Improved Automatic Detection of GPS Satellite Oscillator Anomaly using a Machine Learning Algorithm

Illustration of a decision tree for a three-class classification task

Illustration of the random forest structure consisting of many decision trees, in which each decision tree makes its classification independently. A majority vote is employed to obtain the final predicted label

Scatterplot of L1 vs. L2 phase deviations; an example of oscillator anomalies is shown as black circles; the ratio of phase deviations for each oscillator anomaly event is black line); an example of scintillation is shown as red crosses. The ratio of phase deviation for a scintillation event is approximately red line). From Liu and Morton (2020b)

Illustration of feature importance of the random-forest algorithm. The top 12 most important features are presented. In this example, the random forest is trained using Feature Set #3 with triple-frequency signals.

Comparison of the distributions of L1 maximum phase deviation between random forest and SVM-RBF SVM. The curves above are obtained by the distribution fit to the histograms

A satellite oscillator anomaly that is simultaneously observed by four stations (Alaska, South Korea, Greenland, and Puerto Rico). The anomaly occurred at 6:39 AM, on May 23, 2018 UTC for PRN 10, which is over Pacific Ocean, close to the US coast at that time

Satellite-wise average number of detected satellite oscillator anomaly events per visible day at each station. PRNs are sorted by launch time (PRN 17 is the oldest satellite). Red PRN numbers refer to GPS Block IIRM and blue PRN numbers refer to GPS Block IIF. The anomalies are detected by the random-forest algorithm.