Computers and Electronics in Agriculture 47 (2005) 4158

Automatic velocity control of a self-propelled windrower

Christopher A. Fostera , Rick P. Strosserb , Jeremy Petersb , Jian-Qiao Suna,
a b

Department of Mechanical Engineering, University of Delaware, Newark, DE 19716, USA Electrical Engineering Competency Center, CNH North America LLC, 500 Diller Avenue, New Holland, PA 17557, USA

Received 1 June 2004; received in revised form 23 September 2004; accepted 8 October 2004

Abstract A velocity control methodology is presented for an agricultural vehicle with a hydrostatic drivetrain. A cascade control design incorporating a current control, cylinder position control and velocity control is proposed. The cylinder position controller uses a set of gain scheduled lag compensators in series with a sequence of segmented deadzone inversions. The deadzone inversions compensate for nonlinear valve dynamics. The velocity controller consists of a digital feedback control. A stability analysis of the closed loop system is performed. Experimental results from a self-propelled windrower are used to demonstrate the effectiveness of the control system. 2004 Elsevier B.V. All rights reserved.
Keywords: Variable-structure feedback control; Segmented deadzone inversion; Autonomous vehicles; Vehicle velocity control; Embedded controller

1. Introduction In the past two decades the use of electronic control systems in vehicles has been steadily increasing. Due to economies of scale the auto industry has lead the way in

Corresponding author. E-mail address: sun@me.udel.edu (J.-Q. Sun).

0168-1699/$ see front matter 2004 Elsevier B.V. All rights reserved. doi:10.1016/j.compag.2004.10.001

42

C.A. Foster et al. / Computers and Electronics in Agriculture 47 (2005) 4158

the application of electronic controls on vehicles. Systems such as cruise-control, electronic engine management, climate control and electronic transmissions are now common in automobiles. In the last 10 years manufacturers of agricultural machinery have been developing and releasing more electronic control systems on their equipment. Precision agriculture is an area of research and development making extensive use of control systems (Auernhammer, 2001; Earl et al., 2000; Zhang et al., 2002). Vehicle automation with control systems is an important element of precision agriculture. The material presented in this work demonstrates an automatic velocity control system for a self-propelled windrower. A signicant research effort is being devoted to work on unmanned automated agricultural vehicles (Callaghan et al., 1997; Keicher and Seufert, 2000; Reid et al., 2000; Torii, 2000). A fully automated vehicle has many interacting subsystems and will have several levels of control. The control objectives include vehicle motion, trajectory following, obstacle detection and tool manipulation. Stentz et al. (2002) have developed and tested a semi-autonomous tractor and sprayer for eet applications in orchards. An intelligent autonomous agricultural robot was developed and experimentally tested by Hagras et al. (2002). The robot adjusts to changing and dynamic outdoor operating environments through use of a fuzzy-genetic control system. For this kind of technology to reach the marketplace, the cost has to be reasonable. Manufacturers of agricultural equipment will not t these systems to their products unless they are cost effective. This paper will present a cost effective control system for the windrower. Automatic velocity control, known as cruise-control in automobiles, has been available for quite sometime. Automobile drive-trains exhibit nonlinear dynamic behavior, particularly at lower speeds (Fritz, 1996). At higher speeds the behavior becomes more linear. Classical linear controls have been successful in the higher speed ranges. The emerging technology of automated-highway-systems (AHS) calls for speed control systems that work over the full range of speeds (Rajamani and Shladover, 2001; Rajamani et al., 2000; Rajamani and Zhu, 2002). A variety of nonlinear and adaptive systems have been developed to address this issue (Fritz, 1996; Ishida, 1992; McMahon et al., 1990; Setlur et al., 2003). However, the dynamics of automobile drive-trains are different from those of the windrower hydrostatic drive-train. Due to the larger mass of the windrower and the rougher terrain in which it operates, the control task in this work can be difcult. Electrohydraulics has been a common type of actuation in power machinery. A key element in the windrower propulsion system proposed here is a proportional directional valve. A number of mathematical valve models have been presented in the literature. Vaughan and Gamble (1996) have developed a single-inputsingle-output (SISO) model of a proportional solenoid valve. The model uses solenoid voltage to estimate the spool response. Elmer and Gentle (2001) extended this model into a so called parsimonious valve model by incorporating a nonlinear model of the solenoid. These models require a number of accurate valve parameters. Ferreira et al. (2002) have developed semi-empirical model that predicts ow and pressure of a hydraulic servo-solenoid valve. Jelali and Schwarz (1995) and Schwarz (1998) have studied valve models in observer canonical form in which the spool position is used. In more expensive servo-valves, a linear-velocity-differential-transformer (LVDT) is used to measure the spool position and velocity. In many applications, where cost is a major concern, an LVDT may be too expensive. The models discussed above by Vaughan and

C.A. Foster et al. / Computers and Electronics in Agriculture 47 (2005) 4158

43

Fig. 1. One of the experimental windrowers modied for testing.

