Title: Modeling Vibrating Beam
1Modeling Vibrating Beam
- -using the harmonic Oscillator equation verses
collected data
2Initial Data Collected (unmodified)
-Our initial data collected started with a
zeroing out area
3First Attempt to Model Data
What went wrong with this attempt?
4First Attempt Cont.
- Used fminsearch with initial conditions on C01
and K02 resulted in C-.1163 and K-4.4527 - How can we improve on this model?
- Take out the zeroing out area.
- Shift graph to time0.
- Rerun fminsearch.
5Taking Out Zeroing Area
6Rerunning fminsearch
- CO1 K02 cost6.3941x10-7
- C -.67926355551960
- K -1.5227892111787423x103
- But are these initial conditions the best ones to
estimate C and K?
7Finding better estimates of C and K
- Rewrite code with double for loops to check
initial conditions between -5 and 5 for both C0
and K0 - Stored CO and KO conditions with cost values into
a matrix. - This program ran overnight and did not finish in
time. - Checked the resulting (and incomplete) data for
the smallest cost.
8The Better C and K
- We were able to look at 633 additional different
initial C0 and K0 conditions. - Found C0 -5 and K0-2.41 to have a lower cost
of 6.3875x10-7. - This cost is lower than for C01 and K02 so we
used the resulting C and K of - C -0.68407341326426
- K -1.522633494063630x103
9Determine Standard Error and Confidence Intervals
for C and K
Using the least squares method, we were able to
determine the confidence interval of both C and K.
- Model fitted
- d2 y / dt2 - C dy/dt - K y
- n 5019
- C -0.68407341
- s.e. 0.00979855
- 95 perc. ci ( -0.70367051, -0.66447632 )
- K -1522.633494
- s.e.(K hat) 0.383006
- 95 perc. ci ( -1523.399506, -1521.867482 )
- sigma2 1.2666e-010
10What do our estimates tell us?
We can see that a harmonic oscillator model fits
this data somewhat.
11A closer look
By taking a closer look, we can see that there
are many discrepancies that our model does not
account for.
12Looking at the differences
- The residuals gives us the difference between our
data and the values estimated by the model. - From the stats.m program that we were given, we
have a graph of the residuals vs. time, and
residuals vs. fitted data.
13Residuals vs. Fitted Value
Ideally we expect this plot to be bounded in a
region of residual values and have a cloud-like
spread. Were looking for a constant variance of
the error terms.
14Residual vs. Time
This plot of residual vs. time is used to check
independence and non-independence. Here is we see
that the error terms are non-independent.
15Sample Data vs. Standard Normal
If this were a normal distribution, then our
sample data would agree with the standard normal
line. We see that it only follows the normal line
for awhile and the tails do not.
16Deviation From Normal Distribution
Since there is a deviation from the normal
distribution, we see the tails at the ends of the
above red line.
17So what could we do better?
- A final technique to better model our data was
introduced to our group by Professor Smith as the
beam model. - To look at this, we examined his data and his
beam model which provided a best fit.
18Better Model Fit of Data Example
19A Closer Look
Notice here with this other data set that
multiple nodes are captured by the model.
20Summary
- This analysis took multiple steps in which
careful examination is needed to better improve
our models. - By taking into consideration other variables, we
should be able to better model our own data using
the beam model program.