Control of an automotive electromagnetic suspension system
T.P.J. van der Sande D&C 2011.016
Master’s thesis
Coach(es): ir. B.L.J. Gysen dr.ir. I.J.M. Besselink
Supervisor: prof.dr. H. Nijmeijer
Committee: prof.dr. H. Nijmeijer dr.ir. I.J.M. Besselink dr.ir. J.J.H. Paulides
Eindhoven University of Technology Department of Mechanical Engineering Master Automotive Technology Dynamics & Control
Eindhoven, March, 2011
Acknowledgements
First and foremost, my gratitude goes out to ir. Bart Gysen, my direct supervisor for this project. His ever present support greatly helped me in performing this research. Second, my thanks go out to prof. Henk Nijmeijer, dr. Igo Besselink, dr. Johan Paulides and prof. Elena Lomonova. Their guidance, valuable tips and critical questions often helped me in the right direction. Thirdly, I would like to thank the EPE group for offering me this interesting and challenging research topic in which I could further enhance not only my theoretical but also my practical skills. A special note goes to my roommates, for the interesting discussions. Finally, I would like to thank my family, girlfriend and other friends for their support and encouragement. This made my graduation project much more enjoyable.
i Abstract
The main research goal of this thesis is to determine what performance gains can be achieved with a high bandwidth electromagnetic active suspension. As a baseline vehicle a BMW 530i is used, for which a retrofit electromagnetic suspension consisting of a spring and tubular perma- nent magnet actuator (TPMA) is designed. To design a control system for this actuator, a model of the BMW has been created, which consists of a quarter car model with variable sprung mass, damping coefficient and tire stiffness. As input to this model a road disturbance is used, that was modeled as a white noise source filtered by a first order low-pass filter. To test the performance of the actuator and controllers a full size quarter car test setup is used. As control objectives minimization of the sprung acceleration and dynamic tire compression are used with constraints on the suspension travel and RMS actuator force. The sprung accel- eration is used as an indication for ride comfort and the dynamic tire compression is used as an indication for handling quality. To account for human sensitivity to vibrations, the ISO2631-1 standard is used to filter the sprung acceleration. The suspension travel of the controlled system is limited to the maximum value that the BMW achieved with its spring and damper settings over a given road. Furthermore, the maximum RMS actuator force of 1000 N results from thermal limits. Two control approaches are considered, linear quadratic optimal control and robust control. For the former, a controller is found using a linearized quarter car model. By choosing three weighting factors either comfort or handling can be emphasized. Variations of the plant are accounted for by using robust control. Using frequency dependent weighting, certain frequencies can be emphasized. For instance, human sensitivity to vertical vibrations is incorporated using an approximation of ISO2631-1. By varying this weighting together with the other weighting filters either comfort or handling can emphasized, similar to the linear quadratic control. Measurements on the quarter car setup show that an improvement in comfort of 35 % can be achieved with linear quadratic control. This differs 55 % from the value predicted by simulations. However, this deviation can be explained by friction in the test setup and actuator as well as by uncertainties that were not modeled when designing the LQ controller. In case of the handling controller, measurements do match the simulations better on the smooth road. Dynamic tire compression is stability issues of the controller. With robust control an improvement of 48 % in comfort can be achieved on the setup at the cost of an increase of 99.3 % in dynamic tire compression. In terms of handling, an improve- ment of 17.7 % is achieved, worsening comfort by 10.7 %. Frequency weighting clearly has a desirable effect, as comfort decreases by 6 % for the handling controller on rough road whereas sprung acceleration worsens by 75 %. This means that all vibrations occur outside of the human sensitivity range. Deviations of the measurements from the simulations can be explained by stick slip friction in the suspension actuator as well as vibrations passing through the test setup.
ii Samenvatting
De belangrijkste onderzoeksvraag van deze thesis is wat voor prestatie winst er kan worden be- haald met een hoge bandbreedte electromagnetische actieve ophanging. Als basis voertuig wordt een BMW 530i gebruikt, waarvoor een retrofit tubulaire permanent magneet actuator, bestaande uit een veer en actuator, ontworpen is. Om het regelsysteem van deze actuator te ontwerpen is er een model van de BMW gemaakt dat bestaat uit een kwart voertuig model met variable geveerde massa, dempings coefficient en band stijfheid. Als ingang voor het model wordt een wegverstor- ing gebruikt, bestaande uit witte ruis gefilterd met een eerste orde laag doorlaat filter. Om de prestaties van de actuator en regelaars te bepalen is er een kwart voertuig opstelling gebruikt op ware grootte. Het regeldoel is het minimaliseren van de geveerde acceleratie of de dynamische band in- drukking met als randvoorwaarden de veerweg en RMS actuator kracht. De afgeveerde ver- snelling wordt gebruikt om de mate van comfort te bepalen. De dynamische band indrukking geeft een idee van de kwaliteit van de wegligging. Om rekening te houden met de menselijke gevoeligheid voor verticale vibraties wordt het ISO2631-1 criterium gebruikt. De limiet op de veerweg wordt bepaald door de veerweg van de BMW over dezelfde weg, terwijl de 1000 N actu- ator kracht limiet bepaald wordt door de thermische eigenschappen van deze. Twee controle topologien worden beschouwd, een linear kwadratisch en robuuste regelaar. Voor de eerste geldt dat er een optimale regelaar ontworpen wordt aan de hand van een gelin- earizeerde versie van het kwart voertuig model. Door het kiezen van drie weegfactoren kunnen comfort of wegligging benadrukt worden. Om zeker te zijn dat de regelaar stabiel is met de vari- aties die op kunnen treden in het system wordt een robuuste regeling gebruikt. Deze methode maakt het mogelijk om frequentie afhankelijke weegfilters te gebruiken. Een voorbeeld hiervan is het ISO2631-1 criterium, waarvan een benadering van wordt gebruikt om menselijke gevoeligheid voor verticale vibraties extra te benadrukken. Door dit weegfilter te gebruiken in combinatie met andere weegfilters kan comfort of wegligging benadrukt worden. Metingen op de kwart voertuig opstelling laten zien dat comfort met 35 % verbeterd kan worden met een linear kwadratische regelaar. Dit wijkt 55 % af van de verbetering voorspeld door simulaties. Dit kan echter verklaard worden door wrijving in de opstelling en actuator alsmede door onzekerheden die niet meegenomen zijn in het ontwerpen van de LQ regelaar. De resultaten van de regelaar die ontworpen is voor wegligging komen beter overeen met de simulaties. Een verbetering van 48.5 % kan worden behaald. Dit kon echter niet worden geverifieerd op de ruwe weg door instabiliteit van regelaar. Met de robuuste regelaar kan een verbetering van 48 % worden gehaald in comfort op de test opstelling ten koste van een verslechtering in dynamische band indrukking van 99.3 %. Voor wegligging kan er een verbetering van 17.7 % behaald worden waarbij comfort met 10.7 % verslechterd wordt. De toepassing van frequentie afhankelijke filters heeft een gewenst effect aangezien comfort met maar 6 % wordt verslechterd terwijl de verticale acceleratie met 75 %
iii verslechtert. Dit betekent dat alle vibraties optreden buiten het gebied waar mensen het meest gevoelig zijn. Verschillen tussen de metingen en simulaties kunnen verklaard worden door ’slick- slip’ wrijving in de actuator en in de test opstelling. Verder spelen vibraties die via het frame van de test opstelling naar de sensoren komen een rol.
