Artificial intelligence (AI) technologies, specifically machine learning (ML), are rapidly transforming satellite navigation by substantially increasing its efficiency and accuracy, which is revolutionizing the way consumers and businesses utilize location data. A satellite navigation system with global coverage is called a global navigation satellite system (GNSS). This article discusses the role of AI in satellite navigation systems, specifically GNSSs.
Image credit: Generated using DALL.E.3
AI in Satellite Navigation Systems
In recent years, the rate of adoption of AI-powered satellite navigation systems has increased significantly for different applications in commercial transportation and military operations, as these systems can quickly process and assess data using AI algorithms and ML, ensuring more efficient navigation.
AI can also be utilized to increase the accuracy and reliability of the existing satellite navigation systems. Additionally, AI-powered navigation systems can also decrease the requirement for human labor through greater automation, which reduces costs and increases productivity. AI and ML can be utilized to develop new algorithms that can accurately and quickly analyze satellite images.
For instance, AI-based navigation systems can currently identify several objects in the environment, such as rivers and buildings, and utilize this information to create an accurate map of the area, which makes navigation more efficient, allowing users to reach their destination quickly.
The accuracy of satellite navigation systems can be improved using AI and ML, as these technologies can effectively analyze vehicle movement and precisely predict the most efficient routes. Specifically, AI-based navigation systems predict the best route considering road and traffic conditions by analyzing the patterns and speed of vehicles. AI-based navigation systems can also predict weather conditions accurately by analyzing satellite images.
ML Techniques in GNSSs
In GNSSs, GNSS-denied environments, spoofing and jamming interferences, multipath propagation, and ionospheric effect are the major sources of errors for satellite-based positioning that degrade receiver performance and affect accuracy. Hardware biases, satellite clock errors, and receiver noise and resolution are the other sources of error in GNSSs. ML can be used to address these errors in GNSSs to make these navigation systems more accurate and effective. ML can effectively handle time-series data and can learn time-dependent patterns across several models. Thus, ML can be used to identify hidden and unknown information in the GNSS data.
Several ML techniques can be used in GNSSs, including support vector machines (SVM), extreme learning machines (ELMs), neural networks (NNs), deep learning (DL) models, Gaussian process regression (GPR), naïve Bayes (NB), long short-term memory (LSTM), logistic regression (LR), gradient-boosting decision trees (GBDTs), and decision trees (DTs). These algorithms can be employed for anomaly detection, forecasting, clustering, and classification based on the GNSS application.
ML in GNSS Use Cases
GNSS Signal Acquisition: The signal acquisition process, which is implemented by all GNSS receivers, is realized through a cross-ambiguity function (CAF) evaluation in a discrete-time domain.
The CAF is a two-dimensional (2D) function related to the correlation between the local code and received signal for every possible delay/Doppler pair and possesses specific characteristics that can be used to distinguish the absence or presence of a signal from a specific satellite.
This knowledge can be utilized for training a data-driven model, such as a multi-layer perceptron (MLP) or a convolution neural network (CNN). MLP is a neural network architecture with moderate complexity, while CNN has a much greater complexity than MLPs and can capture complex non-linear phenomena.
Signal Detection and Classification: During the detection and classification of characteristics of an acquired signal, the line of sight (LOS), multipath, and non-line of sight (NLOS) signals must be classified reliably and accurately.
Currently, most solutions that implement multipath detection and mitigation have been designed using spatial geometry modeling, stochastic modeling, advanced techniques in data processing, or new special hardware designs. However, these solutions can become ineffective when an event does not fit the mathematical assumptions utilized to develop the statistical models. ML methods, such as LR, SVM, NB, DT, CNN, and recurrent neural networks (RNNs), can effectively address this issue in the statistical methodology.
For instance, an artificial neural network (ANN) model with the ability to process the structure of the autocorrelation function (ACF) can be used to detect evil waveforms (EWF). Similarly, GBDT and DT-based classifiers can be employed for GNSS signal reception classification using the satellite elevation angle, pseudo-range residuals, and carrier-to-noise-density ratio as the input features to improve the signal classification performance at the receiver.
