Data Analysis of Watch Rate Data

Discussion in 'Watch Repair' started by bbodnyk, Feb 20, 2011.

  1. This site uses cookies. By continuing to use this site, you are agreeing to our use of cookies. Learn More.
  1. bbodnyk

    bbodnyk Registered User

    Aug 14, 2009
    491
    7
    16
    Harrisburg, PA
    Country Flag:
    Region Flag:
    Several months ago I purchased a Microset Watch Timer with accompanying software which I've been using to rate my watches. The watch data recorded by the timer can be written to a file and then imported into Excel which I've been using to analyze the data.

    Attached are a number of images showing the results for an Elgin grade 367 which I have found to be what I consider to be highly accurate; I wore it for a week and it lost about 5 seconds. For the purposes of this test I let the timer run overnight averaging 30 ticks; capturing a rate every 6 seconds.

    The first image is a screen capture of the Microset Window showing about 7 hours of the data. The average rate the Microset software shows is 0.8 sec/month.

    The second image shows the results of using the Descriptive Statistics function in Excel. Note the average rate is 0.199999938 which corresponds to a gain of 5.5sec/month while the Microset software reported a rate of 0.8 sec/month.

    The third image is simply a plot of the data and also shows fitting a line and 6 order polynomial to the data. Note that the yellow line has a slight upward slope which implies the watches rate is decreasing over time which is what I experienced wearing the watch for a week.

    The last image shows the result of a discrete fast fourier transform of the data which shows the major frequencies in the data. Note that the major ticks in this plot are at frequencies of 300, 45, 30, 20 & 15.

    From a diagnostic point of view this implies that there is something happening to the watch every 300, 45, etc. seconds.

    I am thinking that this kind of analysis can be quite useful when interpreting the results of watch rate data as the frequencies highlighted in the FFT should correspond to rates that the wheels turn and indicate possibly where problems are.

    I'm considering writing an article for the Bulletin and would appreciate any comments, thoughts, etc.

    Regards,
    Bruce 84856.jpg 84857.jpg 84858.jpg 84859.jpg
     
  2. sderek

    sderek Registered User
    NAWCC Member

    Jan 8, 2009
    2,681
    3
    38
    Watch repair/ Watch sales
    USA
    Country Flag:
    Hi Bruce,
    I love my Microset. I like it mostly for it's versatility. It can be used to time mechanical watches and accutrons. It can be also used to check the integrated circuit on quartz watches. It will also show the balance wheel amplitude. And it's American made from a small business.
    I'd love to see an article outlining it's uses.
     
  3. le arsi

    le arsi Registered User

    Jul 7, 2008
    461
    2
    18
    Watchmaker,tower clockmaker, watch repair teacher/
    Philippines
    Country Flag:
    "It can be also used to check the integrated circuit on quartz watches."

    Can you please elaborate how microset can be used to check the integrated circuit of a quartz watch? I am just confused.
     
  4. Hayesbd

    Hayesbd Registered User

    Feb 13, 2008
    69
    1
    0
    Engineer
    Newark, OH USA
    Country Flag:
    Region Flag:
    Bruce,

    It's been a while since I used Excel for an FFT but I seem to remember it was a bit limited by using a small data set. I have not use the "Descriptive Statistics" function you mention - I'll have to check that out. For watch tick analysis, I simply use a sound analysis program (the one for my PC has built-in FFT, although I don't use if for watches).

    I'm not quite sure what you are gaining with a polynomial curve fit of the plot of the rate data. The trend you mention is a change to the rate (something you cannot detect from carrying the watch and occasionally checking its gain/loss to a time standard). Isn't this just showing the isochronism as the spring winds down over 8.5 hours? It seems remarkably stable, by the way.

    I believe the 15/20/30/45 Hz spikes are simply harmonics of the basic 2.5 Hz oscillator in the system (your hairspring). The 300 Hz spike may be the natural frequency of the pallet fork or some other component in the escapment. Wouldn't any defects in the train show up at very low frequencies (<2.5 Hz)?

    Interesting stuff.

