Sie sind auf Seite 1von 24

Chapter 2

Optimal Semi-Active Damping of Cables


2.1 Introduction

A closed-form solution exists for the optimal damping of cable vibrations using a linear viscous damper representing the prototype linear passive device [5]. An analytical solution also exists for the addition of a positive spring element in parallel to the linear viscous damper [6]. Exact analytical solutions for other damping devices that are either non-linear passive, or are passive and have time dependent properties, do not exist. In general, the exact solution of a non-linear dierential equation is unknown, and only approximate solutions can be obtained [7]. These kinds of damper are classed as, or emulated by, semi-active devices which are the subject of this investigation. In the literature there are several dierent approaches to obtaining approximate solutions for the optimal operation of semi-active devices. From one direction, damping strategies are developed from the linear passive damper theory based either on energy equivalence [8, 9] or mean forces [10]. From the other direction, active control strategies obtained as numerical solutions [11] are modied to exclude active forces, of which clipped LQR/LQG control is a typical example [12]. There are also studies which consider the optimal tuning of combinations of elements, i.e., a Maxwell

2.2. Problem Formulation

element [13]. In all of the above approaches to nding the best semi-active damping strategy, a predened behavior of the damping device is assumed. For example, the optimal tuning of a pure friction device involves varying the single parameter which denes the dimensions of its rectangular trajectory on the force displacement map [14], not modifying the shape of the rectangle itself. In contrast to these approaches this paper attempts to nd an optimal semi-active damping strategy with no constraints or limitations other than those imposed by the denition of a semi-active device. For this problem an evolutionary algorithm (EA) is employed [15], which takes its inspiration from the evolutionary mechanism of natural selection proposed by Darwin. An EA has not yet been applied to the problem of optimal damping as far as the author is aware, however, its interdisciplinary character means that it has already been applied in such diverse elds as genetics [16], road trac control [17], and structural design [18] to name a few. An EA lends itself well to the problem of optimal damping, in that it is able to search the massive tness landscape involved very eciently. Due to the relatively expensive tness evaluation associated with a numerical cable simulation, this is a major advantage of the EA over other optimization techniques.

2.2
2.2.1

Problem Formulation
The System

The system under consideration is a cable with a transverse damper near to one end. If the cable is modeled as a taut string (Appendix A) the cable damper system can be represented in the following mathematical form T 2 v (x, t) 2 v (x, t) m = f (t) (x a) x2 t2 (2.1)

where T is the cable force, m is the cable mass per unit length, v (x, t) is the transverse cable displacement, and f (t) is the force applied to the cable. Furthermore, a is the location of the damper from the left support,

Chapter 2

Optimal Semi-Active Damping of Cables

x is the distance along the cable from the same support, and is a Dirac function.

x c a L

Figure 2.1: Schematic of cable damper system

In this formulation the semi-active device considered can only remove energy from the cable, i.e., it is limited to the development of dissipative forces only. Such a device can be described mathematically as f (t) = c(t) v (a, t) t (2.2)

where c(t) is the damper viscosity which is an exogenously controllable function of time t and must by denition satisfy the equality c(t) 0. A schematic of the entire system is displayed in Fig. 2.1. For the purposes of this investigation, the internal damping of the cable is neglected. The semi-active device is restricted to purely negative powers P (t) such that P (t) 0, where P (t) = f (t) v (a, t) v (a, t) = c(t) ( ) . t t
2

(2.3)

Following this convention, the force displacement trajectory and force velocity trajectory of a spring with positive stiness and linear viscous damper, respectively, show the behaviors depicted in Fig. 2.2.

