Vous êtes sur la page 1sur 47

Statistical Analysis of Performance Characteristics of

Joint Human-Robot Coupling


Author: Gerald ONeill
Lab: HORC Lab under Dr. Panagiotis Artemiadis
Date: 12/18/12
For: Prof. James Middleton

Executive Summary:
The potential of my novel human-robot coupling to go to market
requires a detailed statistical analysis of the couplings performance
characteristics, specifically the detachment force of the magnetic
safety function which determines use in a given application. This
performance space is analyzed using theoretical maximum pull force,
practical maximum pull force, velocity of pull, and depth of force
through the valid use of multiple regression analysis. I find that only
theoretical maximum pull force, the control I designed for the
couplings main function, is a predictor of the functions measured
variable, practical maximum pull force, whereas velocity of pull is not.
While I produce a model for the performance of the coupling, it is not
suitable for licensing and distribution to the end-user market. To
improve the coupling design and performance model towards this end,
the results of the analysis lead me to recommend investigating the
existence of other driving/predictor variables, as well as to further
investigate and quantify the effect of stacking magnets in series
behind the point of action. As such, this analysis was moderately
successful.

1
Introduction:
Source Note:
It is important to note that because this study uses exclusively
unpublished data, there are very few references. All statistical analysis
methods were derived from material published on ASUs blackboard
website particular to Professor Middletons MAE 394 Topic: Applied
Experimental Statistics Class. All magnetic theory was derived from
Physics for Scientists and Engineers by Serway and Jewett. Besides the
statistical and magnetism theory, and unpublished data, no other
sources were necessary. More detailed information can be found in the
references section.

Problem Statement:
The object of this statistical analysis is to both map and
understand the entirety of the human-robot couplings performance
space. The human-robot performance space consists of human fitting,
robot fitting, and the magnetic safety function. As the human and robot
fitting performance are beyond the scope of this analysis (see design
considerations), the analysis must map the magnetic safety function
and its characterizing parameter of maximum pull force necessary for

2
disengagement. This is to be done through the measurement and
analysis of the following variables: practical maximum pull force,
theoretical maximum pull force, maximum velocity of pull, and
configuration of magnets (depth of force).

Background Information:
I am a member of the Human Oriented Robotics and Control Lab
under Dr. Panagiotis Artemiadis. The HORC Lab focuses on the human-
machine interface, and specializes in the subset of the neural/nervous-
robot interface. This means primary areas of research such as
integration of human control systems with robotic execution systems,
deeper understanding and improvement of the mechanisms and
performance of those same systems as well as all human and robot
systems, and creation and improvement of the cutting edge
technology, tools, and methods needed to perform such research. This
manifests in work on: all aspects of augmentative and rehabilitative
exoskeletons (wearable robots directly controlled via brain signal);
creation of hardware and computational method for reading,
interpreting, and applying such signals; analysis, creation, and
improvement of robotic hardware itself; investigation of the workings
and performance of human muscular, control, and visual error
feedback systems; improvement of collaboration/coordination of
human and robot on joint tasks (whether between healthy arm and
prosthetic, or fully separate human and robot entities); and much
more.
As a large amount of this work involves or requires a human and
a robot entity to be directly, mechanically linked (coupled) during
operation or experiment, fabrication of such a coupling was required.
The lab works with several different robotic entities, multiple humans
with varied characteristics, and many different applications with a wide
range of performance requirements. The coupling had to be adaptable
enough to be used in all or most combinations or permutations of
these entities, humans, and applications.
My research area is in rehabilitative and augmentative
exoskeletons. These are wearable robots that marry the superior
human control system to the superior robotic performance in execution
(stronger, faster, etc). The current, and best, form of exoskeleton
control is using the electrical signals the brain sends to the muscles to
contract them (electromyographic signals, or EMG) to tell the robotic
actuators, or muscles when to flex, essentially substituting or adding
the robotic muscles to the humans. Currently, however, the
transferred information only includes the torque about a joint created
by that muscle. The human arm also has impedance (inertia, damping,
stiffness) characteristics, and it can control the stiffness. I am working
on investigating the human arm impedance characteristics and control

3
system in an effort to add impedance control to the already existing
EMG-based control system to improve it.
As my research is one of the most heavily involved with joint
human-robot operation, I was selected to fabricate the coupling. It had
to achieve the previously mentioned adaptability, as well as include a
solely mechanical (as opposed to electrical or software based) safety
feature, as it would be used with human subjects. I conceived of a
magnetic safety feature, and designed, fabricated, and did initial
testing on this coupling. It is essentially a wrist brace that attaches to
the robot via magnets. The configuration and strength of these
magnets can be varied to obtain different forces needed to break the
connection and disengage from the robot (the safety feature). It
transmits torque and shear force by using a multiple peg-and-slot
configuration. See Appendix A and B for pictures of this coupling.
Initial, brief, qualitative testing showed that the coupling, in its
default magnet configuration, could fit the KUKA robot arm used by our
lab, could fit my forearm, and had adequate maximum pull force for
disengagement for my own experiments. Further testing showed that it
would fit any of our other robots, as well as the limbs of any subject
(within reason). However, no other testing or evaluation was done
regarding the safety function force needed to disengage (hereafter
referred to as maximum pull force, or MPF), the aspect that allowed it
to be optimized for any application.
Not only is the coupling to be used in many of our greatly
differing applications in lab, but also we recently obtained a provisional
patent for the coupling, and AZTE is attempting to license and market
it. Thus, I want to prepare the coupling for as wide a range of
applications as possible. I must provide as much information on all
aspects of the coupling, especially function/performance to prospective
licensors and consumers. Furthermore I want to create a functional
database to provide end users that allows them to input their robotic,
human, and most importantly, application parameters, and will
generate the proper adaptors and, again most importantly, magnetic
configurations for them to buy and use. Finally, a statistical analysis
will provide not only these things, but will help to make inferences as
to the underlying mechanisms/reasons for the behavior/results,
allowing us to manipulate these mechanisms to improve the
performance itself. To these ends, I must map, analyze, and understand
the whole of the couplings performance space.

Design Considerations:
Upon initial consideration the only changeable/controllable
variable is the configuration of magnets. The peg/slot configuration
allows for four slots where magnets can be placed in parallel, i.e.

4
next to each other but all at the immediate point of contact to change
the pull force, as well as the housing configuration in each of the four
peg/slots allow magnets to be stacked in series, i.e. stacked on top of
each other depth-wise to change the pull force. It would seem we
would simply vary the magnet configuration, measure the resulting
maximum pull force, and map the assumed linear space/relationship.
However, as most practical relationships are, it is much more
complicated than the theory would suggest:
To measure the independent variable, we must come up with a
way to characterize a particular magnet configuration. The magnet
provider has rated each magnet for a maximum pull force setup in the
same way as the coupling- two magnets pulling apart. We could use
this to characterize each configuration with maximum theoretical pull
force. Magnetic theory says that magnetic force sums fully in both
parallel and series. Therefore we just sum the manufacturer rated pull
force for each magnet in the configuration to arrive at our first
variable: theoretical maximum pull force.
However, as previously stated, the practical value almost never
reaches the theoretical. Many, many sources contribute to this
reduction, including debris between contact areas, misalignment of
magnets (none are perfectly axially aligned in their grooves, in fact a
few are noticeably misaligned), manufacturing variation between
magnets, angle of pull (I gave the slots additional clearance as it is
difficult to pull in a perfectly axially normal direction), shear forces due
to that same additional clearance, and depth of magnets from contact
point. By finding the general relationship between practical and
theoretical values, we will have a method of predicting the practical
value given manufacturers rating, better informing magnet buying
choice for a given application. So, in addition to the theoretical value,
we arrive at our second variable: practical pull force.
As mentioned above in the reason for the practical value
reduction, depth of magnet is partly responsible for this drop. Theory
says that magnets sum perfectly in both parallel and series, so I ran a
previous experiment to determine the veracity of this theory. The
magnets did indeed sum nearly fully in parallel, but not so in series.
Each successive magnet stacked in series, including the first, added a
lesser fraction of its theoretical pull force to the total. This is because
the magnets are not point masses, i.e. they have depth. So, each
magnet added is further away from the point of contact, and because
magnetic force is a field force inversely related to distance from point
of action, this reduces the added force. Since our configuration is a
combination of both parallel and series, in order to map and
recommend these configurations we must determine the quantitative
effect of depth on configuration force (but only the force of the
magnets in series behind the first). Thus we introduce the third
variable: Depth of force.