iv Used Symbols and Abbreviations
Abbreviations
Name Meaning ABC Active body control DOF Degrees of freedom LQ Linear quadratic LQG Linear quadratic gaussian LQOF Linear quadratic output feedback LQR Linear quadratic regulator NS Nominal stability RC Robust control RS Robust stability RP Robust performance RMS Root mean square TPMA Tubular permanent magnet actuator VAG Volkswagen Audi group
v Symbols
Symbol Meaning ν Cut off frequency of road signal α Side slip angle dr Lateral tire damping ds Sprung damping Croad Road actuator controller 1 Uncertainty matrix Ei Actuator phase back EMF Fact Suspension actuator force Fra Road actuator force Fy Lateral tire force iˆ Current amplitude J LQ control objective kEi Actuator EMF constant kr Lateral tire stiffness kra Road actuator spring stiffness ks Sprung stiffness kt Tire stiffness Li Actuator phase inductance mc Contact patch mass ms Sprung mass mra Road actuator mass mu Unsprung mass µ Structured singular value ns Spatial frequency Q Weighting matrix for LQ control R Controllability matrix Ri Actuator phase resistance ψ Gain that defines road amplitude τp Pole pitch ts Sampletime v Suspension speed Vi Supply voltage Vx Forward velocity w White noise Wi Weighting filter i yc Controlled output yr Lateral tire deflection
vi Symbol Meaning z Suspension travel zr Road displacement zs Displacement of sprung mass zt Tire compression zu Displacement of unsprung mass φ Speed dependent commutation angle
Conventions
dz (t) =z ˙ (t) (1) dt d2z (t) =z ¨ (t) (2) dt 2
vii Contents
1 Introduction 1 1.1 Problem statement and objectives ...... 5 1.2 Literature review ...... 5 1.3 Outline ...... 8
2 Actuator and car model 9 2.1 BMW 530i ...... 9 2.2 Active suspension system ...... 11 2.2.1 Actuator ...... 11 2.2.2 Sensors ...... 16 2.3 Simplified car model ...... 17 2.3.1 Quarter car model ...... 17 2.3.2 Road input ...... 18 2.4 Summary ...... 21
3 Control of the active suspension 22 3.1 Control objectives ...... 22 3.2 Duality of control objectives ...... 23 3.3 Linear quadratic control ...... 25 3.4 Robust control ...... 27 3.4.1 Model ...... 28 3.4.2 Robustness requirements ...... 30 3.4.3 Weighting filters ...... 33 3.5 Summary ...... 38
4 Analysis of simulation results 39 4.1 BMW 530i performance on random road ...... 39 4.2 Linear quadratic control ...... 42 4.3 Robust control ...... 44 4.4 Summary ...... 46
5 Quarter car test setup 48 5.1 Description of the test setup ...... 48 5.2 Control of road actuation ...... 50 5.3 Kalman filter suspension travel ...... 52 5.4 Experimental validation of setup ...... 53 5.5 Summary ...... 55
viii 6 Measurement results achieved on quarter car setup 57 6.1 Linear quadratic control ...... 57 6.2 Robust control ...... 60 6.3 Summary ...... 67
7 Conclusions and recommendations 69 7.1 Conclusions ...... 69 7.2 Recommendations ...... 71
A Tire Model 75 A.1 Vertical stiffness ...... 75 A.2 Relaxation measurements ...... 77 A.3 Magic Formula ...... 79 A.4 Tire parameters ...... 80
B LDIA 2011 digest 81
ix Chapter 1
Introduction
Test drivers usually emerge from the car with their imagination in overdrive. "The greatest single advance in car engineering since the war," the British magazine Car declared on the cover of a recent issue. Car’s editor, Steve Cropley, wrote that one could take the benefit of all other modern automobile developments, "add the up and double the total - and you might come somewhere near the degree to which full active suspension improves a car." [1] Although one should always be sceptical about the enthusiasm expressed during such first time tests, this statement does indicate that active suspension offers the opportunity to change the performance of a car substantially. Ever since, manufacturers have been hard at work to develop systems suited for mass production. Examples of this are the Active Body Control (ABC) [2] by Mercedes, Hydractive [3] from Citroën and air suspension used by up market manufacturers to increase ground clearance in their off-road models and to influence the character of the car (Land Rover, Audi, VW, Lexus, Lincoln etc.). Next to fully active systems, semi-active systems have also been developed. Examples are Delphi magneto-rheological dampers [4] used by Ferrari as well as Cadillac and the VAG group. Alfa Romeo uses a semi active system developed by Magneti Marelli [5] which controls valves in the damper, thereby changing its characteristics. It is obvious that numerous suspension suspension systems are already in production, gener- ally they can be divided into three groups: Passive (Figure 1.1(a)), semi-active suspension (Figure 1.1(b)) and active (Figure 1.1(c), (d) and (e)) systems. The main difference between them is that the former has no possibility of changing the suspension characteristics, whereas the second can vary the amount of dissipative power. The fully active system can not only vary the amount of dissipative power, but can also supply power to the system by means of active force generation. Implementation of the suspension systems is done very differently by various manufacturers. The Mercedes ABC system for instance, works by means of a hydraulic actuator in series with a passive spring-damper combination. Its bandwidth, due to valves and connective hoses is only 5 Hz. It is therefore primarily used to level the vehicle. Due to its 200 bar operating pressure its power demand is in the range of 3-5 kW. Due to this low bandwidth, the suspension becomes virtually rigid from 10 Hz onwards [6], thereby requiring the passive suspension to provide good comfort and roadholding beyond that frequency. The Citroën system as shown in Figure 1.2 uses spheres filled with nitrogen and a hydraulic fluid separated by a rubber membrane to control the ride. When driving over a bump, the fluid is pushed up the suspension strut compressing the nitrogen and thus providing a spring action. The hydraulic fluid is then directed through valves, providing damping. When driving normally, the spheres at the suspension strut are connected to a third sphere increasing the volume of
1 Chassis Chassis Chassis Chassis Chassis d s F ks ks ds ks ds ks F ks ds ds
Wheel Wheel Wheel Wheel Wheel
kt kt kt kt kt
(a) (b) (c) (d) (e)
Figure 1.1: Quarter car representation of (a) passive suspension, (b) semi-active suspension, (c) parallel active suspension, (d) series active suspension and (e) electromagnetic suspension.
the nitrogen and thus providing a lower stiffness and thereby smoother ride. However, when cornering, valves are closed, disconnecting the central sphere. A firmer ride is achieved this way, thereby reducing roll of the car. Continuous pressurization of the system is required, making the power requirement high. A great disadvantage of the system is that when pressure is lost the vehicle will loose ride height and performance will deteriorate.
Figure 1.2: Citroën active suspension system.
The semi-active solutions from Delphi, see Figure 1.3, and Magneti Marelli both influence the flow of the hydraulic fluid inside the damper. The former uses magneto-rheological fluid, which changes viscosity when the fluid is exposed to a magnetic field. According to the manufacturer the damping force is only dependent on the power applied to the magneto-rheological fluid and can be adjusted up to 1000 times a second. A skyhook control algorithm is used to ensure good road to wheel contact with the least impulses to the car body. Due to the semi-active nature of the system, average power is much lower (5 W) compared to the hydraulic suspension systems. Power can, however, not be supplied to the system, limiting the performance gains of the system when compared to a passive system. A novel electro-hydraulic semi-active suspension system is built by Levant Power and is called the GenShock [7]. It operates by means of a hydraulic cylinder connected to a set of valves and a
2 Figure 1.3: Delphi magneto rheological damper.
hydraulic motor that is connected to a generator. When the vehicle drives over a bump, the linear motion of the shock absorber pumps the fluid round. The hydraulic motor connected in the same circuit is then excited by this moving fluid and subsequently excites an electric generator. Electric energy is then stored in the battery. The manufacturer claims a 1-3% increase in fuel efficiency and a reduction in vibrations up to 30 %. An active solution that tries to solve the problem of high power consumption is built by ZF and Volkswagen [8]. It consists of spindle driven by an electric motor in series with a spring and in parallel with a conventional damper. By actively controlling the spindle position the series spring can be loaded, thereby controlling the roll of the vehicle. A skyhook algorithm is further- more included to improve comfort. A clear improvement of vertical acceleration can be observed with the system installed whereas power consumption is 50-65 % less than that of a hydraulic system.
Figure 1.4: Bose Corp. electromagnetic active suspension sytem.
The research group from Bose Corp. recognized the high power demand and low bandwidth limitations of the hydraulic suspension systems and developed an electro magnetic suspension system, as shown in Figure 1.4. Linear electric motors are used, making it possible to achieve a high bandwidth to counteract the effect of road disturbances on the vehicle body [9]. According to the manufacturer, the linear motor is also capable of delivering enough force to counteract roll and pitch during severe cornering and braking maneuvers [10]. Due to the torsional spring to support the vehicle weight and the possibility to regenerate energy, a power consumption (1-
3 1.5 kW) of only one-third of the power of a car’s air conditioner for the full system is claimed by the manufacturer. However, verification of these claims has been impossible to date, since no design details have been released nor has any commercial test been executed.
