We're going to have to reinstate the policy of No food / organic trash *anywhere* in the 40m.  Everyone has been pretty good, keeping the food trash to the one can right next to the sink, but that is no longer sufficient, since we've been invaded by an army of ants:

We are going back to the old policy of Take your trash out to the dumpsters outside.  I'm sure there are some old wives tales about how exercise after eating helps your digestion, or something like that, so no more laziness allowed!

Jamie and Steve got the ETM doors on this morning.

We got the other heavy doors including the ITMs, BS and the access connector in place.

If nobody raises any concerns in reply to this elog, Steve will assume it as a green signal and will start pumping down first thing Monday morning after the final check on the access connector bellow screws.

Steve! 

Ayaka and I got the ITMY and BS door closed at 45foot pounds just now. 

We're still good with the IFO alignment after 7hours.

I found the green still locked in the same state as last night; but no IR (so the arms are stable and the TTs should definitely take the blame).

From last night's observation (elog about drift in TT1), I only moved TT1 in pitch and gained back locking in IR for both the arms

Rana and I traced the cables that ran from c1pem1 to the Weather Station monitor.  We found that the flat blue cable that is plugged into c1pem1 was not connected to the black cable from the Weather Station.  We don't know why they are unplugged, but the Weather Station had been inactive since 2010.  Rana plugged them back in (they are now connected via a sketchy connector that had its pins askew) and now the channels are outputting correct data!  Everything else seems to be in good order and now I can use the data from the Weather Station for the summary pages!

To get the code to run on c1pem1, we had to move the old target back into the /cvs/cds/caltech/target/ directory.  It is in /cvs/cds/caltech/target/c1pem1/.

JoeB had apparently moved it into some other area called 'oldfe' and this was why the weather station has not been running for years.    Joe is at LLO now, but he's not beyond our reach...

Once the code had been moved back I started it up. I also rebooted it from the telnet prompt to ensure that it worked on reboot. It did.

The cable issue that Liz mentions probably happened during the PSL table lifting and cable cleanup. It looks like someone yanked the ethernet cable out of its adapter and broke it...

Before heading back to the 40m to check on the computer situation, I thought I'd check the web screenshots page that Kakeru worked on, and it looks like none of the screens have been updated since June 1st.  I don't know what the story is on that one, or how to fix it, but it'd be handy if it were fixed.

 Apparently I broke this when I added op540m to the webstatus page.  It's fixed now.

AuthType Basic  
AuthName "LVC password"
AuthUserFile /cvs/cds/caltech/apache/etc/LVC.auth
Require valid-user

The nodus restart caused a bit of downtime. The apache configuration files were accidentally deleted the other day, so elog/svn/wikis were just holding on in memory; this fact was unfortunately not elogged. 

Things should be up and running again, except for the 8080->8081 elog redirection which I haven't been able to figure out.

I will also set up the NFS backup to include nodus configuration files from now on

Nodus' /export and /etc directories are now being backed up at /cvs/cds/caltech/nodus_backup

They will be rsync'd over as part of the nightly tape backups (scripts/backup/rsync.backup)

Sorry I was writting the elog, but I had to dive into the chamber (@LHO) before completion.

After Rana pointed me to Yoichi's MEDM snapshot script, I learned how to use Xvfb, which is what Yoichi used to write screens without a real screen.  With this I wrote a new cron script, which I added to Mafalda's cron tab to be run once a day at 6am.

The script is called webview_update.cron and is in /opt/rtcds/caltech/c1/scripts/AutoUpdate/.

#!/bin/bash
DISPLAY=:6
export DISPLAY
#Check if Xvfb server is already running
pid=`ps -eaf|grep vfb | grep $DISPLAY | awk '{print $2}'`
if [ $pid ]; then
        echo "Xvfb already running [pid=${pid}]" >/dev/null
else
# Start Xvfb
echo "Starting Xvfb on $DISPLAY"
Xvfb $DISPLAY -screen 0 1600x1200x24 >&/dev/null &
fi
pid=$!
echo $pid > /opt/rtcds/caltech/c1/scripts/AutoUpdate/Xvfb.pid
sleep 3

#Running the matlab process
/cvs/cds/caltech/apps/linux/matlab/bin/matlab -display :6 -logfile /opt/rtcds/caltech/c1/scripts/AutoUpdate/webview.log -r webview_simlink_update

yoichi, pete

We installed the watchLockLoss script in scripts/AutoDTT/.  This script monitors arm power and uses command line
DTT to save 5 s snapshot of the interferometer when it senses loss of lock.  We ran it on linux and it seemed to
save an xml file about half the time; we'll try it on solaris.  

