Methods for Automatic Hydraulics Calibration in Construction Equipment

(1)

School of Innovation Design and Engineering

V¨

aster˚

as, Sweden

Thesis for the Degree of Master of Science (120 credits) in

Computer Science with Specialization in Embedded Systems

30.0 credits

Methods for Automatic Hydraulics

Calibration in Construction Equipment

Peter Charbachi

pci13001@student.mdh.se

Filippo Ferrario

ffo16001@student.mdh.se

Examiner: Daniel Sundmark

M¨

alardalen University, V¨

aster˚

as, Sweden

Supervisors: Alessandro Papadopoulos & Saad Mubeen

M¨

alardalen University, V¨

aster˚

as, Sweden

Company Supervisor: Lars-Erik Larsson

Volvo Construction Equipment,

Eskilstuna, Sweden

(2)

Abstract

In this thesis we investigate the problem of automatic calibration and control of hydraulic compo-nents in the domain of construction equipment. Methods that are able to remove a costly manual approach in favour of an automatic one are investigated and evaluated. The thesis aims to inves-tigate what methods are available in achieving this goal as well as evaluate the performance and applicability of such methods in the domain of construction equipment.

The literature indicates that a great focus is put on learning a model of the plant at run time in order to provide accurate control. Common approaches to the problem are the Recursive Least Square method and PID controllers for non-linear systems, but other methods are also present, such as the Nodal Link Perceptron Network (NLPN).

The methods chosen to be compared are the existing method of manually calibrating two set points for start and end current and interpolating between them; the use of a PI controller with a static line inverse model; a PI controller with a static curve inverse model; a PI controller with an NLPN adaptive inverse model and lastly, a completely NLPN based control strategy.

The methods were implemented in Matlab Simulink and evaluated in simulations based on data collected from real wheel loaders in the construction equipment domain, produced by Volvo CE. The simulations are performed on data from three machines and were evaluated twice for the adaptive methods in order to evaluate how well the methods improved. The results were then evaluated in terms of average absolute error, as well as a discussion of the behaviour shown in the plots.

The evaluations indicates that the most effective method for control is the PI controller using a static line inverse model. The method produces the smallest average error of both actions evaluated, lifting and lowering of the boom, while the complete NLPN solution provide the worst results.

(3)

List of Figures

1 Wheel loader . . . 5

2 Wheel loader top and side view . . . 7

3 Wheel loader sub system . . . 8

4 Lifting and tilting of GET . . . 8

5 Simplified diagram of the hydraulic system. . . 10

6 Intersection of a 5-way spool valve. . . 11

7 Structure of a general neuron for an artificial neural network. . . 12

8 Structure of a generic artificial neural network. . . 12

9 PID using error feedback. . . 13

10 Polynomial Regression Analysis. . . 14

11 Existing calibration method. . . 16

12 Behaviour of the machine speed with regards to current. . . 17

13 Research methodology. . . 22

14 NLPN Structure . . . 24

15 Simulink representation. . . 25

16 Simulation design of the plant. . . 27

17 PID + Static Inverse . . . 28

18 PID + Adaptive Inverse . . . 30

19 Adaptive NLPN Control . . . 30

20 Evaluation command used to test all solutions. . . 31

21 Evaluation of Existing method on each machine. . . 32

22 Evaluation of PI + Line Inverse on each machine. . . 33

23 Evaluation of PI + Curve Inverse on each machine. . . 34

24 Evaluation of PI + Adaptive Inverse first run on each machine. . . 35

25 Evaluation of PI + Adaptive Inverse second run on each machine. . . 36

26 Evaluation of NLPN method first run on each machine. . . 37

27 Evaluation of NLPN method second run on each machine. . . 38

28 Lifting results of the evaluated methods compared to the baseline. . . 39

29 Lowering results of the evaluated methods compared to the baseline. . . 40

(6)

