You are on page 1of 38

UNIVERSITY of CALIFORNIA SANTA CRUZ TRACTION ELEVATOR DYNAMICS AND CONTROL A thesis submitted in partial satisfaction of the requirements

for the degree of BACHELOR OF SCIENCE in PHYSICS by Ahmed Mahmoud 20 March 2009

The thesis of Ahmed Mahmoud is approved by:

Talal Rabiah, P.E. Technical Advisor

Professor Bruce Rosenblum Thesis Advisor

Professor David P. Belanger Chair, Department of Physics

Copyright c by Ahmed Mahmoud 2009

iii

Acknowledgements
Im endlessly grateful to Tal Rabiah and Bruce Rosenblum, for encouragement and assistance throughout the writing of this paper; and to my parents, who managed to keep me on track through to completion. Also many thanks to all my fellow UCSC physics students for being the thesis woe bandwagon. Where applicable, good luck on your theses, congratulations on nishing them, and best wishes for you in The Real World.

iv

Contents
Acknowledgements 1 Introduction 2 Unforced Traction Elevator Dynamics 2.1 Kinetic Energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Potential Energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3 Traction Elevator Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Controller Design 3.1 Desired Behavior . . . . . . . 3.1.1 Design Goals . . . . . 3.1.2 Design Kinematics . . 3.2 Motor Control . . . . . . . . 3.2.1 Feedback Control . . . 3.2.2 Measurement Schemes 4 Conclusion Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii 1 5 6 8 10 12 12 12 14 25 26 29 32 34

1 Introduction

Moving around on at surfaces is a relatively trivial task Newtons rst law makes sure of that: once you get an object moving, you dont have to do much to keep it moving, unless theres some force being impressed on it that you have to ght against. Unfortunately, moving up, perpendicular to those at surfaces, is much less trivial. Gravity provides an unremitting force that you have to work against whenever you want to get anything up higher than it started. As a result, humans have been inventing machines to help them hoist loads up and down for centuries. While the origin of hoisting machines is unclear, engravings on Hateriis tomb whose construction was completed sometime around 100BC depict a cranelike lift powered by men running in a giant hamster wheel. In 1743, the rst recorded passenger elevator conveyed King Louis XV of France from his balcony at the Palace of Versailles to his mistress room one storey up [2]. The rst passenger elevator enjoyed by more than two people at a time was the ascending room installed at Londons Colosseum by eminent architect Decimus Burton. The ascending room lifted visitors to the roof of the Colosseum to view a panorama of London painted by William Horner (an actual panorama, it is reported, would only have been accessible in the midnight hours even to people who had access to tall buildings because thats when Londons pollution was at its least opaque) [5].

Figure 1.1: Engraving on Hateriis tomb, dated 110 BC. On the left you can see a cranelike, humanpowered hoist.

The big innovation in elevator technology, however, didnt come until Elisha Otis 1853 demonstration at the Crystal Palace Exhibition. Otis demonstrated an elevator that wouldnt plummet to its doom if the cable broke, and he called this revolutionary device the safety elevator. The invention of the safety elevator broke the static friction keeping passenger elevators from becoming mainstream, and by 1857 passenger elevators were deemed trustworthy enough to earn spots in high-trac oce buildings (installed by the new and expanding Otis elevator company). Now that the dam had burst, the ood of innovations came pouring out: patents involving multiple hoist ropes were issued, newfangled electric motors were embraced in 1887, and 1889 marks the birth of the ubiquitous and inexpungible traction elevator.

Figure 1.2: Safety elevator demonstration by Elisha Otis, 1853. Otis demonstrated at the Exhibition by standing on the platform while the rope was cut. His condence paid o: Otis is still one of the largest, most successful elevator companies in the world.

The traction elevator was an improvement over traditional, well-bucket type designs, which put unnecessary strain on the drive motor by spooling up the cable and storing this extra weight on the drum. On a traction elevator, the cable is merely draped over the drum, the car is attached at one free end, and the other end is weighted with a counterweight. Then the motor shifts the whole car/cable/counterweight system back and forth by the traction of the cable against the rotating drum. No spooling necessary! In this paper, we will analyze the dynamics of these traction elevators. This means studying both the unforced dynamics of a traction elevator-like model (a traction elevator without a motor) so that we can see what the natural tendency of the system is; and the way the motor is used to force the system from oor to oor in a useful way. The former is a problem in classical mechanics; and we will spend considerable time on the latter, a problem involving both kinematical techniques

4 and concepts from control theory. In the rst chapter, well model the elevator as a one-dimensional classical system, gure out the Lagrangian function, and solve the Euler-Lagrange equations to get the systems equation of motion. Well integrate these equations to get the motion itself. In the second chapter, well decide what we actually want the elevator to do, translate our vague desires into statements about the kinematics, and then design a controller that will coerce the motor into bringing about these kinematics.