GNSS Navigation and Precise Positioning: In both outdoor and indoor environments, such as automated manufacturing, warehouse management, port operations, supply chain and logistics, autonomous vehicles, and intelligent transportation systems (ITSs), GNSSs are used to achieve precise positioning, which is crucial for safe operations.
Precise positioning is primarily achieved by minimizing errors in both outdoor and indoor positioning. ML can be employed in location-based services (LBS) to improve GNSS navigation and positioning in several situations. For instance, the least-squares support vector machine (LSSVM) technique can be utilized to increase the accuracy of Kalman filtering (KF), to develop the LSSVM-enhanced KF (LSSVM-KF) model. The KF is used extensively as a data-fusion algorithm in navigation.
The LSSVM-KF can adaptively estimate the dynamic modeling bias from the historical information and then use the bias estimate for compensating the dynamic model, as the KF depends on the dynamic model and the quality of observations. Specifically, the LSSVM-trained algorithm considers the dynamic model bias as a time-variant ambiguous function. ML can also be employed in ITSs for GNSS position error estimation by enabling the motion sensor units to provide a more accurate position estimate during periods of blockage/outage of the GNSS signal.
Similarly, an SVM and DT can be implemented to select the most suitable feature to classify the input data at every iteration and to select the most relevant features for location error estimation.
GNSS/Inertial Navigation Systems (GNSS/INS) Integration: Currently, GNSSs are interfaced with INSs along with several filtering techniques to improve the overall positioning system accuracy by reducing positioning errors. Typically, a combination of an INS with different filters is utilized for accurate position calculation.
Although a GNSS blockage for a shorter duration does not significantly impact accurate positioning estimation, an outage for a longer duration can reduce the positioning accuracy substantially or even lead to inaccurate positioning. ML algorithms can effectively maintain positioning accuracy during longer GNSS outages.
In the integration of GNSS/INS, the KF time is increased when a GNSS cannot normally provide measurement updates. In such scenarios, the divergence of the strap-down inertial navigation system (SINS)/INS error occurs quickly without any correction of GNSS information.
A backpropagation neural network (BPNN)-aided integrated navigation method based on vehicle motion learning, RNN, and ELM with an extended Kalman filter (EKF) can be used, or a CNN-based adaptive Kalman filter can be designed to address this issue. Additionally, an NN model can be utilized to design an NN-assisted GNSS/SINS calibration system.
ML can be integrated with three-dimensional (3D) modeled assisted (3DMA) GNSS to realize high positioning accuracy in urban environments. The ML/AI algorithm, coupled with the vast 3D building model database of Google, can be used to model the asymmetric NLOS propagations to correct the NLOS pseudorange errors. This approach can reduce the Wrong-Side-of-Street occurrences in phones from GNSS up to 90%.
Recent Studies
In a paper published in the journal Remote Sensing, researchers proposed a new AI-assisted method for the high-precision integrated INS/GNSS navigation system. In this method, the position increments during GNSS were predicted by the CNN-gated recurrent unit (CNN-GRU). In this process, the CNN was used to extract the multi-dimensional sequence features quickly, while the GRU was used for time series modeling.
Additionally, a new real-time training strategy was proposed by researchers for practical application scenarios, where the GNSS outage duration and the vehicle motion state information were considered. Results from the real road test confirmed the high training efficiency and prediction accuracy of the proposed algorithm.
References and Further Reading
Zhao, S., Zhou, Y., Huang, T. (2022). A Novel Method for AI-Assisted INS/GNSS Navigation System Based on CNN-GRU and CKF during GNSS Outage. Remote Sensing, 14(18), 4494. https://doi.org/10.3390/rs14184494
Siemuri, A., Kuusniemi, H., Elmusrati, M. S., Välisuo, P., and Shamsuzzoha, A. (2021). Machine Learning Utilization in GNSS—Use Cases, Challenges and Future Applications. 2021 International Conference on Localization and GNSS (ICL-GNSS), 1-6. https://doi.org/10.1109/ICL-GNSS51451.2021.9452295.
Satellite Navigation Systems [Online] Available at https://www.sciencedirect.com/topics/earth-and-planetary-sciences/satellite-navigation-systems (Accessed on 07 November 2023)
Frąckiewiczi, M. (2023). The Future of Satellite Navigation with AI and Machine Learning [Link No Longer Active] (Accessed on 07 November 2023)