Content area
This dissertation explores the application of advanced machine learning (ML) techniques for detecting and classifying radio frequency interference (RFI) in global navigation satellite systems (GNSS). It focuses on integrating innovative low-dimensional preprocessing methods, including principal component analysis (PCA), neural network-based latent space projection, and truncated singular value decomposition (SVD), to enhance interference detection capabilities. The study introduces four distinct methods for classifying GNSS interference and three low-dimensional projection approaches, addressing the challenges of processing diverse signal types.
The research comprises four experimental studies. The first two experiments employ unsupervised learning techniques on tabular data from a global network of Continuously Operating Reference Stations (CORS). These experiments use an unlabeled dataset, specifically developed for this research, containing over 2 million observations collected over 700 days. The findings include an support vector machine (SVM) model utilizing a decision boundary and a deep neural network (DNN) autoencoder that detects anomalies with a p-value of 0.0005. Both models have a mean Pearson correlation of 0.824 for days with manually observed GNSS RFI.
The third and fourth experiments utilize a novel truncated SVD approach for preprocessing to reduce noise variance in example datasets. This preprocessing is applied to 5,750 examples extracted from three labeled complex-valued datasets. The third experiment involves a supervised convolutional neural network (ConvNet) with SVD-based preprocessing applied to complex-value radio frequency data from GPS and Galileo spoofing signals, improving the model accuracy from a mean of 0.319 to 0.804.
The fourth experiment adopts an unsupervised ConvNet-based autoencoder for RFI detection within the training set, eliminating the need for labeled examples. This approach enhances model accuracy from a mean of 0.558 to 0.986 with SVD preprocessing. The methodologies developed, particularly in the latter experiments, demonstrate significant potential for broad application in RFI detection across various RF systems, including UAVs, 5G networks, and satellite communications.