Speed Sensor-Aided Navigation Filter for Robust Localization in GNSS-Denied Mining Environments
David Benz, Jan-Jöran Gehrt, René Zweigel, Dirk Abel, Institute of Automatic Control at RWTH Aachen University
Location: Beacon B
Supplying industry with important raw materials while reducing the impact on nature will become increasingly important in the future. A mixed mine operation, i.e. underground and surface mining in a hybrid form, is promising. The automation of all vehicles involved can further increase the productivity, sustainability and safety. Autonomous transport units can operate reliably around the clock, even in areas that are dangerous for employees. Sustainability and safety increases by making more intensive use of existing mines as autonomous vehicles can operate in safety-critical areas.
A robust and high accurate vehicle localization is one of the most challenging task towards autonomous mining, as satellite-based localization methods are reaching their limits in deep mines. Autonomous transport units have high demands on the vehicle state estimation to enable safe travel along the often narrow paths in mines. In addition, automated loading and unloading maneuvers require precise positioning and alignment of the vehicles. Current research focuses on the automation of mining processes, but also on machine-to-machine communication, vehicle localization, and vehicle control.
In this publication, we address the challenge of locating vehicles in the surface area of a hybrid mine as well as in the transition area to underground. Especially in the transition area, it is important that a position solution is available until the localization technologies used underground take over. This must also be guaranteed when there has been no GNSS reception for a longer period of time. For most types of vehicles and applications, coupling inertial measurements and GNSS observables is a suitable approach for self-localization as shown in Konrad et al. (2018). Open-pit mines represent a constantly changing environment. Deep funnel-shaped mines reduce the sky-view angle from a certain position onwards so much that only few to no GNSS signals are received. Uneven surfaces with potholes and the vibration of the vehicle will lead to the fact that a position solution only based on inertial measurements will no longer meet the required accuracies after only a few seconds. The reason for this is the double integration of real and thus noisy acceleration data, which causes even the smallest errors to grow strongly over time and make position determination impossible. Therefore, a localization approach only aided by GNSS cannot satisfy the requirements for guidance and control of mining machines.
One solution is to extend the navigation system by further aiding sensors which achieve sufficient localization accuracy in GNSS-denied environments. Wheel odometry is a possible solution, but is accompanied by several complex uncertainties, as stated in Brossard et al. (2019). The transmission of drive forces via the tires is always associated with slip. Furthermore, the tire size or pressure can change. The measurement of the vehicle speed via rotary encoders on the axles therefore leads to non-negligible measurement errors which have a negative effect on the position determination. This phenomenon can be observed especially on driven wheels or, in the case of articulated vehicles, presumably on all wheels.
In Evans et al. (2021) and Rizos et al. (2011) a constellation of ground-based transmitters, called Locata, is used to augment the navigation solution in open-pit mines. This approach can cover GNSS-denied areas but expects the vehicles to have line-of-sight to at least a certain number of transmitters. As the mining site changes continuously, the position of each transmitter has to be chosen wisely. Furthermore, a position solution based on Locata is no longer available as soon as the vehicle moves underground. The transition zone would therefore have to be short.
Schoelderle et al. (2010) give an alternative approach for precision farming. Although the application is different, the hardware used is interesting for mining applications. A multi-sensor setup consisting of a GPS-receiver, an optical speed sensor a so-called Correvit sensor and a yaw rate sensor is used. The sensor measurements are fused in a Kalman filter. The authors achieve good results but neither the algorithm used is further stated nor is the setup tested in GNSS-denied environments.
The presented approaches are suitable for many applications. However, some do not achieve sufficient accuracy. For constantly changing and extensive mines, the use of purely vehicle-based sensors is a more cost-effective solution for mine operators, at least for a partially automated haulage fleet.
The publication on hand describes the further development of a tightly-coupled navigation filter presented in former publications Konrad et al. (2018), Breuer et al. (2016), and Gehrt et al. (2018). It is part of the developments of the research project ARTUS, which develops an autonomous haulage fleet for use in hybrid mines. The extension of the filter to a multi-sensor navigation filter, which fulfills all requirements resulting from the mining environment, is described in detail.
The filter originally bases on measurements of an IMU and is aided by GNSS observables. The tightly-coupled navigation filter is implemented as Unscented Kalman filter (UKF). To meet the challenges in mining environments, we extend the filter to a multi-sensor navigation filter by additionally aiding the navigation solution by a two-dimensional optical speed sensor (Correvit), providing measurements at 100 Hz. As the Correvit measurements strongly depend on the sensor mounting angle, the misalignment angle is estimated online and considered in the update step of the UKF. The state vector is therefore extended by a new state representing the misalignment angle.
Our new filter approach was validated on a real-time hardware with different test drives with an articulated dumper. We expected the extended filter to maintain position and heading accuracy for several seconds without GNSS. Furthermore, we expected that the new misalignment angle estimation will further increase the filter performance during GNSS outages. In order to verify these expectations, we analyzed three different filters. The first one was the basis UKF, the second one was the basis UKF extended with the Correvit measurement model, and the last one was the Correvit-aided UKF with misalignment angle estimation. The presented results were produced with recorded data in an offline post-processing step, since the misalignment angle estimation was implemented after the measurement campaign. The position and heading solution of the RTK GNSS receiver served as ground-truth.
The multi-sensor navigation filter shows good performance under normal conditions and in GNSS-denied environments. During the 80 s phase with no GNSS the Correvit-aided filters are able to maintain a smooth position with no deterioration in either position, speed or heading. The basis UKF on the other hand drifts away after only a few seconds without GNSS. The filter with misalignment angle estimation can furthermore increase the position and heading accuracy even further. With longer data sets, the difference would become even more obvious. The estimated misalignment angle converges within 35 seconds to its true value. The multi-sensor navigation filter kept the position and heading accuracy over 80 s without GNSS with almost no visible decrease. The mean position error was 0.24 m and the heading error was 0.78° in the test drive with a total duration of 152 s. Due to the good performance, the presented filter is able to provide high-accuracy vehicle state information to the trajectory controller also in the critical transition zone of a hybrid mine. In future work, we will apply the introduced concept to a Bell e30 articulated dump truck with a load capacity of 28 tons and test it in an open-pit mine under real mining conditions.
This paper is structured as follows: First, the basis navigation filter is explained. Subsequently, the optical speed sensor is introduced including the measurement principle, measurement model and misalignment angle estimation. The GNSS preprocessing, filter parametrization and initialization is described afterwards. Then the measurement setup is described and the experimental results are evaluated and discussed. Finally, a conclusion and outlook on future developments is given.