Gamble (1996) and Elmer and Gentle (2001) could be used to estimate the spool position in the absence of an LVDT. A windrower is an agricultural machine used for hay and forage preparation, crop harvesting and crop residue processing. An example of a self-propelled windrower is shown in Fig. 1. The performance of the hydrostatic drive-train can vary considerably depending on the ground and crop conditions. When working in a eld with varying ground and crop conditions the operator has to make many adjustments to maintain vehicle speed and loading. It would, therefore, be desirable to implement a control system to reduce the number of adjustments the operator must make. This will allow the operator to focus more on other systems and processes taking place on and around the machine. The objective of the velocity control system is to improve operator performance, reduce operator fatigue and increase the productivity of the machine. The work presented here documents the development and testing of a cost-effective velocity control system for a self-propelled windrower with a hydrostatic drive-train. The velocity control system is designed to track an operator dened reference and reject disturbances. It is likely that velocity control will become a subsystem of control systems that operate at a higher level in a fully automated windrower. In Section 2.1, a brief description of the experimental hardware is presented. This is followed by a statement of the control objective in Section 2.2. Sections 2.3, 2.4 and 2.6 present the design of control loops of the subsystems and a stability analysis of the controls. Numerical and experimental results of the controls are presented in Section 3.

2. Control designs 2.1. Hardware description The hydrostatic drive-train of the windrower is composed of two primary components, hydraulic pumps and hydraulic motors. The power source is a diesel engine connected directly to the hydraulic pumps. The propulsion system consists of two independent hydrostatic loops, one driving each of the front wheels. Each hydrostatic loop is powered by a

44

C.A. Foster et al. / Computers and Electronics in Agriculture 47 (2005) 4158

Fig. 2. Schematic of the velocity control hardware. Thick lines represent hydraulic connections and thin lines represent electrical connections. Mechanical components are represented by graphics.

variable displacement axial piston pump. The speed of each wheel is proportional to the ow of oil in the corresponding hydrostatic loop. The oil ow in each loop is regulated in two ways, by the rotational speed of the pumps, which is directly correlated to engine speed, and by adjusting the displacement of the pump. Pump displacement is varied by changing the angle of the swash plate. The swash plate angle is adjusted through an external pintel arm. Hence by positioning the pintel arm appropriately, the wheel speed can be controlled. This hydrostatic drivetrain is referred to as the ground drive. The pintel arms are positioned through a mechanical linkage. An electrohydraulic system composed of a spring-centered hydraulic cylinder and a proportional ow control valve is used to equally displace the pintel arms. The hydraulic cylinder position therefore correlates with the average vehicle velocity. The measured average velocity will be the feedback signal in the velocity control design. The hydraulic cylinder is spring-centered to a neutral position where the vehicle becomes stationary. Fig. 2 shows a diagram of the hardware. Two onboard control modules with the Inneon C167 microprocessor are used. The modules communicate with each other over a controller-area-network (CAN) bus. Vehicles with electronic engines have the engine control unit (ECU) connected to the CAN bus. The embedded controllers run at 100 Hz and are used to control a variety of subsystems on the vehicle. 2.2. Control objectives The objective of the velocity control is to follow the command from the operator. Due to the rough ground conditions in which the vehicle operates, the system must be able

C.A. Foster et al. / Computers and Electronics in Agriculture 47 (2005) 4158

45

Fig. 3. Block diagram of the velocity control system.

to sufciently reject disturbances. The characteristics of the drive-train also vary in time. The control system needs to be robust enough to compensate for these disturbances and variations. The velocity control system functions through a set of subsystems. Fig. 3 shows how the velocity control loop encloses the propulsion cylinder position control loop. It should be noted that the cylinder position control loop itself encloses the current control loop. The control structure can be best described as a cascade controller. If the hardware is changed in the future, it will be benecial if the control loops are easy to retune and adjust. 2.3. Current control The current control loop regulates the electrical current output to the electrohydraulic valve. The current through the valve solenoids is sensed on the return lines of the solenoid circuits. The current returning to the control module passes through a second order analog anti-aliasing lter before passing to the analog-to-digital converter. The ow of oil through the valve is proportional to the current and the supply pressure. The solenoid is modeled as a resistor in series with an inductor. The open-loop transfer function is given by Rizzoni (1995) Gs (s) = 500 . s + 500 49.4s + 44444 . s2 + 733s + 44444 (1)