5
Despite all these considerations, a preliminary experiment
revealed that the practical pull force for a given, unchanging
configuration varied considerably- an unforeseen result. Upon
investigation, I found that the velocity of the detachment pull very
significantly affected the maximum pull force needed for detachment. I
wish to gain insight into the reason behind this (to eliminate it if
possible). I also must map the relationship, as this adds a second
parameter to the application- the end user would input both force and
velocity parameters for a given application to determine the necessary
magnet configuration. This is the fourth and final variable: velocity of
pull.
Adaptability to application, governed by the magnetic safety
function, is only one of the three functions of the coupling, the others
being adaptability to human user and to robot. However, the
performance space of both of these is simply fit or doesnt fit, and
the driving variable is the mechanical component on each end of the
coupling. This does not lend itself to a statistical analysis, and has
previously been both investigated and optimized; the coupling is fully
adaptable to any user and any robot, using its hinge mechanism and
modular design. Thus, the performance spaces of these two functions
are beyond the scope of this analysis, which will focus only on the
magnetic disengagement function.
It is inefficient and impossible to experimentally map the entirety
of the couplings possible performance space, as the housing depth,
number of slots, and variation of magnets of different dimensions (and
therefore strengths), allows for a nearly infinite amount of
combinations and permutations of configurations, along with depth of
force and velocity of pull, meaning the coupling is capable of a wide
range of possible maximum pull force, with very fine resolution.
Therefore it is key to the analysis to pick appropriate ranges and
variations for each of the 4 variables, our sample space, to keep the
analysis resource efficient while still adequately representing the
entirety of our performance space.

Procedures:
Sample:
A description of the sample space involves determining the
parameters of each of the variables involved. For each variable, the
following must be found: the method and reason of measurement,
sample size and reason, the nature and source of bias, the nature and
source of error, and the assumed shape of the distribution. All variables
will hereafter be referred to the following variable abbreviations.
Theoretical Maximum Pull Force (T)-

6
Method and reason of measurement- Magnetic theory indicates that
magnetic field forces overlapping at a given point sum. That point
(the point of action) is the point of contact between magnets on the
human side of the coupling and the robot side, i.e. the break point.
The manufacturer of the magnets (K&J magnetics) provides a force
rating for all of its magnets for the case where the point of action is
directly at the end of the magnet, which is identical to our model.
Thus we measure theoretical pull force by summing the
manufacturer ratings of all magnets in the configuration. It is
necessary to measure T because the performance of the coupling
we are interested in is the pull force needed to decouple.
Sample size and reason- this is a theoretical value, and as such
doesnt vary. Thus only one value exists/is given for each magnet
configuration. Due to the large amount of time and effort needed to
run the experiment for each configuration, and the similarly large
amount required to change between configurations, a sample size
of only 10 configurations was chosen, with theoretical pull forces
ranging from around 70 to 220 Newtons at a relatively consistent
step size. Unfortunately, the data from configuration 6 was lost, and
upon reaching configuration 9, some of the key magnets broke,
resulting in the loss of both 9 and 10, and the inability to redo
number 6.
Nature and source of bias- As T is based purely on the
manufacturers ratings, any and all bias would be due to the
manufacturer, and as such, unknowable for the purposes of this
experiment. I conjecture that possible sources of bias would be the
manufacturing process (a malfunctioning machine, etc) or the
testing for rating process (a miscalibrated instrument). Given the
scientific nature and written policies of the manufacturer, I assume
no significant bias for the purpose of this experiment. T is only used
as an index for magnet configuration anyways, so any consistent
bias would not matter.
Nature and source of error- Again all error proceeds from the
manufacturer. As with any fabrication, the magnet will have physical
tolerances within which it varies, causing the actual pull force to
vary slightly from the rating. I assume this error to be entirely
random, independent, and normally distributed, as is standard for
manufacturing tolerances. While its possible something
systematically affected the error (I bad batch of raw metal), T is but
an index.
Assumed shape of distribution- as a theoretical index value that I
control, I picked range of values with a nearly consistent
step/resolution between each value.
Practical Maximum Pull Force (P)-

7
Method and reason of measurement- this is the pull force at which
the coupling disengaged during each trial. The robot used has
internal torque sensors and software that allows it to output the
forces exerted on its end effector (the end point of the robot, to
which the coupling is attached). By recording these, and taking only
the force along the z-axis, I record the force normal to the face of
the end effector and the plane of magnetic contact, thereby
recording detachment force. It is necessary to measure P because
the performance of the coupling we are interested in is the pull
force needed to decouple.
Sample size and reason- I decided, based on previous experience
with HORC lab research, and the principles of experimental design I
learned in stats class, to have around 25 to 30 trials for each
configuration. For this type of experiment, when not limited by
resources, this is the sample size I found generally most quoted as
sufficient. Unfortunately in a stroke of poor decision-making, I did
not standardize the number of trials between configurations. While
each configuration has around that number of trials, the exact
number varies between configurations. However, because of the
very sufficient nature of my trial selection, when required by
statistical test, I can randomly throw out excess trials to match the
configuration with the lowest number.
Nature and source of bias- much like the error below; I calibrated
and zeroed the equipment before use. Given its digital nature, and
that it is the measurement of a variable resulting from a trial, rather
than one pre-picked, there should be no systematic bias. If there
were, the greatest possibility for systematic bias would lie in the
processing method, where a smoothing and filtering process was
used on the raw signal. However, if this does introduce a bias, it
would be a systematic lowering of the magnitude of the values.
Even if this is true (unlikely), after having done the statistical
analysis and given the user and finished model for application, the
relationships would simply be slightly conservative, which is fine as
it is a safety function. It would simply slightly increase the safety
margin.
Nature and source of error- as this value is read directly from the
equipment, with no knowledge of its internal process, all error is
unknown and is dependent on the instrument. Any error is most
likely caused by things like sensor precision. However, I calibrated
the instrument as per its instructions before the experiment, so I
assume all error is random, independent, and normally distributed.
Assumed shape of distribution- as this is likely a dependent
variable, I expect the values to be determined by the inputs, and so
the choice of inputs determines the shape of the distribution.
However for a given choice of inputs, because the error is normally

