An INS/GNSS fusion architecture in GNSS denied environment using gated recurrent unit

Abstract

One of the most used Position, Navigation and Timing (PNT) technology of the 21st century is Global Navigation Satellite Systems (GNSS). GNSS signals are affected by urban canyons that limit line-of-sight and reduce satellite availability to receivers. Smart cities are expected to adopt autonomous Unmanned Aerial Vehicles (UAV) operations for critical missions such as transportation of organs which are time-sensitive. Therefore, higher accuracy for position and velocity information is required. This paper investigates the use of Gated Recurrent Units (GRU) as a suitable technique that can memorize previous information in conjunction with the inputs (consisting of attitude, change in attitude, and change in velocity) to reduce position and velocity error when GNSS is not available. The fusion approach is developed and tested using Spirent's SimGEN GSS7000 hardware simulator which simulates GNSS signals and Spirent's SimSENSOR software that simulates accelerometer and gyroscope stochastic and deterministic errors. GNSS outage is varied between 1 and 20 seconds randomly to affect predicted position and velocity. The data is collected and used to train the GRU to predict the position and velocity error measured by the Inertial Measurement Unit (IMU). From the performance evaluation, a 60% reduction in Root Mean Squared Error (RMSE) is observed compared to Recurrent Neural Networks (RNN). Comparing 95 th percentile with Inertial Navigation System (INS), RNN, and GRU, an 80% reduction is observed between INS and RNN. Furthermore, a 35% drop in the 95 th percentile is observed between RNN and GRU.

Multipath Detection from GNSS Observables Using Gated Recurrent Unit

“…GRU, on the other hand, is more capable of determining non-linear relationships between the features and using previous state information to provide greater accuracy. This is seen in paper [1], where GRU provided better correlation in cases where the relationship between input features and the outputs was highly non-linear. Table 1 compares the mean, standard deviation, and 95% between the original output (Ublox F9P), tree, KNN, SVM, and GRU.…”

“…Since the launch of the Global Positioning System (GPS) in 1983, GNSS has been used in a variety of applications and devices such as navigation for vehicles both on the ground and in the air. [1] [2] Currently, six GNSS systems exist. All these systems work with similar principles to provide positioning information to the receivers.…”

Experimental Evaluation of GNSS and IMU Fusion Using Gated Recurrent Unit

“…The GRU model can be describe as the follow equations [13]: z t " σ pW z ¨rh t´1 , X t s `bz q r t " σ pW r ¨rh t´1 , X t s `br q ĥt " tanh pW ¨rr t d h t´1 , X t s `bq…”

“…The issue with RNN is that it has difficulties solving problems that require learning long-term dependencies due to the gradient of the loss function decaying. [13] applied RNN and its variations long-short term memory network (LSTM) and gated recurrent unit (GRU) for INS navigation when GNSS fails at open sky environment. The results showed that Both LSTM and GRU outperforms RNN with less computational efficiency.…”

A Closed-Loop Output Error Approach for Physics-Informed Trajectory Inference Using Online Data

“…Machine learning has been used for trajectory inference using measurement data [12], [13] with interesting results. Physics informed neural networks [14] uses prior knowledge of the physics of the system with a neural network to infer the trajectory.…”