The transfer function of the anti-aliasing lter is given by Gf (s) = (2)

At the Nyquist frequency of 50 Hz of the control system, the attenuation of the anti-aliasing lter is 15 dB. A digital PI controller is designed by using the method of root locus, Ci (z) = Kpi (z + 0.276) . z1 (3)

The closed current loop has a 5% settling time of 0.03 s and a damping ratio of 0.707. Fig. 4 shows a block diagram of the current control structure. The circuit switch in Fig. 4 determines which of the two solenoid circuits to drive. Fig. 5 shows a root locus of the current control system. The zero of the open loop system is at z = 0.276. The PI gain is chosen to be Kpi = 9.4. The closed-loop poles of the system indicated by the triangles in Fig. 5 are at z = 0.27 j0.25. The system is stable when 0 < Kpi < 56.

46

C.A. Foster et al. / Computers and Electronics in Agriculture 47 (2005) 4158

Fig. 4. Block diagram of the current control loop.

Fig. 5. Root locus of the current control loop: () open loop poles; () open loop zeros; ( ) closed loop poles at z = 0.27 j0.25 when K = 9.4.

2.4. Cylinder displacement control The position of the hydraulic cylinder is directly correlated with the angle of the swashplates in the ground drive pumps. To achieve good velocity control of the vehicle, the hydraulic cylinder must be positioned quickly and accurately. The plant input is the current reference, and the plant output is the position of the hydraulic cylinder, as illustrated in Fig. 6. The hydraulic cylinder position is sensed by a rotary potentiometer (Fig. 2). The spring-centered neutral will be referred to as the zero position. Cylinder position setpoints which extend from the neutral will be assigned positive values, and setpoints which retract from the neutral will be assigned negative values.

Fig. 6. Block diagram of the gain scheduled control for hydraulic cylinder position: rcp is the cylinder position setpoint; ecp is the tracking error of the cylinder position; ri is the valve current setpoint; Q is the oil ow through the valve; ycp is the measured cylinder position; ycpf is the ltered and digitized value of the cylinder position.

C.A. Foster et al. / Computers and Electronics in Agriculture 47 (2005) 4158

47

Fig. 7. Flow vs. current characteristics of the electrohydraulic valve: regions 1A and 1B, proportioanl ow; regions 2A and 2B, dead zone; regions 3A and 3B, external inuence. Letters A and B refer to the ports; P stands for the high pressure supply port; T refers to the tank port at atmospheric pressure; X indicates that the port is blocked; iBT represents the cracking current when the B port begins to open to T port; iBp represents the cracking current when the B port begins to open to P port; iAt represents the cracking current when the A port begins to open to T port; iAp represents the cracking current when the A port begins to open to P port. The blocking currents are iAx = (iAp iAt )/2 and iBx = (iBp iBt )/2. A double solenoid valve is used but a single solenoid valve shown graphically for improved clarity.

The proportional directional solenoid valve used in this work generates a nonlinear relationship between current and cylinder position. The nonlinear relationship between valve current and oil ow through the valve is shown in Fig. 7. The valve has active driving regions (1A and 1B), deadzones (2A and 2B) and a region in which ow is inuenced by current and external forces acting on the system (3A and 3B). Control specications include a 1 s 5% settling time, and damping ratio of 0.707. A digital lag compensator given by Ccp (z) = Kcp (z + bcp ) , (z + acp ) (4)

where 1 acp < bcp 1, is considered in the cylinder displacement control. The sampling time of this subsystem is slowed to 20 Hz. This shifts the plant poles away from the edge of the unit circle, which makes the control design process a little easier. Because of the piecewise nature of the valve response, as shown in Fig. 7, we propose a switching control strategy combined with a segmented inversion of the deadzones in the valve response curve. The switching input parameters are the cylinder position setpoint rcp and the cylinder position error ecp . The switching control is designed in four branches. Each branch is designed for a different combination of regions in the valve response curve. The control for each branch consists of a digital lag compensator, as shown in Eq. (4), and an inversion of a local hardware deadzone, as shown in Fig. 8. In Fig. 8Ccp (n) represents the bank of digital lag compensators and DZinv (n) denotes the bank of deadzone inversions. The switching logic used in the control