1 Introduction

The domain of construction equipment contains a large variety of unique machines and vehicles. All of these machines are tailored to excel at their specific tasks, being excavators for digging, articulated haulers for transporting materials, compactors to flatten the ground or wheel loaders, see Figure1, for moving and loading materials. Every machine has its own purpose and thus a place in the operators toolbox. Even though the machines are tremendously specialised towards their specific task, they are still highly customisable, with a range of different attachments to further tailor them to the current workload. Operating these machines is no easy task [1], an action that seems as trivial as filling the wheel loader bucket with gravel is actually a rather challenging task for a novice. It takes professional operators a lot of time and practise to become familiar enough with the machines to manoeuvre them effectively and smoothly. If the machine becomes unsuitable for the work task at hand and is required to be changed; or if its behaviour changes during its lifetime, the operator might be introduced to a machine with an unfamiliar behaviour. In this case, the operators experience and familiarity loses its relevance, which motivates the need for all machines to display the same behaviour.

Figure 1: Wheel loader with a bucket attachment.

Hydraulics systems, which are a main part of construction equipment, consist of components that posses individual characteristics. This individuality does not only manifest by the manufac-turing of the component, but also by its usage and wear which is known as hysteresis. This in turn results in a range of different machine behaviour which diverges from the behaviour that the operator is familiar to. The described phenomenon does not only reduces operator productivity, but can also negatively impact safety.

The current way to deal with this changing behaviour is done by the use of individual calibra-tion of the hydraulic system, where the aim is to conform machine behaviour as well as improve manoeuvrability. Calibration is a manual process and for this reason it represents a costly pro-cedure for companies. Automatic calibration of the hydraulics is an approach to therefore reduce cost and also improve machine quality.

This thesis focus on the investigation, development and evaluation of automatic calibration methods for construction equipment hydraulics. The study is performed in collaboration with Volvo CE, and is aimed at their wheel loaders machines. The focus will be on the wheel loaders ground engaging tool (GET), also known as boom arm. However, the results of the study are not limited to the wheel loaders’ GET, but could potentially be applied to other parts of the machines hydraulic system such as steering as well as other construction equipment machines.

(7)

1.1 Thesis Goals

This thesis investigates methods used for automatically calibrating and operating the electro-hydraulic system commonly found in construction equipment. The electro-electro-hydraulic system, com-pared to a traditional hydraulic system, makes use of electronic signals to control solenoids which in turn control the hydraulic valve of the system. This design introduces challenges regarding control of the hydraulic system where the control actuator used to operate the valve can be considered distanced from the rest of the system. The solenoid valves are controlled with PWM (Pulse Width Modulation) signals and do not have a perfect translation from signal to the actuation of the valve. Due to various reasons, such as spring strength and areas with varying friction inside the valve, all valves do not conform to the same behaviour. This individuality manifests as different strengths of which the solenoids are required to use in order to achieve the desired flow rate.

In order to achieve sufficient control and manoeuvrability of the system, the current sent to the solenoids needs to be chosen properly. The electronic control actuator, lever, is limited by its area of operation, and thus need to encompass all behaviour of the system while also providing an intuitive response. If the electronic control actuator is operated at 1%, then the output should reflect this. A starting and stopping current are thus used to bound the control area. A current too low and the valves will not open; while a current too high would lead to the solenoid fully opening early and then continuing applying force above and beyond what is needed. While both approaches are able to encompass the desired behaviour, the levers operation area would be filled with dead movement and thus ruin the intuitive response that is desired. By bounding the current to the area that provides any actual behaviour, it is possible to remove any dead movement while improving the intuitiveness of the control.

1.2 Thesis Outline

Section2explains the background necessary to grasp the work. It consists of an explanation of the system as a whole and its components, the existing method used as well as some understanding of methods that are used hereinafter. A further explanation of the problem at hand is then explained in Section 3, where this section also houses the research questions to be answered. Section4provides the reader with some related work present in the area of automatic calibration and control of hydraulic systems. The methodology of the thesis is then explained in Section 5, where the processes used as well as the rationale of their acceptance are explained. A deeper understanding of the specific method studied in this thesis is explained in Section6followed by the explanation of the simulation and the models used in Section7. An evaluation of the study done is presented in Section8which is followed by Section9that describes the threats to validity before being discussed in Section 10. A conclusion presented in Section 11 will follow the discussion, where the results are compiled and used to answer the research questions. The thesis will lastly present the future work in Section12, where all methods and ideas not fully realised are discussed.