    Brian
     
  5. bbodnyk

    bbodnyk Registered User

    Aug 14, 2009
    491
    7
    16
    Harrisburg, PA
    Country Flag:
    Region Flag:
    Brian,

    Excel has a maximum limitation of 4096 data points for the fft. I do believe I may have gotten my frequency incorrect in the fft plot. I'll post a correct one tomorrow.

    The polynomial curve doesn't necessarily tell anything except the rate changes over time. I'm currently testing the same watch over a 24 hr period to see the change in rate as the spring winds down.

    I'm not at all sure how useful this is but certainly careful analysis of the data has to be done to get anything useful out of it. I just downloaded a free 21 day evaluation of a signal analysis program called Sigview. I'm using it to make sure the Excel fft I'm doing is correct.

    I do believe that the lower frequencies would be the most useful. I just have to figure out what the best capture rate over what period of time to use to get them. I must admit I'm sort of flying in the dark here. Signal analysis is not my forte so I'm learning as I'm going.

    Regards,
    Bruce
     
  6. sderek

    sderek Registered User
    NAWCC Member

    Jan 8, 2009
    2,681
    3
    38
    Watch repair/ Watch sales
    USA
    Country Flag:
    I place the quartz watch on the accutron sensor. If the red led "pulses" the ic is good.
     
  7. le arsi

    le arsi Registered User

    Jul 7, 2008
    461
    2
    18
    Watchmaker,tower clockmaker, watch repair teacher/
    Philippines
    Country Flag:
    "I place the quartz watch on the accutron sensor. If the red led "pulses" the ic is good."

    Ah yes, it's the pulses from the motor and not from the IC. So you are indirectly testing the IC and the actual signal is coming from the step motor making the red LED blinks. The accutron sensor will not detect the square wave signals coming from the IC. This is only to clarify things. Thank you.
     
  8. bbodnyk

    bbodnyk Registered User

    Aug 14, 2009
    491
    7
    16
    Harrisburg, PA
    Country Flag:
    Region Flag:
    Here is another set of watch rate data. This time I averaged 60 ticks together giving a capture rate of 1 reading every 12 seconds. I ran it for over 24 hours to get a feeling of the rate change as the mainspring winds down. You can clearly see as time passes, the rate becomes worse and worse. However the average rate comes out to -4.2sec/week.

    The second image is a fft of the data while the third is a magnification of the beginning of the fft results. By increasing the capture rate to 12 seconds I'm hoping to pick up lower frequencies which might indicate problems in the train. The numbers on the third image are the reciprocal of the frequency which is the number of seconds/cycle the peaks represent.

    Again whether this is meaningful is yet to be determined but it's been fun playing around with the data.

    Bruce 85127.jpg 85128.jpg 85129.jpg
     
  9. Hayesbd

    Hayesbd Registered User

    Feb 13, 2008
    69
    1
    0
    Engineer
    Newark, OH USA
    Country Flag:
    Region Flag:
    Bruce,

    I'm not an expert on this kind of analysis, but if I understand correctly, with a capture rate of 1 sample every 12 seconds (1/12 Hz), you are only going to be able to pick out frequencies up to about half that or .0417 Hz. Your first FFT looks like it indeed has an abscissa that goes up to around this frequency. I don't know how you determine that the peaks are such high frequencies.

    Also, if you are sampling the rate and performing the FFT on this, I'm not sure what useful information it would be giving you. In order for the frequency spectrum to be meaningful, I think you would want to sample the raw tick signal as a function of time and do this at a sample rate of twice the expected frequencies. For example, if you want to pick up the 2.5 Hz oscillation of the balance wheel, you would need a sampling rate of at least 5Hz or 5 samples per second.

    It may be easier to leave things in the time domain and simply analyze the raw trace of sound level (voltage) vs. time to determine faults. The shape of these traces has been well documented, along with typical faulty traces and their potential causes.

    And of course plotting the rate over a 24-hour time period is a good measure of the isochronism of the watch and yours is really very good.