To car body
Permanent magnet array
To wheel hub Coil spring Three phase winding
Slotted stator
Figure 1.5: Tubular permanent magnet electromagnetic actuator in parallel with a passive spring.
Considering the low bandwidth and high power demands of hydraulic suspension and limited performance of semi-active suspensions a novel suspension strut has been developed [11]. It consists of a tubular permanent magnet actuator in parallel with a passive spring to support the vehicle mass as is shown in Figure 1.5. The tubular structure gives it the capability of delivering large direct drive forces in a small volume. Furthermore, its bandwidth is in the order of hundreds of hertz, which is larger than required to improve comfort and handling. As a safety feature, aluminum rings are installed in the stator. These rings provide fail-safe damping by means of Eddy current damping. Power consumption is lower than that of a hydraulic system since no continuous pressurization is required. Energy can even be recuperated, depending on the amount of fail-safe passive damping and controller design [12]. All the favorable properties of the novel tubular actuator give rise to very different control possibilities and challenges compared to currently available systems. For example the bandwidth is great for improving comfort and handling, but might also cause resonances that other actuators are not capable of exiting. These favorable features and potential problems make the control of this actuator an interesting topic of this thesis.
4 1.1 Problem statement and objectives
Given the limitations in the aforementioned active suspension systems a tubular direct drive electromagnetic active front suspension for a BMW 530i is developed. In this research, the control of this active suspension will be considered. The goal is to improve the comfort or handling of the vehicle by means of proper control of the active suspension. The main research question can therefore be formulated as:
What performance gains in comfort and handling can be achieved with a high bandwidth electromag- netic active suspension given the constraints of maximum actuator force and suspension travel?
To answer this question, a controller for this active suspension system has to be developed. Fur- thermore, several research objectives have to be fulfilled to answer this question:
• Given the parameters of the suspension system and the BMW, a simulation model has to be created that represents the system correctly.
• The comfort and handling performance criteria have to be defined to objectively measure improvements.
• A suitable control structure has to be developed such that, the best performance is achieved.
• Real life measurements are better proof than computer simulations. A setup, will be devel- oped that facilitates this.
• Using this test setup, measurements have to be performed and will be compared with simulations.
1.2 Literature review
Ever since the 1950’s manufactures have been looking at active suspension systems such as Cit- roëns hydropneumatic suspension. Control of these active suspension systems was picked up by authors from the 1970’s onwards [13, 14]. Since then numerous papers have been written on the control of active suspension systems. This section aims at giving an overview of the control topologies and results achieved with active suspensions. In Michelberg et al. [15], a quarter car model is used to show the disadvantage of LQG control. The author states that if the parameters of the system differ from the nominal parameters, the controlled system can behave worse than the original plant. To counteract this the author uses state-feedback H∞ control in the presence of structured (parametric) uncertainties to find the robust LQG controller. The paper shows that a trade off has to be made between the robustness and performance requirements. To find a suitable controller one or two parameter iterations are necessary. De Jager [16] makes a comparison between various Linear Quadratic (LQ) controllers and H∞ control, based on γ -iterations, using a quarter car model. Weighting filters are included to account for human sensitivity to vertical acceleration. The author finds that H∞ control is not suitable for the design of an active suspension controller using simple weighting filters. The bad √ performance is mainly caused by the invariant point ωi = kt /mu which severely limits the loop shaping possibilities the author argues. The Linear Quadratic Regulator (LQR) controller is found
5 to be more robust over Linear Quadratic Output Feedback (LGOF) or Linear Quadratic Gaussian (LQG), the LQR controller was, however, not found practical due to the large amount of sensors required. Yamashita et al. [17] applies robust control in order to reduce vertical acceleration and im- prove handling for a full car, 7 degrees of freedom (DOF) model. A multiplicative uncertainty is introduced to account for changing body mass and actuator uncertainty. It is assumed that the full state is observable in the design of the controller. However, states are estimated by means of integration and algebraic calculations in the vehicle. Experimental validation is done using a four post shaker where all four wheels can be excited individually. A comparison is made between a nominal performance controller and a robust controller, showing that, under the influence of perturbations the nominal controller performs worse than the passive case. The robust controller performs similar in the perturbed case as in the nominal case, in both cases the performance is better than the passive vehicle. Real driving tests showed similar results as the simulations. Fuzzy control has been applied by Sharkawy [18]. Its performance is better than standard LQR control, however finding the output surface of the fuzzy controller is achieved by trial and error. To overcome this, an active fuzzy controller was developed based on Lyapunov direct method resulting in fast convergence of the parameter vector. The author, however, never discusses the stability of this method. Hrovat [19] discusses various models for controller synthesis. From the one DOF model using an LQ controller he concludes that the control topology, using the suspension travel and vertical vehicle speed as state variables, functions as a "skyhook damper". The author argues that a real "skyhook" is physically not possible, an active device replacing the "skyhook damper" is therefore a suitable replacement. Better performance is achieved when the active device emulates a damper connected to a smooth inertial ground, thereby facilitating larger damping rates, but preventing the transmissibility of road vibrations normally associated with large damping values. Using a two DOF model, Hrovat points out that when limiting tire deflection to the tire deflection of a passively suspended vehicle, an improvement of 11 % can be achieved. He argues that this little improvement might not even be felt by most drivers, however, using the full potential of an active suspension, in which adaptive tuning is used, an improvement of 67 % in comfort can be reached. Depending on the driving conditions, the constraints on either comfort or handling can be loosened as a function of for instance the lateral acceleration or steering angle. Hrovat continues with a 2D model, with which he evaluates the effectiveness of preview. He finds that a preview time of one second can reduce sprung accelerations by 50 % to 70 %. Since this time is too long for practical implementation, the author considers a 50 ms preview, which already results in a reduction of 30 % in tire deflection. Hrovat also discusses the stability of LQG controllers, his conclusion, which corresponds to Doyles’ conclusion [20], is that no guaranteed margins exist. In one example the LQG margins are only 0.2 d B and 18◦ whereas these were ∞ d B and 100◦ for the LQ case. Venhovens [6] uses optimal control technology to enhance the comfort of a vehicle. First, he uses a quarter car model to synthesize a controller. Both full and limited state feedback are tested. Venhovens argues that limited state feedback is favorable due to the lower amount of states that have to be determined. Furthermore, an integrator to eliminate the steady state error that occurs with full state feedback can be avoided. The duality of emphasizing both road holding and comfort is discussed, the author finds that not much gain can be achieved compared to a well tuned passive system, however, he notes that the benefit of an active system is the adaptability. Due to the high power demands of full active suspensions, Venhovens also investigates a semi active suspension system, in which the damping value can be changed. This results in large tire
6 deflections when "skyhook" damping is used. Even with a large range of possible damping values, no improvement in tire deflection can be achieved compared to the passive system. Venhovens has also considered adaptive control of the suspension system, in which a predefined limit on the tire deflection is used to tune the controller in the direction of handling or comfort. Full car simulations showed a good correlation with the quarter car model. A benefit of the full car model is the possible use of preview for the rear wheels, especially improving the dynamic tire compression and suspension travel of the rear wheels. Exchange of state information between the four corners does not improve performance noticeably.
Experimental verification of a robust controller in combination with an hydraulic actuator is done by Lauwerys [21]. The actuator consists of a hydraulic cylinder in which continually variable valves are places to control the flow. Furthermore, a pump is added capable of delivering a force. RMS power consumption of this system is approximately 500 W and a delay of 6 ms is consid- ered due to the hydraulic tubing and electrodynamic valves [22]. The dynamics of the quarter car test setup are determined using a frequency domain approach in which the parameters and uncertainties of the model are estimated from measurements. For this, integrated white noise is used that represents the road disturbance. The measured outputs are the accelerations of the sprung and unsprung mass. The nominal model is a linear approximation of the quarter car test rig, whereas the uncertainties caused by sensor noise, non-linearities and unmodeled dynamics are of the multiplicative type. The controller design is aimed at reducing the body vibrations at 1.5 Hz without amplifying the body acceleration or tire force in other frequency regions. A bandpass filter is used in which the cut-off frequencies can be chosen to reach the desired per- formance. The performance gain in body acceleration was found to be 50 % without serious drawbacks in dynamic tire load variations.