8
distributed, and there is no bias, the resulting values would
inherently be normally distributed. For example, if a user was to
make average or normal pulls in the course of operation, I would
expect the resulting values to be normally distributed.
Velocity of Pull (V)-
Method and reason of measurement- this is the velocity at which
the coupling is pulled/detached. I used a motion tracker/capture
system able to sense Cartesian coordinates of strobing markers in
conjunction with a marker placed on the end of the coupling. This
allowed me to record the position of the coupling (accurate to
0.1mm) and its derivative, the velocity of the coupling. Using this, I
found the velocity at the moment of decoupling (the maximum
velocity during the whole pull), thereby getting a measure of how
fast the user pulled the coupling. I chose to measure this as a
previous perfunctory study showed that the pull force was very
likely related to the velocity of the decoupling.
Sample size and reason- as each trial resulted in the recording of P
and V, I had the same sample size and reasons as for P (see above).
Nature and source of bias- much like the error below, I calibrated
and zeroed the equipment before use. Given its digital nature, and
that it is the measurement of a variable resulting from a trial, rather
than one pre-picked, there should be no systematic bias. If there
were, the greatest possibility for systematic bias would lie in the
processing method, where a smoothing and filtering process was
used on the raw signal. However, if this does introduce a bias, it
would be a systematic lowering of the magnitude of the values.
Even if this is true (unlikely), after having done the statistical
analysis and given the user and finished model for application, the
relationships would simply be slightly conservative, which is fine as
it is a safety function. It would simply slightly increase the safety
margin.
Nature and source of error- as this value is read directly from the
equipment, with no knowledge of its internal process, all error is
unknown and is dependent on the instrument. Any error is most
likely caused by things like sensor precision and offset. However, I
calibrated the instrument as per its instructions before the
experiment, so I assume all error is random, independent, and
normally distributed.
Assumed shape of distribution- as this is likely a user chosen input
(the user chooses how fast to pull), I expect the values to be
determined by the user choice, and so the choice determines the
shape of the distribution. However for a given choice of velocity,
because the error is normally distributed, and there is no bias, the
resulting values would inherently be normally distributed. For
example, if a user was to make average or normal pulls in the

9
course of operation, I would expect the resulting values to be
normally distributed.
Depth of Force (D)-
Method and reason of measurement- this is a measure of, for
magnets stacked in series (as opposed to parallel), how far behind
the point of contact/action the magnet is. Rather than create a
complicated calculation to find this value (ex: point of contact is the
front of the first magnet, but the force is modeled as acting from the
middle of the magnet, so we must add the depth, etc), we will
simply model each magnetic field as centered at the front of the
magnet. Thus for the front line of magnets, the D value is zero, and
for the 1/8 tall magnets stacked behind the first, the D value is 1/8,
and so on. As this is simply an index, and I assume the relationship
between distance and magnetic force wont change, this should be
fine. In reality we could choose almost any index of depth and, as
long as we stayed constant with this choice, the relationship would
hold. I measure this because magnetic theory says that magnetic
field forces decrease with distance. While our T value assumes all
the magnets to be point masses at the same point, in practicality
they have size, so adding them in series (but not parallel) moves
them away from the point of contact/action, theoretically reducing
their force contribution. Thus I measure the depth to both confirm
and model this mechanism.
Sample size and reason- this sample size is the same as T, as we
have one set of D values for each magnet configuration. While this
was necessary due to resource limitations, and then enforced when
the magnets broke, it is unfortunate, as it forces me to work with a
very small sample size; only 4 configurations had depth, and 3 with
no depth as the control. Thus any results should be taken with a
grain of salt. However, the existence of the proposed relationship is
strongly indicated by physics, so I go ahead with the analysis and
put stock in its results anyways.
Nature and source of bias- Once again, these values (physical
dimensions) are taken directly from the manufacturer. The possible
sources of bias and error have already been discussed under T (see
above), but once again I will assume no bias, and random,
independent, and normally distributed error.
Nature and source of error- See above
Assumed shape of distribution- Once again, I choose the values. As
the population is whatever I choose (which magnets I buy), the
distribution of the D value simply depends upon my choices. Just as
before, however, for a given model of magnet, for example the 1/8
deep magnets, the actual D value for this set should be normally
distributed, as the error is normally distributed.

10
Matlab was used heavily in the gathering of the final dataset. Both
the velocity and force data was gathered as a continuous reading
sampled at 1000 Hz. This signal was very noisy, occasionally had bits
missing, and showed the entire experiment, rather than just the values
at the moment of detachment. Matlab was used to process this data in
order to get solely the velocity and force at the moment of detachment
(the maximum values, during each pull). Matlab was also used to
conduct the entirety of the following statistical analyses. See appendix
script F for full Matlab script.
The type of magnets (large, medium, and small), as well as their
depth and force rating, are given by appendix table C. The
configuration magnet setup is also given by appendix table C. Only two
rows were used, so there were never more than 2 magnets in series.
There were four slots, so for either row, up to 4 magnets could be in
parallel. The large magnets were only ever in the front row, and the
medium and small magnets were only in the second (in series) row.
Finally, my data set in its entirety is given by appendix table D.

Statistical methods used:


I first must define how I wish to investigate the performance
space of the coupling. The purpose and hopeful outcome of this
analysis is to be able to provide potential licensors and all end users
with a model allowing them to provide their application parameters
and be supplied with an appropriate magnet configuration with which
to equip their coupling. To accomplish this, my analysis should answer
the following:
What phenomenon (hint- velocity and theoretical force of pull,
V and T) affect/change the P, the practical pull force needed
to disengage the coupling?
Having determined these controlling variables, can I
quantitatively model said relationship with P?
How does the depth of force, D, relate to the forces
contribution to P?
Using the relationships between V, T, D, and P (and any others
found to contribute), can I build a model that allows an end
user to input his application parameters and so choose an
appropriate magnet configuration?
Finally, what has the analysis revealed about the driving
phenomena and mechanisms that I can use to improve future
iterations of the coupling?

Predictors of P, practical pull force:


To decide what to include in an analysis of the phenomena
driving/determining the practical pull force needed to disengage the

11
coupling in any one instance, I will list all possible factors, then discuss
the reason for their inclusion or exclusion from the analysis-
T, theoretical magnetic force. As the coupling was designed to
not only operate based on magnetic theory, but to vary its
performance through manipulation of the same, it is a prime
candidate for inclusion (I would hate to think that the main
design control is a complete failure). The physics of magnetism
state that a magnet creates a magnetic field that exerts a force
on any magnetic object placed in that field- attractive or
repulsive based on pole- in this case, attractive. The strength of
this field at a given point is based on distance from the magnet
and strength of the magnet, and the magnetic fields at any given
point sum. Thus, theoretically, by increasing the magnets
strength or placing other magnets such their fields overlap in a
complimentary direction (i.e. as close to the same direction as
possible), one increases the force on a magnetic object at that
point. So by adding magnets in parallel or series (the physical
setup of the coupling ensures the fields exert force in the same
direction) to the coupling configuration, one theoretically varies
the attractive force, and therefore the force needed to overcome
and decouple. So, of course, theoretical magnetic force will be
included as a predictor.
V, the velocity of pull during decoupling. Theoretically, if the pull
force is exerted perfectly normal to the plane of contact (i.e.
along the same axis as the attractive force), and there are no
other forces involved, the magnetic attraction force should be
the only predictor of force necessary to decouple. However,
practical performance is almost never exactly the same as the
theoretical model. While performing early, entirely qualitative
testing using only my own senses, I noticed that without varying
the magnetic configuration, the strength I needed to decouple
seem to vary considerably. So, I performed a very brief test using
the robots sensors and confirmed this. The only variable I
seemed to be able to vary, the only one I had control over) was
how quickly I attempted to pull the coupling during decoupling.
Thus I added the motion tracker to record velocity of pull, and did
a preliminary test, which resulted in the data and fit seen in
appendix figure E. This evidence, along with the need to map the
performance space based on configuration as well for the end
user, prompted this analysis. Thus, despite not yet knowing why
it affects pull force, or whether it does not affect it but is simply
an indicator of something that does, I include it. I do, however,
have an inkling of something that might be this reason:
Theta, the angle between the axis of magnetic force (normal to
the plane of action- the axis of the cylindrical magnets) and the