I managed to get up to arm power of about 20 a couple of times.  IFO lost lock a couple of times after turning
off moving zero.  MC2 would often get tripped by lock loss and need resetting.  Maybe we will try to stiffen the
op levs.

[Jenne / Kiwamu]

We spent approximately an hour for the weekly Wednesday cleaning.

This time we moved onto an area where a desk and optics shelf reside along the Y arm.

We will continue cleaning up there in the next time too.

[Jon, Yehonathan, Gautam, Aaron, Shruti, Koji]

We get together on Wednesday afternoon for cleaning the lab. Particularly, we collected e-wastes: VME crates, VME modules, old slow control cables, and other old/broken electronics. They are piled up in the office area and the cage outside rioght now (Attachments 1/2). We asked Liz to come to pick them up (under the coordination with either Gautam or Koji). Eventually this will free up two office desks.

Also, we made the acromag components organized in plastic boxes. (Attachment 3)

[Liz, Gautam, Chub, Jordan, Koji]

We removed a significant amount of e-waste from the lab. The garbage was moved to the e-waste station in WB SB and are waiting for disposal.

This past week I have completed the following tasks:

1. Built a trigger and power box for the camera GC 750M (06058) and took some test images to see whether the trigger box really works. Result: It is doing fine!

2. Went over the setup that is already sitting on the table. Ref: Aidan's elog entry

3. Attended seminars and talks given by Alan, Jahms, Koji and Rana.

4. Attended the mandatory laser safety training by Peter.

Expected task for this week (could be more):

1. Work out analytical expressions of the power of the carrier and sidebands going to the camera in the setup. (As suggested by Rana and Joe)

2. Work on producing beat signal to the camera using the He-Ne laser setup.

3. Move,if possible, to the Nd:YAG setup.

4. Go over the codes and paper by the past SURFers on the phase camera experiment.

I spent the week reading up on filter algorithm theory, particularly Wiener filtering. I have also learned how to get data from specific channels at specific times, and I've been getting myself acquainted with Matlab (which I have not previously used). Finally, I started messing around with the positioning of the accelerometers and seismometers in order to try to find the setup that yields the best filtration.

I've spent most of the last week doing background reading; fourier transforms, shm, e&m, and other physics that I didn't cover at school.  I also read a few chapters in Saulson, especially the chapter on noise and shot noise.  To get a better grip on what I'm going to be doing I read through the polarization chapter in Hobbs' "Optics" text, mostly on wave plates since that's a large part of this readout.  Since then I've been working up to calculating the shot noise, starting with the electric field throughout the new interferometer readout.

I spent the last week working a lot with the differences between a basic Michelson readout and the new one as a displacement sensor.  The new one (w/ wave plates) ends with two differently polarized beams and should have better sensitivity; I've also been going through noise/sensitivity calculations for each, although that hit a road block when I had to start the 1st SURF progress report, which has taken up most of my time since Saturday.

The last week I've spent mostly working on calculating shot noise and other sensitivities in three michelson sensor setups, the standard michelson, the "long range" michelson (with wave plates), and the proposed EUCLID setup.  The goal is to show that there is some inherent advantage to the latter two setups as displacement sensors.  This involved looking into polarization and optics a lot more, so I've been spending a lot of time on that also.  For example, the displacement sensitivity/shot noise on the standard michelson is around 6:805*10^-17 m/rHz at L_=1*10^-7m, as shown in the graph. 

I've spent most of the last week working on finishing up the UCSD calculations, comparing it to the EUCLID design, and thinking about getting started with a prototype and modelling in MATLAB.  Attached is something on EUCLID/UCSD sensors.