Another experimental test setup was built by Lee [23]. A tubular brushless permanent magnet motor with a peak force of 29.6 N was developed and installed on a scaled quarter car test setup. The sprung mass of the test setup was scaled down by a factor 150 (up to 2.299 kg), whereas the unsprung mass was scaled down by a factor 20 (to 2.278 kg). The road excitations are provided by a cam, thereby fixing the shape of the road profile. Both the sprung and unsprung accelera- tion were measured as well as the suspension travel. Using these measurements, three types of controllers were tested: A lead-lag, LQ and fuzzy controller. The LQ controller did not perform as well as the other two controllers due to the errors in the estimated state. However, performance gains of up to 64 % were still achieved compared to the system not in operation. The fuzzy and lead-lag controller achieved performance gains of 77 % and 73 % respectively requiring equal RMS currents. For both the fuzzy and lead-lag controller, the suspension travel sensor was not used.
Many authors have considered LQ control in theory. This thesis will discuss the performance of the LQ control topology including measurements performed on a full size quarter car test setup. Furthermore, since a car always has variations in its parameters, this thesis will consider robust control. The actuator is a novel tubular electromagnetic suspension system, that offers great benefits over hydraulic systems due to its efficiency, high force density and bandwidth. Since only the front suspension is replaced by the active suspension and measuring the road is deemed unsuitable due to its high sensitivity for errors, no preview will be used.
7 1.3 Outline
In Chapter 2 the properties of the BMW will be introduced. After this, the suspension actuator and sensors are discussed in detail. Having defined the actuator and BMW, a quarter car model is created that can be used for simulations and controller development. Comfort and handling performance criteria are defined using the quarter car model. Having defined the performance criteria for comfort and handling, Chapter 3 defines the objectives for control. The limitations defined by the equations of motion are shown after this. The first control topology discussed in this chapter is linear quadratic control which finds an optimal controller given a linearized plant. To account for uncertainties not incorporated in the linear model, robust control is discussed. The results of these two control topologies are determined and discussed in Chapter 4. Fur- thermore, the performance of the BMW is shown. In Chapter 5 the test setup is introduced to verify the results obtained in simulations. All sensors fitted to the test setup are discussed, as well as control of the industrial actuator that facilitates the road disturbances. A solution to a sensor measurement error is also presented. Using this test setup, the results of the linear quadratic and robust controller are discussed in Chapter 6. Finally, conclusions and recommendations are drawn in Chapter 7.
8 Chapter 2
Actuator and car model
In this chapter the model of the actuator and car will be discussed. First the base car, a BMW 530i, will be introduced. After this, the retrofit active suspension system will be discussed. This sus- pension system is equipped with three sensors used for control of the active suspension, why these sensors were chosen and what their properties are is shown in the next section. After this the simplified car model that will be used for controller synthesis is discussed. Finally, the mathematical description of the road disturbance to this model is given.
2.1 BMW 530i
The BMW 5-series, see Figure 2.1, is a German built executive saloon car well known for its sporti- ness, agility and comfort [24]. It is available with a range of engines, from a two liter four cylinder up top a five liter V10. In this report, a BMW 530i will be considered, which has a three liter in- line six cylinder engine. With its aluminum engine, bonnet and front quarter panels it achieves a near to perfect 50.9/49.1 front to rear weight distribution [25]. The front suspension, which will be replaced by the active suspension, is a MacPherson strut. This system uses two suspension arms that are connected to the bottom part of the hub and provide lateral and longitudinal fixa- tion of the wheel. The top of the hub is attached to a suspension strut which consists of a spring and damper in parallel as Figure 2.2 shows. Furthermore, a rebound spring is fitted inside the damper, lowering the spring stiffness in rebound.
Figure 2.1: BMW 5 series.
9 Figure 2.2: MacPherson suspension system.
Table 2.1: Technical data of the BMW 530i. Parameter Value Unit Unloaded vehicle mass 1546 [kg] Maximum vehicle mass 2065 [kg] Unsprung mass front (left+right) 96.6 [kg] Unsprung mass rear (left+right) 89.8 [kg] Spring stiffness 30.01e3 [N/m] Tire vertical stiffness min-max 3.1e5 − 3.7e5 [N/m] Weight distribution front-rear 50.9 − 49.1 [%] Maximum compression (bump) 0.06 [m] Maximum extension (rebound) 0.08 [m]
Figure 2.3 shows measurements performed on the suspension strut by Janssen [25]. The effect of the rebound spring is clearly visible from the change in gradient at -0.015 m stroke. The damper force is clearly asymmetrical in the bump and rebound region. In compression as little damping as possible is desired, such that the vehicle is capable of absorbing bumps. Kinematic limitations, however, require a certain amount of damping, thereby limiting suspension travel. Generally more rebound damping is applied to prevent ’abruptness’ in the suspension [26]. This means that the motion of the wheel stops suddenly, thereby increasing the jerk (derivative of acceleration) on the vehicle body. Tires are generally considered to be non-linear both in vertical as well as cornering stiffness. Vertical stiffness measurements have been performed on the Dunlop SP Sport 225/50R17 94W tires on a flat plank tire tester [27], see Appendix A. This showed that, given nominal operating conditions, the tire stiffness varies between 3.1e5 and 3.7e5 N/m. This, together with the other car parameters is summarized in Table 2.1.
10 3000 5000
2000 4000 1000 Bump ← 3000 Bump ← 0 ] ] N −1000 N 2000
−2000 → Rebound 1000 Force [ Force [ −3000 0 −4000 −1000 → −5000 Rebound
−6000 −2000 −0.1 −0.05 0 0.05 0.1 −2 −1 0 1 2 Stroke [m] Speed [m/s] (a) Spring Force. (b) Damper Force.
Figure 2.3: Spring and damper characteristics of the BMW 530i front suspension.
2.2 Active suspension system
2.2.1 Actuator To generate a force for suspension control an electro-magnetic actuator has been designed [28, 29]. The actuator has been designed such that it is a retrofit for a BMW 530i McPherson front suspension strut. Performance specifications have been derived from measurements performed on the Nürburgring in Germany. There it was found that a peak and RMS force of respectively 4000 N and 2000 N was necessary to eliminate the vehicle roll angle. The author also com- mented that these driving conditions are not very common. A duty cycle of 50% is therefore proposed, resulting in an RMS force of 1000 N. The electro-magnetic actuator is a tubular slotted three-phase permanent magnet actuator [29]. A graphical representation is given in Figure 2.4 with a detailed view in Figure 2.5. It can be seen that a quasi-Halbach array has been chosen. This topology offers the highest force density [30] by focussing the magnetic field into the actuator. An external magnet array has been chosen to increase magnetic loading. Furthermore, copper losses are reduced due to the smaller circum- ference of the coils. Another benefit is the absence of moving wires since the power electronics are situated on the sprung mass. In this stator, angular coils are fitted, such that, according to Lorentz Z Z
FE = JE × BdVE ⇒ Fz = Jθ BvdV, (2.1) an axial force is generated. The aluminum rings are fitted such that Eddy currents are induced when the actuator moves. This provides passive damping since the Eddy current creates a force opposing the original movement. The spring that is placed in parallel with the actuator compen- sates for the mass of the car such that no continuous power is required to levitate the car. Fur-
11 Sprung acceleration sensor To car body
Coil spring Aluminium ring
Sliding bearing Halbach magnet array
Ls Laser Sensor Unsprung acceleration sensor
Ri
Rm Three phase winding
Bump stop
Rr
Ro
Figure 2.4: Electro-magnetic actuator cross section.
thermore, the stroke of the actuator is chosen such that is equal that of the passive suspension strut. The linear guidance of the stator is done by means of a linear sliding bearing that is fitted over the entire length of the magnet array. Relevant parameters are summarized in Table 2.2. Figure 2.6 shows the electric motor model composed of a voltage source (Vi ), resistor (Ri ), in- ductor (Li ) and back-EMF (Ei ). The differential equation that describes this model is formulated as d I V = E + R I + L i (2.2) i i i i i dt where the subscript i denotes phase a, b or c, furthermore, Ri denotes the resistance per phase and Li is the inductance. The current in these phases is given by πz Ia = iˆ sin + φ (2.3) τp πz 2π Ib = iˆ sin − + φ (2.4) τp 3 πz 4π Ic = iˆ sin − + φ (2.5) τp 3
12 θ z Flux lines r Aluminium rings
Hallbach magnet array
Figure 2.5: Electro-magnetic actuator detailed view of three phases.