12
direction of pull. The peg-and-slot design was meant to restrict
pulls to the perfectly axial direction. However, humans lack the
capability to move their arm very quickly in a prescribed
direction that they must derive visually from the small plane of
the robot end effector, whether in emergency or not. This pull is
often skewed to an incorrect direction by a small but significant
degree. This was causing the sharp-angled peg to scrape or even
catch on the slot wall, creating a force that was preventing pull.
To counteract this, I increased the clearance between peg and
slot, consequently allowing the user to pull in a direction not
perfectly axially aligned, albeit very small. By pulling in a non-
fully-normal direction, you change the force needed to overcome
the attractive force, and, more importantly, introduce shear
forces at the surface of magnetic contact, as well as begin to
peel the magnets apart as opposed to perfectly axial
separation, all of which could greatly affect the necessary
detachment force. I would assume that as the velocity the user
pulls his arm increases, his directional control decreases, and
theta increases. However, I have decided not to include this as a
predictor for three reasons: Firstly, this is a very difficult thing to
measure. While I might be able to do so from the
position/velocity data (angle between velocity vectors), the very
small range of angles, combined with the briefness/speed of the
action, the sensitivity of the instrumentation, and the vibration of
the coupling during detachment would likely make this measure
very inaccurate. Secondly, we would still be unable to control
theta, as evidenced by the need for me to change the clearance
in the first place. As such, we might gain knowledge as to its
effect, but such knowledge would be of no aid to the end user.
Thirdly, the relationship with velocity that lead to this theory
shows that velocity is an index of theta anyways, and we are
already including it; no need to include multiple measures of the
same phenomena and introduce co-linearity and codependence
issues when we already have one.

Determination of the relationship between predictors T and V and


outcome P:
Having identified the likely predictors, I will attempt to model
their relationship with the outcome variable. As magnetic theory
indicates a likely linear relationship between them, I will perform a
multiple regression analysis. Upon performance, I will test for
significance of the relationship, hopefully validating my choice of
predictor variables. This analysis will also allow me to see which of the
predictor variables has a bigger impact (if any) on the outcome. A
residual analysis will allow me to determine if a linear model was an
appropriate choice, which I suspect it will be, at least for T as magnetic

13
fields sum linearly in theory. I will use 95% confidence interval (an
alpha of 0.05), as I want to detect small differences. While increasing
my ability to detect small differences increases risk of a type I error, a
false positive would simply result in a more conservative safety
function, and so does not put the users health at risk- it simply would
make their application slightly less convenient. For example, if I made
a type I error by saying increasing magnetic configuration increases P,
and so the user added more magnets to increase decoupling force, the
decoupling force would simply remain low, allowing him to still easily
detach- a minor inconvenience, but with no health risk. Finally I
examine the assumptions necessary for my regression analysis to be
considered valid: The sample space covers the majority of the relevant
performance space (a user likely wouldnt need a force greater or
lesser than the ones covered), so the assumption that the sample is
representative of the population for the inference prediction holds. As
discussed in the sample section of the procedure, the error is a random
variable with a mean of zero conditional on the explanatory variables.
Finally, as discussed in the predictors section, given my exclusion of
the theta predictor, and that the user has control over the velocity of
pull and magnet configuration separately, the predictors are linearly
independent (unless the user decides to pull only at certain speeds for
a given configuration, which I most certainly did not, instead pulling at
pseudorandom speeds). Finally, I note that D, the depth of some of the
acting magnets, is a possible confounding variable, as they are slightly
farther away from the point of action, and magnetic theory states that
magnetic force decreases with distance. As such, I will perform the
regression on only those configurations with no second row magnets (D
= 0).

Determination of the effect of D, the depth of second row magnets:


Having previously excluded those configurations with D values,
as they were a confounding variable, I must determine their effect on
the outcome variable, as they are an integral part of the design,
allowing for more range and resolution of magnetic force. To this end, I
again perform a multiple regression analysis. The same validation,
significance and impact testing, residual analysis, alpha, and
assumptions hold. To do this, I will separate the theoretical force into
front row force (still T), and back row force, D, and use them as
separate predictor variables. Since the magnetic coupling will only ever
include a one second row, and never any additional rows behind that, I
will treat it as a binary, categorical variable, either second row or not,
as opposed to a continuous variable classified by measured depth.
Finally I compare the beta coefficients for T and D (if significant) using
a t test. If they are significantly different, (the coefficient for D would
theoretically be less), then I know moving the magnets to the second
row has an effect, and have quantified it. If they are not significantly

14
different, I can conclude that putting magnets in series in the second
row does not significantly affect their force contribution.

End user model:


Having determined whether putting magnets in series in the
second row has a significant effect, I can create my final model. I
perform one last multiple regression (all aspects hold once again), this
time on all the configurations. If I conclude D is significant, I include it
as a predictor (D = 0 for those with no back row magnets), if not, I
dont. The results of this regression are my final model. By isolating P
(previously the outcome variable) with V on one side, and leaving T
and D (or just T if depth has no effect) on the other, the user can input
his applications required practical force and maximum velocity, and
will receive the necessary configuration of magnets he must use. This
will conclude my statistical analysis.

Results:
I conducted four multiple regression analyses. For each I
produced a scatter plot of real and predicted values (real values in
blue, predictions in red), the coefficients of the regression model, an
r^2 value for the amount of variance in the output explained by the
regression model, a t test statistic for the significance of each
coefficient, a critical t value to which to compare them, a normplot of
the residuals to determine normal distribution of the residuals, and
residual plots for each of the predictor variables to determine
appropriateness of linear fit. I did this for only the first row
configurations (T and V vs. P), for only the configurations with magnets
in the second row (T-D, D, and V vs. P), for all configurations (T and V
vs. P), and for all configurations (T-D, D, and V vs. P). Thus I perform all
the previously mentioned regression analyses and t-tests, all at 95%
confidence level (5% chance of type I error), and all validated by their
assumptions. They will be presented in the following order- regression
plot (real values in blue, predictions in red), normplot, residual plots,
table of coefficients, and table of t-results. All figures are labeled as:
Fig. analysis#.plot#.
They are as follows:

First row configurations- T&V vs. P:

15
16
17
Table 1.1
B0 29.3592
B1 0.2468
B2 -9.6916
R^2 0.1748

Table 1.2
t statistic for T 4.482
t statistic for V -1.951
Critical t statistic 1.664

Second row configurations- T-D&D&V vs. P:

18
19
20
Table 2.1
B0 -17.5581
B1 0.3023
B2 -0.0166
B3 15.0128
R^2 0.2436

Table 2.2
t statistic for T 5.623
t statistic for D -.274
t statistic for V 2.211
Critical t statistic 1.660

All configurations- T&V vs. P:

21
22
23
Table 3.1
B0 14.3033
B1 0.1530
B2 2.9019
R^2 0.2069

Table 3.2
t statistic for T 6.905
t statistic for V 0.699
Critical t statistic 1.653

All configurations- T-D&D&V vs. P:

24
25
26
27
Table 4.1
B0 12.9760
B1 0.2068
B2 0.0995
B3 1.1867
R^2 0.2135

Table 4.2
t statistic for T 5.295
t statistic for D 2.571
t statistic for V 0.287
Critical t statistic 1.653

Conclusion:

Table 5.1-
Summarized
Results
1 (front 2 (second
Analysis # row) row) 3 (all-T) 4 (all-T/D)
B0 29.3592 -17.5581 14.3033 12.976
B1 0.2468 0.3023 0.153 0.2068
B2 N/A -0.0166 N/A 0.0995
B3 -9.6916 15.0128 2.9019 1.1867
R^2 0.1748 0.2436 0.2069 0.2135
t statistic for 4.482 5.623 5.295
T 6.905
t statistic for N/A -0.274 2.571
D N/A
t statistic for -1.951 2.211 0.699 0.287
V
Critical t 1.664 1.66 1.653 1.653
statistic

Before deriving useful information from the analysis, I must


confirm its validity. The norm plots for all four regressions were very
close to normal, indicating that the foundational assumption of a
normal distribution of error was relatively correct. Also, the residual
plots against each variable for every regression showed no distinct
pattern whatsoever, validating our assumptions of independent error
and homoscedasticity, and most importantly, validating the use of a
linear fit (as opposed to hyperbolic, exponential, etc).

28
Having confirmed the validity of the results, I can examine the
results themselves. The first standout result is the coefficient of
determination, r^2. This represents how much of the variation in the
dependent variable is explained by variation in the independent
variable, or goodness of fit. The r^2 values for the regressions are
rather low, ranging from 0.15 to 0.24. While r^2 values of 0.3 are often
deemed acceptable in real world application, I have a tough time
accepting them for a controlled application designed to only have one
driving factor. This means the designed predictor of magnetic
configuration, in conjunction with what I thought to be an unexpected
predictor in velocity, only explains at most 24% of the variance in the
practical pull force- not optimal.
Before examining the coefficients of the predictors, it is
necessary to see which of these were statistically significant.
Thankfully, the coefficient for T, theoretical magnetic force, was
significant, meaning the t statistic was greater than the critical t value,
for all four regressions. This means that the variable I designed to be
the control for the outcome variable does indeed do so. However the
other variable I chose in an effort to explain and model the practical
pull forces behavior, velocity, was only significant in the first two
regressions, which only used parts of the sample. In fact, in the
regressions that used the full sample, velocity was not statistically
significant, which I will further delve into upon examination of the
coefficients. Finally, the significance of the D variable, depth of force,
was slightly confounding. While significant for the regression making
use of the whole sample space, it was not so for the set examining only
those configurations with magnets in the second row, i.e. D values. A
look at the coefficients themselves might aid in understanding these
issues.
The coefficients for the theoretical pull force were very
consistent, at least in comparison to the rest of the results. This along
with the coefficients being significant in every case, validates the
magnetic theory behind the design. The variance in theoretical
magnetic force (by varying the configuration) always accounted for
variance in the practical pull force, did so across all the sample space,
and did so in a way consistent with the design theory. Despite my
preliminary studies and conjectures, however, the coefficients for V,
velocity of pull, were not nearly so consistent. In fact, they were
strongly negative for one of the partial regressions, strongly positive
for the other, and largely neutral (and insignificant) for both full
regressions. This means that velocity is likely not, in fact, a predictor. I
confirmed this by plotting velocity against practical pull force
individually for each configuration- the general trend not only varied in
severity, but sign as well randomly from configuration to configuration.
Finally, the coefficient for D was also inconsistent. For the partial
regression involving only configurations with D values, the coefficient

29
was negative- entirely unexpected, as magnetic theory would predict
that magnets in series would still add to the practical pull force, just
not as much as those on the plane of contact. However, this coefficient
was insignificant. Instead, the significant value derived from the
regression of the entire sample showed more expected results. The
coefficient was about half that of the coefficient for T, those in parallel
(i.e. on the plane of action). While this reduction of force contribution
was more severe than expected, it is consistent with, and somewhat
validates, the magnetic design theory. Magnets stacked in series
behind the first would seem to contribute to practical pull force in a
lesser manner. However, the significance is still not entirely consistent,
and it is puzzling that the D coefficient would be insignificant for the
regression involving ONLY those configurations with D values, as
opposed to the full regression involving both those with and without.
This is perhaps attributable to the extremely small sample size and
variation. While the number of trials for practical pull force and velocity
was large and quite sufficient, I was only able to do seven different
magnet configurations, and of those only four with second row
magnets, and of those very small variation in depth of force.
It is interesting to note that the coefficients for T are around .2,
meaning that the practical pull force is only a small amount of the
theoretical pull force. As I tested the relation of practical pull force to
theoretical in a much simplified setup identical to the manufacturers
test conditions before this analysis, and found that the rated
theoretical value was indeed very near to the practical, this means that
there is something responsible for systematic bias (reduction) from the
rated pull force. This results in a predicted range of only around 30-50
N of practical pull force, despite a theoretical range of 70-210 N.
Furthermore, as seen by the R^2 values, my designed control,
magnetic configuration, only accounts four about of the variance in
practical pull force. This is unfortunate because while I might be able to
control the pull force somewhat by varying magnetic configuration, for
any one given emergency detachment, the pull force necessary could
vary drastically. Not only could this be potentially dangerous, but also
makes it very difficult to plan for a given application. In fact, upon
examination of the regression scatter plots, I see several phenomena:
Even with the strongest magnetic configuration, practical pull forces
were experienced below that predicted by the weakest configuration;
While both the predicted and upper range of pull forces experienced
increased with magnetic configuration strength, the lower range and
variation stayed relatively consistent (and very low, 10 N). This means
that even if I plan for an application involving greater forces by
increasing the magnetic strength of the configuration, I still risk
detachment at very low force.
The coefficient for intercept is also interesting. It is very high-
anywhere from 25 to 43% of the predicted value. Theoretically, we

30
expect the intercept to be zero, as without magnets, the coupling does
not stay together. However, even without magnets, since the user does
not pull in a perfectly axial direction, forces from the slot wall will still
be experienced. Additionally, given the poor correlation, large
variance, and small slope of the main predictor, a high intercept is to
be expected and not interpreted in real application.
I can now fulfill the goal of this analysis and answer the
questions for which it was intended:
What phenomenon (hint- velocity and theoretical force of pull, V and
T) affect/change the P, the practical pull force needed to disengage
the coupling?
As expected due to the theoretical design mechanisms, T is in fact a
predictor of coupling performance. Despite what my perfunctory
conjectures and studies suggest, however, V is not. The results
indicate the likelihood of another predictor, as T explains only a
small amount of the variance in P. I believe this missing predictor is
theta. In this case velocity would not be an index of theta- the user
is not capable of fine control of theta regardless of a slow or fast
pull. With large theta, the walls of the slots would change the
necessary force, and the pull would introduce shear force and
peel at the magnet surface. As the direction and magnitude of
theta would be completely random, so too would be the introduced
variance in P, which is consistent with the results. Either way, the
lack of correlation and evidence means I will need to perform
another in depth experiment and analysis to determine or validate
any other predictors. For the time being however, the user can still
only control magnetic configuration, so even if I found this other
predictor, it would not affect anything without a change to the
coupling itself.
Having determined these controlling variables, can I quantitatively
model said relationship with P? I can and have. However, as
previously discussed, while this model provides a prediction of
expected practical pull force for a given pull, the variance in results
is such that the model may not be particularly useful to the end
user.
How does the depth of force, D, relate to the forces contribution to
P? The results indicate that I need to do more experimentation and
analysis specifically on this relationship, as the sample size was
much too small for this analysis and the confounding variables too
many. However, I can tentatively conclude that magnets in series,
i.e. at a distance away from the point of action, contribute a lesser
amount of their theoretical force to practical pull force, and that this
reduction is proportional to distance and is quantifiable. I only feel
comfortable in saying this, however, because of the concrete