48

C.A. Foster et al. / Computers and Electronics in Agriculture 47 (2005) 4158

Fig. 8. Block diagram of the switching control for the cylinder position.

is explained below and presented in Fig. 9. The lag compensators used for each branch are tuned slightly differently but retain the structure shown in Eq. (4). Three thresholds 1 , 2 and 3 used in the deadzone inversions are dened below. Threshold 1 denes a cylinder position error deadband in branches 1, 2, and 3. 2 denes a small control deadband in Eq. (7) for branches 2 and 3. 3 denes a control deadband for branch 4 in Eq. (8). Branch 1 is used when |ecp | > 1 . This branch uses an aggressively tuned controller in combination with a deadzone inversion using piecewise linear interpolations of the valve response in regions 1A and 1B to drive the system quickly toward setpoint. The lag compensator for this branch is designed as Ccp (z) = 35.1(z + 0.6) , (z 0.8) (5)

The deadzone inversion of this branch is given by uc > uc max iAp max , uc (iAp max iAp ) + iAp , 0 uc uc max u c max iref = uc iBp + iBp , uc min uc 0 i u c min Bp max iBp max , uc < uc min

(6)

Branch 2 is used when |ecp | < 1 and rcp > 0. The controller of this branch is tuned to overdamp the system so that it smoothly approaches the setpoint with no overshoot. The control for this branch uses the valve dynamics in region 1A to push away from neutral and the spring-centering passive action to bring the cylinder toward neutral in region 3A.

Fig. 9. Flowchart of the switching logic for cylinder position control. n is the index of the selected control.

C.A. Foster et al. / Computers and Electronics in Agriculture 47 (2005) 4158

49

The deadzone inversion is given by iAp max , (uc 2 ) (iAp max iAp ) + iAp , (u c max 2 ) iref = iAx , (u + ) c 2 (iAt iAt min ) + iAt , (u 2 ) c min iAt min ,

uc > uc max 2 uc uc max |uc | < 2 uc min uc 2 uc < uc min (7)

Notice that a small deadband exists to lock the cylinder in position using region 2A when the cylinder position error is sufciently small. This deadband is adopted because the cylinder can only be positioned with a nite precision due to quantization errors, sensor noise, and hardware limitations. Branch 3 is used when |ecp | < 1 and rcp < 0. This branch is similar to branch 2. It includes regions 1B, 2B and 3B. The lag compensator is tuned slightly differently to achieve performance similar to that in branch 2. This is because the volume owrates of the oil associated with equivalent motion of the cylinder in each direction differ due to the volume occupied by the connecting rod on one side of the piston. The deadzone inversion in Eq. (7) is used with the B current values substituted for the A values. Branch 4 is used when rcp = 0. The controller is tuned to bring the vehicle to a halt quickly. The lag compensator used is the same as shown in Eq. (5). The small deadband uses the passive action of the spring to center the cylinder gently as it approaches the neutral position. The inversion mapping with a small deadband is given by uc > uc max iAp max , (uc 3 ) (iAp max iAp ) + iAp , 3 uc uc max (u c max 3 ) |uc | < 3 iref = 0, (8) (u + ) c 3 (iBp max iBp ) iBp , uc min uc 3 (u c min 3 ) iBp max , uc < uc min The cylinder position error deadband, 1 , is used to switch from branch 1 to either branch 2 or 3. The reason for this is discussed in more detail in the remark below. The value of 1 used in our experiments is 3 mm. The control deadbands, 2 and 3 , are to stop the control uc when it drops below a certain level. The smaller the size of the deadband, the more accurately the controller will try to position the cylinder. A small deadband can also lead to chatter about the setpoint. The selection of the deadband size therefore has to balance these two considerations. The value of 2 and 3 used in our experiments is 1.5 mm. In Eqs. (6)(8), uc is the output of the lag compensator as shown in Fig. 8. The parameters iAp , iAt , iBx , iBp , iBt and iBx are dened in Fig. 7. The parameters iAp max and iBp max represent the upper limits of the valve currents on the respective solenoids. The parameters iAt min and iBt min represent lower limits of the valve current on the respective solenoid. The parameters uc max and uc min are the limits of the control. All four of the inversion mappings have