Over the past week, I have been focusing on the issues I brought up in my last ELOG,  6956.  I spent quite a while attempting to modify the script and create my own spectrogram function within the existing code.  I also checked out the channels on the PSL table for the PSL health page and produced a spectrogram plot of the PMC reflected, transmitted, and input powers, the PZT Voltage and the laser output power.  When I was entering these channels into the configuration script, I came across an issue with the way the python script parses this.  If there were spaces between the channel names (for example: C1:PSL-PMC_INPUT_DC, C1:PSL-PMC_RFPDDC... etc) the program would not recognize the channels.  I made some alterations to the parsing script such that all white spaces at the beginning and end of the channels were stripped and the program could find them.

 The next thing that I worked on was attempting to see if the microphone channels were actually stopping the program or just taking an extraordinarily long time.  I tried running the program with shorter time samples and that seemed to work quite well!  However, I had to leave it running overnight in order to finish.  I am sure that this difference comes from the fact that the microphone channels are fast channels.  I would like to somehow make it run more quickly, and am thinking about how best to do this.

I finally got my spectrogram function to work after quite a bit of trouble.  There were issues with mismatched data and limit sets that I discovered came from times when only a few frames (one or two) were in one block.  I added some code to  ignore small data blocks  like those and the program works very well now!  It seems like the best way to get the right limits is to let the program automatically set the limits (they are nicely log-scaled and everything) but there are some issues that produce questionable results.  I spent a while adding a colormap option to the script so that the spectrogram colors can be adjusted!  This mostly took so long because, on Monday night, some strange things were happening with the PMC that made the program fail (zeros were being output, which caused an uproar in the logarithmic data limits).  I was incredibly worried about this and thought that I had somehow messed up the script (this happened in the middle of when I was tinkering with the cmap option) so I undid all of my work!  It was only when I realized it was still going on and Masha and Jenne were talking about the PMC issues that I figured out that it was an external issue.  I then went in and set manual limits so that a blank spectrogram and redid everything.

The spectrogram is not operational and the colormap can be customized.  I need to fix the problem with the autoscaled axes (perhaps adding a lower bound?) so that the program does not crash when there is an issue.

Yesterday, I spoke with Rana about what my next step should be.  He advised me to look at ELOGs from Steve (6678) and Koji (6675) about what they wanted to see on the site.  These gave me a good map of what is needed on the site and where I will go next.

I need to find out what is going on with the weather channels and figure out how to calibrate the microphones.  I will also be making sure there are correct units on all of the plots and figure out how to take only a short section of data for the microphone channels.  I have already modified the tab template so that it is similar to Koji's ELOG idea and will be making further changes to the layout of the summary pages themselves.  I will also be working on having the right plots up consistently on the site.

The last week I've started setting up the HeNe laser on the PSL table and doing some basic measurements (Beam waist, etc) with the beam scan, shown on the graph.  Today I moved a few steering mirrors that steve showed me from at table on the NW corner to the PSL table.  The goal setup is shown below, based on the UCSD setup.  Also, I found something that confused me in the EUCLID setup, a  pair of quarter wave plates in the arm of their interferometer, so I've been working out how they organized that to get the results that they did.  I also finished calculating the shot noise levels in the basic and UCSD models, and those are also shown below (at 633nm, 4mw) where the two phase-shifted elements (green/red) are the UCSD outputs, in quadrature (the legend is difficult to read).

Chris,

Some comments:

0. Probably, you are working on the SP table, not on the PSL table.

1. The profile measurement looks very nice.

2. You can simplify the optical layout if you consider the following issues
  A. The matching lenses just after the laser:
      You can make a collimated beam only with a single lens, instead of two.
      Just put a lens of f0 with distance of f0 from the waist. (Just like Geometrical Optics to make a parallel-going beam.)

      Or even you don't need any lens. In this case, whole optical setup should be smaller so that your beam
      can be accomodated by the aperture of your optics. But that's adequately possible.

  B. The steering mirrors after the laser:
      If you have a well elevated beam from the table (3~4 inches), you can omit two steering mirrors.
      If you have a laser beam whose tilte can not be corrected by the laser mount, you can add a mirror to fix it.

  C. The steering mirrors in the arms:
      You don't need the steering mirrors in the arms as all d.o.f. of the Michelson alignment can be adjusted
      by the beamsplitter and the mirror at the reflected arm. Also The arm can be much shorter (5~6 inches?)

