Progress in Centralized Monitoring of the International GPS Service Network Angelyn W. Moore Peter N. Jeziorek Eric W. Richardson Ruth E. Neilan IGS Central Bureau
4 quantities from the teqc summary L1 multipath L2 multipath Number of observationsSlips (x1000) per observations
Changes in these parameters can be sudden or gradual L1 multipathSlips/obs
Compare value & variance against the rest of the IGS Slips/obs
Change point analysis Cumulative sum of the differences between the values and the mean S 0 = 0 S 1 = S 0 + X 1 – X mean S N = S N-1 + X N - X mean “Bootstrap” (randomly reorder) the data set and check whether peak of cusum is higher or lower. Repeat a bunch of times. Confidence level of the change point is the fraction of times the bootstrapped set’s cusum is flatter
Outliers? We don’t really want isolated outliers flagged, but we do want significant changes to be found. When outliers are detected, we use the rank of the data point, instead of the value. Original data Same data; ranks instead of values This decreases the impact of the outliers on the cumulative sum, but real changes are still detected.
Some examples
How are we using this? As a screening tool to decide what a human should look at more closely. We’re gathering data on what patterns in the time series correspond to what kinds of real events. No automatic notification is sent to operators at this time.
Conclusions 1.I have lots of time-series data to examine 2.Maybe you do too 3.The computer can help by making a first pass through the data, using cumulative sum change-point analysis to decide what deserves a closer look from a human.