50

C.A. Foster et al. / Computers and Electronics in Agriculture 47 (2005) 4158

saturation limits. Critical parameters of an individual valve are obtained from a calibration procedure. Remark. The reason for switching control from branch 1 to either branch 2 or 3, as the system approaches the setpoint, is to eliminate oscillations about the setpoint that exist under the control for branch 1. This phenomenon results from the fact that regions 1A and 1B reside on opposite sides of Fig. 7. When the system in branch 1 overshoots, it switches from region 1A to 1B. As this transition occurs, the system passes through all states in between. Recall that the hydraulic cylinder is spring-centered. As the system passes through regions 3A and 3B on its way to region 1B, the cylinder will passively move toward the neutral. When the system crosses setpoint again, it switches back and hence once again passes through regions 3A and 3B. This causes the system to oscillate near the setpoint. Branches 2 and 3 are implemented to prevent this from happening. 2.5. Cylinder position control stability analysis 2.5.1. Regions 1A and 1B When the valve is operating in regions 1A and 1B the hydraulic cylinder effectively integrates the oil ow coming from the valve. The oil ow can be approximated as a linear function of the valve current minus the cracking current (iAp or iBp ). As shown in Fig. 7 this is due to the fact that only a small range of the valve capacity is used. An approximate model of the plant in these regions is given by, ycp TKA1 uc (9)

where the dot denotes differentiation with respect to time, and KA1 is a combined gain including the effects of the deadzone inversion, the current control and the cylinder position control. Branches 1 and 4 cover regions 1A and 1B. The output of the compensator uc can have both negative and positive values. Fig. 10 shows a root locus plot of the control system in regions 1A and 1B using Eq. (5) as the controller. The system is stable when 0 < Kcp < 808. The closed-loop poles as shown in Fig. 10 have a damping ratio 0.707. The other branches have minor gain and polezero changes in the lag compensator but retain the same basic root locus prole. Hence, the system is stable in regions 1A and 1B. 2.5.2. Regions 2A and 2B In regions 2A and 2B the oil ow is blocked, and the cylinder is locked in position. In this state, we have ycp = 0. (10)

In other words, ycp remains constant. The constant value of ycp is determined by the initial condition upon entering either region 2A or 2B. The system in this state is marginally stable since the hydraulic cylinder has hardware limits that prevent ycp from becoming unbounded.

C.A. Foster et al. / Computers and Electronics in Agriculture 47 (2005) 4158

51

Fig. 10. Root locus of the cylinder position control operating in regions 1A and 1B of Fig. 7: () open loop poles; () open loop zeros; ( ) closed loop poles at z = 0.85 j0.13 when K = 35.1.

2.5.3. Regions 3A and 3B When the valve is operating in regions 3A and 3B, the valve is cracking one of its ports open to tank. The current determines how large the opening to tank will be. The spring force determines the pressure of the oil in the cylinder. The oil ow is therefore a function of both valve current and cylinder position. Due to the passive nature of this action, the spring always pushes the cylinder toward the neutral. The plant can therefore be approximately modeled by ycp = KA3 ycp uc , (11)

where KA3 > 0 represents a combined gain of the effects of the valve current, and spring force. When in regions 3A or 3B, the controller operates on either branch 2 or 3. To be in region 3A, the controller must be on branch 2 with ycp > 0. On branch 2, the controller will be in region 3A only if uc < 0. Under these conditions, Eq. (11) is stable. A similar argument can be made to prove that Eq. (11) is also stable in region 3B. Hence, the system is stable. 2.6. Velocity control The basic structure of the velocity control system is shown in Fig. 3. The input to the single-inputsingle-output plant is the cylinder position reference. The output of the plant is the average ground speed of the vehicle. The plant model incorporates the dynamics of the current control loop, cylinder position control loop and the drive-train. The plant is modeled using two methods. The rst is step response tests. This approach estimates the time delay of the system. The second method involves collecting data from the system under persistent excitations. The plant parameters are estimated by using a modied recursive-least-squares (RLS) algorithm (Dumont and Huzmezan, 2002; Salgado et al., 1988). A rst order model with time delay is developed for the velocity control. The input of the model is the cylinder reference position and the output is the ltered velocity of the vehicle. This resulted in the model of Eq. (12) used in the control design. Due to the cascade structure of the control system this model captures the dynamics of the vehicle drive-train

52

C.A. Foster et al. / Computers and Electronics in Agriculture 47 (2005) 4158

as well as the closed-loop dynamics of the cylinder position and current control loops. Gv (s) = 6e0.15s . 0.85s + 1 (y y)2 (y y )2 (12)

