How to tell if the noise level has changed significantly?

[joefowler]

Original report by Galen O'Neil (Bitbucket: oneilg, GitHub: oneilg).

---

I've noticed in a few files that the pretrig_rms cuts calculated from noise are too strict for the actual pretrig_rms values seen in pulses. This suggests that either the quiescent noise has changed, or that crosstalk is causing an increase in the noise level. I would like a way to calculate if this has happened, and this issue is for discussing possible methods. Below I'll propose a few and talk about some pros and cons.

• Observe the distribution of pretrig_rms in pulses, and see if the peak is in a different location than it was in the noise. Cons: need to learn over many observations, if you want to observe only "good" (aka no tail from a previous pulse) pulses, you need some cuts on pretrig_rms and your calculated cuts could be wrong. Maybe use a time difference cut?

• Fit for a decaying function in the pretrig rms, and inspect residual??

• Notice that a large fraction of pulses are being cut?

• Take noise triggers (portions of data long after an edge trigger). Inspect the noise trigger rms.

[joefowler]

Original comment by Joseph Fowler (Bitbucket: joe_fowler, ).

---

Galen, this raises the more general issue of "change detection". Brad has just been working on this very issue for the 2011 high-rate data.

This is an important topic, but I'm not sure what action it might possibly turn into.

[joefowler]

Original comment by Joseph Fowler (Bitbucket: joe_fowler, ).

---

Incidentally, I've noticed some cases where auto-cuts "fail". Technically, they run, but they fail in the sense that 10% of pulses pass the auto-cuts (whereas 90% were passing my hand-tuned cuts). An example is the Tupac 20160506 data set that Young took with 9 transition metal calibrators.

I still don't know how to proceed on this issue, but I at least see what you're talking about.