Table 2.2: Actuator parameters Parameter Value Unit Description Rs 26.925 [mm] Stator radius Rm 28 [mm] Inner magnet radius Rr 36 [mm] Outer magnet radius Ro 39 [mm] Outer translator radius Ls 400.4 [mm] Stator length τp 7.7 [mm] Pole pitch Lb 60 [mm] Bound stroke Lrb 80 [mm] Rebound stroke ks 30.01 [N/mm] Spring stiffness FRMS 1000 [N] Maximum RMS actuator force mtrans 7 [kg] Actuator translator mass mstat 8 [kg] Actuator stator mass
13 Where, z = zs − zu is the displacement of the actuator (suspension travel), τp the pole pitch, iˆ the amplitude of the current and φ speed dependent commutation. The EMF Ei is given by
Ei = kEi v (2.6) with kEi the EMF gain and v = vs − vu the speed of the actuator. Assuming that iˆ and v are independent, the force delivered by the actuator is given by:
Fact = Fcurrent + Fdamp = kI iˆ + dv (2.7) with kI the force gain and d the damping coefficient. The measured force as a function of current can be seen in Figure 2.7. Also visible is a linear approximation of this measured force
Fact = ki iˆ = 115iˆ. (2.8)
This approximation is valid up to 2000 N which means ki can be used in simulations. The amplifier, using a 2 k Hz current control loop, makes sure that this force is really generated. Figure 2.8 shows the contribution fail-safe electromagnetic damping. Due to final inductance of the rings, the damping has a regressive character. Also visible is an approximation of this non-linear damping, this can can be formulated as
Fdamp = −1341 arctan (0.985v) . (2.9)
Using simulations, the occurrence of a certain damping value is tested when driving over a cer- tain road. As Figure 2.9 shows, large damping values occur much more than small values, this corresponds to relatively small suspension speeds, v. The average damping value is 1450 Ns/m.
Ri Li
Vi Ei
Figure 2.6: Model of an electric motor.
14 3500
Measured 3000 F = 115 i
2500
] 2000 N [
act 1500 F
1000
500
0 0 5 10 15 20 25 30 Current [A]
Figure 2.7: Actuator force vs current, measured and approximated with a Ki of 115 N/A.
1400 Measured 1200 1341arctan(0.985v)
1000 ]
N 800 [
600 damp F
400
200
0 0 0.5 1 1.5 Speed [m/s]
Figure 2.8: Eddy current damping, measured and approximated with Fdamp = −1341 arctan (0.985v).
15 1
0.9
0.8
0.7 ] − 0.6
0.5
0.4
Occurrence [ 0.3
0.2
0.1 Mean ds →
0 1000 1100 1200 1300 1400 1500 1600 ds [Ns/m]
Figure 2.9: Damping occurrence on rough road.
2.2.2 Sensors
As Figure 2.4 shows, three sensors are fitted to the actuator. This set of sensors is most commonly used in literature [31]. The sprung acceleration sensor measures the acceleration of the vehicle body, this sensor is fitted, because the acceleration of the vehicle body is a direct measure for comfort. The second sensor fitted is a laser sensor that is used to measure the suspension travel. Since the suspension travel is directly coupled to the commutation of the actuator, it is important to measure this value directly. Moreover, since suspension travel is limited, this sensor gives a good indication of the state of compression. Appropriate action can be taken if the system gets close to its limits. The third sensor fitted is a 50 g acceleration sensor, this is used to measure the acceleration of the unsprung mass. Since it is impossible to measure the absolute compression of a tire due to the unpredictable nature of a road surface, this measurement is the most convenient to determine the state of the tire. This unpredictable nature is best expressed by the example of driving over a brick or a carton of milk. Both appear to be solid for a sensor, whereas the carton of milk will be compressed easily by the tire, the brick might cause damage. Obviously, measurements are noisy. Each sensor has a certain noise level, based on mea- surements and manufacturer specifications, the noise levels have been determined as Table 2.3 shows. Implementation in simulations is done by multiplying a white noise signal by a gain, Wni , to achieve the sensor noise as Figure 2.10 shows. As Van de Wal [31] indicated, good control is also possible with only the acceleration sensors. In theory, the suspension travel could be estimated through integration, however, due to possible sensor drift and the high importance of correct commutation the suspension travel is measured directly.
16 Table 2.3: Sensor noise. Sensor Deviation Sprung acceleration sensor ± 0.024 m/s2 RMS Sprung acceleration sensor ± 0.178 m/s2 RMS Suspension travel sensor ± 0.002 m RMS
White noise Sensor noise Wni
Figure 2.10: Implementation of sensor noise.
2.3 Simplified car model
Having defined the baseline car and tubular actuator, a model has to be constructed that repre- sents these parts. For this, the actuator is installed in the BMW as Figure 2.11 shows. Only one quarter of the car will be used as simulation model, this will be explained in more detail in the next section. Section 2.3.2 will discuss the road input to the model.
Figure 2.11: Actuator installed in BMW.
2.3.1 Quarter car model Models to asses the vertical dynamics of a vehicle exist at various different levels. From very simple, 1 DOF models [19] up to non linear large number of DOF models [32]. It has, however, been shown [33] that a 2 DOF car model represents the vertical dynamics of a vehicle accurately enough to predict the comfort and tire compression of the vehicle. A quarter car model represents one corner of the vehicle for which only the vertical dynamics are considered. Figure 2.12 shows a graphical representation of the quarter car including an actuator. Here, ms is the sprung mass of the vehicle, mu is the unsprung mass, this is usually made up out of the weight of the rim, tire, brake and part of the suspension. The stiffnesses and
17 damping are denoted by ks, kt and ds respectively, with kt the vertical tire stiffness. The degrees of freedom are the displacement of the sprung (zs) and unsprung mass (zu). The displacement of the road zr is prescribed by the road profile as is discussed in more detail in section 2.3.2. Finally, the actuator force is denoted by Fact . The equations of motion are given by:
ms z¨s = −ks (zs − zu) − ds (z˙s −z ˙u) + Fact (2.10)
ms z¨u = ks (zs − zu) + ds (z˙s −z ˙u) − kt (zu − zr ) − Fact (2.11)
The vertical acceleration (z¨s) is a good indication of the ride comfort of a car, humans, however, are only sensitive to vibrations up to a certain frequency. To take this into account, a weighting filter according to ISO2631-1 will be used [34]. Figure 2.13 shows the frequency dependent weight- ing function. As can be seen, humans are most sensitive to vibrations in the 4-10 Hz range, with fast decreasing sensitivity beyond this range. At lower frequencies humans are also less sensitive, however, motion sickness occurs at roughly 0.125 Hz for humans that are sensitive for this.
zs, z˙s, z¨s
ms
k s Fact z, v ds zu, z˙u, z¨u
mu
k t zt zr
Figure 2.12: Quarter car model.
Figure 2.12 also shows the tire compression (zt ), this is a good measure for handling, since side force can be maximized when the vertical force changes remain minimal. This is due to relaxation effects being minimized when dynamic tire compression is minimized. The third important performance parameter indicated in Figure 2.12 is the suspension travel (z). This is defined by the available space in the suspension system, for the BMW this is 0.06 m in compression and 0.08 m in extension. One can imagine that if more space is available, a more comfortable car can be built, since the distance before the suspension hits the bump stops is larger en therefore less damping or a lower stiffness can be chosen. This parameter is therefore of high importance to ensure a fair comparison between the passive and active suspension.