(8)

2 Background

This background section introduces the key elements necessary to better understand the work presented in this thesis. First, a general knowledge of the machines and how they are operated is provided. Following this, an explanation on their working hydraulics, spool valves and the calibration process needed is given. The section is concluded with an introduction of the techniques used in the thesis.

2.1 Wheel Loader Machines

Wheel loaders are machines that consist of four wheels, a driver cabin and ground engaging tool (GET), with an operating weight ranging from about 5 metric tonnes to over 50, see Figure2for a side and top view of a typical wheel loader. They exist in various sizes which in turn introduce them to a broad set of applications. Small to mid-range machines see more work in construction and material handling while the larger ones are more commonly used in mining areas [2].

Figure 2: Volvo L260H wheel loader design in side and top view.

A typical wheel loader is powered by a diesel engine which is responsible for not only driving the wheels but also powering the hydraulic pump. Power to the wheels is transferred through a torque converter and mechanical transmission, similar to the design found in cars and trucks. The hydraulic pump, which is also powered by the diesel engine, handles task such as steering, braking and controlling of the GET. Figure 3 shows how the diesel engine is the sole power source and how the power travels to the tires as well as get transferred to the hydraulic pump. The hydraulic pump then utilise that power to operate brakes, steering and GET, while both brakes and steering are also used in the control of the tires.

Compared to a normal car, a wheel loader does not steer by turning the front wheels or even the front axis. The front and rear axis are fixed, with the rear axis only able to flex a bit vertically to keep as many wheels on the ground as possible in order to increase stability. Wheel loaders have hydraulically controlled articulated steering, where the vehicle’s whole front section is connected to the rear by an articulation joint. The rear axis supporting the engine is connected to the front axis through this articulation joint and allows for power to be transferred in order to provide power to all four wheels. The articulated steering allows the front axis of the vehicle to turn independently of the rear, causing the front and rear tires to follow in each others tracks, see the top view in Figure

2. This is achieved with the use of two single acting hydraulic cylinders mounted on each side of the articulation joint. To turn the vehicle, one of the hydraulic cylinders is pressurised to push the vehicles front section around the joint. The single acting design of the hydraulic cylinder permits only one of the cylinders two actions to be used, where only the extension action is allowed. This requires the other cylinder to be left unpressurised permitting the oil to flow out of the cylinder in

(9)

Brake Hydraulic Power Source

(Diesel Engine)

Torque Converter +

Mechanical Transmission Tires

Machine Traction & Speed

Tool Force & Speed Implements

Hydraulic Ground Engaging Tool (GET)

Steering Hydraulic Hydraulic Pump

Powertrain

Figure 3: Wheel loader sub system, based on the work of Cobo et al. [2].

order to perform the retraction of the piston.

The wheel loader’s hydraulically controlled GET is an iconic part of the machine. It is able to be modified with a copious amount of different attachments, all aimed for various tasks. These attachments could vary from buckets specially made to work with gravel or rock, forks following a standardised pallet size, claws for carrying trees, or brushes for cleaning the road. The wheel loader GET is the part of the machine responsible for engaging the environment and does this according to the operator. These GET actions are often operated by the use of four levers that control the hydraulic actuators. The levers responsible for controlling the GET varies between two designs, the first is the hydraulic control, where the lever is a mechanical control to operate the hydraulic valves. This approach requires hydraulic tubing to be connected inside the cabin, making it only practical as an option for smaller machines. The other approach makes use of electric levers, where the signal from the lever is transported to actuate a solenoid to open a valve. This approach is exclusive for the control of larger machines, where laying hydraulic tubing into the cabin is costly and impractical.

Figure 4: Lifting and lowering of the GET is shown on the left while tilting in and out is shown on the right.