  D. The lenses and the mirrors after the PBS:
      You can put one of the lenses before the PBS, instead of two after the lens.
      You can omit the mirror at the reflection side of the PBS as the PBS mount should have alignment adjustment.

The simpler, the faster and the easier to work with!
Cheers.

This week, I made several modifications to the Summary page scripts, made preliminary Microphone BLRMS channels and, with Rana's help, got the Weather Station working again. 

I changed the spectrogram and spectrum options in the Summary Pages so that, given the sampling frequency (which is gathered by the program), the NFFT and overlap are calculated internally.  This is an improvement over user-entered values because it saves the time of having to know the sampling frequency for each desired plot.  In addition, I set up another .sh file that can generate summary pages for any given day.  Although this will probably not be useful in the final site, it is quite helpful now because I can go back and populate the pages.  The current summary pages file is called "c1_summary_page.sh" and the one that is set up to get a specific day is called "liz_c1_summary_page.sh".  I also made a few adjustments to the .css file for the webpage so that plots completely show up (they were getting cut off on the edges before) and are easier to see.  I also figured out that the minute and second trend options weren't working because the channel names have to be modified to CHANNEL.mean, CHANNEL.min and CHANNEL.max.  So that is all in working order now, although I'm not sure if I should just use the mean trends or look at all of them (the plots could get crowded if I choose to do this).  Another modification I made to the python summary page script was adding an option to have an image on one of the pages.  This was useful because I can now put requested MEDM screens up on the site.  The image option can be accessed if, in the configuration file, you use "image-" instead of "data-" for the first word of the section header.

I also added a link to the final summary page website on the 40 meter wiki page (my summary page are currently located in the summary-test pages, but they will be moved over once they are more finalized).  I fleshed out the graphs on the summary pages as well, and have useful plots for the OSEM and OPLEV channels.  Instead of using the STS BLRMS channels, I have decided to use the GUR BLRMS channels that Masha made.  I ELOGged about my progress and asked for any advice or recommendations a few days ago (7012) and it would still be great if everyone could take a look at what I currently have up on the website and tell me what they think!  July 22 and 23 are the most finalized pages thus far, so are probably the best to look at.

https://nodus.ligo.caltech.edu:30889/40m-summary-test/archive_daily/20120723/

This week, I also tried to fix the problems with the Weather Station, which had not been operational since 2010.  All of the channels on the weather station monitor seemed to be producing accurate data except the rain gauge, so I went on the roof of the Machine Shop to see if anything was blatantly wrong with it.  Other than a lot of dust and spiders, it was in working condition.  I plan on going up again to clean it because, in the manual, it is recommended that the rain collector be cleaned every one to two years...  I also cleared the "daily rain" option on the monitor and set all rain-related things to zero.  Rana and I then traced the cabled from c1pem1 to the weather station monitor, and found that thy were disconnected.  In fact, the connector was broken apart and the pins were bent.  After we reconnected them, the weather station was once again operational!  In order to prevent accidental disconnection in the future, it may be wise to secure this connection with cable ties.  It went out of order again briefly on Tuesday, but I reconnected it and now it is in much sturdier shape!

The most recent thing that I have been doing in relation to my project has been making BLRMS channels for the MIC channels.  With Jenne's assistance, I made the channels, compiled and ran the model on c1sus, made filters, and included the channels on the PEM MEDM screen .  I have a few modifications to make and want to .  One issue that I have come across is that the sampling rate for the PEM system is 2 kHz, and the audio frequencies range all the way up to 20 kHz.  Because of this, I am only taking BLRMS data in the 1-1000 Hz range.  This may be problematic because some of these channels may only show noise (For example, 1-3 and 3-10 Hz may be completely useless).

The pictures below are of the main connections in the Weather Station.  This first is the one that Rana and I connected (it is now better connected and looks like a small beige box), located near the beam-splitter chamber, and the second is the c1pem1 rack.  For more information on the subject, there is a convenient wiki page: https://wiki-40m.ligo.caltech.edu/Weather_Station

Over the past week, I have continued refining the summary pages.  They are now online in their final home, and can be easily accessed from the 40 meter Wiki page!  (It can be accessed by the Daily Summary link under "LOGS").  I have one final section to add plots to (the IFO section is currently still only "dummy" plots) but the rest are showing correct data!  I have many edits to make in order for them to be more intelligible, but they are available for browsing if anyone feels so inclined.