31
underlying physics theory backed up by massive weight of evidence
throughout physics history.
Using the relationships between V, T, D, and P (and any others
found to contribute), can I build a model that allows an end user to
input his application parameters and so choose an appropriate
magnet configuration? While I do have a model allowing the user to
input his applications practical pull force (but not velocity, as it is
not a predictor) and find the necessary configuration, I feel that it is
far from adequate for an end user. The lack of range of 30-50
Newtons is not sufficient for the range of applications demanded by
the market. Furthermore the massive variance from the models
predictions makes the model not particularly useful because the
nature of the couplings safety function is expected to be relied
upon for single trials, or one-time events (as opposed to
performance only required over a longer operation time or number
of uses).
Finally, what has the analysis revealed about the driving
phenomena and mechanisms that I can use to improve future
iterations of the coupling? The coupling suits our particular needs
for the HORC lab, as we deal with low normal force applications, but
is currently is not suitable for licensing. Fortunately, while the
analysis has not provided specific design improvements, it has
pointed me at the analyses I should conduct next to derive them: It
has evidenced the existence of another (or more than one other)
predictors, and has hinted at its nature. While I believe this may be
the angle of pull, I must conduct further research to this regard (as
well as hope it is a user-controllable variable). Furthermore, I must
perform further research to confirm and quantify the relationship of
depth from point of action and theoretical magnetic pull force
(which, though given by theory, is likely slightly different for my
practical situation). Finally, it has confirmed and validated the use of
magnetic configuration as a control for the safety features
detachment function.

References:
Jewet. , & Serway, (2010). Physics for scientists and engineers.
(8th ed.). Brooks/Cole.

Middleton, J. (2012, August 01). Mae 394: Applied experimental


statistics blackboard. Retrieved from
https://myasucourses.asu.edu/webapps/portal/frameset.jsp?
tab_group=courses&url=/webapps/blackboard/execute/content/fi
le?
cmd=view&content_id=_6963381_1&course_id=_235395_1&fra
mesetWrapped=true

32
Appendix:
Figure A- Human joined to robot via coupling.

Figure B- Expanded SolidWorks view of coupling

33
Figure E- Preliminary data suggesting velocity as a predictor

Table C- Experimental Parameters


# of
Configuration # of Large Medium # of Small Front line Back line
# Magnets Magnets Magnets Theor. Force force force
1 2 0 0 69.4 69.4 0
2 2 0 2 88.2 69.4 18.8

3 3 0 0 104.1 104.1 0
4 3 0 2 122.9 104.1 18.8
5 4 0 0 138.8 138.8 0

34
6 4 0 4 176.4 138.8 37.6

7 4 2 2 196.8 138.8 58

8 4 4 0 217.2 138.8 78.4


Large magnet rated force- 34.7 N
Medium magnet rated force- 19.6 N
Small magnet rated force- 9.4 N

Table D- Raw Data


T(without
config V P T D D)
2.06061081
1 6 41.54279524 69.4 0 69.4
2.16216703
1 6 36.48718379 69.4 0 69.4

2.44083513
1 3 15.10426528 69.4 0 69.4
1.93347337
1 7 13.58991452 69.4 0 69.4
1.88867865
1 4 12.11751157 69.4 0 69.4
1.70456469
1 2 23.95328385 69.4 0 69.4
1.92885061
1 8 14.66100247 69.4 0 69.4

1.66877677
1 6 34.20253059 69.4 0 69.4
1.78554472
1 2 33.71765537 69.4 0 69.4
2.28677138
1 6 11.47990599 69.4 0 69.4
1.67408821
1 4 34.79954133 69.4 0 69.4
1.95141414
1 4 35.9369521 69.4 0 69.4
1.28096392
1 7 41.0735417 69.4 0 69.4
1.43923434
1 7 42.95006143 69.4 0 69.4
1.84503068
1 5 16.57648633 69.4 0 69.4
1 0.80347902 27.80866579 69.4 0 69.4

35
1
1 1.57172397 8.245677473 69.4 0 69.4
2.10796665
1 9 7.990729732 69.4 0 69.4
2.19318712
1 4 7.390135001 69.4 0 69.4
2.07944438
1 3 40.48554494 69.4 0 69.4
1.82130574
1 1 33.32386475 69.4 0 69.4
1.52561978
1 4 44.30056216 69.4 0 69.4
1 1.54211678 47.30806934 69.4 0 69.4
1.64551997
1 2 32.60246425 69.4 0 69.4
2.07052180
1 2 24.3482951 69.4 0 69.4
1.51352978
1 4 36.22005847 69.4 0 69.4
2 1.91924389 36.54294389 88.2 18.8 69.4
2 1.71894373 30.85447487 88.2 18.8 69.4
1.98162003
2 9 20.57354556 88.2 18.8 69.4
1.86279362
2 1 46.98677648 88.2 18.8 69.4
1.96893360
2 8 42.14803983 88.2 18.8 69.4
1.90526125
2 1 27.71962653 88.2 18.8 69.4
1.92747952
2 2 32.22407631 88.2 18.8 69.4
1.73607956
2 7 36.83834524 88.2 18.8 69.4
2.14850817
2 7 30.66220246 88.2 18.8 69.4
1.43620639
2 8 16.87733909 88.2 18.8 69.4
1.67848893
2 8 9.914928315 88.2 18.8 69.4
1.73181887
2 6 43.54900967 88.2 18.8 69.4
1.66561342
2 5 41.81605927 88.2 18.8 69.4
2 1.9735323 16.87339604 88.2 18.8 69.4
2.36109907
2 1 21.52975894 88.2 18.8 69.4
1.51678933
2 4 42.71687078 88.2 18.8 69.4
1.64802669
2 1 30.82224828 88.2 18.8 69.4
1.76916562
2 5 43.33190316 88.2 18.8 69.4

36
1.65098131
2 2 46.440392 88.2 18.8 69.4
2.48941826
2 6 11.80271608 88.2 18.8 69.4
1.95671942
2 3 17.49888795 88.2 18.8 69.4
1.60426847
2 5 48.30015493 88.2 18.8 69.4
2.16496871
2 5 46.94382738 88.2 18.8 69.4
1.82777409
2 6 29.28963228 88.2 18.8 69.4
1.97041913
2 2 14.75491024 88.2 18.8 69.4
2.02339943
2 2 43.04479462 88.2 18.8 69.4
1.92447727
2 6 19.96678556 88.2 18.8 69.4
1.56094937
3 6 36.07446073 104.1 0 104.1
3 1.73880951 43.75172136 104.1 0 104.1
1.78641812
3 4 47.76517976 104.1 0 104.1
1.82798431
3 7 22.96281419 104.1 0 104.1
3 1.83061304 13.6545313 104.1 0 104.1
1.47162741
3 1 44.22008488 104.1 0 104.1
1.87719944
3 6 50.59583963 104.1 0 104.1
1.94472883
3 5 45.85000837 104.1 0 104.1
3 2.33968485 41.48378796 104.1 0 104.1
1.53474912
3 9 44.01782663 104.1 0 104.1
2.34461515
3 1 44.74423937 104.1 0 104.1
1.86726732
3 9 31.51857318 104.1 0 104.1
1.65736776
3 2 62.37748954 104.1 0 104.1
1.93297555
3 9 47.66251019 104.1 0 104.1
2.05804360
3 4 32.57742145 104.1 0 104.1
1.98445794
3 2 35.38000008 104.1 0 104.1
1.92390073
3 4 37.56488196 104.1 0 104.1
1.63462895
3 5 25.96777662 104.1 0 104.1
1.91625395
3 6 41.58177937 104.1 0 104.1

