GNSS Karıştırma Tipi Sınıflandırmada Makine Öğrenmesi Yaklaşımı
Özet
Today, GNSS signals are signals that are evaluated in determining position, time and speed depending on satellite sets. In order to obtain relevant data many of the military and civilian elements use these signals from systems such as GPS, GLONASS, BEIDOU, which are parts of the GNSS system. This information aims to ensure that all manned/unmanned platforms reach the target they want to achieve in the most accurate way with the lowest deviation. However, the approach of jamming these signals and preventing from accessing the data the user they need to obtain from this signal is one of the approaches frequently used today. Platforms and systems can be adjusted to their position or speed and they can be neutralized by losing their consciousness about it with the intervention of jamming. Situations such as seizing, breaking, disappearing, and inability to perform tasks as desired are among the main outputs of jamming activities. In order to get rid of these negative effects of jamming, the user aims to identify the mixer and show some approaches to reduce its effect.
There are certain types of jamming. The separation of these cases from each other and from the state without jamming is considered a critical situation in reducing the jamming effect. The correct classification enables the most accurate solution to be applied to reduce negative effects of jamming. It is not possible to make this classification at high speeds by human hands. For this reason, using one or more of the machine learning approaches in solving these classification problems in order to obtain the most accurate classification, has been discussed in many different studies.
There are four cases as AM, FM, NB and Chirp jamming and fifth situation are determined as a case without jamming In this study which aims to detect GNSS jamming and to classify the jamming types correctly. We had a totally of 50000 signals are produced as 10000 signals for each cases. The generated signals are made more realist depending on certain noise ratios and mixing ratios.
The obtained signals are provided as input to the machine learning models firstly as raw data directly, then by using the features extraction by short-time Fourier transform (STFT), and then by using the features extraction with the Wigner-Ville Distribution. It has been analyzed which of the preferred STFT and Wigner-Ville analysis methods performs better in feature extraction proccessing and how does the performance obtained in studies conducted directly with raw data without feature extraction compare to the studies conducted by feature extraction.
Raw data and features extracted by time-frequency analyzes were given to five different machine learning models determined as CNN, SVM, RF, MLP and NN and learning performances were obtained. Results with all data are labeled as AM, FM, NB, Chirp, and NJ cases. The NJ case represent the absence of jamming, while the AM, FM, Chirp and NB cases represent the cases in which the presence of the relevant jamming is detected.
It has been observed that the best performance from the preferred machine learning models can be achieved as a result of the CNN learning model carried out by using the features extracted with STFT, and the classification performance is %98.5.