I also spent quite a while formatting the pages so that the days are in PDT time instead of UTC time.  This process was quite time consuming and required modifications in several files, but I tracked my changes with git so they are easy to pinpoint.  I also did a bit of css editing and rewriting of a few html generation functions so that the website is more appealing.  (One example of this is that the graphs on each individual summary page are now full sized instead of a third of the size.

This week, I also worked with the BLRMS mic channels I made.  I edited the band pass and low pass filters that I had created last week and made coherence plots of the channels.  I encountered two major issues while doing this.  Firstly, the coherence of the channels decreases dramatically above 40 Hz.  I will look at this more today, but am wondering why it is the case.  If nothing could be done about this, it would render three of my channels ineffective.  The other issue is that the Nyquist frequency is at 1000 Hz, which is the upper limit of my highest frequency channel (300-1000 Hz).  I am not sure if this really affects the channel, but it looks very different from all of the other channels.  I am also wondering whether the channels below 20 Hz are useful at all, or whether they are just showing noise.

The microphone calibration has been something I have been trying to figure out for quite some time, but I recently found a value on the website that makes the EM172 microphones and has a value for their sensitivity.  I determined the transfer factor from this sensitivity as 39.8107 mV/Pa, although I am not sure if all of the mics will be consistent with this.

Over the past week, I have been working on my progress report and finalizing the summary pages.  I have a few more things to address in the pages (such as starting at 6 AM, including spectrograms where necessary and generating plots for the days more than ~a week ago) but they are mostly finalized.  I added all of the existing acoustic and seismic channels so the PEM page is up to date.  The microphone plots include information about the transfer factor that I found on their information sheet (http://www.primomic.com/).  If there are any plots that are missing or need editing, please let me know!

I also modified the c1_summary_page.sh script to run either the daily plots or current updating plots by taking in an argument in the command line.  It can be run ./c1_summary_page.sh 2012/07/27
 or ./c1_summary_page.sh now to generate the current day's pages.  (Essentially, I combined the two scripts I had been running separately.)  I have been commenting my code so it is more easily understandable and have been working on writing a file that explains how to run the code and the main alterations I made.  The most exciting thing that has taken place this week is that  the script went from taking ~6 hours to run to taking less than 5 minutes.  This was done by using minute trends for all of the channels and limiting the spectrum plot data.

The summary pages for each day now contain only the most essential plots that give a good overview of the state of the interferometer and its environment instead of every plot that is created for that day.

I am waiting for Duncan to send me some spectrogram updates he has made that downsample the timeseries data before plotting the spectrogram.  This will make it run much more quickly and introduce a more viable spectrogram option.

Today's Summary Pages can be accessed by the link on the wiki page or at:

https://nodus.ligo.caltech.edu:30889/40m-summary/archive_daily/20120808/

Hey, the pages got significantly nicer than before. I will continue to give you comments if I find anything.

So far: There are many 10^-100 in logarithmic plots. Once they are removed, we should be able to see the seismic excitation during these recent earth quakes?

Incidentally, where the script is located? "./" isn't the absolute path description.

This week, my work fell into two categories: Artificial Neural Networks and lab-related projects.

Artificial Neural Networks

- I played around with radial basis functions and k-means classification algorithms for a bit in order to develop an algorithm to pick out various features of seismic signals. However, I soon realized that k-means is an extremely slow algorithm in practice, and that radial basis functions are thus difficult to implement since their centers are chosen by the k-means algorithm in practice.

- Thus, I moved on to artificial neural networks. Specifically, I chose to implement a sigmoidal neural network, where the activation function of each neuron is f(u) = 1/ (1 + e-u/T), T constant, which is nice because it's bounded in [0, 1]. Classification, then, is achieved by generating a final output vector from the output layer of the form [c1, c2, c3, ..., cN] where N is the number of classes, ci = 1 (ideally) if the input is of class i, and ck = 0 otherwise.