(10)

The first and second levers are responsible for the lifting and tilting actions of the GET respec-tively, where the actions can be seen in Figure4. These commands are essential to the functionality of the machine, and for many attachments, the only actions applicable. They work by controlling typically three hydraulic cylinders for larger wheel loaders or possibly two for smaller ones. One or two hydraulic cylinders are used to perform the lifting action with the remaining hydraulic cylinder being used for tilting. In order to provide both directions of actions, lifting and lowering, tilting in and out, the hydraulic cylinders are required to be double acting. This means that they are able to fill both chambers separated by the piston with hydraulic fluid, thus allowing them to be operated in both directions. This design is different from the one regarding steering, where the hydraulic cylinder’s retraction is dependant on the other cylinders extraction. The third and fourth levers are different, they provide general purpose hydraulic actions which are dictated by the attachment. Example of such actions can be road cleaning brush attachments which requires hydraulic fluid in order to spin and operate. Other examples can be special fork attachments which have one height modifiable fork arm, or bucket attachments that have the additional functionality to tip sideways. All attachments and their actions are dependant on the GET which in turn makes them dependant on the hydraulic system operating it.

2.2 Hydraulics

The increased digitalisation which has affected a magnitude of fields, has not advanced without making an impact on the construction equipment domain [2]. Advances in autonomous machines has even made its way into the construction equipment domain [3], with Volvo CE even introducing prototype autonomous machines [4]. Other advances such as electric or hybrid machines has also been making an appearance in order to reduce noise, fuel consumption and emissions. An example of this is the EX2 fully electric excavator introduced by Volvo CE as part of their electromobility product line [5]. This product line consist of vehicles and machines that are able to utilise an electric motor to perform either the machines propulsion or main function [6].

The hydraulic system has also gotten changes, however more discrete. One such change affects the control of the hydraulic actuators, which involve the introduction of electronic control by the use of pulse-width modulation (PWM) signals and solenoids. The process and connections can be seen in the simplified diagram shown in Figure 5. The operator commands the machine by the use of the levers located in the cabin. These are then processed by an electronic control module which converts the actions into PWM signals. The signals are then sent to operate the solenoid valves inside the pilot valves controlling the hydraulic fluid, which in turn controls the main valve that then controls the hydraulic cylinder. The solenoid valves operating the hydraulics are of the proportional design, which allows the operator to gradually specify the size of the orifice of which the hydraulic fluid exits. This design affects the rate of which the hydraulic fluid travels through the valve to the hydraulic cylinder which translates into its operating speed. The use of pilot valves allows for weaker valves and is able to better utilise power already trapped in the system, it does however allow for more steps in between, which depending on their individual characteristics can affect the overall behaviour.

2.2.1 Spool Valves

Hydraulic valves are used for controlling operations regarding hydraulics in a system by restricting the flow of fluid (hydraulic oil) transported into the actuator. There are various kind of valves, such as spool valves, servo valves and poppet valves, where electronically controlled variations also exist [7,8]. The focus in this thesis is on spool valves, which are the main valve used in the system considered for evaluation. Spool valves consist of a ridged spool housed in a cylinder body with various ports depending on configuration. Figure 6 shows a 5-way spool valve, where both tank ports are connected internally, which in turn results in only four external ports. This valve is similar to the “Main Valve” in Figure5, whereas the “Pilot Valve” is a 3-way spool valve.

Control of the spool valve can be done with various methods, such as the previously mentioned electronic and hydraulic control. The control strategy moves the spool inside the cylinder body to open up or close off different paths in the valve to direct the hydraulic oil inside towards the intended ports. The proportional versions of the valves also exist, which extend the functionality of

(11)

Lever Tank Pump 1 Pump 2 Electronic Control Module (ECM) PWM Operator Hydraulic Cylinder Tilt Lift Hydraulic Cylinder Main Valve T P T O O Main Valve T P T O O P T O Pilot Valve P T O Pilot Valve P T O Pilot Valve P T O Pilot Valve

Figure 5: Simplified diagram of the hydraulic system.

the mentioned valves by allowing an increased precision of the spool’s location inside the housing. These valves are able to produce finer control of flow rate, which in a wheel loader hydraulic system would be translated into the actions operating speed.

Due to the design of the spool valve, it is able to perform two actions at once, which intuitively makes it a good choice for a hydraulic cylinder. It allows pressurised oil to travel to the direction of choice, while allowing a path for the oil in the other cylinder to escape. Without a way to release the oil in the other cylinder, vacuum and compressed oil would prevent the system from performing as intended. Such a two action design, while keeping control of the system simple, also makes adjustment to the control coarse, causing fine control of the system difficult. Methods that handle control of the in-take and out-take separately do exist and are able to improve control precision. This “independent metering” design, does not only increase control difficulty, but also price, as it imposes the need for separate valves for every port, practically doubling the amount of valves.

