GNSS pre-correlation waveform feature extraction methods for multipath-afflicted signal classification

Prime goal of this research is to propose and test novel algorithms for GNSS multipath environment classification on the receiver Digital Signal Processing (DSP) stage, but at an earlier processing point than it is used usually; at the generation of the digitized samples of the RF signal. Towards this direction, a detailed study behind the theory and modelling of multipath has been conducted.

Multipath interference is the result of a signal’s reception via two or more paths due to reflection or diffraction of the transmitted signal. Physically, the path distance travelled by non-line-of-sight (NLoS) signals is larger than the line-of-sight (LoS) one, therefore all multipath components incident to the antenna arrive with a delay with respect to the corresponding LoS signal. Essentially, the composite signal is a superposition of the LoS and NLoS components which sum constructively and destructively; consequently, segments of the composite waveform are either amplified or attenuated.

Different feature extraction methods were studied and assessed according to their suitability to characterize multipath-contaminated waveforms. Namely, the methods of Generalized Hurst Exponent, Detrended Fluctuation Analysis, Correlation Dimension, Fuzzy Entropy and Recurrence Period Density Entropy were proposed and tested in numerical simulations implemented in MATLAB.

For testing and simulation purposes, the GPS L1 C/A signal structure was selected, as it represents the most fundamental GNSS signal, being the simplest in structure, and potentially allows for further extensions considering more complex-structured signals. The test data are numerically generated by custom MATLAB scriptsrepresenting different multipath-afflicted signals. The validity of the selection of feature extraction algorithms was performed by further simulations using off-the-shelf classifier estimators such as LDA and SVM. Finally, a combination of different features is tested to specify the optimal solution to the classification problem.