine learning classifier for each direction (X, Y, and Z), one for detecting if there is angular acceleration or not, and the other for detecting if there is velocity or not. The output of this two classifiers is not crisp output it's either 1 or 0, if the output from the classifier is 0, this means there is no acceleration/velocity in this given direction, while if the output is 1, this means there is acceleration/velocity in this given direction. After choosing the appropriate inputs that enable the classifier to detect these states the second challenge arises which is how this information from the classifiers will aid to enhance the INS solution during GNSS outage? or how it's going to be fused with the IMU through EKF?. So the outputs from the two classifiers are combined together in each direction and it is dealt with as a truth table with four possible outcomes, (1,1), (1,0), (0,0), and (0,1), each outcome represents a certain state and according to this state different information is fed to the EKF to be fused with INS solution, like if (0,0) is the outcome that means zero velocity mode along that axis, so force the velocity to be zero with some uncertainties to represents the motors fluctuations.
To clearly identify what is meant by the relation between the repeated patterns and actuators lets consider the Z-axis as an example, for the quadcopter to be in zero velocity mode in this given direction the total thrust of the four motors (T) must be equal to the total weight of the quadcopter (m) multiplied by the gravity (g), while for the quadcopter to accelerate the total thrust must be greater than the weight multiplied by the gravity and keep increasing (accelerating mode), but for the last mode (constant velocity mode) the thrust of the motors must be maintained greater than the total weight multiplied by the gravity without increasing.
In order to test the proposed approach, different trajectories were created to serve as a learning data for the ML algorithm, then a totally diverse test trajectory is created to evaluate the proposed algorithm in two different outage places with different noise amplitudes to prove the ability and the robustness of the proposed approach to enhance the navigation solution during GNSS outage. The solution of the proposed approach is compared to the solution of two IMU grades (Tactical and low-grade MEMS) created with the aid of IMU signal simulator, which is a IMU-Simulator to generate different IMU grades raw measurements with the ability to simulate different types of sensor errors, as bias instability, random walk, sensors biases, sensors scale factors, non-orthogonality, misalignment and a combinations between all of them.
Two experiments were performed in different trajectory segments each of 100 seconds outage to evaluate the ability of the ML-VM to enhance the navigation solution during the outage. The first outage experiment includes a sharp maneuver, while the second experiment includes two successive compound maneuvers that were not included in the learning data to prove the robustness of the proposed approach.
To evaluate the performance of the proposed approach the RMSE during the outage period was compared to the RMSE of the low-cost and tactical grade IMU. The results showed that the proposed algorithm was able to enhance the navigation solution during the first outage period with 2D RMSE ranging from 1.28 m to 7.5 m according to the noise level, and during the second outage period the 2D RMSE was confined between 1.1 m to 3.6 m.
Although the low-cost IMU 2D RMSE is 513 m and 452.8 m in the first and second experiment respectively. while the tactical-grade IMU 2D RMSE is 4.9 m and 5.1 m first and second test respectively.
Finally, small and micro-UAV contingent on GNSS system mainly to bound the drift in the IMU to reach an acceptable estimate of the states (position, velocity, and attitudes). Without GNSS availability the INS solution will be deteriorated massively lead to the inability of the UAV to navigate properly during this period. Different aiding sensors were utilized to replace the GNSS system during the outage to limit this drift, although the small scale of these micro UAVs imposes a limitation on the aiding system, due to the limited space, power, weight and cost of this kind of UAVs. Vehicle modeling as an aiding approach considerate this limitation but it requires special equipment/software and massive effort to be able to model every single part of the UAV to be able to reach an appropriate model to enhance the navigation solution, also lapse in the modeling parameters may lead to massive drift resembles the IMU drift during GNSS outage.
To be able to benefit from the knowledge that can be gained from the VM to enhance the navigation solution without the need for special equipment/software, massive effort, and time, machine learning vehicle model approach is adopted in this paper to recognize and relate the repeated dynamic patterns during flight to the actuators and benefits from these repeated patterns during mission to aid the INS solution if GNSS signal is blocked through EKF. This approach mainly depends on utilizing two machine learning classifiers to find a relation between the actuators and the repeated dynamic patterns for quadcopter UAV during mission (like zero-velocity, constant velocity, and acceleration/deceleration) along a given axis, and according to the detected states different information can be used to enhance the INS solution and act as odometer (velocity update).
To verify the proposed approach (ML-VM) and assess its ability to enhance the navigation solution during outage periods of 100 seconds, different test scenarios where conducted, and the RMSE was compared to low-cost and tactical grade IMU. During tests the ML-VM approach proved its ability to mitigate the INS drift and enhance the navigation solution during GNSS outage periods, such that the 2D positioning error decreased from 512 m and 452 m during the first and second outage period when using low-cost IMU to be 7.5 m and 3.6 when utilizing the ML-VM approach.



Previous Abstract Return to Session A5 Next Abstract