2 Unforced Traction Elevator Dynamics

Traditional cable elevators from the mid-1800s consisted of a drum with an attached cable that hung down into the hoistway and connected to the top of the elevator car. iThe var was moved by turning the drum with a motor and allowing the cable to wrap or unwrap if the car was being lifted or lowered, respectively. This method was extremely tough on the hoisting machinery: you needed a powerful motor because you had to apply an enormous torque to directly lift the car, and you needed a big drum to hold all the cable youd spooled up. The common case of steel cable was particularly troublesome, because you cant bend steel cable around an extremely tight angle or it will wear severely. So old-fashioned elevators used a hefty drum with a large diameter to provide the gentle curvature required. Traction elevators were invented to make elevator machines (the lifting mechanism itself) far more graceful and ecient. With a traction elevator, the cable is draped over the drum, the car is attached to one end, and a counterweight is attached to the other end. The counterweights mass is tuned to balance the mass of the cars structure (typically, its chosen to be as massive as the car plus 45% of a full loads mass) [4]. The motor then simply shifts the cable back and forth over the drum using friction hence the name traction elevator. With this method, the motor only needs to provide enough torque to overcome the dierence in weight between the car and counterweight, which is comparably small, and the drum is relieved of its storage responsibility since none of the cable gets spooled up as in the old-fashioned designs.

Figure 2.1: Traction elevator vs. Traditional hoist. The hoist drum in the traditional design had to spool up the cable, while the traction elevator just needs to shift the cable back and forth across the drum.

There are two parts to the analysis of the dynamics of these traction elevators. The rst part is an analysis of the unforced dynamics: if the motor was disconnected and the drum was allowed to move freely, how would the car move? The second part is a treatment of the forced dynamics, i.e. the controller design. If we want the system to move a certain way, what signals do we have to send the motor to achieve that motion? The rst part is a problem in classical mechanics, and the second part is a problem in control theory. Lets start by obtaining the unforced dynamics.

2.1

Kinetic Energy
Abstractly, a traction elevator is a one-dimensional Lagrangian system consisting of four

parts: a car, a counterweight, a drum, and a connecting cable. The system is one-dimensional because once you specify the height of the car above the ground, you know the position of the counterweight, the net rotation of the drum, and the position of all the points on the cable because

7 the cable is taut under its load1 . Of course for any conguration of the system, the height is easy to extract. So we have a one-dimensional parametrization of the traction elevator: the height of the car above the ground. Then the kinetic energy is T = Tcar + Tcw + Tdrum + Tcable (2.1)

1 2 and Calling the height of the elevator above the ground h, we can write Tcar = 2 mcar h

Tcw =

1 2 2 mcw h .

Tdrum is slightly more involved. Tdrum is the rotational kinetic energy of the where I is the drums moment of inertia and coordinatizes the rotation of

drum, given by

1 2 2 I ,

the drum. We have to write in terms of h. Luckily this isnt too hard. Dening = 0 to be the initial angle of the drum (that is, = 0 when h = 0), for a drum of radius R, the angle that generates a change in car height of h is =
h R. d h 2 So Tdrum = 1 2 I ( dt R ) = 1 2 2R 2 I h .

The linear speed

of the cable is the same as the linear speed of the car. We will write the cables mass as a linear
1 2 . Adding it all together, Lh mass density times the length of the cable L to obtain Tcable = 2
1 assuming an ideal cable that doesnt stretch or contract in response to external degrees of freedom- this is moreor-less true in practical applications, at least, true to within reasonable tolerances; then the cable wont have any wiggles in it, making the position of any point along its length well-determined

8 the kinetic energy for our traction elevator is T = I 2 L 2 mcar 2 mcw 2 h + h + h + h 2 2 2R2 2 (2.2)

2.2

Potential Energy
The total potential energy is the sum of the gravitational potential energy of each compo-

nent. U = Udrum + Ucar + Ucw + Ucable (2.3)

Since the drum is symmetric as it rotates and its height is xed, it makes no contribution to the potential energy. That is, we can take Udrum = 0 (2.4)

The car contributes Ucar = gh(mcar ). The counterweight contributes Ucw = g (mcw ), where is the height of the counterweight above the ground. Lets name the height of the center of the drum H . We know that the distance from the ground beneath the car, all the way up the cable, around the top of the drum and back down to the ground beneath the counterweight is 2H + R. So, if you start adding distances from the ground beneath the car, you pick up an h to get to the bottom of the cable, an L for the length of the cable up and around the drum, and then you need more distance to get from the bottom of the cable (where the counterweight is connected) to the ground. So we can write h + L + = 2H + R, or = 2H + R L h and so the potential energy from the car and counterweight together is Ucar + Ucw = gh(mcar ) + g (2H + R L h)mcw = gh(mcar mcw ) + g (2H + R L)mcw (2.5)

Like with the drums energy, we can neglect the second term in the expression above because its a constant, leaving us with: Ucar + Ucw = gh(mcar mcw ) (2.6)

9 To analyze the contribution the cable makes to the potential energy, we split the cable into three parts. The rst part is the semicircle that always sits on top of the drum. If we impose the reasonable requirement that the car is never pulled all the way over the drum to the other side, a portion of the cable will always be in contact with the upper half of the drum. We can neglect the potential energy from this portion of the cable because its constant. The other two parts dangle on either side of the drum. Their potential energy will be given by the weight of the entire dangling portion times the height of that portions center of mass above the ground. Since were assuming the cable has a uniform mass density, the center of mass of each dangling portion will be located halfway up that portions length.