The accuracy of the model can be evaluated using the index dened below = 100 1 . (13)

where lies in the range from 0 to 100, y is the measured plant output, y is the simulated plant output from the model, y is the mean value of the measured output. When = 100, the model perfectly represents the physical plant. The value of for the model in Eq. (12) is 78.9, which suggests a moderately high quality of t. The ZeiglerNichols tuning rule is used to generate an initial set of proportional-integralderivative (PID) control gains (Ogunnaike and Ray, 1994). The control gains are tuned further through simulation and again during implementation. It should be noted that the derivative control implemented uses the output derivative only, known as the tachometric derivative, excluding the unwanted derivative of the reference signal, thus allowing nonsmooth references. The result is that a zero is removed from the closed-loop transfer function. This changes the stability margin as the relative degree of the closed-loop system is increased by one. Since the transfer function in Eq. (12) captures the dyanmics of the cylinder position and current control loops, the stability of the velocity control loop implies the system stability. The control is given by u(k) = up (k) + ui (k) + ud (k) up (k) = Kp e(k) Kp Ts (e(k) + e(k 1)) + ui (k 1) 2i Kp d (yv (k) yv (k 1)) Ts (14a)

(14b)

ui (k) =

(14c)

ud (k) =

(14d)

where e(k) represents the error signal, yv (k) is the measured averaged velocity, Ts is the sample time, Kp is a control gain, i is the integral time constant and d is the derivative time constant. To prevent integral windup due to actuator saturation, the following conditional integration (CI) algorithm is used. if(un = us ) and [(ui > ui max ) or (ui < ui min )] and 0, ev (un u) > 0, e(k) = (14) ev (k), otherwise where ev (k) is the velocity error, us = sat(un , umin , umax ) is the control with the saturation limits umin and umax , un is the nominal control, u = (umin + umax )/2, ui is the integral

C.A. Foster et al. / Computers and Electronics in Agriculture 47 (2005) 4158 Table 1 Implemented velocity control gains Proportional gain (Kp ) Integral time constant (d ) Tachometric derivative time constant (d ) 0.15 1.0 0.4

53

control, and ui max is the upper limit of the integral control. The closed-loop poles of the PID control with tachometric derivative are located at z = {0.44, 0.96 0.017} showing that the system is stable. Table 1 lists the implemented control gains.

3. Control simulations and experimental results The control systems have been implemented and tested on prototype windrowers. Here, we present the simulation and experimental results on one windrower with propulsion-bywire capability. Some relevant parameters of the vehicle are listed in Table 2. 3.1. Current control A PWM amplier operating at 500Hz is used as the current driver. A circuit is designed in such a way that when the current reference is negative, the PWM driver for solenoid A is activated, and the PWM driver for solenoid B is deactivated, and vice versa when the current is positive. A dither signal is added to the PWM output to minimise stiction on the valve spool. This dither signal is responsible for the ripples in the measured current, e.g., in Figs. 11 and 12. Fig. 11 shows simulation results. Fig. 12 shows the corresponding experimental results. The current reference used in the experiments is taken as the reference signal in the simulation. The agreement between the simulation and the experiments is generally excellent. The notable discrepancy is that the overshoot in the experiments is larger than in the simulations. A possible reason is that the control is tuned to have a very fast rise, as required by the cylinder position control in the following stage. We also observe that the feedback control eliminates steady-state errors. In the future, we shall allow the model to adapt in real time to account for varying operating conditions including changing temperature in order to further improve the performance of the control system. 3.2. Cylinder position control The parameters in the deadzone inversions are carefully identied off-line by following the calibration routine for each valve. Experimental data of hydraulic cylinder displaceTable 2 Parameters of the vehicle used in the experiments Total vehicle mass (tractor + header) (kg) Front wheel radius (mm) Wheelbase (mm) Engine power rating (kW) 7853 1493 3485 168

54

C.A. Foster et al. / Computers and Electronics in Agriculture 47 (2005) 4158

Fig. 11. Simulations of the current control loop. The chatter in the response reects the PWM dither signal. Current setpoint represented by dashed line. Current measured represented by solid line.

ment control are shown in Figs. 13 and 14. The system tracks the reference signal with 1 mm steady-state error. The response of the system is smooth. There are no discernible discontinuities in the response as the system transitions from one branch to another. However, Fig. 14 shows large and sharp changes of the control as the system moves across the deadzones. Recall that the control signal of the cylinder position system is the valve current setpoint (ri ). The large and rapid changes of this quantity show the effect of the deadzone inversions. 3.3. Velocity control The ultimate control objective is the velocity of the vehicle. Figs. 15 and 16 show plots of the simulation and experimental results. The velocity reference signal from the experiments is used as the reference signal for the simulation. The experimental velocity control test was designed to evaluate the performance of the vehicle on smooth surfaces, rough ground,

Fig. 12. Experimental results of the current control loop on the vehicle. The PWM dither can be seen from the small chatterings in the ltered response. Current setpoint represented by dashed line. Current measured represented by solid line.