- First, I built a network with randomly generated weights, ten neurons in the one hidden layer, and two output neurons - to simply classify [1, 0] (earthquake) and [0, 1] (not an earthquake). I ran this on fake input I generated myself, and it quickly converged to error 0. Thus, I decided to built a network for real data.

- My current network is a 2-layer, 10 neuron / 2 neuron sigmoidal network that also classified earthquake / not an earthquake. It trains in roughly 80 - 100 iterations (it's learning curve on training data it attached). It decimates full data from DataViewer by a factor of 256 in order to run faster.

- Next steps: currently, my greatest limitation is data - I can use US Geological Survey statistics to classify each earthquake (so that N = 10, rather than 2, for example), but I would like definite training data on people, cars, trucks, etc. for supervised learning, in order to develop those classes. Currently, however, the seismometers are being used for mine and Yaakov's triangulation project, so this may have to wait a few days.

Lab-Related Projects

- I apologize for all of the E-logs, but I changed the filters in the RMS system (to elliptic and butterworth filters) and changed the seismic.strip display file.

- I repositioned the seismometers so that Yaakov and I can triangulate signals and determine seismic noise hot-spots (as a side-project).

Right now I'm going to try for more classes based on USGS statistics, and I will also explore other data sources Den suggested.

Thanks for your help, everybody in 40m!

This week, I continued to work with my Artificial Neural Network. Specifically, I implemented a 3-hidden layer sigmoidal, gradient-descent supervised network, with 3 neurons in the final output layer, since I have introduced a new class, trucks. I have overcome my past data limitation, since I observed that there is a multitude of trucks that comes between 9 and 10 am, and thus I have observed a bunch of trucks after the fact (their seismic patterns are rather distinct, and thus there could prove to be a very large supply of this data - I have gathered data on the past 50 days so far, and have gathered 60+ truck patterns).

With 3 classes, the two-layer network converges in ~200 epochs, while the 3-layer network takes around ~1200 (and more time per iteration). Since the error gradients in the stochastic gradient descent are recursively calculated, the only real time limitation in the algorithm is just lots of multiplication of weight / input vectors, lots of computation of sigmoidal functions, and lots of data I/O (actually, since the sigmoidal function is technically an exponentiation to a decimal power and a division, I would be curious to know if theory or Matlab has any clever ways of computing this faster that can be easily implemented - I will look into this today). Thus, the networks take a long time to train. I'm currently looking at optimizing the number of layers / number of neurons, but this will be a background process for at least several days, if not the next week. In the greater scope of things, however, training time isn't really a problem, since the actual running of the algorithm requires only one pass through the network, and the network should be as well-trained as possible. However, due to the fact that I am only here until the end of August, it would be nice to speed things up.

As far as other classifications, I can simulate signals either by dropping the copper block from the Stacis experiment, or by applying transfer functions to general seismic noise. However, I would like more real data on noise sources, but the only other one plausible to LIGO that I can currently think of (cars don't show up very well) is the LA Metro. Perhaps I will take a day to clock trains as they come in (since the schedule is imprecise) and see if there is any visible seismic pattern.

I also, with the help of Yaakov, Jenne, and Den, now have three working, triangulated seismometers, which can now begin taking triangulation data (the rock tumblers are still working, so there should be opportunities to do this), both to find hot-spots as Rana suggested, and to measure the velocity and test out my algorithm, as Den suggested.

 A lot of my time this week was spent struggling with implementing my neural network code in Simulink in order to experiment with using neural network control systems, as Rana suggested. Perhaps I'm inept at S-Functions, but I decided try to use the Model Reference Controller block in Simulink instead of my own code, and experiment with using that to control a driven, damped harmonic oscillator with noise. The block consists of two neural networks, one which is trained on a model of the plant to simulate the plant, and one that it trained on a reference model (which, in my case, is just input -> gain 0 -> output since my desired signal for now is 0. So far, I have managed to adjust the parameters enough to stop the neural network controller from outputting too much force (this causing the amplitude of the oscillator to increase with each iteration), but outputting enough to keep the plant oscillating with a maximal displacement of 2 m/s (with 30 neurons, 100 reference time delays, and 100 plant output time delays). I will continue to work on this, especially with added noise sources, and see how feasible it is to train such a controller to actually perform well.