Figure 2.2: The three cable segments: dangling above the car (green), the semicircular portion on top of the drum (blue), the dangling portion above the counterweight (red).

Recall that the height to the center of the drum is H . This height corresponds to the elevations of the top ends of the two dangling cable portions, since the dangling portions start where the constant semicircular portion ends. If the lengths of the two hanging portions are labeled lcar

10 and lcw , the potential energy contributed by the cable is lcar lcw Ucable = (lcar )g (H ) + (lcw )g (H ) 2 2 (2.7)

Since H = lcar + h because the bottom end of the cable portion is connected to the car, and lcar + lcw = L R, since R is the length of the neglected semicircular portion of the cable, we can solve for the ls in terms of h and constants, then remove pesky constant terms in the usual way, to obtain the potential energy of the cable: gh (3H + R L h) Ucable = 2 (2.8)

2.3

Traction Elevator Dynamics


So the Lagrangian, L = T U for our whole traction elevator becomes L= 1 2 gh(mcw mcar ) gh (3H + R L h) (mcar + mcw + I/R2 + L)h 2 2 (2.9)

The unforced dynamics of our one-dimensional system will obey the single Euler-Lagrange equation: d dt L h = L h (2.10)

Which for our particular Lagrangian becomes


= g (mcar mcw ) + 2 (3H + R L + 2h) h (mcar + mcw ) + L + I/R2

(2.11)

These are the unforced dynamics of our system. This is an inhomogeneous linear dierential equation with constant coecients: = h + h where =g (mcar + mcw ) + L + I/R2 (2.13) (2.12)

11 and = g (mcar mcw ) + 2 (3H + R L) (mcar + mcw ) + L + I/R2 (2.14)

We can integrate this to get the unforced kinematics: h(t) =


+ C1 e t + C2 e t

(2.15)

where C1 and C2 are arbitrary constants determined by initial conditions. There are 2 constants since this was a second order dierential equation.

12

3 Controller Design

Lets now turn our attention to the problem of controller design. If the traction elevator will naturally obey the dynamics obtained in the previous section, what inuence does the motor need to exert in order to get the elevator to go where we want it to go?

3.1

Desired Behavior
In order to specify how we need to manipulate the motor to force the desired motions,

we should discuss what the desired dynamics actually are. There are certain design criteria that a good elevator ride should achieve. Chiey, the ride should be smooth and comfortable, take as little time as possible, and should reach the desired oor with good positioning accuracy. With a little consideration, we can translate these stated goals into a kinematical description of the desired dynamics.

3.1.1

Design Goals
Elevator passengers dont like to sit around idly waiting for the elevator to arrive at their

oor, nor do they like to stand around awkwardly in the moving car for very long. A slow elevator can cause a lot of frustration, especially if the passenger works or lives very high up. So an important design goal is speed: the elevator should get from oor to oor as quickly as possible so that the inescapable elevator ride occupies the smallest possible portion of a busy passengers day. This isnt that big a problem for our controller, its really just a problem for whoever has to pay for the motor!

13 So we wont spend much time with this important but trivially satised criterion. For concreteness, a reasonably fast elevator goes about 2 m/s at top speed[4]. Although we want the elevator to go as fast as possible from oor-to-oor, there is a limit on how fast were allowed to accelerate to that speed. It would be uncomfortable for most passengers if the elevator took o like a rocket, plastering the riders to the oor, or if the elevator ducked down from underneath them, leaving them oating weightlessly. Generally, the acceptable acceleration levels for passengers are gentle, about 1.5 m/s2 [1], or about 0.15g , where 1g is the acceleration due to gravity. This value was found through the experience of elevator designers through the ages, after they eventually realized that some of the older elevators accelerated too ercely. While acceleration is an important factor to consider when designing for passenger comfort, the dynamical quantity that has the greatest eect on the rider is jerk, the rate of change of acceleration. Jerk is what gives you that feeling in your stomach when you crest a hill on a rollercoaster. An elevator that jerks too severely will leave passengers feeling nauseous or disoriented. A reasonable maximum jerk has been found, again experientially, to be about 2 m/s3 . Additionally, to minimize discomfort, the jerk shouldnt endure very long, and should be kept roughly constant [1]. Finally, a good elevator is required to level extremely accurately. That is, the elevator should stop with the oor of the car lined up with the oor of its destination. Typical leveling accuracies come within a few millimeters, roughly 10 mm [1]. This is a tricky thing to do, especially since the accuracy is determined by how well the elevator stops at the oor. So the elevator controller needs to level accurately without exceeding the acceleration and jerk limits by stopping too abruptly. A common strategy is to decelerate the car nearly to a stop when it comes somewhat close to the oor, and then coast slowly to the desired position. The ecacy of this solution still relies on the controllers ability to decelerate the car into a region very near the desired position so that the time spent adjusting the level is minimized wasting time readjusting the cars position would violate the design goal of minimizing the travel time! These are the most important design goals. To summarize, our elevator ride needs to satisfy

14 these constraints: 1. The jerk should be brief, and shouldnt exceed jm = 2 m/s3 . 2. The acceleration shouldnt exceed am = 1.5 m/s2 . 3. The velocity should be as high as possible without violating the jerk or acceleration constraints. Elevators generally can go roughly vm = 2 m/s. 4. The position of the car after deceleration should be as close as possible to the position required to level with the destination oor.

