---

links to sections and plots:
the problem

• rfifind plot for cband .pdf

setup
processing
average spectra/acf's

• the spectra averaged over each file (.ps) (.pdf)

dynamic spectra

• Dynamic spectra for Cband data (.png)
• Dynamic spectra for lband data (.png)

total power vs time

• gpu power vs time (.ps) (.pdf)

summary

Problem resolution:

• 180430: found the IF cables going into puppi (from the receiver room) were very loose.. Both polA and PolB.
  
• The sma connectors at puppi's panel were tightened.
• The cables could have been touched for the hsa/frb runs when more attenuation was needed.
• 3-5apr18 run.. no attenuation added.
  
• This  explains the standing wave in polA.
  
• The comb in polB is still outstanding.
  

---

The problem:

    P3054 took coherent search data on the frb at cband and lband using pupping on 21apr18. Jason reported that rfifind returned excessive power for both cband and lband (see rfifind plot for cband .pdf). I looked at two of the many files generated to see what was happening:

• puppi_58229_C0531+33_0289_0005.fits
• 214 seconds, (41GB), cband
• puppi_58229_C0531+33_0291_0001.fits
• 214 seconds (41GB) lband wide
  

---

Experiment setup:

Puppi setup:

•  coherent search, 512 channels, 800 MHz bandwidth, 10.24 usecond sampling, dm=557,  polA,B,CrossReal,CrossImg
• coherent search will split the 512 channel into 8 sets of 64 channels and send each set to a gpu to do the coherent dedispersion.
• The blocks are later combined back into a 512 channel spectra.

Receiver setup:

• cband:
• 4101.56  to 4900 MHz
• Dana's up,down converter is used, so the entire  800 MHz bw is sent to puppi.
• lband:
• 918.562 to 1780 MHz
• Dana's up,down converter is not used.
• the upper edge of the rf filter and the lower edge of the 1-2 GHz IF filter define the bandwidth
• puppi's input is 800 1600 MHz IF (center at IF=1200MHz --> 1380MHz rf).
  
• the 1Ghz IF filter cutoff will start to cut off the rf below (1380-200)=1180 MHz rf.
  
• So there should be very little power from rf freq 918 to 1150MHz  because the 1GHz IF filter will cut it out.
  

---

 Processing:

    The spectra in each file were first averaged over 1 row giving:

• cband:  650 x .330 second spectra
• lband :  706 x .304 second spectra.
• These were then averaged further to give 1 average spectra for each file

The row averaged spectra were used to make dynamic spectra, and total power by gpu vs time.
The file averaged spectra were plotted, and used to compute the acf.

---

Average spectra, acf

The first set of plots show the spectra averaged over each file (.ps) (.pdf)

• Page 1: spectra averaged over file (214 seconds)
• color usage:
• Black is polA, red is polB
• the blue dashed lines show the 64 channel boundaries of the gpu processing.
• the green dashed line shows where the IF Filter cuts off the lower part of the lband spectra.
  
• Top: cband spectra
• There is a large ripple in PolA
• polB has some smaller spikes (hard to see before blowing up the picture)
  
• Bottom: lband spectra
• PolA has the same large ripple
• this means the ripple is not in Dana's up,down converter
• the ripple is just as strong in the 918-1150 MHz region where the IF filter has cutoff the IF power
• So the ripple is after the IF 1-2 GHz filter
• PolB shows spikes below 1150 MHz (where the IF filter is cutting off the power).
  
• These spikes probably continue across the band, but get swamped by the rf power.
• existing below 1150 also says that these are coming after the 1-2 GHz IF filter.
• Summary:
• ripple in PolA is coming after the IF 1-2 GHz filter, it is not from Dana's box .
  
• Page 2: ACF shows shows distance to reflection for polA
  
• Black is polA, red is polB
• The fft of the spectral density function gives the acf.
  
• Any ripples in the spectra will become spikes in the ACF at the delay for the reflection.
• The x axis is labeled: C/2 *.7 *delay.  I'm using .7 for the velocity in the cables (could be different).
• The polA reflection is over a distance of 7.8 meters ( 1/800 MHz should give a resolution of .13 meters.. about 5 inches).
• Page 3: band pass corrected spectra to shows comb in PolB.
  