2.2.2

Optimization Problem

The optimization problem can be formulated as follows: Given a cable with initial energy E , being statically displaced in a single mode n, assuming

2.2. Problem Formulation

10

1 Force (N) 0 1 1 Force (N) 0 Displacement (m) 1

1 0 1 1

0 Velocity (m/s)

Figure 2.2: Force displacement trajectory of spring with positive stiness (left) and force velocity trajectory of linear viscous damper (right)

a sinusoidal mode shape v (x, 0) = sin (nx/L) copt (t) is sought such that E (copt (t)) E (c(t)) [0, tp ] (2.5) (x, 0) = 0, t (2.4)

for all c(t) in the search space, where E is the energy remaining in the cable after a time tp , where 2 tp = n mL2 . T

(2.6)

The time tp dened in (2.6) corresponds to one time period of the mode under consideration. The problem is formulated so that the solution is as close as possibly optimal for a single mode of vibration. To this end, the initial conditions dictate that the energy of the cable is completely in one mode n. However, when applying a nonlinear damper to a cable with energy in a single mode, spillover of energy to higher modes occurs, resulting in a distribution of energy over several modes [8]. By limiting the time tp to a single period, the eects of spillover on the desired conditions of the optimization are minimized. Reducing the number of time periods considered also minimizes the computation time needed to solve the problem when it is reformulated for an EA in the next part of the investigation.

11

Chapter 2

Optimal Semi-Active Damping of Cables

2.3

Methodology

The two main components of the optimization method are an EA and a numerical cable simulation. The following subsections will describe how to adapt the problem formulated in the previous section for an EA, and detail how the two components of the method should be implemented thereafter.

2.3.1

Chromosome Encoding

The time dependent viscosity c(t), given as the candidate solution of the optimization in the problem formulation, is not directly applicable in its continuous form to the encoding of a chromosome in the EA. However, by discretizing c(t), a nite array of real numbers can be obtained which is of a suitable form. [ck ] = [ck = c(kTs )k = 0, 1, 2, ..., Nc 1] (2.7)

Equation (2.7) describes the chromosome as a series of Nc discrete viscous levels with time steps Ts over the optimization time tp .

2.3.2

Cable Simulation

For the simulation of the cable and damper response, the partial dierential equation is dicretized with the spatial sampling interval x and put into the standard form Mv (t) + C v (t) + Kv (t) = f (t) (2.8)

where M is the mass matrix, C is the structural damping matrix, and K is the stiness matrix. From this, the standard state space form can be constructed z ( t) = [ 0 M 1 K 0 I ] f ( t) ] z (t) + [ M 1 M 1 C z (t) = [ v (t) ] v (t) (2.9)

(2.10)

2.3. Methodology

12

where is the shape function which denes the point at which the damper force f (t) acts on the cable. The integration method of choice to solve (2.9) and (2.10) is a fourthorder accurate method based on integration by parts [19]. This method is implemented with the sampling time t.

2.3.3

Evolutionary Algorithm

Although the choice of chromosome has already been discussed, there are many more design parameters to consider for the EA. An overview of the EA ow diagram is presented in Fig. 2.3.

Generate initial population

Test fitness of individuals in population

Terminate algorithm

Is Eq. (11) true?

Select two individuals to reproduce

Create two new variations of selected individuals

Replace two individuals in population

Figure 2.3: Flow diagram

13

Chapter 2

Optimal Semi-Active Damping of Cables

Fitness Function The tness function is slightly dierent to the optimization criterion given in the problem formulation, i.e., the energy E remaining in the cable after the time tp . Instead, the tness function is dened as the total energy removed by the damper from the cable over the time tp , E - E . This is done so that the tness increases towards an optimal solution following convention.

Initial Population and Selection The initial population is created as a set of random individuals based on a Wiener process. An individual of the initial population is given by [ck0 ] = [ck0 = c0 (kTs )k = 0, 1, 2, ..., Nc ] c0 = W (t) min(W (t)) 2 ) W (t) W (s) t sN (0, initial W (0) = 0 [0, tp ] 0 s < t tp (2.11) (2.12) (2.13) (2.14)

where W is a random variable, and N is a normal distribution with standard variance initial . The selection method chosen is rank selection. This means that the probability that a creature is selected is f /F , where f is the ranking of the creature (1 being the least t) and F the sum of the rankings of the entire population. This method outperforms other common methods with tness bias such as roulette selection. Its advantage over the latter method is its independence of tness bias with changing tness gradients [20].

Replacement The method of choice here is rank replacement. For this method the probability that a member of the population will be replaced by a child is D/d, which depends inversely on the ranking of the population member

2.3. Methodology

14

where d is the inverse ranking of a creature and D is the sum of the inverse rankings of the entire population. This method is chosen in favor of other methods such as random replacement, roulette replacement, absolute tness replacement, and locally elite replacement for the same reason as in with the selection method [20].