As far as classification, I took up that project again, and realized that I have been approaching it the wrong way the whole time - instead of using a neural network to classify an entire time series, I could just classify it online using a recurrent neural network. This, however, meant that my data (which used to be in packets of 900 seconds) had to be parsed through to generate time-dependent desired output vectors. I did this last night, and have been trying various combinations of neuron numbers, time delays, and learning parameters this morning. Below is my current best result for mean square error with time, obtained with 1 neuron in the hidden layer, 50 time delays (so that there are actually 51 neurons feeding into the hidden layer, and a subnetwork of 50 neurons connected to 50 neurons, and learning parameter 0.7). The peak is due to the fact that a large amount of sharp earthquakes occur around that time, essentially giving the neural network a surprise, and causing it to rapidly learn. However, I suspect this sharp rise would decrease if I were to stop decimating my data by a factor of 256, and using all of the inputs as they come in (this, however would be drastically slower). Currently, I have a massive loop running which tries different combinations of neurons and time delays.

In terms of other lab stuff, Jenne and I ordered parts to make a cable for Gurlap-1, and I updated the pin map for Gurlap-1. Also, I wrote my progress report.

The main thing that I did this week was write a C block that, given static weights, would classify seismic signals (into one of three categories - Earthquake, Truck, Quiet). I have successfully debugged the C block so that it works without segmentation faults, etc, and have made various versions - one that uses a recurrent neural network, and one that uses a time-delayed input vector (rather than keeping recurrent weights). I've timed my code, and it works very fast (to the point where clock() differences in <time.h> are 0.000000 seconds per iteration). This is good news, because it means that operations can be performed in real-time, whether we are sampling at 2048 Hz, or, as Rana suggested, 256 Hz (currently, my weights are for 256 Hz, and I can decimate the incoming signal which is at 2048 Hz right now).

In order to optimize my code, since at the core it involves a lot of matrix multiplications, I considered how the data is actually stored in the computer, and attempted to minimize pointer movement. Suppose you have an array in C of the form A[num_row][num_col] - the way this array is actually stored on the stack or heap is row_1 / row_2 / row_3 / ... / row_num_row, so it makes sense to move across a matrix from left to right and then down (as though reading on a page). Likewise, there's no efficient algorithm for matrix multiplication which is less that O(N^2) (I think), so it's essentially impossible to avoid double for loops (however, the way I process the matricies, as mentioned before, minimizes this time).

The code is also fast because, rather than using an actual e^-u operation for the sigmoidal activation function, it uses a parametrized hyperbola - this arithmetic operations are the only ones that occur, and this is much faster than exponentiation (which I believe is just computer by Taylor series in the C math library..)

The weight vectors for this block are static (since they're made from training data where the signal type is already known). I am not currently satisfied with the performance of the block on data read from a file, so I am retraining my network. I realized that what is currently happening is that, given a time-dependent desired output vector, the network first trains to output a "quiet" vector, then a "disturbance" vector, and then retrains again to output a "quiet vector" and completely forgets how to classify disturbance. Currently, I am trying to get around this problem by shifting my earthquake data time-series, so that when I train in batch (on all of my data), there is probably an earthquake at all time points, so that the network does not only train on "quiet" at certain iterations. Likewise, I realized that I should perform several epochs (not just one) on the data - I tried this last night, and training performance MSE decreased by a factor of 1 per iteration (when on average, it's about 40, and 20 at best).

After I input the static weight vectors (which shouldn't take long since my code is very generalized), the C block can be added to the c1pem frame, and a channel can be made for the seismic disturbance class. I've made sure to keep with all of the C block rules when writing my code (both in terms of function input/output, and in terms of not using any C libraries).

As for neural networks for control, I talked to Denis about the controller block, and he realized that we should, instead of adding noises, at first attempt to use a reference plant with a lower Q pendulum and a real plant with a higher Q pendulum (since we want pendulum motion to be damped). I've tried training the controller block several times, but each time so far the plant pendulum has started oscillating greatly. My current guess at fixing this is training more.

Also, Jenne and I made a cable for Guralp 1 (I soldered, she explained how to do it), and it seems to work well, as mentioned in my previous E-log. Hopefully it can be used to permanently keep the seismometer all the way down the arm.