2.3.2 Road input A vehicle is subjected to a lot of disturbances while driving. Typically, two types of disturbances can be identified: stochastic irregularities and deterministic disturbances. Stochastic irregular-
18 1.1
1 ]
− 0.9
0.8
0.7
0.6
0.5
0.4 Magnitude ISO2631-1 [ 0.3
0.2
0.1 −1 0 1 2 10 10 10 10 Frequency [Hz]
Figure 2.13: Human body sensitivity to vertical vibrations, ISO2631-1.
ities describe normal driving conditions, from speedbumps to random vibrations reflecting the surface quality of the road. Measurements [35] have shown that these stochastic irregularities can be represented accurately by colored noise resulting from the application of a first order filter to a white noise signal w [6].
1 z˙r + zr = w (2.12) νVx
Here zr is the vertical road input, Vx is the forward speed of the vehicle. The parameter ν defines the cut-off frequency and thereby the shape of the road irregularities. Figure 2.15 shows measure- ments performed on a smooth and a rough road. Table 2.4 shows the parameters that have been used to create the simulated road profiles. Here, the first order low-pass output is multiplied with ψ to achieve the correct road amplitude. This gain is based on a sample time (ts) of 0.001 s in Matlab-Simulink and has to be changed when this sample time is changed.
Table 2.4: Typical road parameters. Road type ν [rad/m] Vx [m/s] ψq[−] Smooth 0.2 30 0.05 0.001 q ts Rough 0.8 7.5 0.125 0.001 ts
19 0.05
0.04
] 0.03 m [ r
z 0.02
0.01
0 0.5 1 1.5 2 x [m]
Figure 2.14: Profile of the speed bump used as deterministic disturbance.
−2 10
−3 10
−4 10 ] /Hz 2 −5 m [ 10 Very Good r z Average Very Poor
PSD −6 10 Poor Good
−7 Measured smooth road 10 Measured very rough pavement Simulated smooth road Simulated rough road ISO8608 classifications −8 10 −2 −1 0 1 10 10 10 10 Spatial frequency [1/m]
Figure 2.15: Power spectral densities of different road types.
20 Figure 2.15 shows this first order low-pas filter including the parameters compared to the measurements in spatial frequency. This spatial frequency is determined as:
f ns = (2.13) Vx , with f the frequency and Vx the forward velocity. As is visible, the measured smooth road is generally classified as a very good road according to the ISO8608 [36] with a large peak at small spatial frequencies, most likely caused by a very low frequency wave in the road surface. The simulated smooth road has been chosen such that it is also classified as a very good road and that it matches the measured smooth road in at higher spatial frequencies. For the simulated rough road, the maximum capabilities of the test setup are considered, as will be discussed in Chapter 5, this results in a road that is classified as a good to average road by the ISO8608 criterion. As deterministic disturbance a 30 mm high speed bump is chosen as Figure 2.14 shows. It has a 45◦ degree angle relative to the road surface. Tire enveloping behavior over this road disturbance will not be taken into account for simplicity reasons.
2.4 Summary
In this chapter the BMW 530i is introduced as the base vehicle. Its spring and damping character- istics are found to be non-linear. This is also true for its tire properties, which have been measured on a flat plank tire tester. The active suspension that will be used consists of a tubular electromag- netic actuator in parallel with a passive spring to support the vehicle weight. Aluminum rings are fitted to provide passive fail-safe Eddy current damping. To asses improvements in comfort and handling a quarter car vehicle model is used. For the comfort, vertical acceleration of the sprung mass will be used. To account for human sensitivity the ISO2631-1 criterion is used. The quality of handling is expressed by dynamic tire compres- sion. An important constraint that limits the amount of comfort that can be achieved is the suspension travel. As input to the model both random and deterministic disturbances will be used. The random road is represented by white noise filter by a first order low-pass filter. As deterministic input a three centimeter high speed bump is used.
21 Chapter 3
Control of the active suspension
Given the high bandwidth and low power consumption of the active suspension introduced in the previous chapter a controller has to be developed that makes uses of these favorable properties. This chapter will deal with the design of the controller, based on the quarter car model with a random road as input. First the control objectives will be formulated. After this limitations for control of the active suspension will be given. Finally, two different control approaches will be explained, being Linear Quadratic control (LQ) and Robust Control (RC).
3.1 Control objectives
For both control topologies, the same objectives and constraints hold. The ride comfort objective is defined by the RMS vertical acceleration z¨s weighted by the ISO2631 criterion. The smaller this value the better.
Ride comfort = RMS (WISO2631 ·z ¨s) = RMS (z¨sw) (3.1) Here, the ISO2631 weighting filter is approximated by a fifth order transfer function as suggested by Zuo [37] 87.72s4 + 1138s3 + 11336s2 + 5453s + 5509 W = . (3.2) ISO2631 s5 + 92.69s4 + 2550s3 + 25969s2 + 81057s + 79783
The handling objective is defined by the RMS dynamic tire compression zt , minimization of this variable maximizes the lateral and longitudinal forces.
Handling = RMS (zt ) = RMS (zu − zr ) (3.3)
Furthermore, suspension travel z is limited due to constraints of the suspension. The allowed suspension travel of the active system is limited to the maximum value of the passive BMW suspension to make a fair comparison.
max (z Active) ≤ max (z BMW ) (3.4)
Finally, as was indicated in Section 2.2, the actuator has been designed for a maximum RMS force of 1000 N. So when designing the controller the RMS value should stay below this value.
RMS (Fact ) ≤ 1000 N (3.5)
22 3.2 Duality of control objectives
Independent of the chosen control topology, constraints exist that limit the performance of any active suspension [38]. To make this clear, consider the equations of motion of the quarter car, see (2.10) and (2.11). Converting them into Laplace domain and summing them results in
2 2 ms zss + mu zus = −kt (zu − zr ) , (3.6) in which no suspension forces can be seen, the consequences of this will be discussed later on. The transfer functions between the input zr and outputs z¨s, zu − zr and zs − zu are defined as k (d s + k ) s2 H = s2 H = t s s . (3.7) zs¨ /zr zs/zr 2 2 2 mss + dss + ks mus + kt + mss (dss + ks) Furthermore, 2 2 2 mus mss + dss + ks + mss (dss + ks) H = − (3.8) (zu−zr)/zr 2 2 2 mss + dss + ks mus + kt + mss (dss + ks) k m s2 H = − t s . (3.9) (zs−zu)/zr 2 2 2 mss + dss + ks mus + kt + mss (dss + ks) Now (3.6) can be rewritten as
2 2 mss Hzs/zr + mus + kt Hzu/zr = kt . (3.10) If s = jω, it can be seen that if ω equals s kt ωWH = , (3.11) mu this reduces to kt −ms Hzs/zr = kt . (3.12) mu This is frequency is graphically illustrated in Figure 3.1 where the transfer functions of a con- trolled and passive quarter car model can be seen. The amplitude of Hzs/zr can now be derived as being mu Hzs/zr |s= jωWH = (3.13) ms and the amplitude of kt Hzs¨ /zr |s= jωWH = − . (3.14) ms This frequency (3.11) is called the wheel-hop frequency [39]. Rewriting (3.6) relates the suspension travel to the sprung motion 2 (mu + ms) s + kt Hzs/zr − kt H = . (3.15) (zs−zu)/zr 2 mus + kt
From this equation it follows that s kt ωRS = . (3.16) ms + mu
23 The amplitude of (3.15) becomes
mu + ms H(zs−zu)/zr |s= jωRS = − . (3.17) mu
This is called the rattle space frequency [39] and can not be influenced by the suspension force as (3.6) showed. Again, this frequency is illustrated in Figure 3.1. Equation (3.6) can also be
Sprung acceleration 10 10 ]
dB ωW H [ 5 ← 10 ¨ zs/zr
H 0 10 0 1 10 10 Suspension travel 5 ] 10
dB ω [ RS → 0 /zr ) 10 zu − zs ( −5
H 10 0 1 10 10 Dynamic tire load 15
] 10
dB [
10 /zr
) 10
zr Passive −
zu Controlled ( 5
H 10 0 1 10 10 Frequency [Hz]
Figure 3.1: Invariant points in sprung acceleration and suspension travel. rewritten into a transfer function that relates tire deflection and the body motion transfer function 2 s mu + ms Hzs/zr H = − . (3.18) (zu−zr)/zr 2 mus + kt
There is no point in frequency that can not be influenced by an active suspension, except at ω = 0, where H(zu−zr)/zr = 0. These invariant points thus show that, independent of the control approach, certain points can not be influenced. Equations (3.10), (3.15) and (3.18) furthermore show that if one were to optimize one perfor- mance variable, concessions have to be made in terms of the other variables. Hedrick [38] points out by considering the derivative of (3.6) that, although the transfer functions are coupled, at low
24 frequencies an improvement in both ride comfort and tire deflection can be made. At high fre- quencies, the author found that when requiring an improvement in tire deflection, the increase in vertical acceleration is large. The second point that Hedrick discusses, is that in general, for lower vertical acceleration, a large increase at low frequencies and near the wheel hop frequency can be observed. A physical explanation for this phenomenon is that an improvement in vertical acceleration requires a low spring stiffness, which results in a large suspension travel.