3.1.2

Design Kinematics
Instead of tackling the problem of generating a motor control for the entire interoor

motion in one fell swoop, well break the motion up into phases where the elevator dynamics are qualitatively dierent. It will be clearer later, but this division of the motion is more than just a useful convenience for doing the math: dierent phases necessitate dierent control methods. The elevator motion is, at rst glance, extremely straightforward: accelerate from rest to the maximum velocity, coast until the car is near the destination oor, and then decelerate. While this is a good general outline, the details of the kinematics are a bit more complicated. For one, we cant just immediately start accelerating at our maximum acceleration am , because that means that we jerked to that acceleration instantaneously, violating the constraint that the jerk have magnitude smaller than jm . Another complication is that it takes a nite distance to decelerate the car to a stop, so we need to anticipate that distance when we get near the destination oor. There are three qualitatively dierent phases of motion for the elevator: accelerating, coasting at a constant velocity, and decelerating. But now the details enter: the accelerating phase is made up of three subphases: ramping up, holding the acceleration, and ramping back down. The decelerating phase has the same subphases, but with oppositely-signed jerks and accelerations. To simplify our analysis, well start by analyzing the kinematics of the accelerating phase, and then show

15 that we can treat the de celerating phase as a mirrored version of the ac celerating phase. Proving this will let us take a step back and study the entire elevator motion simply as a constant velocity phase bookended by complementary accelerations that we fully understand.

Accelerating and Decelerating We didnt set a limitation on how quickly were allowed to ramp up the jerk, so we assume that were capable of instantaneously maxing the jerk out to jm . We want to keep the jerk as constant as possible once weve obtained this value, though, because thats one of our design constraints. This will cause the acceleration to increase linearly. Another constraint dictates that we want the jerk to endure for the shortest duration possible. So once the elevator car reaches its maximum acceleration am , we drop the jerk back to zero. This ends the rst subphase of an overall acceleration. For the second subphase, the jerk remains zero, and the acceleration is held at its maximum value am . We allow this acceleration to continue as long as possible (precisely how long that is will be addressed later because it turns out to be kind of a tricky issue!) before the beginning of the third and nal subphase. The last subphase is a mirror image of the rst one: we slam on the brakes by setting the jerk instantaneously to jm , and dont switch the jerk o until the accelerating phase terminates when the acceleration reaches zero. At this point, a nite amount of time will have passed, the car will have picked up some residual velocity and it will have displaced a little bit. Thats great ramping up the velocity is the entire point of accelerating in the rst place. An accelerating phase then looks like this:

16

Figure 3.1: The subphases in a single acceleration phase: ramping up (green), holding (blue), ramping down (red)

The subphases are characterized by the following kinematics: j1 (t) = jm , j2 (t) = 0, a2 (t) = am , j3 (t) = jm . Well work with the positive acceleration for concreteness, but the negative kinematics can be obtained by simply switching the signs on jm and am . In order to understand the accelerating kinematics, we need to know where the subphase boundaries are. Well start by nding the boundary times, tk , and it turns out to be more convenient for our control system to convert these times into boundary positions, sk . The rst boundary at t1 occurs when the acceleration reaches its maximum, i.e. a1 (t1 ) = am . The [t0 , t1 ] range has a constant, maximum jerk. So then the acceleration obeys a1 (t) = jm t and we want to nd t1 so that am = a1 (t1 ) = jm t1 which means t1 = am jm (3.3) (3.2) (3.1)

17 In fact, for the range [t2 , t3 ], the kinematics are extremely similar. The range is dened to have a constant, maximally negative jerk: j3 (t) = jm . Hence a3 (t) = a2 jm (t t2 ) (3.4)

where a2 is shorthand for at (t2 ), the acceleration at the end of the second subphase, which becomes the initial acceleration in the third subphase, so that a3 (t3 ) = a2 jm (t3 t2 ) meaning that we want t3 t2 = a2 jm (3.6) (3.5)

but the acceleration is constant and equal to am for the entire second phase, so that the third phase starts o with a2 = am and t3 t2 = am = t1 jm (3.7)

Evidently, the duration of the ramping down phase is the same as the duration of the ramping up phase. This shouldnt be surprising, since it makes sense that if you push in one direction for some amount of time and then push back in the opposite direction for the same amount of time, you undo the eect of pushing in the rst place. There are some situations where this intuition fails, however, so we will be careful to prove these kinds of symmetry results instead of assuming them to be true. These kinds of results are extremely useful, though, and a similar but slightly more complicated version of the above symmetry will be key when we analyze the overall elevator motion. This proves that the ramping time is well-dened, so we can distinguish it with its own symbol: tr t1 = t3 t2 (3.8)

The duration of the holding subphase ([t1 , t2 ]) cant be obtained without more knowledge of the overall picture, so we wont try and gure it out at the moment. Instead, well just dene a convenient label for its duration because its extremely important: th t2 t1 (3.9)

18 The next pressing question we need to answer is: what is the velocity achieved at the end of the accelerating phase? We can nd the velocity with the following nifty trick: v 3 = v0 + a t3 (3.10)

where v0 is the initial velocity and a is the average acceleration over the range [t0 , t3 ]. We can break this up: v3 = v0 + a 1 t1 + a 2 (t2 t1 ) + a 3 (t3 t2 ) (3.11)

Since in the [t0 , t1 ] subphase the acceleration is increasing uniformly, the acceleration spends equal time being higher and lower than half the maximum acceleration obtained. Since this maximum acceleration is am at t2 , and further the acceleration is constantly am from t1 until t2 , the average accelerations are given by a 1 = am /2 a 2 = am a 3 = am /2 (3.12) (3.13) (3.14)

where the value of third average acceleration can be justied with the same argument as the rst. Then the nal velocity is v3 = v0 + t1 am /2 + (t2 t1 )am + (t3 t2 )am /2 = v0 + tr am /2 + th am + tr am /2 = v0 + am (tr + th ) = v0 + am t2 (3.15) (3.16) (3.17) (3.18)

It turns out that the holding time th is the same for an accelerating phase bringing the velocity up from 0 to v3 or a decelerating phase bringing the velocity down from v3 to 0. v3 th = = 0 + am (tr + th ) v3 tr am (3.19) (3.20)

19 while for an ending velocity of zero, starting velocity of v3 , and acceleration of am , you get 0 th = v3 am (tr + th ) = v3 tr am (3.21) (3.22)

Behold! this checks out: apparently, accelerating to cruising speed takes the same amount of time as stopping the car once its at the cruising speed! This symmetry is the rst step towards showing the most important result of this section: that accelerating to cruising speed and stopping a car going at cruising speed covers the same distance. With this result in hand, we can greatly simplify our picture of the overall elevator motion. But lets not get ahead of ourselves. The bridge to proving that result is an acquisition of the position kinematics for the entire course of the acceleration phase1 .

Acceleration Kinematics For each subphase, well integrate the characteristic kinematics until we get the position function for the whole acceleration. The initial conditions used in the integration of each subphase will be inducted recursively from the subphase before it. Well maintain the convention that a function of motion written without an argument, like s2 , is dened to be the value of that function evaluated at the subphases end time, like s2 (t2 ). The jerk is what we start with: j1 (t) = jm j2 (t) = 0 (3.23) (3.24) (3.25)

j3 (t) = jm Integrating to get acceleration: a1 (t) = jm t

(3.26)

1 Theres probably a really slick, tricky way to prove this, but we need to derive the kinematics anyway, so we might as well use them and make the proof messy but straightforward

20 a2 (t) a3 (t) = am = a2 jm (t t2 ) = am jm (t t2 ) (3.27) (3.28) (3.29)

Here it should be noted explicitly that the time argument is the absolute time in the acceleration phase. So for subphases 2 and 3, durations spent in the subphase are given by the dierence between the time argument and the absolute time that marks the beginning of the subphase. Now velocity: 1 v1 (t) = v0 + jm t2 2 v2 (t) = v1 + am (t t1 ) v3 (t) = v2 + am (t t2 ) = 1 jm (t t2 )2 2 (3.30) (3.31) (3.32)

Finally, integrating one more time we get the position: s1 (t) s2 (t) s3 (t) Since t1 = tr =
am jm

1 = v0 t + jm t3 6 1 = s1 + v1 (t t1 ) + am (t t1 )2 2 1 1 = s2 + v2 (t t2 ) + am (t t2 )2 jm (t t2 )3 2 6

(3.33) (3.34) (3.35)

is a known constant, and v0 is a parameter that we specify, the only

thing we dont know about the acceleration phases motion is the holding time. Unfortunately we still need to defer that discussion. But there is also good news: were now ready to derive the main fact for this section: Let s+ k (t) denote the position function for the k th subphase when jm and am are chosen to be positive. Similarly, s k (t) will be used for the position function when jm and am are negative. Then
+ s+ 3 (t3 ; v0 = 0) = s3 (t3 ; v0 = v3 )

(3.36)

+ is the residual velocity at the end of the motion modeled by s+ where v3 3 (t3 ; v0 = 0). Weve already + derived that v3 = am t2 when v0 = 0.

21 So, we already know that 1 1 2 3 s+ 3 (t3 ) = s2 + v2 (t3 t2 ) + am (t3 t2 ) jm (t3 t2 ) 2 6 1 1 3 = s2 + v2 tr + am t2 r jm tr 2 6 (3.37) (3.38)

Expanding s2 and v2 , and then recursing and expanding s1 and v1 when they show up, we end up with s+ 3 (t3 ) = 1 1 1 1 1 1 2 2 3 2 3 jm t3 r + jm tr th + am th + jm tr + am th tr + am tr jm tr 6 2 2 2 2 6 (3.39)

Which simplies to s+ 3 (t3 ) = For the decelerating version s 3 (t3 ) 1 1 3 = s2 + v2 tr am t2 r + jm tr 2 6 (3.41) 1 1 2 jm t2 r t2 + am t2 2 2 (3.40)

But this time, when the initial conditions are expanded, we pick up factors of the initial
+ velocity, v3 = am t2 , so the expansion this time is

s 3 (t3 ) = am t2 tr

1 1 1 1 2 2 2 jm t3 r + am t2 th jm tr th am th am th 6 2 2 2 1 1 1 2 3 + am t2 tr jm t3 r am th tr am tr + jm tr 2 2 6

(3.42)

+ Simplifying, and remembering that t2 = tr + th , it is indeed found that when v0 = v3

s 3 (t3 )

1 1 2 jm t2 r t2 + am t2 2 2

(3.43)

The Coasting Phase Now we know essentially everything about the accelerating and decelerating phases, including the fact that over the course of our elevator motion, both accelerating and decelerating take the same amount of time and cover the same distance. The only thing we need to know is th : how long we should hold the acceleration at its maximum value. This is a somewhat tricky problem because there are two cases that have dierent expressions for the holding time.

22 To meet the design goal of making the elevator ride as short as possible, we should get the car going as fast as possible. Then the holding time should be whatever time it takes to make v3 = vm . Recall that th = v3 tr am (3.44)

so then if we want to get to top speed, the holding time should be th = vm vm am tr = am am jm (3.45)

This is the value of the holding time in one of the two cases. But if we accelerate all the way to vm , the combination of accelerating and decelerating phases takes up quite a lot of space. Substituting th into our expression for the displacement due to acceleration, we get 2s3 = jm t2 r recalling tr =
am jm ,

v2 vm + am m am a2 m

(3.46)

2s3 =

v2 am vm + m jm am

(3.47)

which for our reasonable maximum values of am = 1.5m/s2 , jm = 2.0m/s3 , vm = 2m/s, is more than 4m! Were unable to move just one oor well overshoot it! A faster elevator will overshoot even more horribly, potentially missing multiple oors. So there is a second case where we need to nd th : the case where the elevator never reaches its maximum velocity and starts the decelerating phase immediately after the accelerating phase. In this case, the distance the elevator travels is exactly 2s3 , so we can nd th given the distance we need to go in order to get to the desired oor. Say that the interoor distance is H . Then H = 2s3 = 1 a2 m t2 + am t2 2 jm 2 (3.48)

We can use the quadratic equation to solve for t2 : t2 = am + jm a2 m + 2am H 2 jm (3.49)

23 Since t2 = tr + th , then th is given by th = am + jm am + jm am a2 m + 2am H 2 jm jm a2 m + 2am H 2 jm (3.50)

= 2

So, in case H is too small to accommodate two full acceleration/deceleration phases,


m choose th = 2 a jm +

a2 m 2 jm

+ 2am H . Otherwise, choose th =

vm am

am jm ,

and reach maximum

velocity. In the second case, the two accelerating phases always cover a constant distance. To get to oors farther away than this distance, the elevator needs to undergo a phase where it coasts at its maximum velocity far enough to make the whole trip cover H . In particular, the coasting time needs to be tc = and then the deceleration phase should begin. H 2s3 vm (3.51)

Global Goal Kinematics Now we have all the information about the desired elevator kinematics! Weve already mentioned, though, that phase boundaries should be labeled by their positions, not times, because positions are more useful for the controller that well be talking more about later. Heres a diagram of the acceleration vs. distance. The transitions are clear on this diagram, and are labeled pk .

24

Figure 3.2: The overall acceleration prole. The coasting phase (blue) may not be present of p5 is small.

p1 is given by the distance covered in the ramping up phase: p1 = s+ 1 = 1 a2 m 6 jm (3.52)

p2 is farther from p1 by the distance covered in the holding phase. Well leave th as th , with the understanding that the applicable th is the shorter one. 1 1 a2 1 m 2 p2 = s+ = p + v t + a t = p + th + am t2 1 1 h m 1 h h 2 2 2 jm 2 (3.53)

p3 is another ramping phase after the holding phase. Its displacement is the same as the displacement in p1 , so p3 = p2 + p1 (3.54)

25 p4 marks the end of the coasting phase: p4 = p 3 + vm t c = p3 + H 2 1 am 1 a2 m am ( + th ) + am ( + th )2 2 jm jm 2 jm (3.55)

From here on, its just another decelerating phase. This one is easy to label because the displacements in this phase match the displacements in the accelerating phase! p5 = p4 + p1 (3.56)

p6 = p4 + p2

(3.57)

p7 = p4 + p3

(3.58)

3.2

Motor Control
So now that we know the way we want the elevator to move, we need to gure out what

signals we have to send the motor via the controller in order to achieve that motion. To do this, we need to understand some of the basic notions of controller design, and we need to be more specic about what sensors and actuators are available to us in this particular application. Once weve got these tools in hand, we can apply them to the goal kinematics derived above and come up with our controller design. Fundamentally, what we need is a method that will drive the sensor measurements to their desired values by controlling the actuators. The control signals we actually send the actuators will be based on the dierence between the desired sensor values and the current sensor values. Before we do anything, though, we have to assume controllability : that the actuators have enough inuence over the elevator car that we can eventually drive the car to a state that exhibits the desired sensor measurements just by controlling the actuators2 . Assuming
2 Actually, what were assuming is reachability. Controllability is usually dened to mean that the output can be driven to zero rather than any desired state. The dierence being somewhat formal, the word controllability is often abused to mean reachability, and is a more evocative term. So, like much of the literature, well use controllability here to really mean reachability.