• to look at weaker, narrow spikes, a  band passbandpass correction was computedided into the polB avg spectra.
• a 43 order harmonic was fit to the average band pass and then divided info the average spc
• This will  leave interference that is narrower than 67 MHz.
• Top Frame: cband PolB
• where is a comb across the entire band
• Bottom:  lband PolB
• It is not as easy to see the comb (because of other rfi).
• The comb is very strong on the left part of the band (where the IF filter cuts off any rf/If power.
• The normalization then kicks up the comb amplitude.
• Page 4:polB  Bandpass corrected spectra vs puppi channel number
• The band pass corrected spectra are plotted vs puppi channel number (0..511).
• Black is the Cband spectra, Red is the lband spectra.
  
• Green shows the comb spacing (8 channels 12MHz)
• Top Frame: channels 0 to 255 are plotted
• Bottom Frame: channels 256  to 511 are plotted.
• You can clearly see that the comb lines up at 8 channel spacing for both lband and cband.
• Since both receivers are using different IF setups, this comb is in puppi, rather than something upstream.

---

Dynamic spectra:

    Dynamic spectra were generated from the .3 sec averaged spectra.

• A median band pass was used to flatten each dataset
• This removes the PolA ripple (since it is stable in time).
• The PolB comb is mostly removed.. although some of it can still be seen (which implies that the amplitude to not perfectly stable in time.
• For each image, polA is shown on the left, and polB on the right
• Vertical red lines show the gpu boundaries
• the gpu sections are labeled at the bottom

Dynamic spectra for Cband data (.png)

• The polA ripple is removed by the median band pass correction.
  
• You can see a little of the 8 channel comb on the right side of polB

Dynamic spectra for lband data (.png)

• The red lines show that
  
• gpu 6 (cfr 1430 MHz) has trouble starting around 185 seconds
• gpu 1 (cfr 1730 MHz) has large changes for more at least the 2nd half of the observation
• The problems occur in polA and polB
  
• The problem stays within a gpu, so the problem is in puppi rather than the receiver/ if/lo
• PolA radar saturation.
• The radars are saturation polA much more than polB.
• I wonder if the strong standing wave in polA is contributing to this?
• 918-1160 MHz data Has been filtered out by the IF 1-2 GHz filter.
• so what we see is coming down stream from the filter
• You can see the 8 channel comb in polB more easily (since there is not rf power to cover it up).
• The polA radar saturation is also strong. This Probably means the problem is at or after the digitizer.

---

Total power vs time

    The average power in each gpu band was computed (at the .3 second row time resolution) and then plotted vs time.

gpu power vs time (.ps) (.pdf)

• Each gpu band is plotted using a different color.
• The top Frames are polA, the bottom frames are PolB
  
• The power values are the average value for the spectra over each band.
  
• A value of < 10 implies low power.
• Page 1 Cband power vs time.
• each plotted power is labeled with the gpu name and the center frequency of the band
• gpu's 1,2,3 have values < 15 . I thing this comes from the band pass shape of Dana's up / down converter . (see the average spectra).
• Page 2: lband power  vs time
• Top frame polA, bottom frame polB
• gpu1 started to have trouble around second 50. This occurred in polA and PolB
• gpu6 started having trouble around sec 185.
• Page 3: blowup lband power vs time
• Top is polA, bottom is polB
• the + show the .3 second avged samples
• The problem with gpu1 shows a dropout every 5 samples  (or rows in fits file) .. (about every 1.5 seconds)
• Around 130 seconds the 5 sample dropout started to change.
  

Summary:

• the p3054 data taken 21apr18 did not have a large external rfi causing problems in cband and lband.
• It looks like the problems are localized to puppi.
• PolA - there is a large standing wave in polA data.
• This is present with and without the puppi mixer.
• If it is an reflection in a cable , then the reflection distance is 7.8 meters (assuming velocity in the cable is .7*C).
• The reflection is occurring after  1-2 GHz  if filter.
• PolA shows much more saturation when the radars hit than polB. I wonder if the large standing wave is contributing to this?
• coherent search does predetection dispersion. The frb dm is around 500?.. i'd expect the standing to get smeared out some by this.
  
• PolB - there is an 8 channel comb in polB.
  
• It is present in the lband and cband data (and stays in the same puppi channels).
• this is most likely coming from puppi.
  
• It is spreads across all 512 channels (it is not isolated to a set of gpus).
  
• I wonder if the polyphase filter bank is have trouble when it splits the data into 8 sets for the 8 gpu's?
• gpu 1,6   both had dropout problems during the lband portion of the run.
  
• ToDo:
• see if we can see the standing wave at the input to puppi using a spectrum analyzer.
  
• Do this before trying to tighten any cables.
• then start tightening cables.
• Not sure how to attack the 8 channel comb.
  
• should look in fast4k mode (without added polarizations if that's possible) to see if it is present.
  

processing: x101/180421/p3054rfi.pro