Crossover When creating two children from two parents, crossover enables the parents genes to be combined to create the childrens genes. This facilitates a more global search of all data structures accessible to the EA. The method chosen here is two-point crossover. This means that two random loci are generated on each of the childrens genes, whereby the rst parents genes are copied to before and after the two loci, and the second parents in between.

Mutation Mutation diers as an operator for diversication from crossover in that it facilitates a more local search of data structures. It does this by making small random changes to the children created by their parents. The chosen method of mutation here is probabilistic mutation with rate , where the rate is usually equal to the reciprocal of the gene length. This means that a mutation occurs at every point on the gene with a probability . The type of mutation that occurs is a Gaussian mutation, which means the addition of a random variable, this variable having a normal distribution with mean 0 and standard variance .

Population Size and Model of Evolution The population size is a trade-o between diversity and computing time. Smaller populations may not contain enough diversity or randomness to nd a good solution to the problem posed to the EA, however, too large a population may lead to the burning of excessive randomness and hence

15

Chapter 2

Optimal Semi-Active Damping of Cables

much larger computation times. The model of evolution used for this investigation is a steady-state evolutionary algorithm. For this model every act of selecting a pair of parents and replacing (or not) two members of the population with children is counted as one generation.

Termination Condition Without knowing the solution to the optimization problem in advance, it is dicult to devise a good termination condition. However, with an idea of the computation time needed to run one generation of the EA, a reasonable condition is to say, stop the algorithm when mating events = 10000 (2.15)

This end condition is reasonably good if the EA is run many times with dierent initial conditions. From these multiple runs, the probability that the nal solution is globally optimal will be greatly increased.

2.4
2.4.1

Results
Parameters of Cable Simulation

A cable is considered with the following set of ctitious properties: tension force T = 300N , mass per unit length m = 2kg /meter, cable length L = 4 meters, and damper position a = 0.02L (2% of cable length). For the cable simulation the number of discrete increments of viscosity Nc = 66, the cable is divided into 100 segments whereby X = L/100, the rst mode is considered n = 1, and the sampling time T = Ts /2.

2.4.2

Parameters of Evolutionary Algorithm

When looking at the evolution of the population over time, the ttest individual in the population of each generation was seen to converge asymptotically to a solution. For dierent parameters of the evolution, this

2.4. Results

16

convergence was quicker or slower. It was found that a population of 600, initial = 3000, = 100/n, and = 1/Nc yielded good results. Figure 2.4 and Fig. 2.5 show the best tness evolution and standard deviation evolution, respectively, for all of the EA runs. The run which performed the best was selected as the solution and is depicted in both plots with a thicker line. Figure 2.6 shows a tness histogram for the selected run, where the frequency is the number of population individuals within each bin.
1 0.9 0.8 Best Fitness 0.7 0.6 0.5 0.4 Best run Other runs

2000

4000 6000 Mating Events

8000

10000

Figure 2.4: Best tness

2.4.3

Results of Evolutionary Algorithm for a Single Cable

The ttest individual in the nal generation is shown in Fig. 2.7. The solution shows two large peaks which represent periods of time in which the damper eectively tries to hold the cable. It is at the lower viscosity levels between these peaks where most of the damper displacement and energy dissipation occur. A consequence of this is that there is still some randomness at the higher levels of the prole, due to its insignicance with respect to the damper performance, and a much smoother prole at lower viscosities for the contrary reason. Like many optimal solutions

17

Chapter 2

Optimal Semi-Active Damping of Cables

0.14 0.12 Fitness Standard Deviation 0.1 0.08 0.06 0.04 0.02 0 Best run Other runs

2000

4000 6000 Mating Events

8000

10000

Figure 2.5: Standard deviation

Figure 2.6: Population tness histogram evolution with 50 bins

2.4. Results

18

the damper holds the cable at certain points in time, not because this optimally removes energy at this time instant, but because it will improve the potential of the damper to remove energy more eciently at later time instants. Figure 2.7 also shows the damper power which illustrates this point.
x 10 Viscosity (Ns/m) 2 1 0 0 0 Power (J/s) 2 4 6 8 0 0.1 0.2 0.3 Time (s) 0.4 0.5 0.6 0.1 0.2 0.3 Time (s) 0.4 0.5 0.6
4