3.3 Linear quadratic control
A time invariant linear system can always be stabilized by a linear feedback if it is fully control- lable [40]. By choosing the poles far in the left half of the complex plane, infinitely fast conver- gence can be achieved. To make this possible, large control amplitudes are necessary. Since this is physically impossible, there exists a limit on how far the poles can be moved to the left. To find an input that suffices both the requirement of fast control and does not require infinite control power, an optimization problem has to be solved. A very useful criterion is the quadratic integral criterion Z t1 T J = lim yc (t) Qyc (t) dt (3.19) t1→∞ t0 with yc the controlled variable and Q a diagonal non-negative weighting matrix containing the weighting factors. Given the objectives of dynamic tire compression zt , sprung acceleration z¨s and suspension travel z the output, yc, can be formulated zu − zr zt yc = z¨s = z¨s = Cx + Du. (3.20) zs − zu z
With x the state vector T x = zs z˙s zu z˙u zr (3.21) and u the input Fact , C and D can be written as: 0 0 1 0 −1 ks ds ks ds C = − − 0 (3.22) ms ms ms ms 1 0 −1 0 0 0 1 D = . (3.23) ms 0 Now consider (3.20), substituting this into (3.19) results in Z t1 J = lim (Cx + Du)T Q (Cx + Du) dt = t1→∞ t Z 0 t1 C T QCC T QD x T uT x u dt (3.24) DT QCDT QD t0
25 this is abbreviated as Z t1 T T Qc Nc J = lim x u T x u dt. (3.25) t1→∞ N R t0 c c
Calculus of variations leads to the state feedback
u (t) = −K x (3.26) with K −1 T T K = Rc Nc + B P (3.27) and P (P = P T > 0) the solution of the Riccati equation
−1 T −1 T T −1 T −1 T P(A − BRc Nc ) + (A − BRc Nc ) P − PBRc B P + Qc − Nc Rc Nc = 0, (3.28) with A and B being the state matrices defined by (2.10), (2.11) and (2.12) as 0 1 0 0 0 ks ds ks ds − − 0 ms ms ms ms A = 0 0 0 1 0 , (3.29) ks ds (ks + kt ) ds kt − − mu mu mu mu mu 0 0 0 0 −av
T 1 1 B = 0 0 − 0 . (3.30) ms ms The foregoing analysis has omitted the fact that a white noise disturbance is present as road input. Consider the state equation x˙ = (A − BK ) x + w (3.31) with w the white noise disturbance. It can then be proven [40] that the solution does not alter, except to increase the minimum value of (3.19). A necessary requirement is that the system is controllable. Controllability can be checked by determining the rank of R = BABA2 B ... An−1 B (3.32) if this rank is equal to n, with n the size of the state, the system is controllable. The system defined here is not fully controllable, which is expected, since the actuator cannot influence the road displacement zr . A suitable controller will, however, still follow from the above Riccati equation as the system is still stabilizable. Since LQ control only considers a linear model, constant parameters have to be chosen. Ta- ble 3.1 shows the parameters selected. The sprung mass, ms, is based on the empty weight of the car plus two passengers and half a tank of fuel. It is assumed that the weight of the passengers and fuel is evenly distributed over the car. For the damping value, ds, the average value as was de- termined in Section 2.3.1. Finally, for the tire stiffness, kt , the total load of ms + mu is considered. From this the stiffness is derived from Figure A.2.
26 Table 3.1: Car parameters considered for LQ control. Parameter Value Unit Description ms 395 [kg] Sprung mass mu 48.9 + mtrans [kg] Unsprung mass + actuator mass ks 30.01e3 [N/m] Spring stiffness ds 1450 [Ns/m] Damping kt 3.37e5 [N/m] Tire stiffness
Given the three output variables, a weighting matrix Q can be defined as q1 0 0 Q = 0 q2 0 (3.33) 0 0 q3 with q1, q2 and q3 emphasizing tire deflection, sprung acceleration and suspension travel re- spectively. These weighting factors are solved by means of a constrained nonlinear optimization algorithm, fmincon. The objective function minimized is ! RMS (z¨ ) RMS (z ) O = 0.5 ζ s + (1 − ζ ) t (3.34) RMS z¨sp RMS ztp where comfort (ζ ) or handling (1 − ζ) is emphasized depending on the choice of ζ . Here, zsp is the performance of the BMW. Suspension travel and actuator force are used as constraints for the optimization. In this section it is assumed that the full state is measurable. On the quarter car test setup this will not be a problem since the full state is measurable. On a real car, this will, however, be a problem. The state will therefore have to be estimated. The resulting problem is the Linear Quadratic Gaussian (LQG) control problem.
3.4 Robust control
As Doyle has shown [20], stability margins can not be guaranteed with LQG control. It is therefore necessary to explore other control topologies, that can guarantee stability, even with an uncertain plant. H∞-control seems to be able to guarantee this stability [41]. Using the structured singular value, defined as
µ(M)−1 ≡ min{σ ¯ (1) | det (I − M1) = 0 for structured 1} (3.35) 1
DK-iteration can be performed to synthesize a µ-optimal controller. Here, M is considered the part of the plant connected to the uncertainty matrix 1. The idea is to find a controller that minimizes the peak value over frequency of the upper bound µ (N) ≤ min σ¯ DND−1 (3.36) DD
27 namely, −1 min min |DN (K ) D |∞ . (3.37) K DD
Here, K is an H∞ -controller that is synthesized while D is kept fixed. D is a matrix that is found by minimizing σ DND−1 ( jω) with N fixed. Finally, each element of D( jω) is fitted to a stable and minimum phase transfer function D(s). The matrix N is the generalized plant defined as the lower fractional transformation of P and K , with P the plant and K the controller.
−1 N = Fl (P, K ) ≡ P11 + P12 K (I − P22 K ) P21 (3.38)
−1 The iterations continue until |DND |∞ < 1 or the H∞-norm no longer decreases. The order of the controller resulting from this process is equal to the number of states in the plant plus the number of states in the weighting filters plus twice the number of states in D(s) [42].