37
1.87606315
3 9 37.04390962 104.1 0 104.1
2.02412958
3 8 37.16706133 104.1 0 104.1
2.13128235
3 3 44.33863965 104.1 0 104.1
1.90145324
3 8 46.30250099 104.1 0 104.1
2.11831659
3 1 52.90151296 104.1 0 104.1
1.92449322
3 4 43.95401635 104.1 0 104.1
1.93091980
3 1 44.66587706 104.1 0 104.1
2.17948888
3 4 44.27461279 104.1 0 104.1
1.60378403
3 8 48.36785487 104.1 0 104.1
1.97498054
3 5 16.51173542 104.1 0 104.1
1.84553889
3 2 19.60821485 104.1 0 104.1
2.19389211
3 9 21.79520976 104.1 0 104.1
1.92251368
4 7 47.55274345 122.9 18.8 104.1
1.84543808
4 4 51.28965187 122.9 18.8 104.1
1.94848802
4 2 23.3952045 122.9 18.8 104.1
1.73037956
4 2 19.9501568 122.9 18.8 104.1
1.53075478
4 3 22.50558037 122.9 18.8 104.1
2.15663442
4 4 43.79291616 122.9 18.8 104.1
1.61805533
4 1 55.05500249 122.9 18.8 104.1
4 1.87057427 45.52614796 122.9 18.8 104.1
1.96846448
4 6 47.50006662 122.9 18.8 104.1
1.70535108
4 5 42.94464605 122.9 18.8 104.1
1.61622291
4 8 46.55192599 122.9 18.8 104.1
1.90994889
4 6 44.31185383 122.9 18.8 104.1
1.40976239
4 6 49.92017409 122.9 18.8 104.1
1.65219951
4 5 20.03359923 122.9 18.8 104.1
1.90593456
4 5 13.03524082 122.9 18.8 104.1

38
1.90004596
4 8 22.36515128 122.9 18.8 104.1
4 1.43811685 48.84850693 122.9 18.8 104.1
1.68728639
4 3 52.0782236 122.9 18.8 104.1
1.40279417
4 9 51.78523336 122.9 18.8 104.1
2.21567867
4 3 26.24255587 122.9 18.8 104.1
1.69337795
4 5 52.43011793 122.9 18.8 104.1
1.75922320
4 7 46.69310522 122.9 18.8 104.1
1.70570846
4 7 15.99792758 122.9 18.8 104.1
1.70793322
4 5 47.54860782 122.9 18.8 104.1
1.70032567
4 6 49.32360968 122.9 18.8 104.1
2.06745790
5 1 60.73390673 138.8 0 138.8
2.27796325
5 5 60.8799555 138.8 0 138.8
5 1.60251886 39.86982785 138.8 0 138.8
2.93399318
5 5 31.8832345 138.8 0 138.8
2.19205621
5 9 46.81418942 138.8 0 138.8
2.51096633
5 8 57.1763109 138.8 0 138.8
2.06662443
5 3 47.7372315 138.8 0 138.8
1.90857582
5 1 59.02079924 138.8 0 138.8
2.13044137
5 6 60.48804656 138.8 0 138.8
2.24222832
5 1 28.05571446 138.8 0 138.8
1.80465259
5 8 22.70835604 138.8 0 138.8
2.20211497
5 6 23.29434305 138.8 0 138.8
1.92113099
5 1 53.61902232 138.8 0 138.8
2.16156966
5 7 53.32808027 138.8 0 138.8
2.12603417
5 3 54.42737922 138.8 0 138.8
2.15081053
5 3 17.38897499 138.8 0 138.8
2.02285433
5 7 56.89359225 138.8 0 138.8
5 1.75992429 42.01775141 138.8 0 138.8

39
8
2.14128575
5 4 17.11514681 138.8 0 138.8
2.41631036
5 9 16.43806912 138.8 0 138.8
1.86723947
5 5 17.70488773 138.8 0 138.8
1.86841925
5 6 16.91823482 138.8 0 138.8
2.13096234
5 7 57.77336908 138.8 0 138.8
1.86846115
5 5 63.71962359 138.8 0 138.8
2.40287728
5 6 37.1944242 138.8 0 138.8
2.46955095
7 7 76.71786065 196.8 58 138.8
2.30010274
7 6 77.51793977 196.8 58 138.8
2.07366075
7 8 32.67117623 196.8 58 138.8
1.68510269
7 3 25.83405362 196.8 58 138.8
1.83360896
7 6 63.89367986 196.8 58 138.8
2.07806198
7 7 50.57740491 196.8 58 138.8
7 1.97377727 50.66437657 196.8 58 138.8
1.86740750
7 8 60.94945657 196.8 58 138.8
1.85273782
7 4 68.91893147 196.8 58 138.8
1.71283610
7 5 36.23174991 196.8 58 138.8
1.66479337
7 9 66.65402891 196.8 58 138.8
1.75855527
7 9 59.50769958 196.8 58 138.8
1.83303701
7 1 46.72279808 196.8 58 138.8
7 1.67725246 44.34017442 196.8 58 138.8
1.82129169
7 8 22.09263919 196.8 58 138.8
1.95988352
7 4 65.99648525 196.8 58 138.8
1.81010679
7 8 54.30637851 196.8 58 138.8
7 2.42091147 79.81192301 196.8 58 138.8
1.93869004
7 9 68.14438483 196.8 58 138.8
1.69801153
7 3 24.82211701 196.8 58 138.8

40
2.14409795
7 2 75.48628935 196.8 58 138.8
1.88883641
7 8 51.42724299 196.8 58 138.8
1.77989582
7 2 13.31976374 196.8 58 138.8
1.42440077
7 1 17.20889311 196.8 58 138.8
2.10119128
7 7 74.61203099 196.8 58 138.8
1.89635365
7 8 66.9550146 196.8 58 138.8
8 2.00544168 67.14417944 217.2 78.4 138.8
1.92515643
8 4 60.68767428 217.2 78.4 138.8
1.65170178
8 2 50.08282111 217.2 78.4 138.8
1.53505283
8 4 46.8580906 217.2 78.4 138.8
1.99867240
8 7 31.69023181 217.2 78.4 138.8
1.83328661
8 8 38.74293633 217.2 78.4 138.8
1.84545367
8 2 33.64283482 217.2 78.4 138.8
1.66177728
8 4 59.03608676 217.2 78.4 138.8
1.86697270
8 1 64.98376698 217.2 78.4 138.8
8 1.87733057 21.03974067 217.2 78.4 138.8
1.99234364
8 3 64.96189583 217.2 78.4 138.8
1.84077294
8 2 48.4041533 217.2 78.4 138.8
1.83417198
8 5 37.54353404 217.2 78.4 138.8
1.65237317
8 4 18.31531595 217.2 78.4 138.8
1.22082606
8 4 29.1079063 217.2 78.4 138.8
1.68547698
8 1 24.82404549 217.2 78.4 138.8
1.54458827
8 3 20.52552257 217.2 78.4 138.8
1.54846392
8 9 37.75489608 217.2 78.4 138.8
1.69753137
8 9 47.5081129 217.2 78.4 138.8
1.69763460
8 5 56.4487294 217.2 78.4 138.8
1.30789130
8 1 64.86445952 217.2 78.4 138.8
8 1.76888555 64.68071722 217.2 78.4 138.8