Figure 2.7: Viscosity prole of solution (above) and damper power (below)

In Fig. 2.8, the solution is plotted as force against displacement at damper position, where the path of the plot proceeds in an anticlockwise direction with time due to (2.2). Characteristics emerge from the plot which are remarkably similar to those of a friction device with a negative stiness [4]. A negative stiness acts contrary to a positive stiness in that as deformation occurs an active force is developed which helps the deformation to proceed. The reason that this phenomenon can be observed here, where the damper is limited to passive forces and is clearly not unstable, is that the friction device acts in parallel, keeping the summed forces of the combined elements from becoming active. The negative power seen in Fig. 8 also conrms the dissipative nature of the simulated semi-active damper. A gap in the force displacement loop is present due to the dierence between the theoretical time period for a cable without an external damper used in the simulation, and the actual time period when applying the nonlinear device.

19

Chapter 2

Optimal Semi-Active Damping of Cables

A qualitative explanation can be given for the presence of a negative stiness in the EA result. It is known that for energy dissipating elements, such as viscous dampers and friction devices, the optimal tuning is a compromise between the maximum force developed and amplitude of response, i.e., the maximum dissipation of energy occurs when the enclosed area of the force displacement trajectory is maximized. A negative stiness, as mentioned above, develops a force in the direction of displacement. This means that as the damper goes towards its maximum displacement, the negative stiness can be viewed as acting to increase the amplitude of response, thereby increasing the total energy dissipated.

40 20 Force (N) 0 20 40 60 Cross indicating step in control viscosity 0.008 0.006 0.004 0.002 0 0.002 0.004 0.006 0.008 Displacement (m)

Figure 2.8: Damper force against displacement

A comparison by numerical simulation is now made with two popular damping strategies. These are optimal linear viscous damping and clipped LQR control assuming an ideal semi-active device, i.e., no dynamics and constraints. The weighting of the cost function in the LQR solution is optimally tuned so as to remove the maximum energy from the cable over the optimization time tp . In practise, an observer must be used with fewer states than the continuous cable on which it is implemented. This makes the LQG controller unreliable, due to the unobserved modes. However, this is not the case for a semi-active device which is limited to passive forces [21].

2.4. Results

20

3.8 3.6 3.4 Energy (J) 3.2 3 2.8 2.6 0 0.1 optimal viscosity found with EA optimal linear damping clipped LQR with optimal weighting 0.2 0.3 0.4 Time (s) 0.5 0.6

Figure 2.9: Comparison of damping strategies

Figure 2.9 shows the simulated change in energy of the cable over time for the three dierent strategies. The linear viscous damper, clipped LQR controller, and time varying viscosity found with the EA remove 12.7%, 21.6% and 24.9%, respectively, of the total initial cable energy E over the time tp . Assuming the energy to be completely in the rst mode, the change in energy can be expressed in terms of the logarithmic decrement and the damping ratio , as shown in (2.16) and (2.17). = ln A /A = ln E /E (2.16) = / 4 2 + 2 (2.17) The performances of the respective strategies are equivalent to damping ratios of 1.08%, 1.94%, and 2.28%.

2.4.4

Results for Single Cable at Higher Modes

An identical procedure to that carried out in the previous section was applied to higher cable modes. Table 2.1 shows the damping ratios for all control strategies considered for the rst three cable modes. Although the

21

Chapter 2

Optimal Semi-Active Damping of Cables

Table 2.1: Comparison between damping ratios from EA results with those from optimal viscous damping and clipped LQR control for rst three cable modes

Mode 1 2 3

Optimal Linear Viscous 1.08% 1.05% 1.00%

Clipped LQR 1.94% 1.83% 1.65%

EA Result 2.28% 1.99% 1.85%

performance of the EA result decreases with increasing mode number, it still outperforms the clipped LQR strategy which also shows a decrease in performance with increasing mode number.