2.3 Calibration

Calibration is the process done in order to conform machine behaviour towards a desired uniform behaviour. For the electronically controlled hydraulic systems, this process is manually done by setting lower and upper ranges for the PWM signal of which the lever operates within. This is achieved by a technician that is required to travel to the machine in person. The process of calibration then requires the use of specialised equipment in order to analyse various hydraulic machine actions and set the ranges for the PWM signals. These actions include lifting and tilt-ing of the GET, which are operated and evaluated until deemed satisfytilt-ing. The hydraulics can

(12)

Tank Pressure A B P→B T→A P→A T→B

Figure 6: Intersection of a 5-way spool valve.

be considered satisfying when the control is consistent and smooth, i.e. no sudden uncontrolled movements or jerking actions, and in line with the uniform behaviour. Such a calibration method does not take into consideration any behaviour occurring whilst within the specified range, but only aims to encapsulate it. To constantly achieve the desired behaviour of the control, repeatedly calibrating the hydraulics is a necessary task. This becomes more apparent when major adjust-ments to machine configuration are made, such as machine repairs. The changes in attachadjust-ments can also affect this, which due to the varying weight of the attachments are able to make existing calibration inadequate. In order to avoid the need of manually calibrating the system to respond to such changes, an automatic approach is investigated as a way to allow the machine to respond and adapt itself. The calibration process can be seen as an optimisation problem, where the aim is to minimise the difference between desired and actual machine behaviour. Optimisation strategies work by constantly aiming to improve the current solution to a problem by maximising or min-imising a property. This is done by modifying a set of parameters and evaluating whether the new configuration provides a solution closer to the optimal result. This process is then repeated until an acceptable solution is provided. The acceptable solution in these outcomes are often within a predefined tolerance range of the optimal result. It is not impossible for optimisation problems to be excessively large and unmanageable, which makes finding a perfect solution impractical. In these cases a tolerance value is used to make the problem feasible.

2.4 Artificial Neural Network

Artificial Neural Networks (ANNs) are computational units that has been partially inspired by the observation of biological systems as explained by Mitchell [9]. These biologically inspired learning systems are built as complex webs of interconnected units also known as neurons or perceptrons to mimic the biological connections found in the brain. A neuron follows a simple design that consists of a number of real valued inputs, real valued weights, an activation function and a single output. The amount of weights are mirrored by the amount of inputs plus the bias. The weight of each input specifies its importance towards the activation function and in turn towards the neuron as a whole. A visual description of the structure can be found in Figure 7. While the design of artificial neural networks is inspired from the biological neural systems, there are still many biological complexities that are neglected within ANN systems. The architecture used when designing a typical ANN is built upon three kind of layers: input, hidden and output layer, as depicted in Figure 8. The input layer acts as the neural networks interface, thus containing one neuron per specified input to the system. Following the input layer is one or many hidden layers, where every layer can have an independent and arbitrary number of neurons defined by the architecture. For every hidden layer, each neuron’s input is connected to all the outputs from

(13)

∑

X0 = 1 W0 Activation Function Output X1 X2 X3 Xn W1 W2 W3 Wn

Figure 7: Structure of a general neuron for an artificial neural network.

neurons in the previous layer. The network ends with an output layer where the amount of neurons is chosen depending on the task. For classification problems, the output layer usually contains one neuron per category, whilst approximation problems often only use one neuron. The data sent to such a network can be visualised as passing through a pipeline where every step is a network layer. The training of a neural network is done with a technique called backpropagation. In this phase the aim is to minimise a target scalar cost function (e.g., the average quadratic error) using a set of training data samples. This is achieved by computing the gradient of the error with respect to the network’s weights and using a minimisation technique.