26 controllability allows us to get down to business in earnest without worrying about whether or not our control scheme will be a Sisyphean failure: without controllability, controlling the system is as hopeless as screaming Latin directions louder and louder to someone who only speaks English.

3.2.1

Feedback Control
There are two types of controller: open and closed loop. An open loop controller simply

directs the actuators, and assumes that theyre responding in exactly the way predicted by the model of the system. Surprisingly, open loop controllers can be used in the real world for some simple systems, but their utility cant compare with those of closed loop controllers. Closed loop controllers actually measure the response of the system to the control signals, making them capable of reconsidering the best signal to send based on what the response actually was. Well construct a closed loop, or feedback controller for our elevator. The mathematical model of a control system is traditionally written in the form: x = f (x, u) y = g (x) (3.59) (3.60)

where x parametrizes the systems state, u parametrizes the control signal (the input), and y parametrizes the sensor measurements (the output). These equations mean that the sensor value depends on the actual state of the system, and the state changes according to what control signal you give it while its in a particular state. The function f is called the system map, and we assume it to be dierentiable. Then we can Taylor expand the system map to get a local linear approximation: x (x, u) = f (x0 , u0 ) + = x (x0 , u0 ) + where
f (x,u) (x0 ,u0 )

f (x, u) (x, u) (x0 ,u0 ) f (x, u) (x, u) (x0 ,u0 )

(3.61) (3.62)

is the Jacobian of f at the point in question. This local linearity means that as

long as we keep our adjustments small, the state will be pushed around in a neighborhood of its starting state it wont jump o to a completely dierent state unless you drive it really hard. (If

27 the Jacobian has a massive determinant, you might need to be extremely gentle with the system. In practice, systems that jumpy usually have their hardware redesigned so that the controller doesnt have to be so careful.) So by assuming local linearity and controllability, we have some idea of how to control this system: nd the dierence between the goal measurements and the current measurements, and apply the actuators in such a way that the current measurements inch towards the goal by a small portion of the dierence. Now, recompute the dierence and iterate this whole process until the dierence goes to zero. This type of controller is called a feedback controller, because the response of the system to the actuators is measured and fed back into the controller so that it can consider the most recently measured deviation from the desired state.

Figure 3.3: A feedback controller. The goal is fed into the system, but the sensor measurements from the elevator are subtracted so that the error continues into the controller. The control signal is then fed to the elevator, and the process starts again.

This reduces our primary question from how do we nd the appropriate actuator signals from the sensor measurements? to how do we nd the appropriate actuator signals from the error between the current measurements and the desired ones? In the simplest case, the output

28 signals to the actuators will be proportional to the error. Keep in mind that there are generally multiple sensor measurements describing a state, and multiple actuators to control. This means that the input and output signals are vectors and the constant of proportionality is then a matrix. So thinking geometrically, in order to drag the state closer to the desired state, the actuator signal should look like the dierence vector between the two states, but stretched or squashed, and maybe rotated3 . In practice, a proportional controller (often abbreviated P Controller) is usually too simple to be a good control system. While the system might reach the goal state, theres no guarantee that it will stay there and not oscillate around the goal forever, or it might stabilize to a constant value very quickly, but might not stabilize exactly where we want it to. These problems are corrected by extending the P controller to a PID controller: a Proportional, Integral, Derivative controller. With a PID controller, not only is the actuator signal proportional to the state discrepancy, but it also takes into account a weighted sum of the integral and derivative of the state discrepancy. That is, the derivative part contributes to the actuator signal as the error keeps oscillating wildly, and so will tend to stabilize oscillation and overshoot. The integral part inuences the actuator signal while the system spends too much time in a state dierent from the desired one, so it will tend to increase the steady-state accuracy of the controller. Each of these contributions are combined in a weighted sum, so a PID controller is specied with its three weights: kp , the proportional weight, or gain ; ki , the integral weight, or reset ; and kd , the derivative weight, or rate.
3 This isnt strictly a reasonable intuition, because while lists of sensor measurements look like Euclidean vectors because theyre a list of real numbers, they dont necessarily transform like Euclidean vectors. So we need to be wary about conclusions suggested by our geometric intuition of these states.

29

Figure 3.4: This is a zoomed-in version of the controller portion of the feedback system. The outputted control signal is a weighted sum of the integral, derivative, and value of the discrepancy between the actual and goal states.

Tuning these controller weights isnt easy, however, but there are many dierent methods with their advantages and disadvantages. PID controller tuning is often done empirically, or with specialty software. For the purposes of this paper, we just need to know that the PID controller is the natural extension to the almost obvious P controller, and that it can be tuned for our system. For more detail, see [3].

3.2.2

Measurement Schemes
Now that we have a general idea of how a feedback controller should work, we need to

know what measurements we need to take over the course of the elevator ride. Since all of our goals are kinematical, all the measurements we need can be determined from the position of the elevator. With a timer, we just take dierence quotients of position measurements at dierent points in time, and we get velocity, acceleration and jerk. Luckily, an extremely clever instrument called an encoder can be used to measure position.

