About The SETI@home Screensaver
DATA ANALYSIS
This is where all the action takes place. whereas the other two text panes remain fixed while the data is processed, this section is dynamically updated as your computer works. This section of the screen contains a wealth of information about what your computer is doing at any given moment during the analysis of your work-unit. Keeping a watchfull eye on this pane will help you understand what SETI@home is doing with all the data.
What is the screensaver doing NOW?
The top line tells you what the program is currently doing. It can say any of several things. We'll list them below and explain what each means.
Scanning Result Header File
When the SETI@home kicks in (or when you launch it manually) the screensaver must somehow recollect where it was when it last left off doing its calculations. To find this information it reads a file that we've stored on your hard disk. The screensaver then resumes its work exactly where it left off with all the data on the screen intact.
Connecting To Server
When you see this, the screensaver is trying to contact the SETI@home data server.
Receiving Data
The SETI@home data server is sending you data when you see this. We send you about 350 kbytes of actuall radio telescope data and about another 1k that describes the data (time data was taken, where in the sky, base frequency of this work-unit, etc...) This shouldn't keep your internet connection open very long (less than 4 minutes for even a 28.8 kbaud modem).
Doing Baseline Smoothing
When you receive a new work-unit from the server at Berkeley, there are signals af all kinds mixed in. We are only interested in looking at the narrow bandwidth signals. These narrowband signals are what we believe an alien civilization would use to communicate. On the other hand, broadband signals are most likely due to natural astronomical processes. To reject broadband noise, the screensaver does a sort of "average" through the data that eliminates this broadband noise and brings all the other narrow bandwidth signals down (or up) to a common "baseline" level. Also, over the 107 seconds the signal sometimes gets slowly louder and/or softer. Baseline smoothing brings it all to the same level. This is the first thing that is done to the data after you've received your work-unit and it's usually only done once. Certain clients (like the Mac client) do not keep the smoothed data in RAM and must re-compute this whenever the screensaver is started. A progress bar appears to the right lets you know how far your computer has gotten through this process.
Computing Fast Fourier Transform
This is where all the work gets done. The data supplied to you from the telescope is a signal that varies with time - like a line on an oscilloscope that wiggles up and down in response to your voice through an attached microphone. In this case, time runs along the horizontal x-axis and signal strength (air pressure) along the vertical y-axis. The raw radio telescope signal is not very useful to us. What we would like to see is if there are any constant (and loud) "tones" within the signal. We would rather be looking at a graph with frequecy running along the horizontal x-axis, and power along the vertical y-axis. Any spike in this graph would be a loud signal at a single frequency.To turn a set of time-based data into a set of frequency-based data, we apply a relatively complex mathematical operation called a "fast fourier transform" or FFT. For more information on the FFT, please consult a book on digital signal processing.
The result of this processing is the graph produced in the lower frame of the screen saver. You may notice a few interesting things about the FFT. At the beginning of a work-unit, we do 15 different FFT's, each looking at the data with varying accuracy. We start looking for details as small as .07 Hz wide. There are tradeoffs when you are doing this kind of analysis. If you want to be very accurate in frequency, you have to observe the data for a longer time. You will notice that at the 0.075 Hz frequency resolution, we must look at chunks of data 13.42 seconds in length. To completely analyze our 107 second sample, we need to do 8 of these FFT's. When we reduce the frequency resolution to 0.14 Hz we only have to look at a 6.7 second sample of data. We now have less frequency resolution, but we have more time resolution. We have to look at twice the number of these (16 of them) to cover our 107 seconds of data! We look at 15 different frequency resolutions (0.075, 0.15, 0.3, 0.6, 1.2, 2.5, 5, 10, 20, 40, 75, 150, 300, 600, and 1200 Hz) in our analysis. With each halving of the frequency resolution we must do twice the number of FFT's to cover our 107 seconds of data. The amount of number crunching is dizzying!
Again, the progress bar that appears to the right lets you know how far your computer has gotten through each set of FFT's. You can also watch the FFT's accumulate in the graph in the bottom section.