Neural networks have been able to provide robust approaches for approximating real, discrete and vector target functions. Additionally, ANNs were also deemed the most effective methods for certain types of problems, such as the ability to interpret complex real world sensor data [9]. Their design makes them well suited for problems involving different kind of sensors, where the training data is complex, noisy and derived from devices producing information about a specific environment. After training, the ANN will eventually have constructed a set of weights unique to its architecture that model a solution to solve the problem. These weights are not always simple to be interpreted by a human but represents the key for an ANN to achieve its goal.

Input Layer Hidden Layer 1 Output Layer Hidden Layer 2

(14)

2.5 Proportional Integral Derivative

Proportional Integral Derivative (PID) control is the most common technique for feedback control in engineering systems according to ˚Astr¨om and Murray [10]. The name of this method derives from the combinations of the three components it is composed of. Their purpose is to produce a correction based on the error, computed as a difference between the desired setpoint and the measured one.

∑

System

1 r

e

Controller

u

y

e(τ)dτ

K

i

_∫

t 0

K

d

de

_dt

e

K

p

Figure 9: PID using error feedback.

Figure9illustrate the structure of a PID which compute an error value e as difference between the setpoint r and the measured process variable y. Corrections to this method are applied ac-cording to the three components, where P represent the proportional part of the error, can be seen as the “present” error. I is the integral part of it and represents the accumulation of the “past” error. And finally the last part, D can be considered as the prediction of the “future” error. In fact, as stated by Araki [11], the PID controller can be understood as a controller that takes the present the past and the future of the error into consideration. Thus, the control action is the sum of all the three terms. The system tries to minimise the error e by adjusting over time the control variable u.

The input/output relation of the constants in a PID controller with error feedback, as the one illustrated in Figure9, can be described according to Araki [11] as follow:

u = Kpe + Ki Z t 0 e(τ )dτ + Kd de dt (1)

where all the controller parameters are visible: the proportional gain Kp, the integral gain Kiand

the derivative gain Kd.

The PID is the most widely used controller in process control until today, according to Araki [11]. Investigations performed in Japan, showed that 90% of the controllers used by the industries are PID controllers or modified version of this technique.

(15)

2.6 Regression Analysis

Regression analysis is a statistical technique used to identify a correlation between data in a dataset and aims to construct a model that represents predictors relationship [12]. The resulting model can then be used to estimate values in ranges the data set did not cover.

A model from such a regression analysis can be seen in Equation2, that shows the form of a model from a simple linear regression analysis. The β are the models unknown parameters that the regression analysis aims to estimate whilst the variable x is often known as “predictor” or “regressor” and y goes by the name “response”. The is added to the model as the difference between the observed value of y and the estimated result produced by the regression model and is known as the “error”, which is modelled as a random variable with a given probabilistic distribution. The error is seen as the red lines in Figure10 and can also be seen as the distance between the estimated result and the observed value.

This simple linear regression model plots a line through the data set and does this with one regressor x, these two facts are what gave it its name simple linear regression model.

y = β0+ β1x + (2)

Other models also exist that consist of multiple regressors, known as multiple linear regression models, or designs that does not produce a line, but a polynomial known as polynomial regression models. Polynomial regression models aim to estimate a curve through the data set and can do this with a curve of a predetermined arbitrary order. They are thus able to provide a better representation for non-linear system, as compared to their linear counterparts. A general rule for polynomial regression models is the use of as low order as possible for the regression model, as transformations are preferred before increasing the model complexity. Model estimation for polynomial regression models outside the range of the data set, known as extrapolation, as well as inside the data set known as interpolation, can also be very hazardous, where polynomial roots can appear at unexpected places.

A way to handle such methods is the piecewise regression models, such as the Spline method or Piecewise Cubic Hermite Interpolating Polynomial. By splitting the variable space at different locations, known as knots, it is possible to assign each section with their own function. This can be useful to model real system, such as hydraulic valves. A polynomial will have difficulties capturing the flat behaviour seen in Figure12, whilst a piecewise method would be more appropriate.

Figure 10: Polynomial Regression Analysis.

2.6.1 Least Squares Estimation Method

The least squares method is a prediction method capable of function estimation [13]. It is possible to perform both population based as well as sample based and has therefore seen common use in literature, see Section4.3. The latter approach, namely sample based, is also known as recursive least squares, and operates with the same goal as the regular one but continuously adapts the model to the ever growing data set.