30 An encoder is an apparatus with a laser pointed at a photoreceptor. This apparatus is placed on the outside of the elevator car, and a perforated strip of metal running up and down the length of the hoistway is arranged so that it is in the path of the laser beam on the car. As the car moves, the photoreceptor either sees the laser or doesnt see the laser, depending on whether the car is passing a gap in the perforated strip, or just the solid metal. Then the output from the photoreceptor will look like a series of pulses. By counting the number of pulses, you know the number of gaps the elevator has passed on its trip. The gaps are evenly spaced, so if we know the number of photoreceptor pulses, we know the distance that the car has traveled! Combined with a sensor keeping track of which direction the elevator is moving, and a timer so that we can take derivatives of the displacements measured by the encoder, we are capable of measuring all the kinematical quantities that wed like to control. By the very nature of a PID controller, the quantity that we are measuring will be regulated so that it maintains its goal value. But the controller will do everything in its power to regulate that measurement, and so if we arent careful about which quantities were controlling, other design goals might get violated. For example, if we are trying to maintain a constant velocity, but there is some snag in the hoistway that drops the velocity pretty severely, the controller will notice this discrepancy and work really hard to ramp the velocity back up even if it means violating the acceleration limits! This problem cant be avoided entirely without getting extremely fancy, but it can be mediated by being careful about which quantity we control during each phase. Of course, because the position of the car is what tells the controller when to change kinematics (since there are slightly dierent kinematics in each phase!), we need to measure position constantly. Ideally there will be a microcontroller or something that is constantly polling the position and resetting the controller goal values appropriately when each phase of motion begins. What we need to decide is which quantity should be accountable for meeting its goal measurement during each phase. The measuring schemes are as follows:

31 Phase [0, p1 ] [p1 , p2 ] [p2 , p3 ] [p3 , p4 ] [p4 , p5 ] [p5 , p6 ] [p7 , p8 ] Releveling Goal Quantity j a j v j a j s

The nal phase doesnt appear in our earlier analysis. Our analysis assumes that the controller gets the car to stop right at its destination. In reality, the controller wont do a perfect job, especially since we spent the last phase measuring the jerk : the acceleration is probably correct by the end of that phase, but the velocity will likely be o by a bit, and the position will probably be o more than our design goals will allow. If our controller is reasonably good, the position should at least be very close to within tolerances, but we should clean it up by controlling exactly on position. As long as the elevator car isnt way o, the controller will probably have to make very small adjustments, so the acceleration and jerk limits wont be exceeded, and then by the end of the whole ride we can be completely sure that the passenger is where they wanted to be. We now have a complete description of the controller for the elevator, and a complete knowledge of the kinematics that the elevator should achieve. We also know how the elevator would have behaved without the controller from the previous section. Our analysis of the kinematics of a traction elevator has come to its end. DING!

32

4 Conclusion

In this paper, weve explored the dynamics forced and unforced of the most common elevator in use: the electric traction elevator. We found that its unforced behavior was essentially that of a complicated Atwood machine, and if even slightly unbalanced, the whole system would crash to the ground exponentially fast, and the car with the passengers onboard would be obliterated. We then discussed the design of a device for making the elevator obey our demands on its kinematics (instead of just crashing to the ground like the unforced system would!). We settled on a PID feedback controller that controlled dierent kinematical variables during dierent phases of the elevators motion: accelerating, coasting at constant speed, and decelerating. We found the the accelerating and decelerating phases were themselves made up of ramping up, holding, and ramping down subphases, since the instantaneous rate of change of acceleration was constrained by our kinematical demands. Furthermore, the accelerating and decelerating phases were completely antisymmetric, making analysis of the global motion much simpler because the two inevitable phases both drove the car the same distance. With buildings getting taller and taller, and with computers becoming more and more prevalent in every mechanical system we build, the advance of elevator control technology is necessary and inevitable. For example, huge skyscrapers like Taipei 101 and the Burj Dubai have elevators that are so fast and so tall that their cables experience vibration signicant enough to severely aect the ride quality[6]. The modeling and control of these systems involves dierential equations of fourth order, and involves measuring equipment much more sophisticated than the tried-and-true

33 encoders utilized in this paper. Despite the complex diculties that arise when trying to control these new, cutting-edge elevators, the general machinery remains the same. Like sharks, which are so perfectly tuned to their environment that they havent changed much in a hundred million years, the electric traction elevator will likely continue to serve the towering structures of the future.

34

Bibliography
[1] GC Barney. Elevator Electric Drives. Ellis Horwood Limited, 1990. [2] A Goetz. Up, Down, Across: Elevators, Escalators and Moving Sidewalks. Merrell, 2003. [3] Aidan ODwyer. Handbook of Pi And Pid Controller Tuning Rules. Imperial College Press, 2 edition, February 2006. [4] G Strakosch. The Vertical Transportation Handbook. Wiley, 1998. [5] J Timbs. Curiosities of London. Longmans, Green, Reader and Dyer, 1868. [6] S. R. Venkatesh and Y. M. Cho. Control systems for ultra-high rise elevators. Journal of Dynamic Systems, Measurement, and Control, 123(4):687690, 2001.

You might also like