C.A. Foster et al. / Computers and Electronics in Agriculture 47 (2005) 4158

55

Fig. 13. Experimental results of the cylinder position control. Cylinder position setpoint represented by dashed line. Measured cylinder position represented by solid line.

Fig. 14. Control signal (valve current setpoint) of the cylinder position control corresponding to the case shown in Fig. 13.

Fig. 15. Simulation results of undisturbed PID velocity control of the vehicle. Velocity setpoint represented by dashed line. Measured velocity represented by solid line.

56

C.A. Foster et al. / Computers and Electronics in Agriculture 47 (2005) 4158

Fig. 16. Experimental results of the velocity control under the inuence of disturbances due to unmodeled dynamics and road conditions. The data has been collected from the vehicle on a terrain with transitions from level ground, through a hollow and up a hill. Velocity setpoint represented by dashed line. Measured velocity represented by solid line.

uphill grades, downhill grades, as well as rapidly changing and sustained ground conditions. We have found a path that traverses a at asphalt surface, a section of gravel and grass land, and dips into a small gully before rising up a hill to another at asphalt surface. This path is sketched in Fig. 17. By climbing and descending on this path, various ground conditions are tested. Fig. 15 shows the simulation results where the plant is undisturbed and there is no measurement noise in the sensors. It can be seen from the gure that the system tracks relatively well with less than 3% overshoot. The steady-state error is zero and the system responds smoothly. The experimental results in Fig. 16 indicate how well the system operates on the vehicle traveling on the chosen path discussed above. The vehicle is turned around and goes back down the hill along the same path to complete the data set. The most notable feature of the experimental results when compared with the simulation is the oscillations in the measured velocity. These oscillations can be accounted for by unmodeled dynamics in the plant, external disturbances, and sensor noise. The experimental results show some signicant undershoot that is absent in the simulation. This is most likely due to the changes of the system parameters between acceleration and deceleration.

Fig. 17. Illustration of the ground conditions along the path for testing the velocity control. The sketch is not to scale.

C.A. Foster et al. / Computers and Electronics in Agriculture 47 (2005) 4158

57

Overall, the vehicle tracks a velocity command of 11.3 km/h with an error of approximately 0.5 km/h, or 4%. The uncontrolled system with a constant cylinder position providing a velocity of 11.3 km/h on smooth level ground, gives an average tracking error of approximately 2.4 km/h, or 21% when following the same path.

4. Conclusion This paper has presented an automatic velocity control system for a self-propelled windrower. The control system has a cascade structure involving several subsystems. In particular, the control system includes a hydraulic valve with strongly nonlinear characteristics. A segmented inversion method of the valve nonlinear dynamics such as deadzone has been developed for designing the cylinder position control. Stability of the control system has been proven. The control system has been tested and validated experimentally. The experimental results show that the velocity control system is stable, tracks the reference and rejects disturbances sufciently well. The studies show that the techniques developed in this work provide an effective control solution for vehicles with nonlinear electromechanical subsystems. The control designed in this study is based on a static model of the plant. In the future the plant model will be allowed to adapt in real time to account for varying operation conditions. In this regard, this paper provides a foundation of such an extension. Finally, we point out that while this study has not considered wheel slippage, it can be a potential problem, and will be addressed in the future.

Acknowledgements The authors would like to thank John Posselius, Leader of the Research and Development Team at the CNH Technical Center, New Holland, Pennsylvania, for his support throughout the project. This work would not have been possible without the assistance of many people at the CNH Technical Center. We would especially like to thank members of the Electrical Engineering Group and the New Generation Windrower Design Team for their help. The help of Steve Paulin of Eaton hydraulics has been instrumental. Finally, we also recognize the early contribution of Dr. James Glancey of the University of Delaware in the formation of the project.

