Signal Processing?
Post Reply
 
Thread Rating:
  • 0 Votes - 0 Average
  • 1
  • 2
  • 3
  • 4
  • 5
06-04-2017, 07:06 AM (This post was last modified: 06-04-2017 07:15 AM by GirlyMan.)
Signal Processing?
Anybody got any signal processing experience?

I got some longitudinal data collected from an accelerometer attached to 39 subjects wrists for 3-4 months with a 1 minute sample rate. I want to see if I can use that data to augment other biometric indicators to classiify individuals. How should I go about doing this?

What I've done so far is generate periodograms (initial period estimate = 1440 minutes = 24 hours = human circadian rhythm) for the spectral density and average activity diagrams (based on the actual period result from the periodogram = 1427-1479 minutes). Can I use the average activity to classify new accelerometer data? How should I split the data into training and test sets? Each participant has a different number of samples so I'm not sure how to use n-fold cross-validation here. Split each participants' samples into half? Take a random sample form a random participant? A random interval won't be aligned. How many samples should I take to ensure the peridocity is apparent? Can I compare the periodograms themselves directly instead? Use both for classification? The different number of samples result in periodograms of different resolution (see below). Currently I'm using a bunch of different similarity measures (Pearson correlation, MSE, MAE, cosine difference, etc.). How do I know to prefer one over the other?

I don't know dick about signal processing other than what I've learned in the last week working on this problem.

Thanks!

[Image: p00aacp_zpsdpoxzfnq.png][Image: p00aac_zpsh3f5zwgi.png]
[Image: p13aacp_zpskjrtqhuc.png][Image: p13aac_zpsa3vfcvd7.png]

#sigh
Find all posts by this user
Like Post Quote this message in a reply
[+] 2 users Like GirlyMan's post
06-04-2017, 07:36 AM
RE: Signal Processing?
The third image down reminds me of past mornings in my youth.

“I am quite sure now that often, very often, in matters concerning religion and politics a man’s reasoning powers are not above the monkey’s.”~Mark Twain
“Ocean: A body of water occupying about two-thirds of a world made for man - who has no gills.”~ Ambrose Bierce
Find all posts by this user
Like Post Quote this message in a reply
[+] 1 user Likes Full Circle's post
06-04-2017, 08:22 AM
RE: Signal Processing?
I don't know shit about what you're doing but it looks fun Smile

We'll love you just the way you are
If you're perfect -- Alanis Morissette
(06-02-2014 03:47 PM)Momsurroundedbyboys Wrote:  And I'm giving myself a conclusion again from all the facepalming.
Find all posts by this user
Like Post Quote this message in a reply
06-04-2017, 09:40 AM
RE: Signal Processing?
I'm not clear enough on your application here to make detailed suggestions, but I know some signal processing. One thought I had was that maybe your validation test set should come from participants that are not in the training set. Otherwise you might be biasing your results.
Find all posts by this user
Like Post Quote this message in a reply
06-04-2017, 09:45 AM
RE: Signal Processing?
I think your deeper questions are targeting stochastic signal analysis, not so much signal processing per se.
I am definitely not versed in high end signal processing or stuff like that, but i think i maybe can help you on fundamental understanding of looking at signals in the time domain vs frequency domain (aka fourier analysis).

PM me if you may have some issues with basic stuff, or we maybe can arrange a skype talk.

@morondog: Actually i liked the topic at university, since it seemed ....to show how beautiful nature can be from a scientific perspective (think we have a whole thread about the beauty of science)

Ceterum censeo, religionem delendam esse
Find all posts by this user
Like Post Quote this message in a reply
[+] 1 user Likes Deesse23's post
06-04-2017, 09:50 AM
RE: Signal Processing?
(06-04-2017 07:06 AM)GirlyMan Wrote:  Anybody got any signal processing experience?

I got some longitudinal data collected from an accelerometer attached to 39 subjects wrists for 3-4 months with a 1 minute sample rate. I want to see if I can use that data to augment other biometric indicators to classiify individuals. How should I go about doing this?

What I've done so far is generate periodograms (initial period estimate = 1440 minutes = 24 hours = human circadian rhythm) for the spectral density and average activity diagrams (based on the actual period result from the periodogram = 1427-1479 minutes). Can I use the average activity to classify new accelerometer data? How should I split the data into training and test sets? Each participant has a different number of samples so I'm not sure how to use n-fold cross-validation here. Split each participants' samples into half? Take a random sample form a random participant? A random interval won't be aligned. How many samples should I take to ensure the peridocity is apparent? Can I compare the periodograms themselves directly instead? Use both for classification? The different number of samples result in periodograms of different resolution (see below). Currently I'm using a bunch of different similarity measures (Pearson correlation, MSE, MAE, cosine difference, etc.). How do I know to prefer one over the other?

I don't know dick about signal processing other than what I've learned in the last week working on this problem.

Thanks!

The Army Distributing Learning Center has people that look at this kind of data for training metrics. They might not be the people you need to talk to, but they might be able to tell you who to contact.

Save a life. Adopt a greyhound.

[Image: anigrey.gif]
Find all posts by this user
Like Post Quote this message in a reply
06-04-2017, 09:53 AM
RE: Signal Processing?
[Image: ichi-5-tturnber.png]

These are good 'uns. Big Grin

Insufferable know-it-all.Einstein God has a plan for us. Please stop screwing it up with your prayers.
Find all posts by this user
Like Post Quote this message in a reply
06-04-2017, 10:09 AM
RE: Signal Processing?
(06-04-2017 09:40 AM)John Derderian Wrote:  I'm not clear enough on your application here to make detailed suggestions,

Given a new series of accelerometer data with at least 2 periods long (intuition seems like it's gotta be at least 2 without a guarantee that the data begins at the start of a period) worth of samples, which participant did the new data most likely come from?

#sigh
Find all posts by this user
Like Post Quote this message in a reply
06-04-2017, 10:39 AM
RE: Signal Processing?
(06-04-2017 10:09 AM)GirlyMan Wrote:  Given a new series of accelerometer data with at least 2 periods long (intuition seems like it's gotta be at least 2 without a guarantee that the data begins at the start of a period) worth of samples, which participant did the new data most likely come from?

Okay, it sounds like the new data is always assumed to have come from one of the known participants, and you want to say what you can about which one. It seems like you can maybe do the following:

Define some goodness-of-fit metric that can be calculated between any two samples (like a simple correlation, maybe). Compile statistics of this metric value over all pairs of samples in your training set, and separate into two classes: when the samples are from the same participant, and when they are from different participants. Make distributions of the metric value for each class.

When you try a new sample from the training set, look at its metric values vs each participant and pick the participant that gives the largest. Compare the value against the two distributions (the "correct" and "wrong" distributions) and judge from this whether it is consistent with the "correct" distribution. A tool for doing this kind of judgement is the ROC curve.
Find all posts by this user
Like Post Quote this message in a reply
07-04-2017, 10:57 AM
RE: Signal Processing?
This could be a good reference for you, Girly.

http://www.ece.rutgers.edu/~orfanidi/int...s-i2sp.pdf

Good primer on signal processing.

"IN THRUST WE TRUST"

"We were conservative Jews and that meant we obeyed God's Commandments until His rules became a royal pain in the ass."

- Joel Chastnoff, The 188th Crybaby Brigade
Find all posts by this user
Like Post Quote this message in a reply
Post Reply
Forum Jump: