Diagnostics and trends¶
To judge a run beyond its acceptance check: the residuals against a normal distribution and per half-hour segment, every estimate with its uncertainty, their correlations, the iterations, and the prediction of the run before against the uncertainty it predicted. A satellite’s trends set its runs side by side, figure by figure.
diagnostics = run.diagnostics() # a finished run's
print(diagnostics.radial.sd, diagnostics.along_track.sd) # residuals over sigma: near 1 is right
drag = diagnostics.parameter("drag_coefficient").values[0]
print(drag.value, drag.sigma)
print(diagnostics.correlations.strongest[:3])
for horizon in diagnostics.prediction_check: # the run before's prediction, over its sigma
print(horizon.after_hours, horizon.position_error_m, horizon.z)
for point in diagnostics.prediction_series: # the same every five minutes, in metres
print(point.epoch, point.error_m.along_track, point.sigma_m and point.sigma_m.along_track)
for row in client.runs.trends(satellite, limit=60): # oldest arc first
print(row.arc_end, row.weighted_rms, row.drag_coefficient, row.at(24))
The residuals over sigma, per axis of the orbit: their standard deviation is near 1 for fixes weighted as they are; above 1, the fixes are noisier there than assumed (or the model misses a force); below 1, weighted too loosely.
The estimates: one whose uncertainty after the fit is near the one it had before (
a_priori_sigma) was not determined by the data. Correlations near ±1 name estimates the data cannot tell apart.The prediction check: each error of the prediction before over the uncertainty it predicted (
z) is mostly within ±1 and almost always within ±3 where the predicted uncertainty is honest.
Both are features of a tier: an organisation without them gets
FeatureNotIncluded.
As data frames¶
With pandas (pip install 'lunaspace[pandas]'), lunaspace.frames turns them into
data frames, in SI units with UTC time indexes:
from lunaspace import frames
trend = frames.trends(client.runs.trends(satellite)) # one row per run, by arc end
trend[["weighted_rms", "drag_coefficient"]].plot(subplots=True)
residuals = run.residuals() # every fix, metres, by epoch
residuals[residuals.used].groupby(residuals.index.hour)["along_track_m"].std()
diagnostics = run.diagnostics()
frames.parameter(diagnostics, "along_track_acceleration") # value, sigma, a priori
frames.correlations(diagnostics) # the matrix, named
prediction = diagnostics.prediction() # the run before's prediction against this fit
prediction[["along_track_m", "sigma_along_track_m"]].plot() # the error and its 1 sigma