References
Auernhammer, H., 2001. Precision farmingthe environmental challenge. Comput. Electron. Agric. 30 (13), 3143. Callaghan, V., Chernett, P., Colley, M., Lawson, T., Standeven, J., Carr-West, M., Ragget, M., 1997. Automating agricultural vehicles. Ind. Robot 24 (5), 364369. Dumont, G.A., Huzmezan, M., 2002. Concepts, methods and techniques in adaptive control. In: Proceedings of the American Control Conference, vol. 2, Anchorage, Alaska, 810 May, pp. 11371150. Earl, R., Thomas, G., Blackmore, B.S., 2000. Potential role of GIS in autonomous eld operations. Comput. Electron. Agric. 25 (1), 107120.

58

C.A. Foster et al. / Computers and Electronics in Agriculture 47 (2005) 4158

Elmer, K.F., Gentle, C.R., 2001. A parsimonious model for the proportional control valve. Proc. Inst. Mech. Eng. Part C: J. Mech. Eng. Sci. 215 (11), 13571363. Ferreira, J.A., Gomes de Almeida, F., Quintas, M.R., 2002. Semi-empirical model for a hydraulic servo-solenoid valve. Proc. Inst. Mech. Eng. Part I: J. Syst. Control Eng. 216 (3), 237248. Fritz, H., 1996. Neural speed control for autonomous road vehicles. Control Eng. Pract. 4 (4), 507512. Hagras, H., Colley, M., Callaghan, V., Carr-West, M., 2002. Online learning and adaptation of autonomous mobile robots for sustainable agriculture. Autonomous Robots 13 (1), 3752. Ishida, A., Takada, M., Narazaki, K., Ito, O., 1992. A self-tuning automotive cruise control system using the time delay controller. SAE Paper No. 920159. Jelali, M., Schwarz, H., 1995. Nonlinear identication of hydraulic servo-drive systems. IEEE Control Syst. Magazine 15 (5), 1722. Keicher, R., Seufert, H., 2000. Automatic guidance for agricultural vehicles in Europe. Comput. Electron. Agric. 25 (1), 169194. McMahon, D.H., Hedrick, J.K., Shladover, S.E., 1990. Vehicle modeling and control for automated highway systems. In: Proceedings of American Controls Conference, vol. 1, San Diego, CA, pp. 279303. Ogunnaike, B., Ray, W.H., 1994. Process Dynamics, Modeling, and Control. Oxford University Press, New York. Rajamani, R., Shladover, S.E., 2001. Experimental comparative study of autonomous and co-operative vehiclefollower control systems. Transportation Res. Part C: Emerging Technol. 9 (1), 1531. Rajamani, R., Tan, H.-S., Law, B.K., Zhang, W.-B., 2000. Demonstration of integrated longitudinal and lateral control for the operation of automated vehicles in platoons. IEEE Trans. Control Syst. Technol. 8 (4), 695708. Rajamani, R., Zhu, C., 2002. Semi-autonomous adaptive cruise control systems. IEEE Trans. Vehicular Technol. 51 (5), 11861192. Reid, J.F., Zhang, Q., Noguchi, N., Dickson, M., 2000. Agricultural automatic guidance research in North America. Comput. Electron. Agric. 25 (1), 155167. Rizzoni, G., 1995. Principles and Applications of Electrical Engineering. McGraw-Hill, Burr Ridge, IL. Salgado, M.E., Goodwin, G.C., Middleton, R.H., 1988. Modied least squares algorithm incorporating exponential resetting and forgetting. Int. J. Control 47 (2), 477491. Schwarz, H., 1998. Output feedback stabilization of hydraulic drives based on bilinear approximations and canonical observers. Proc. Inst. Mech. Eng. Part I: J. Syst. Control Eng. 212 (2), 7985. Setlur, P., Wagner, J.R., Dawson, D.M., Samuels, B., 2003. Nonlinear control of a continuously variable transmission (CVT). IEEE Trans. Control Syst. Technol. 11 (1), 101108. Stentz, A., Dima, C., Wellington, C., Herman, H., Stager, D., 2002. A system for semi-autonomous tractor operations. Autonomous Robots 13 (1), 87104. Torii, T., 2000. Research in autonomous agriculture vehicles in Japan. Comput. Electron. Agric. 25 (1), 133153. Vaughan, N.D., Gamble, J.B., 1996. The modeling and simulation of a proportional solenoid valve. Trans. ASME: J. Dynamic Syst. Measurement Control 118, 120125. Zhang, N., Wang, M., Wang, N., 2002. Precision agriculturea worldwide overview. Comput. Electron. Agric. 36 (23), 113132.