2.4.5

Results of Evolutionary Algorithm for a General Cable

The procedure for nding the optimal time dependent viscosity prole was applied to cables with dierent properties. All of the cable properties were varied by plus and minus 50% from the bench mark cable properties, including the tension force T , length L, mass per unit length m and damper position a. Mode 2 is also considered with the benchmark cable properties. In Fig. 2.10, the results from the EA for each variation of cable properties are plotted as force against displacement. It is clear from this plot that the characteristics of the solutions plotted as force against displacement are very similar to those of Fig. 2.8, i.e., a friction element with a negative stiness. It is now interesting to see how the EA results described as a relationship between force and displacement for cables with dierent properties depend on the properties of the cable itself. A qualitative analysis of the results in Fig. 2.10 leads to the redenition of the axes of the plot to a normalized force f (t) against a normalized displacement v (a, t) at damper position. f (t) = f (t)L Tn (2.18)

2.4. Results

22

100 80 60 40 Damper Force (N) 20 0 20 40 60 80 100 0.015 0.01 0.005 0 0.005 Damper Displacement (m) 0.01 m m/2 a a/2 T T/2 m + m/2 T + T/2 a + a/2 L L/2 L + L/2 n+1 0.015

Figure 2.10: Solutions plotted as force against displacement at damper position for cables with dierent properties and modes before normalization

23

Chapter 2

Optimal Semi-Active Damping of Cables

v (a, t) =

v (a, t)L an

(2.19)

Figure 2.11 shows plots of the normalized damper force against normalized damper displacement for cables with the aforementioned varying parameters. It is clear from the close convergence of the dierent trajectories to a single trajectory that the solutions are strongly dependent on the mode number n, and the cable properties length L, tension force T , and damper position a used in (2.18) and (2.19), and weakly dependent on the remaining parameter mass per unit length m.

0.8 0.6 Normalised Damper Force (m) 0.4 0.2 0 0.2 0.4 0.6 0.8 0.004 0.002 0 0.002 Normalised Damper Displacement (m) 0.004 m m/2 L L/2 a a/2 T T/2 m + m/2 L + L/2 a + a/2 T + T/2 n+1

Figure 2.11: Solutions plotted as normalized force against normalized displacement at damper position for cables with dierent properties and modes

2.4.6

Characterization of Solution

As mentioned earlier, the results clearly suggest that the optimal behavior of the semi-active damper pertains to that of a friction device in parallel

2.4. Results

24

with a spring element with a negative stiness, whereby a control law can be dened as f (t)F N = kEA sign(x (t))X (t) + kEA x(t) (2.20)

where X (t) is the displacement envelope of the vibration at the damper position and kEA is dened as kEA = T /a (2.21)

where is the single parameter required for the control law and is independent of the cable properties. To ensure that the desired force is always dissipative when there are errors in the amplitude envelope estimation, the following condition is added f (t)F N if f (t)x (a, t) 0 f ( t) = { (2.22) 0 if f (t)x (t) > 0.

2.4.7

Fitting of Single Parameter to results and Estimation of Displacement Envelope

The displacement envelope X (t), used in the control law of (2.20), is not estimated in real-time within the simulations described subsequently. Instead, an iterative method is used to nd the amplitude peaks at the damper position, where the continuous envelope X (t) is derived by the tting of a smooth function to these peaks. Amplitude estimation in realtime based on a model predictive approach is the topic of future work. The single parameter can now be tted to the mean of the normalized EA results for cables with dierent properties as depicted in Fig. 2.11. The value of which results in the closest match between the force from the control law and the force given by the EA result is = 0.88. In Fig. 2.12, the control law and EA result are compared by plotting the damper force against displacement for both. Figure 2.13 shows the amplitude estimation X (t) and the damper displacement over the simulation time tp . As can be seen in Fig. 2.14, there is almost no dierence between the performance of the control law and the EA solution in terms of the energy removed over the simulation time tp .