    Brian
     
  10. Hayesbd

    Hayesbd Registered User

    Feb 13, 2008
    69
    1
    0
    Engineer
    Newark, OH USA
    Country Flag:
    Region Flag:
    Bruce,

    Here's and interesting read: http://nvl.nist.gov/pub/nistpubs/jres/045/3/V45.N03.A05.pdf

    This is essentially what you are doing when collecting the rate data over a long period of time. You could take the rate data (which is the slope of the curves in Figure 7) and integrate them to produce an isochronism curve as shown in Figure 7. Note that the isochronism curve does not plot the rate, but the actual deviation from the correct time from the start. This will show how a watch deviates from the correct time as the spring winds down. Figure 8 is rather interesting as an example of a bad job of watch repair! I guesstimated the initial slopes for the various positions and this watch would probably have shown the following initial timing:

    Pend right: -27 sec/day
    Pend left: +2
    Pend down: -25
    Pend up: 0
    Dial up: -7
    Dial down: 0

    Seeing how the rates (slopes) of Figure 8's vertical positions change over time illustrates a balance probably out of poise.

    Regards,
    Brian
     
  11. bbodnyk

    bbodnyk Registered User

    Aug 14, 2009
    491
    7
    16
    Harrisburg, PA
    Country Flag:
    Region Flag:
    Brian,

    Thanks! I'll have to print the pdf out. Attached is a spread sheet where I've calculated the frequencies of the different wheel rotations. I believe I've got them right.

    I also show some different capture rates of the Microset timer and what the corresponding capture frequency and resulting max fft frequence for that data capture rate.

    Bruce 85279.jpg
     
  12. Tinker Dwight

    Tinker Dwight Registered User

    Oct 11, 2010
    13,666
    65
    0
    Calif. USA
    Hi
    I don't think I made the right impression on what
    I was saying earlier.
    Your method of sub-sampling will fold higher frequency
    elements into your spectrum, causing spurious peaks
    that are not useful and misleading.
    In order to look at low frequency things, you can not just
    drop samples ( decimating ) without first filtering.
    You need to first sample every tick and then filter it
    with a brick wall type filter before decimating to a lower
    sample rate.
    Tinker Dwight
     
  13. bbodnyk

    bbodnyk Registered User

    Aug 14, 2009
    491
    7
    16
    Harrisburg, PA
    Country Flag:
    Region Flag:
    The attached is a plot of rate vs time for an Elgin 466 I have capturing a rate every 2 seconds for about 1/2 hour. Every 5 minutes or so the rate changes abruptly. For this particular plot the cannon pinon was removed so nothing from the dial side would be affecting the rate.

    I am trying to figure out what would be happening every 5 minutes inside the watch to cause rate changes like this. In analyzing a number of watches I have seen this 300 second pattern occur before.

    Any ideas on what combination of events could cause a cycle of 300 seconds?

    Thanks!
    Bruce 86105.jpg
     
  14. John Runciman

    John Runciman Registered User
    NAWCC Fellow NAWCC Member

    Aug 13, 2003
    257
    8
    18
    Seattle, WA
    Country Flag:
    Region Flag:
    Bruce

    My normal procedure when I'm servicing a watch is to place it on the timing machine before I clean it to look for problems and to compared to the after it's cleaned. The watch that I was servicing is a 16 size 21 jewel Burlington. On my Witschi Watch Expert II the before showed a slow variation in timekeeping. Sometimes you see this when the watch is really gummy and especially if the mainspring is really gummy. After cleaning I had the exact same problem and changing the mainspring didn't solve the problem.

    Changing to a different timing machine I ran a 30 min. plot that looks very similar to the image that you have. Repeatable patterns like this can usually be traced to power fluctuations in the gear train someplace. Because watches are influenced by isochronal errors power variations in the gear train will cause amplitude changes which change the rate of the watch.

    I find really old timing machines came with really wonderful manuals. The second image came from a timing machine manual showing the effect the gear train has on timekeeping. The left-hand column shows the time in seconds for the passage of a pair of teeth in each mesh; the right-hand column shows the revolution times the individual components. Note each watch will be a little bit different as there very likely are a different number of teeth in the gears.

    In this watch the problem can be traced to the tooth passage of the mainspring barrel with the center wheel pinion. One of the things that bothered me when I was assembling the watch was the mainspring barrel didn't seem right. But visually the watch look like it was running fine it's just didn't look fine on the timing machine. Fortunately with the Internet I could find a good picture of another Burlington and the mainspring barrel definitely was not right. Replacing the mainspring barrel with the correct mainspring barrel now produces a reasonably straight 30 min. plot and on the timing machine is now showing consistent rate with no variation.