Chirping Data
It's quite unlikely that an alien planet will be at rest with respect to our Earth. You may remember that humankind is whizzing along on a rotating planet which is revolving around the Sun, which itself is orbiting the center of our Milky Way galaxy. We can assume that our extra-terrestrial friends are likewise situated.There is an interesting effect that all this motion will have on a signal emitted from a moving source and/or received on a moving planet. This is the doppler effect. You are undoubtably familiar with this if you've heard a car honking its horn as it passes you. The frequency, or pitch, of the sound changes as the car passes. You can go out and try this yourself. Stand at the side of the road and listen as a friend drives by with the horn blasting. You could also drive by a stationary car honking its horn and you will also hear the pitch change. It's the relative velocity that's important.
Although our remote friends aren't honking their horns at us, they are sending waves (electromagnetic waves) at us. Their signal will be distorted by the mutual motions of our two systems in much the same way that the car horns are distorted. To disentangle this the SETI@home screensaver analyzes the data many times over trying a great variety of possible doppler accelerations. Actually, the screensaver first takes the raw data and mathematically "undoes" a specific doppler acceleration or "chirp". It then feeds the resulting "de-accelerated" data to the FFT (Fast Fourier Transform) routines. This is called "De-chirping" the data. SETI@home tries to do this at many points between -50 Hz/sec to +50 Hz/sec. At the finest frequency resolution of 0.075 Hz we check for 5409 different chirp rates between -10 Hz/sec and +10 Hz/sec!
Searching For Gaussians
As explained briefly in the FFT section, when the frequency resolution is coarser, the time resolution is finer. When the time resolution is high enough, we can look at the data to see if signals get stronger and weaker in the 12 seconds it takes them to pass through the telescope's field of view. This is an excellent test to tell if the signal is "out there" rather than a simple source of interference somewhere here on the earth. An earthbound signal would not get louder and then softer over the 12 second period we are interested in. This curve fitting checks to see if the signal gets louder and softer over this 12 second period. The test is only applied for frequency resolutions greater than or equal to 0.59 Hz.Since we are looking for these 12 second "gaussians", your 107 seconds of data overlaps the previous and next blocks of data by 15 seconds. In this way we make sure that we don't miss an important signal by splitting it in the middle.
Searching For Pulses / Triplets
A new feature in the SETI@home client looks for repeated pulses in the radio signals. Our alien neighbors may not be sending out a nice even tone for us to detect. They may be sending a series is closely or widely spaced set of pulses. This is much more economical powerwise at their transmitter if they are doing it intentionally (and who knows what ther are sending if it's unintentional!) For all frequency resolutions greater than or equal to .59 Hz, the screensaver will do a search for repeating pulses and triplets of pulses.That finishes the first line of the Data Analysis panel. Phew!
Doppler Drift Rate
The second line of the Data Analysis panel contains the current "Doppler drift rate". The first tests that are made on the data assumes a drift rate of 0 Hz/sec. These unaccelerated signals are more likely to be sources of radio frequency interference (RFI) from earth based terrestrial transmitters. Between the drift rates of -10 Hz/sec and +10 Hz/sec we try all 15 frequency resolutions and increment the doppler drift rate by 0.002 Hz/sec between FFT's. Between +-10 and +-50 we increment by 0.296 Hz/sec.
Frequency Resolution
The second line also tells us the current frequency resolution (bandwidth) we are using in our calculations. You will notice that most of the time we will be calculating FFT's with a frequency resolution of 0.075 Hz. Every 4 FFT's we'll do one with a frequency resolution of 0.14 Hz. Every 16 FFT's we'll do one with a frequency resolution of 0.29 Hz. Every 64 FFT's... well, you get the idea. Remember that there are 15 different frequency resolutions (0.075, 0.15, 0.3, 0.6, 1.2, 2.5, 5, 10, 20, 40, 75, 150, 300, 600, and 1200 Hz). We drop the two finest frequency resolutions (0.075 Hz and 0.15 Hz) when the doppler drift rate is greater than 10 Hz/sec or less than -10 Hz/sec.