3.4.1 Model The quarter car model used to design the robust controller is similar to the model introduced in section 2.3.1, however, various uncertain parameters and weighting filters are now included in the model as is shown in Figure 3.2. Uncertainties in the model can have several origins [41]; • There are always parameters in the linear model that are only known approximately or are simply wrong. Furthermore, parameters can vary due to non-linearities (such as the damping coefficient) or changes in operating conditions (such as changing tire stiffness, as a function of load and inflation pressure, and sprung mass). • Measured signals are imperfect, sensor noise and discretization errors can cause the signal to deviate from its real value. This can give rise to uncertainty in the input. For the three sensors present in the test setup the noise levels are summarized in Table 3.2 together with the parametric uncertainties. • At high frequencies the structure and model order are unknown. Therefore uncertainties will always surpass 100 % at some frequency. Good examples of this are the chassis reso- nances beyond 30 Hz and the natural frequencies of the tire, which typically start at 35 Hz [43] and beyond. • A simpler model can be chosen in favor of a very complex model, the neglected dynamics can be incorporated as uncertainties. • Controller implementation may differ from the one obtained by solving the synthesis prob- lem. To account for controller order reduction, one may include some uncertainty. • Output uncertainty can influence the performance of the system. Particularly deviations in the actuator introduced in Section 2.2 such as hysteresis and temperature dependency can influence the performance of the actuator and thereby the performance of the system. In Figure 3.2 the uncertain sprung mass, tire stiffness and damping are included in the per- turbed plant as uncertain parameters that can vary within a certain range. It is furthermore as- sumed that beyond 30 Hz the dynamics of the system are not known completely. A multiplicative uncertainty is therefore included
Pp = P (I + WUnmod 1I ) (3.39)
28 Controlled outputs
z Wo1 t Weighted dynamic Wi1 Pertubed plant, Pp White noise 1 tire compression s/av+1 zr Wo3 z¨s F ISO2631 Weighted sprung act acceleration
z¨u
W z − z o4 s u Weighted suspension Wn1 travel noise
Wn2 Wn3 noise noise
Controller Controller inputs
Wo2 Fact Weighted actuator force
Figure 3.2: Model used for DK-synthesis.
Pp wUnmod ∆I Parameteric perturbed Unmodeled dynamics plant
Figure 3.3: Unmodeled dynamics.
29 Table 3.2: Uncertainties of the quarter car model. Parameter Type Mean value Deviation Sprung mass Parametric uncertainty 395.3 kg −42.77 +75.38 kg Tire stiffness Parametric uncertainty 3.4e5 N/m ±0.3e5 N/m Damping coefficient Parametric uncertainty 1450 Ns/m −550 +250 Ns/m Sprung acceleration sensor Sensor noise - ± 0.024 m/s2 RMS Sprung acceleration sensor Sensor noise - ± 0.178 m/s2 RMS Suspension travel sensor Sensor noise - ± 0.002 m RMS
as Figure 3.3 shows. Here WUnmod is defined as
1 s2 + 2·0.707 s + 1 (2π30)2 2π30 WUnmod = . (3.40) 1 2 2·0.707 (2π400)2 s + 2π400 s + 1
Sensor noise is included in the form of additive uncertainties to the measured sprung acceler- ation, unsprung acceleration and suspension travel. This additive uncertainty is in the form of white noise multiplied by a weighting function Wni , with i ranging from 1 to 3. The four weighted and controlled outputs, dynamic tire compression, actuator force, sprung acceleration and sus- pension travel are used in the DK-synthesis. Inputs to the controller are the sprung acceleration, unsprung acceleration and suspension travel. Table 3.2 summarizes the uncertainties. The choice of the weighting filters will be discussed in more detail in section 3.4.3.
3.4.2 Robustness requirements The main requirement of the controlled system is performance, however, stability is also of im- portance. This stability requirement can be divided into nominal stability (NS) and robust sta- bility (RS). Nominal stability can be shown by determining the poles of the controlled system with 1 = 0. This is shown in Figure 3.4 together with the pole plot of the allowed perturbations. It can be seen that all poles are in the left half plane, which means that all perturbations of the un- controlled system are stable. Varying sprung mass or tire stiffness results in a shift of the poles in vertical (imaginary-axis) direction. Changing damping results in the real value of the poles changing. A damping, ds, of zero will results in poles on the imaginary axis, however, this will not occur in practice. Furthermore, it is assumed that 1 is stable. Robust stability means that the controlled system is also stable for all perturbed plants. For this the N1 structure is considered as is shown in Figure 3.5. The transfer function from exogenous inputs u to outputs yc is defined as −1 Fu (N, 1) = N22 + N211 (I − N111) N21. (3.41)
With N11 the coupling of the plant to the disturbances 1 and N22 the nominal plant. Nominal stability already proves that the whole of N must be stable, therefore the only source of instability −1 can be the feedback term (I − N111) . Thus when the system is nominally stable, the stability of the perturbed system is equal to the stability of the M1-structure shown in Figure 3.6 with M = N11. The stability of the M1-structure can be proven by applying the Nyquist criterion. This results in the requirement that the M1-structure is stable for all allowed perturbations with
30 σ(1) ≤, ∀ω if and only if µ (M ( jω)) < 1, ∀ω, (3.42) with µ the structured singular value.
1st pole pair 2nd pole pair 100 10 ←→ ←→ 80 8 m 60 6 m ] ] ←→ ds s s − 40 − 4 and ←→ ds and 20 2 Perturbed plants k 0 0 k t Nominal plant t −20 −2
−40 −4 Imaginary-axis [ Imaginary-axis [ −60 −6
−80 −8
−100 −10 −20 −15 −10 −5 0 −2.5 −2 −1.5 −1 −0.5 0 Real-axis [−] Real-axis [−]
Figure 3.4: Pole plot of quarter car model and perturbations.
If all stability requirements are satisfied, the robust performance (RP) demands have to be fulfilled. These demands indicate whether the controlled system achieves better performance than the uncontrolled system under the influence of the uncertainties. For this the worst case gain from exogenous inputs w to outputs z is calculated over all frequencies for the controlled and uncontrolled plant. Robust performance is then achieved if
µ (N ( jω)) 1ˆ c < 1, (3.43) µ1ˆ (Nu ( jω)) where µ is calculated with respect to the matrix 1 0 1ˆ = , (3.44) 0 1P
1 contains the true uncertainties and 1P is a full complex matrix with the same size as the number of outputs of P stemming from the H∞-norm performance specification. Robust performance, however, is not required for all outputs. If for instance sprung acceler- ation is emphasized, tire deflection does not have to perform robustly. It is only required to be stable.
31 ∆ u ∆ y∆
N11 N12 u yc N21 N22 N
Figure 3.5: N1-structure used for robust performance analysis.
∆ u ∆ y∆ M
Figure 3.6: M1-structure used for robust stability analysis.
32 3.4.3 Weighting filters Weighting filters can be used to shape input signals, such as the road disturbance discussed in section 2.3.2 or set performance goals for the output. Depending on the shape and amplitude of performance filters, the frequency response of the outputs can be influenced. For instance by choosing the ISO2631 criterion as a weighting filter for the sprung acceleration, frequencies between 4 and 10 Hz are emphasized much stronger than frequencies outside of this range. Below, the individual weighting filters will be discussed in more detail.
Sprung acceleration Humans are most sensitive for vertical vibrations between 4 and 10 Hz [34]. The ISO2631-1 standard has been created to take this into account when evaluating suspension performance. For simulation purposes, this frequency dependent weighting has to be converted into a continuous time transfer function, this has been done by Zuo et al. [37] and is shown in Figure 3.7. As can be seen, up to fifth-order fits have been created, however, to keep the controller order as low as possible, it has been decided to use the second order fit which is expressed as
86.51s + 546.1 W (s) = w , (3.45) zs¨ zs¨ s2 + 82.17s + 1892 with wzs¨ a gain that determines the importance of this weighting filter. A problem that occurs with the use of this weighting filter is that at high frequencies the filter has a very low gain, thereby allowing the vertical acceleration to be extremely large and causing instability. To prevent this,
Wz¨s is multiplied by a first order PD-filter at 200 Hz. The deviation from the ISO2631-1 standard at low frequencies is not considered to be a problem, since this is not the most important region. Figure 3.11 shows the normalized sprung acceleration weighting filter together with the other weighting filters.
Dynamic tire load In literature [33, 44] dynamic tire compression is often used as an indication for the quality of the cars road holding. This is a valid assumption as vertical load influences the lateral force a tire can develop [43]. However, due to tire relaxation effects, this quick variation of vertical tire compression might not have such a major influence on lateral tire force. To investigate this, a tire model proposed by Pacejka [45] will be used. In Figure 3.8 the model is shown. The contact patch of the tire is connected to the rim via the lateral stiffness, kr . This contact patch is only allowed to 0 move in y-direction with respect to the rim, therefore, Vx is assumed to be equal to Vx . The force Fy in the contact patch is calculated using the Magic Formula tire model