On 4/28/2013 1:44 PM, Mary wrote:
> dpb <none@non.net> wrote in message <kljore$g4v$1@speranza.aioe.org>...
>> On 4/28/2013 11:41 AM, Mary wrote:
>> ...
>>
>> >> W/O far more actual knowledge about the experiment and the data I have
>> >> know idea about whether it's "good" or not, but what about just
>> >> standaradized all of the time series--ie, subtract mean, divide by
>> >> estimated st dev.
>> >>
>> >> And, I'd ask why you can't do something similar over the duration--or
>> >> use a normalized accumulative time over a duration length as a
>> >> slightly different presentation of same idea...
>> >>
>> ...
>>
>> > The data is good it's just difficult to line up over time duration. The
>> > data is the change in angle from the elbow to shoulder originating from
>> > motion capture data of a person raising and lowering their arm. (each
>> > set ends up looking similar to a bell curve (the angle increases then
>> > decreases), but because each person is different, and because there
>> were
>> > no triggers added to the motion capture software, each data set has a
>> > bell curve starting and ending at different times.
>> >
>> > I would like to compare these angles across sets of data- but I'm
>> having
>> > trouble figuring out the best way to scale them so that I can take an
>> > average of all of them to draw conlcusions from.
>> >
>> > Example: Trial A's set has the person moving starting at 3.76 seconds
>> > and ending at 9.87 seconds Trial B's set has the person starting at
>> 2.39
>> > seconds and ending at 8.342 seconds
>> > I would like a way to normalize set A and B so that I can average them.
>> > Then use the std deviations between the sets to create an error bar.
>>
>> I wasn't talking about the data; I was talking about whether the
>> suggested normalization method was meaningful or not...
>>
>> But again it seems essentially throwing darts at a wall blindfolded to
>> no so little about a research project and expect to make meaningful
>> assessments of how the data should be analyzed.
>>
>> These are the questions that should have been asked/addressed _before_
>> the data were collected, obviously, of how it was to be processed in
>> order to draw whatever conclusions it was that were intended to be
>> made from the experiments.
>>
>> At this point, having apparently collected a brown paper bag of
>> happenstance data, it appears you're trying to make something of the
>> hash.
>>
>> I'd guess one should be able to fit some sort of general model to each
>> channel of measured positions vs time. Start then by generalizing that
>> to include another term that is the time shift and include it in the
>> model as another term to estimate.
>>
>> It's not clear at all what is important wrt to the time variable
>> here--the comparison of the position at a given time since the
>> beginning of the movement or what? Or is time important at all and
>> what angle is it that is being measured? There's just too little
>> background for somebody here to be able to do anything of meaning
>> methinks. Talk to your thesis advisor and see if can get some help out
>> of the morass you've created...
>>
>> --
>
> Hi thank you for the help and commentary- it's much appreciated.
>
> As for the data collected - this collection type is extremely normal,
> and was a very well thought out process. The goal of finding an average
> change in angle from individual over multiple trials is extremely
> helpful in biomechanics research and not something that I am attempting
> to cobble together as evidence for a made-up thesis.
> Despite the fact that I am still struggling with my normalization, my
> current code processes motion capture data (x,y,z coordinates per
> marker) and changes it into vector of angle between two points over time
> ( a single vector derived from mechanics). It then marks in each vector
> the start and stop time of the movement with triggers.
>
> Be it biomechanics, electronics, or mechanical engineering, real time
> data should be something that needs to be normalized across trials for a
> true average, and I don't believe I'm asking anything far out of the
> ordinary by wanting to equalize vector lengths
Well, I'd say if you new ahead of time what you were after there should
have been a methodology in place to know how you would process the data.
I've done a tremendous amount of vibration and other motion testing and
indeed, multiple channels are time-correlated. _BUT_, it's done a
priori by setting trigger levels for asynchronous events or by a
reference channel trigger if delay, etc., is important, not by trying to
do something w/ just sampled data ex post facto.
Again, you're arguing for the validity of your setup rather than
providing any additional help on what it is you're actually trying to
infer that could help anybody else in trying to provide ideas.
What's wrong w/
a) using the previously suggested approach, or
b) as suggested in the previous postings which never saw any response to
on finding the starting times, apply a trigger level to each channel and
use that to align them.
Again, you've not described what it is you mean by normalization
precisely nor what you mean by equalizing vector lengths. As mentioned
in earlier posting in this thread you _could_ certainly force the time
durations to match but what does it mean in the end if you do? That all
depends on what it is that you're actually trying to determine and that
is totally unclear; you've not given enough background to understand
what the situation is well enough to have any realistic ideas of what to
tell you. Which is why I suggest consult your advisor or presuming this
isn't absolutely some unique measurement/analysis, research what has
been done by previous researchers w/ similar data.
