Navigation and CNS Research,
Development of an artificial neural network to improve the detection of ionospheric spatial gradients
GNSS is the generic term that refers to satellite constellations that can provide a positioning service (ie, GPS, GALILEO…). GNSS has been introduced in aviation in the 90s, starting by the less demanding flight phase in terms of required performance: the En Route where aircraft are flying at high altitude. From the 2000s, Ground Based Augmentation Systems (GBAS) are being developed and deployed in order to augment the GNSS performances and allow its use for more demanding flight phases where aircrafts fly closer to the ground like the landing phase in adverse weather conditions.
GNSS is based on the triangulation principle: the GNSS receiver measures its distance from the satellite well-known positions and derives its position from these ranging measurements. Besides the GNSS space segment, the GBAS system consists of a GBAS ground subsystem and a GBAS aircraft subsystem. The GBAS ground system precisely determines its position using the GNSS satellites signal and calculates range corrections for these satellites. These range corrections are broadcasted to the GBAS aircraft subsystem via a radio link which allows the aircraft subsystem to compute a position with the required accuracy and integrity to support precision landing in adverse weather conditions.
One of the main sources of GNSS ranging error is the propagation delay induced by a particular layer of the earth’s atmosphere: the ionosphere. This layer is mainly composed of charged ions that interact with the GNSS radio signals and induce a propagation delay, thus a ranging measurement error. This propagation delay is taken into account in the range corrections broadcasted by the GBAS ground system and is fully corrected under nominal ionosphere condition. However, under anomalous ionosphere condition, the ionospheric conditions can have a significant spatial variation and the aircraft can experience a different ionospheric delay than the GBAS ground system. The range corrections broadcasted by the ground station become inappropriate and the aircraft will experience an unusual position error. To mitigate this risk, local ionosphere threat model that bounds the maximum ionospheric delay spatial gradient needs to be developed.
The detection of ionospheric delay spatial gradient is complex and challenging: a significant amount of GNSS raw data needs to be processed and the current processing techniques provide a large number of “fake” gradients coming from measurement artefacts. As a result, a manual validation of the processing outcome is required, which represents a significant amount of effort from skilled operators. So far, daily GNSS raw data from the last four years have been processed, and a total of 60716 potential gradients have been manually validated, out of which 1457 were true ionosphere gradients. These validated gradients compose a valuable database that can be used to further develop and improve the processing techniques.
Artificial Neural Networks have shown valuable results when applied to differentiation problems, like image recognition for instance. The challenging aspect of ionospheric delay spatial gradients detection is also a differentiation problem where true ionosphere measurements needs to be separated from measurement artefacts. Given that a large database of validated gradients has been developed and could be used to train an Artificial Neural Network, the use of Artificial Neural Network seems to be a promising technique to further improves the GNSS raw data processing and ionospheric delay spatial variation detection.
Purpose of the traineeship
The purpose of this traineeship is to evaluate the use of neural network in order to improve the GNSS raw data processing and better separate the true ionosphere spatial variation from the measurement artefacts.
- Review of the different possible neural network architecture: Given that GNSS raw signal are temporal signal of variable length, the use of recurrent neural network may be appropriate and Long Short-Term Memory Units (LSTMs) should be considered. However, other architecture may be evaluated as well
- Preparation of the training database: the large amount of validated ionosphere gradients provides a valuable training database. However this database is biased: out of 60716 gradients, only a small subset (1457) does not come from measurement artefact. Some preparation would be required to avoid biased neural network training.
- Development of the neural network using MATLAB: the GNSS raw data are processed using a MATLAB based tool. The use of MATLAB language for the development of the neural network would ease the integration with the current processing. However, if required for performances reasons, other programming language may be used (python, C++).
- Training of the neural network and optimization of the hyper parameters
- Integration of the neural network within the current processing tool: if the neural network has shown significant performance improvement, it will be integrated within the current processing tool and be used for the quarterly processing of the GNSS raw data.
- The candidate should follow a master degree in mathematics, electrical engineering or computer science engineering
- A good knowledge of MATLAB programming or C-programing is required. A previous experience (educational project and/or previous traineeship) in software development using MATLAB would be an advantage.
- A good knowledge of Neural Network (especially convolutional and recurrent neural network) is required. A previous experience (educational project and/or previous traineeship) in the use of Neural Network would be an advantage.
- A knowledge of GNSS would be an advantage
- Good communication skills and the ability to work in a multi-cultural environment are required
- The working languages of the Agency are English and French. For this particular post, candidates must have excellent level of English. A working level and/or the knowledge of an additional language would be an advantage.