Analysis Results
The next part of the data analysis panel displays intermediate results about the best gaussian, pulse and triplet found so far. This part of the panel alternates between all three, but only when a significant result is available. For instance, if there are no significant triplets, you will not see triplets displayed.
Best Gaussian
If a signal is above the average noise and also gets stronger and then weaker in a "gaussian" fashion as the object passes through the telescope beam, we're interested!The number labeled "power" tells us how strong the signal is relative to the baseline power calculated above. The number labeled "fit" is a measure of how well the rising and falling signal fits an ideal gaussian (bell curve) profile. A lower "fit" number means a better fit. (It's actually a chi-square fit, ie. how far the data departs from an ideal gaussian.) Even if you see a strong peak and a low fit number, do not call the press or announce to the world that you have discovered the aliens. Any strong signal must be verified (several ways) to rule out sources of radio-frequency-interference (RFI) before it becomes "official". Since noise can sometimes randomly simulate a gaussian, we've set a threshold to avoid being overwhelmed with trivial results. If the signals are stronger than 3.2 times the average noise level that have a fit better (less than) 10, they are returned by the screensaver client to our server in Berkeley.
The graph below the power and fit numbers displays the curve fitting analysis as it is happening and also displays the best gaussian so far for this work unit. Note: If the telescope is slewing across the sky too slowly or too quickly during the observation, no graph is drawn.
The red line shows the actual data - power at a given frequency, as seen over time. This view is a back to front slice of the big chart at the bottom of your screensaver display. This aspect of the graph changes each time the gaussian fitting moves to a new frequency. The white line shows best fit gaussian for that data, i.e. what your client is actually calculating!
At each data point we try a new fit. You see this as the white line changing very quickly. If the analysis were not happening so quickly you would see the gaussian (the bump in the white line) move from left to right across the graph as we try to best fit your data.
Best Pulse
In order to look for a series of weak repeated pulses, the SETI@home screensaver applies a special test called a "fast folding algorithm." If the routine finds a set of repeating pulses, it will display them with statistics describing what it found.
The number labeled "power" tells us how strong the pulses are relative to the baseline power calculated above. The number labeled "period" is a measure of how far apart the pulses are in seconds. Because both RFI and random noise can simulate a pulsed signal, we've set a threshold here, too. This threshold is calculated dynamically and depends upon the period and the number of times the data has been folded. (For you math nerds, it involves inverting a function known as the "incomplete gamma function".) The score value for a pulse is the ratio of the pulse amplitude to this threshold value. A pulse with a score of greater than 1 will be reported when your screensaver client returns a result to Berkeley.
The graph below the power, period and score numbers displays the pulse analysis as it is happening and also displays the best pulse so far for this work unit. Note: If there are no significant pulses found, no graph is drawn.
Like the gaussian above, red line shows the actual data - power at a given frequency, as seen over time. Unlike the gaussian, this graph probably won't cover the entire 107 seconds of data, but rather will cover two periods of the pulse (twice the "period in the line above the graph...). You should see two spikes sticking up out of the noise. The right-hand and left-hand sides of the graph are the same. Showing two periods makes it easier for you to see the pulses.
For a more technical description of the SETI@home data analysis see the The SETI@home Sky Survey paper.
Best Triplet
The SETI@home client does one more test for pulses. This one looks for three equally spaced pulses. To do this, the screensaver looks at every pair of pulses that are above a certain threshold power. The client then looks for a pulse precisely between the two pulses. If one is found, it is logged and sent back to Berkeley.
If a triplet is found a line is displayed showing the power of the pulses (relative to the noise baseline) and the time between pulses (the period) in seconds.
The graph below the power and period can display the best triplet found so far for this work unit. The three pulses will be marked with short yellow tick marks. Note: If there are no significant pulses found, no graph is drawn.