--
> dpb <none@non.net> wrote in message <kljore$g4v$1@speranza.aioe.org>...
>> On 4/28/2013 11:41 AM, Mary wrote:
>> ...
>>
>> >> W/O far more actual knowledge about the experiment and the data I have
>> >> know idea about whether it's "good" or not, but what about just
>> >> standaradized all of the time series--ie, subtract mean, divide by
>> >> estimated st dev.
>> >>
>> >> And, I'd ask why you can't do something similar over the duration--or
>> >> use a normalized accumulative time over a duration length as a
>> >> slightly different presentation of same idea...
>> >>
>> ...
>>
>> > The data is good it's just difficult to line up over time duration. The
>> > data is the change in angle from the elbow to shoulder originating from
>> > motion capture data of a person raising and lowering their arm. (each
>> > set ends up looking similar to a bell curve (the angle increases then
>> > decreases), but because each person is different, and because there
>> were
>> > no triggers added to the motion capture software, each data set has a
>> > bell curve starting and ending at different times.
>> >
>> > I would like to compare these angles across sets of data- but I'm
>> having
>> > trouble figuring out the best way to scale them so that I can take an
>> > average of all of them to draw conlcusions from.
>> >
>> > Example: Trial A's set has the person moving starting at 3.76 seconds
>> > and ending at 9.87 seconds Trial B's set has the person starting at
>> 2.39
>> > seconds and ending at 8.342 seconds
>> > I would like a way to normalize set A and B so that I can average them.
>> > Then use the std deviations between the sets to create an error bar.
>>
>> I wasn't talking about the data; I was talking about whether the
>> suggested normalization method was meaningful or not...
>>
>> But again it seems essentially throwing darts at a wall blindfolded to
>> no so little about a research project and expect to make meaningful
>> assessments of how the data should be analyzed.
>>
>> These are the questions that should have been asked/addressed _before_
>> the data were collected, obviously, of how it was to be processed in
>> order to draw whatever conclusions it was that were intended to be
>> made from the experiments.
>>
>> At this point, having apparently collected a brown paper bag of
>> happenstance data, it appears you're trying to make something of the
>> hash.
>>
>> I'd guess one should be able to fit some sort of general model to each
>> channel of measured positions vs time. Start then by generalizing that
>> to include another term that is the time shift and include it in the
>> model as another term to estimate.
>>
>> It's not clear at all what is important wrt to the time variable
>> here--the comparison of the position at a given time since the
>> beginning of the movement or what? Or is time important at all and
>> what angle is it that is being measured? There's just too little
>> background for somebody here to be able to do anything of meaning
>> methinks. Talk to your thesis advisor and see if can get some help out
>> of the morass you've created...
>>
>> --
>
> Hi thank you for the help and commentary- it's much appreciated.
>
> As for the data collected - this collection type is extremely normal,
> and was a very well thought out process. The goal of finding an average
> change in angle from individual over multiple trials is extremely
> helpful in biomechanics research and not something that I am attempting
> to cobble together as evidence for a made-up thesis.
> Despite the fact that I am still struggling with my normalization, my
> current code processes motion capture data (x,y,z coordinates per
> marker) and changes it into vector of angle between two points over time
> ( a single vector derived from mechanics). It then marks in each vector
> the start and stop time of the movement with triggers.
>
> Be it biomechanics, electronics, or mechanical engineering, real time
> data should be something that needs to be normalized across trials for a
> true average, and I don't believe I'm asking anything far out of the
> ordinary by wanting to equalize vector lengths
Well, I'd say if you new ahead of time what you were after there should
have been a methodology in place to know how you would process the data.
I've done a tremendous amount of vibration and other motion testing and
indeed, multiple channels are time-correlated. _BUT_, it's done a
priori by setting trigger levels for asynchronous events or by a
reference channel trigger if delay, etc., is important, not by trying to
do something w/ just sampled data ex post facto.
Again, you're arguing for the validity of your setup rather than
providing any additional help on what it is you're actually trying to
infer that could help anybody else in trying to provide ideas.
What's wrong w/
a) using the previously suggested approach, or
b) as suggested in the previous postings which never saw any response to
on finding the starting times, apply a trigger level to each channel and
use that to align them.
Again, you've not described what it is you mean by normalization
precisely nor what you mean by equalizing vector lengths. As mentioned
in earlier posting in this thread you _could_ certainly force the time
durations to match but what does it mean in the end if you do? That all
depends on what it is that you're actually trying to determine and that
is totally unclear; you've not given enough background to understand
what the situation is well enough to have any realistic ideas of what to
tell you. Which is why I suggest consult your advisor or presuming this
isn't absolutely some unique measurement/analysis, research what has
been done by previous researchers w/ similar data.
--