v0.06: AO, LS and TC detector based on apriori differences in time

[DavorJ]

v0.06 extends v0.05 with temporal change (TC) detection, and also fully makes use of apriori hydrostatic pressure information. The latter allows for dynamic threshold determination based on any specified time-interval.

Results: inbo and geotech

A comparison with tsoutliers and previous models: geotech and inbo.

TC is defined as an outlier with an exponential decay. In contrast to tsoutliers and Chen's suggestion to use 0.7 for decay, here we use a variable decay and consider the TC as a two-parameter model which is tested against the one-parameter AO and LS models. Best examples are within INBO data:

There are cases where it failes:

The decay is very slow, and failes the significance test.

The top spikes are not outliers, so they are OK. The right lower spike has an extra low point, before it starts the exponential return.

The Geotech data doesn't seem to have good TC examples. The model sometimes gets confused with tidal flows and decays. Here an example:

I see two ways to get around this: either (1) disable 'TC' detection, or (2) model the tidal flow. The latter is currently planed for +v0.07.

Compared to tsoutliers we have:

• Speed: with TC detection ON, it usually takes less than a second or 2. There are few pathological cases: DYLP029A_S4054 and born_pb 112d_151202090850_T4180.MON that take 2 minutes to complete.
• No need for aggregation
• Variable decay
• Apriori information
• Variable time-intervals