    John

    View attachment 4847

    View attachment 4848
     
  15. bbodnyk

    bbodnyk Registered User

    Aug 14, 2009
    491
    7
    16
    Harrisburg, PA
    Country Flag:
    Region Flag:
    John,

    Thanks!

    Your little diagram got me thinking about the rate of each tooth passage rather than the rate an entire wheel rotates. Attached is a new chart showing the frequency of the wheel rotations and teeth passage.

    Now some of the frequencies that show up in the fast fourier transform analysis of my watch data start to make sense. In fact all of the frequencies listed tend to show up to one extent or another. Now I need to understand harmonics and their effect the ultimate rate.

    I may not be getting anywhere but all this is making me think! ;)

    Regards,
    Bruce 86424.jpg
     
  16. bbodnyk

    bbodnyk Registered User

    Aug 14, 2009
    491
    7
    16
    Harrisburg, PA
    Country Flag:
    Region Flag:
    I thought I'd post some new plots of my watch rate data analysis project. Again I'm not entirely sure how useful this is but it is giving me a good understanding of what can affect the rate of a watch.

    I've attached three images showing results for three different watches. Each shows 3 rate plots taken at 12, 6 and 2 second capture intervals. To the right is the corresponding fast fourier transform of rate data. The peaks correspond to frequencies which are the most prevalent in the rate data. The inverse of the frequency is seconds per cycle.

    The first for an Elgin 466 shows patterns in the rate occuring at different frequencies with 0.0033 & 0.166 hz or 300 sec & 6 sec per cycle being the most dominant and the 300 sec pattern being the major one.

    The second image for a Waltham is almost exact to the Elgin 466. The third for an Elgin 181 displays the same frequency patterns but here the 6 sec pattern is the most dominant.

    Regards,
    Bruce 87346.jpg 87347.jpg 87348.jpg
     
  17. DaveyG

    DaveyG Registered User

    Mar 21, 2005
    2,460
    104
    63
    Male
    Retired military aircraft engineer
    North Wales, UK
    Country Flag:
    I don't think you're the only one who has stimulated grey matter Bruce! I'm an old timer and some of the analysis tools leave (FFT?) me behind but the core principal is probably something that many of us had in the back of our minds yet were unable to work out just how to apply it.

    Fascinanting stuff, thank you.

    daveyG
     
  18. bbodnyk

    bbodnyk Registered User

    Aug 14, 2009
    491
    7
    16
    Harrisburg, PA
    Country Flag:
    Region Flag:
    I believe there is a saying of some sort that goes, "if you don't want to see the dirt, don't look under the covers." This set of data kind of shows this.

    The first image is data from an Elgin 453 run for 38 hours along with the fast fourier transform of the entire data set. The larger plot shows the results of doing a fft on sets of data through time. Note that the larger fft plot is fairly consistent across the entire 38 hour period which implies to me the watch is running extremely consistantly across the entire 38 hours until the mainspring winds down.

    The second image shows the results for a Keystone-Howard Series 11 run for 28 hours. Note how initially the larger fft plot starts off being relatively "clean" then gets kind of nasty, gets clean again and then finally goes crazy. One of the most distinct peaks occuring at 0.033 hz disappears at some point and then comes back.

    Until I started looking deeply into this rate data I had always assumed my Howard was one of my better running watches and is one I tend to carry quite a bit. Comparing the two however, the Elgin 453 is clearly a better performing timepiece.

    (I'll still keep wearing my Keystone-Howard!)

    Regards,
    Bruce 87484.jpg 87486.jpg
     
  19. Tinker Dwight

    Tinker Dwight Registered User

    Oct 11, 2010
    13,666
    65
    0
    Calif. USA
    Hi
    I'd suspect it is more a problem of poor contact
    with the pickup. A change there may make the difference.
    It looks to be mostly background noise.
    Tinker Dwight
     
  20. bbodnyk

    bbodnyk Registered User

    Aug 14, 2009
    491
    7
    16
    Harrisburg, PA
    Country Flag:
    Region Flag:
    Tinker,

    The pickup I use is an alligator clip which I usually fasten on the bow. I ran the Howard again for over 8 hours and got similar results of the major frequencies changing over time. So I don't believe it would be a problem with the pickup.