 The Past Week

Attempted to design coupling telescope, turned out waist measurement was still off. Took another waist measurement, this time more reasonable.

Used recent waist measurement to actually design a coupling system to couple NPRO light into Panda PM980 fibers (see recent elog)

The Next Week

Assemble fiber coupling system

Measure coupling efficiency, ensure it's at least 60%

Begin measuring Polarization Extinction ratio

Materials

PLCX lens with f = 0.25m ------> status: here

Fiber Coupled Powermeter//PD ------> status: unknown (have any laying around?)

Quarter Wave Plate, Polarizing Beamsplitter, Photodiodes ------> status: here

other components from original razorblade measurement  setup

 The Past Week

In the past week, I have improved the coupling in the fiber testing setup on the SP table to up to ~45%

I also measured the input/output modes of the fiber with collimators.

Manasa, Q and I have designed, and redesigned a setup to measure Polarization Extinction Ratio introduced by fibers.

I have also partially assembled the box that will hold the frequency counters and RPi for FOL.

Today (Tuesday) I measured waists of PSL and AUX, at dumped light from the SHG's for use in designing coupling telescopes for FOL.

Next Week

In the next week, I will design and couple light from PSL and AUX (Y arm) into fibers for use in testing FOL.

Once that's done, I will continue testing fiber characteristics, starting with Polarization Extinction Ratio.

Items Needed

Power cord for Raspberry Pi (ordered)

AD9.5F collimator adapter (ordered)

 Last Week:

-I continued to struggle with the razorblade beam analysis, though after a sixth round of measurements, and a lot of fiddling around with fit parameters in matlab, there seems to be a light at the end of the tunnel.

Next Week:

-I plan to check my work with the beamscan tomorrow (wednesday) morning

-Further characterize the light from the fibers, and set up the collimator

-Design and hopefully construct the telescope that will focus the beam into the collimator

Materials:

- Razorblade setup or beamscan (preferably beamscan)

- Fiber Illuminator

- Collimator (soon to be ordered)

- Lenses for telescope (TBD)

Last Week

Took first round of PER measurements after a long setup.

Started setting up to take measurement of the other polarization--ran into issues with mounts again. (Spinning of their own free will again.)

Devised a new scheme for taking more robust measurements of PER--still in progress.

Next Week

Finish data analysis of these latest PER measurements

Hopefully finally move on to frequency noise characterization

Materials Needed

None for PER

Unknown for frequency noise

 Attached is the weekly work plan / equipment requirement / lab expert's presence needed for the upcoming week.

I tracked the tendency for ezcaPut to fail and sometimes seg-fault in the camera code to a conflict between the camera API and ezca, either on the 
network level or the thread level.  Since neither are sophisticated enough to provide controls over how they handle these two things, I instead 
separated the call to ezcaPut out into a small, separate script (a stripped down ezcawrite), which the camera code calls at the system level.  This is a 
bit hacky of a solution, but its the only thing that seems to work.

I've developed a transformation based on Euler angles that should be able to take the 4 OSEMs in a picture of the end mirror and use their relative 
positions to determine the angle of the camera to the optic.  This would allow the position data determined by the fitting software to be converted 
from pixels to meaningful lengths, and should aid any servo-ing done on the beams position.  I've yet to actually test if the equations work, though.

The servo code needs to have slew rate limiters and maximums/minimums to protect the mirrors written in to it before it can be tested again, but I 
have no idea what reasonable values for these limits are.

Joe and I recently scanned the PMC by driving C1:PSL-PMC_RAMP with the trianglewave script over a range of -3.5 to -1.25 (around 50 to 150 volts 
to the PZT) and read out C1:PSL-ISS_INMONPD to measure the transmission intensity.  This included slightly under 2 FSRs.  For slow scans (covering 
the range in 150 to 300 s), the peaks were very messy (even with the laser power at 1/6 its normal value), and it was difficult to place where the 
actual peak center occurred.  For faster sans (covering the range in 30 seconds or so), the peaks were very clean and nearly symmetric, but were 
not placed logically (the same peak showed up at two very different values for the PZT voltage in two separate runs).  I don't have time to put 
together graphs of the scans at the moment; I'll have that up sometime this afternoon.