41
1
1.95483261
8 8 77.71734226 217.2 78.4 138.8
8 1.80676404 62.90519809 217.2 78.4 138.8
2.00495409
8 7 67.99705824 217.2 78.4 138.8
1.90406876
8 6 75.29879476 217.2 78.4 138.8

Script F- Matlab script used for entirety of analysis:


clearall
closeall
clc

%loadingdataandcreatingmatrices
data=load('/Users/Tre/Documents/Academics/Subjects/Fall12/MAE
394/Statsproject/Data/data.mat');
data=struct2cell(data);
data=cell2mat(data);
config=data(:,1);
V=data(:,2);
P=data(:,3);
T=data(:,4);
D=data(:,5);
TnoD=data(:,6);
s_index=[1262753548485109110134135160161186];

%Firstmultipleregression
V1=[V(1:26);V(54:84);V(110:134)];
P1=[P(1:26);P(54:84);P(110:134)];
T1=[T(1:26);T(54:84);T(110:134)];
[l1,trash]=size(V1);
unos1=ones(l1,1);
X1=[unos1,T1,V1];
Y1=P1;
[b1,bint,r1,rint,stats1]=regress(Y1,X1,0.05);
b1
stats1
Yhat1=Y1r1;
count1=1:l1;
scatter(count1,Y1)
holdon
scatter(count1,Yhat1,'r')
title('Fig.1.1regressionofonlyfirstrowconfigurations')
xlabel('trial(#)')
ylabel('practicalforce(N)')
%firstsignificancetests
SSresid1=sum(r1.^2);
SEresid1=sqrt(SSresid1/(l12));
SET1=sqrt(sum((T1mean(T1)).^2));
SEbT1=SEresid1/SET1;

42
tT1=b1(2)./SEbT1
SEV1=sqrt(sum((V1mean(V1)).^2));
SEbV1=SEresid1/SEV1;
tV1=b1(3)./SEbV1
tcritical1=tcdf(.05,l12)
%firstresidualplots
figure
normplot(r1)
title('Fig.1.2normplot')
figure
scatter(T1,r1)
title('Fig.1.3residualplotofT')
xlabel('T')
ylabel('residual')
figure
scatter(V1,r1)
title('Fig.1.4residualplotofV')
xlabel('V')
ylabel('residual')

%Secondmultipleregression
figure
V2=[V(27:53);V(85:109);V(135:160);V(161:186)];
P2=[P(27:53);P(85:109);P(135:160);P(161:186)];
D2=[D(27:53);D(85:109);D(135:160);D(161:186)];
TnoD2=[TnoD(27:53);TnoD(85:109);TnoD(135:160);TnoD(161:186)];
[l2,trash]=size(V2);
unos2=ones(l2,1);
X2=[unos2,TnoD2,D2,V2];
Y2=P2;
[b2,bint,r2,rint,stats2]=regress(Y2,X2,0.05);
b2
stats2
Yhat2=Y2r2;
count2=1:l2;
scatter(count2,Y2)
holdon
scatter(count2,Yhat2,'r')
title('Fig.2.1regressionofonlysecondrowconfigurations')
xlabel('trial(#)')
ylabel('practicalforce(N)')
%secondsignificancetests
SSresid2=sum(r2.^2);
SEresid2=sqrt(SSresid2/(l22));
SETnoD2=sqrt(sum((TnoD2mean(TnoD2)).^2));
SEbTnoD2=SEresid2/SETnoD2;
tTnoD2=b2(2)./SEbTnoD2
SED2=sqrt(sum((D2mean(D2)).^2));
SEbD2=SEresid2/SED2;
tD2=b2(3)./SEbD2
SEV2=sqrt(sum((V2mean(V2)).^2));
SEbV2=SEresid2/SEV2;

43
tV2=b2(4)./SEbV2
tcritical2=tcdf(.05,l22)
%secondresidualplots
figure
normplot(r2)
title('Fig.2.2normplot')
figure
scatter(TnoD2,r2)
title('Fig.2.3residualplotofTwithoutD')
xlabel('T')
ylabel('residual')
figure
scatter(D2,r2)
title('Fig.2.4residualplotofD')
xlabel('D')
ylabel('residual')
figure
scatter(V2,r2)
title('Fig.2.5residualplotofV')
xlabel('V')
ylabel('residual')

%Thirdmultipleregression
figure
[l3,trash]=size(V);
unos3=ones(l3,1);
X3=[unos3,T,V];
Y3=P;
[b3,bint,r3,rint,stats3]=regress(Y3,X3,0.05);
b3
stats3
Yhat3=Y3r3;
count3=1:l3;
scatter(count3,Y3)
holdon
scatter(count3,Yhat3,'r')
title('Fig.3.1regressionofallconfigurationswithoutD')
xlabel('trial(#)')
ylabel('practicalforce(N)')
%thirdsignificancetests
SSresid3=sum(r3.^2);
SEresid3=sqrt(SSresid3/(l32));
SET=sqrt(sum((Tmean(T)).^2));
SEbT=SEresid3/SET;
tT=b3(2)./SEbT
SEV=sqrt(sum((Vmean(V)).^2));
SEbV=SEresid3/SEV;
tV=b3(3)./SEbV
tcritical3=tcdf(.05,l32)
%thirdresidualplots
figure
normplot(r3)
title('Fig.3.2normplot')

44
figure
scatter(T,r3)
title('Fig.3.3residualplotofT')
xlabel('T')
ylabel('residual')
figure
scatter(V,r3)
title('Fig.3.4residualplotofV')
xlabel('V')
ylabel('residual')

%Fourthmultipleregression
figure
[l4,trash]=size(V);
unos4=ones(l4,1);
X4=[unos4,TnoD,D,V];
Y4=P;
[b4,bint,r4,rint,stats4]=regress(Y4,X4,0.05);
b4
stats4
Yhat4=Y4r4;
count4=1:l4;
scatter(count4,Y4)
holdon
scatter(count4,Yhat4,'r')
title('Fig.4.1regressionofallconfigurationswithD')
xlabel('trial(#)')
ylabel('practicalforce(N)')
%firstsignificancetests
SSresid4=sum(r4.^2);
SEresid4=sqrt(SSresid4/(l42));
SETnoD=sqrt(sum((TnoDmean(TnoD)).^2));
SEbTnoD=SEresid4/SETnoD;
tTnoD=b4(2)./SEbTnoD
SED=sqrt(sum((Dmean(D)).^2));
SEbD=SEresid4/SED;
tD=b4(3)./SEbD
SEV=sqrt(sum((Vmean(V)).^2));
SEbV=SEresid4/SEV;
tV=b4(4)./SEbV
tcritical4=tcdf(.05,l42)
%fourthresidualplots
figure
normplot(r4)
title('Fig.4.2normplot')
figure
scatter(TnoD,r4)
title('Fig.4.3residualplotofTwithoutD')
xlabel('T')
ylabel('residual')
figure
scatter(D,r4)
title('Fig.4.4residualplotofD')

45
xlabel('D')
ylabel('residual')
figure
scatter(V,r4)
title('Fig.4.5residualplotofV')
xlabel('V')
ylabel('residual')

46

Vous aimerez peut-être aussi