    Bruce
     
  21. le arsi

    le arsi Registered User

    Jul 7, 2008
    461
    2
    18
    Watchmaker,tower clockmaker, watch repair teacher/
    Philippines
    Country Flag:
    Thats it! Fastening the alligator clip to the bow is not reliable! It must be clipped to the movement itself. If I am using any external part of the watch as the contact point then I am ready to receive unwanted pulse and analyze it quick.
     
  22. bbodnyk

    bbodnyk Registered User

    Aug 14, 2009
    491
    7
    16
    Harrisburg, PA
    Country Flag:
    Region Flag:
    I was going to start a new thread called "Why you want to clean a watch" but decided to continue this one instead. I recently cleaned a Waltham Vanguard 23j watch. I recorded the watch rate using my Microset timer before I cleaned it and afterwards.

    The four images below show the before and after results. The first is a plot of the watch rate over about 3 hours. Note the standard deviation is about 3 1/3 sec. which corrollates to the variation in rate. The second image shows the results of a fast fourier transform analysis. Compared to the 4th image there are no frequencies that standout over the others.

    The third image is a plot of the rate after cleaning. Note the standard deviation is just over 1/2 sec. Also the fft plot , the 4th image is much cleaner and you start seeing frequencies you might expect to see. These were there before they were just being masked by the noise.

    Amazing what a simple cleaning will do!

    Regards,
    Bruce 90162.jpg 90163.jpg 90164.jpg 90165.jpg
     
  23. bbodnyk

    bbodnyk Registered User

    Aug 14, 2009
    491
    7
    16
    Harrisburg, PA
    Country Flag:
    Region Flag:
    In case anyone with Microset timer data would care to perform a frequancy analysis similar to the types I've posted throughout this thread, I've added the capability to my Time Lord Watch Database to upload a Microset timer file on the fly and have the system generate several plots; one of the rate data and the second a plot of a fast fourier transform which highlights the major frequencies in the rate data.

    I've attached several screen shots. The first shows the form to specify the Microset data file you want analysed. The make, serial # and comment are optional but they will show up on the pdf file so it can be useful to enter them.

    The second and third are screen shots of the results. A slider underneath the rate plot allows you to zoom into a portion of the data and re-analyse that portion.

    The fourth image is a screen capture of the pdf that is produced. This is a 8 1/2" x 11" document so will print nicely for record keeping purposes.

    Note: The Microset data file should be generated with the Microset computer software showing seconds per beat. If you record beat per hour it will plot but give you some strange rates like gaining a year a day. The filename of the Microset file should have a "mst" extension. Lastly, avoid attempting to upload data sets with 10,000 points or more. The system takes too long to process that much data and will timeout. If you stick with a data capture less then that you shouldn't have any problems.

    If you want to try it go to www.timelordwatches.net and login with a name and password of "nawcc". This does NOT mean your NAWCC password, this means the text string "nawcc" for both. Once you are past the login screen select the "Tools" button at the top of the screen and then the "Quick Rate Analysis". This will bring up the page shown in the first screen shot. From there you should be able to figure it out. Once you select the "Upload & Analyse" button the system will "go away" for a while so wait a bit before deciding nothing is happening. The system is doing a lot of processing and it can take awhile. If you get a "timeout" message than either try a smaller data set or try a different time of day when the server is o as loaded.

    If you don't happen to have a data file there is a "Plot Watch Rate" button in the "Tools" section that would allow you to examine several sets stored with watches in the database.

    (While you are there feel free to check out the web site.)

    I haven't tested it with rate data from a clock. I suppose the limitation is I only have a fixed set of beats per hour in the Beat per Hour selection list and if a particular clock has a unique one it would show the incorrect results for the average and standard deviation. Interestingly enough the fft results would be correct.

    Currently the software only supports Microset time data files because that's what I have personally.

    At the bottom of each web page is a "Questions & Comments" link. If you find this useful send me a message and let me know! Or you can respond here!

    Regards,
    Bruce 90990.jpg 90991.jpg 90992.jpg 90993.jpg
     

Share This Page