25

Chapter 2

Optimal Semi-Active Damping of Cables

1 Normalised Force (m)

control law optimal viscosity found with EA

0.5

0.5 0.4 0.3 0.2 0.1 0 0.1 0.2 0.3 Normalised Damper Displacement (m) 0.4

Figure 2.12: Comparison of damper force against displacement for control law and EA result over simulation time

Displacement (m)

0.005 0 0.005 0.01 0 0.1 0.2 0.3 Time (s) 0.4 0.5 0.6
X(t) v(a,t)

Figure 2.13: Displacement envelope X(t) found by iteration (not in real-time)

2.4. Results

26

4 3.8 Energy (J) 3.6 3.4 3.2 3 2.8 0.1 0.2

control law optimal viscosity found with EA

0.3 0.4 Time (s)

0.5

0.6

Figure 2.14: Comparison of cable energy with control law and EA result over simulation time

2.4.8

Performance of Simplied Solution with Fitted

A simple control law has been dened in (2.20). It has been shown to be almost identical in behavior and performance to the results of the EA over a single time period tp . Now the control law will be applied to a more realistic vibration situation, a free vibration decay test. In this test, the cable is excited to a steady state condition by an excitation force with a distribution and frequency aimed at the rst cable mode. The excitation is then stopped, and the cable is allowed to decay freely. From the decay of the cable, the damping ratio of the system in the given mode of vibration can be measured. This test diers from the problem formulation of the EA in that the cable energy is not restricted to only the target mode of excitation. Due to the nonlinearity of the damper multiple modes of vibration are excited, and it is of interest to assess the control law performance in this more realistic situation. Figure 2.15 shows the displacement at damper position and amplitude estimated by the iterative method mentioned earlier for the decay test. The damping ratios for each successive period of vibration, once the excitation force is stopped, are plotted against time in Fig. 2.16. It can be seen that at the start of decay, the damping ratio is equal to 1.9%, this is 17 percent less ecient than the performance predicted by the EA,

27

Chapter 2

Optimal Semi-Active Damping of Cables

which is an extrapolation of the energy decay found over one period. As the decay proceeds with time the damping ratio drops further to around 1.6%, equivalent to a 30% drop in eciency. The reason for this loss in performance is the presence of multiple modes of vibration, which become more prevalent once the excitation force is stopped. The same eect can be observed in Fig. 2.17 which shows the energy stored in the cable over the decay time as a percentage of the maximum energy over the whole decay test. In this gure, the clipped LQR controller is included for comparison. In contrast to the control law derived from the EA, the clipped LQR controller performs better when multiple modes of vibration occur, approximately equalling the performance of the EA control law for the rst period of decay and then outperforming it as the decay proceeds since the EA solution is optimized for single mode vibrations.

0.001 Displacement (m) 0.0005 0 0.0005 0.001 0.0015 2 4 6 Time (s) 8 10

X(t) v(a,t)

12

Figure 2.15: Damper displacement and displacement envelope for excitation test

2.4. Results

28

Damping Ratio (%)

1.9 1.8 1.7 1.6 1.5 6 7 8 9 Time (s) 10 11 12

Figure 2.16: Damping ratio against time during decay

100
Control law Eq. (15) EA single mode performance Clipped LQR

Energy Percentage (%)

80

60

40

20

9 10 Time (s)

11

12

13

Figure 2.17: Percentage of energy at excitation turn o over decay time

29

Chapter 2

Optimal Semi-Active Damping of Cables

2.4.9

Performance of Control Law for Multiple Mode Vibrations

Very often it is observed that multiple modes of vibration occur in cables. Therefore, it is of interest to test the proposed control law for multiple mode vibrations. As before, a decay test is considered for the assessment of the control law performance. However, instead of a single mode excitation, an excitation consisting of modes 1 and 3 will be used where the maximum excitation force for each mode is identical. Figure 2.18 and Fig. 2.19 show the simulation results from this test where it can be observed that the presence of mode 3 has no detrimental eect on the damper performance compared to the performance of the control strategy for single mode excitations, in fact there is even a slight improvement.

0.0015 0.001 Displacement (m) 0.0005 0 0.0005 0.001 0.0015 0 2 4 6 Time (s) 8 10

X(t) v(a,t)

12

Figure 2.18: Damper displacement and displacement envelope for mixed mode excitation test

2.4. Results

30

Damping Ratio (%)

2 1.9 1.8 1.7 1.6 6 7 8 9 10 Time (s) 11 12

Figure 2.19: Damping ratio against time during mixed mode decay

Das könnte Ihnen auch gefallen