The least square method attempts to estimate values for the regression models parameters, such as β0and β1, in order find the parameters that produce the “least squares”. The least squares is

the criteria of minimising the sum of residual squares, where the residuals are also known as the error . For more on Least Squares, we direct the reader to the work by Montgomery et al. [12].

(16)

3 Problem Formulation

In order to describe the problem addressed in this thesis, a first introduction of the working prin-ciples and the existing method is presented. The problem formulation is described right after the initial introduction. Concluding this section are the research questions, representing the funda-mental core of this thesis.

3.1 Calibration Principles

In order to better understand the purpose of this thesis it is necessary to describe the current calibration procedure. The current calibration method of hydraulic systems in construction equip-ment is manually conducted, which makes it both a laborious and costly process. The customers need to request and hire a specialised technician to travel to the machine in order to assess and calibrate it. This process can take a long time, especially when considering the machine’s location. Omitting calibration reduces the responsiveness and control of the machine which in turn affects the safety and productivity of the system. Other than the cost introduced to the customer, the calibration process is also costly for the company. It demands the need for educated technicians and specialised hardware. Besides maintenance of the systems, the process also becomes costly during manufacturing. Before delivering a system, the company needs to make sure that it respects their quality guidelines. This includes hydraulic manoeuvrability and response, which require the company to take the role of a customer and hire technicians to perform calibration. Depending on the number of systems manufactured and the time taken for calibration, this cost could quickly escalate.

A solution able to online calibrate the system while improving control would help keep the cost down by reducing travelling and training time for the technician as well as the time they would take to perform the calibration.

The problem with existing solutions lay in the use of electronic signals for control of the hy-draulic operations. The electronic signal is acquired from the electrical levers located in the cabin which basically outputs a percentage of the amount the lever was tilted in a certain direction. This results in an output space of -100% to 100%, where the sign represent a different action, for example lifting or lowering. Currently, the lever percentage value is mapped to a current between two set calibration points. The set points are located in the input space of the solenoid pilot valves, which is in the form of a PWM signal with a current ranging from 0 to one large enough to cover the behaviour of all valves as specified by the manufacturer. These two set points are calibrated or “set” accordingly to encompass the range in the solenoid valves input space where it produces an actual machine response. The “start current” or 0% set point is set slightly below the current where the valve is just beginning to open, where the “end current” or 100% set point marks the current where the valve is fully open. Failing to properly set the start and end points would result in similar scenarios as depicted in Figure 11(a) and 11(b). If the set points are placed too far from the range of the actual machine response, it will lead to portions of the lever input space being translated into a current that produces no actual machine response, see Figure11(a). These ranges cause what is known as deadpan, which reduce the manoeuvrability of the machine. It will appear not only in the form of useless lever movements, but also introduce coarseness to the control by compressing the active range. If the set points instead are set to trespass the range of actual machine response, other problems arise, see Figure11(b). Here the starting value is set so high that the machine would produce too much current at the beginning of the lever movement, thus causing jerking actions. By setting the end set point too low there is a limitation imposed on the machine which restrains the machines potential which cause sluggishness. An ideal calibration of the set points is shown in Figure 11(c). Here only the range of the actual machine response is encompassed, removing any deadpan and jerking action while allowing the machine to perform optimally.

3.2 Existing Method of Calibration

The current method of finding these set points varies depending on the action to be calibrated and point under calibration, but the method is however still mainly done with the use of manual testing.

(17)

A)

No Current 0% set point 100% set point Max Current

6,7% 13,3% 20,0% 26,7% 33,3% 40,0% 46,7% 53,3% 60,0% 66,7% 73,3% 80,0% 86,7% 93,3% 100,0%

0% 5% 10% 15% 20% 25% 30% 35% 40% 45% 50% 55% 60% 65% 70% 75% 80% 85% 90% 95% 100%

B)

10,0% 25,0% 40,0% 55,0% 70,0% 85,0% 100,0%

Actual machine respone

0% 5% 10% 15% 20% 25% 30% 35% 40% 45% 50% 55% 60% 65% 70% 75% 80% 85% 90% 95% 100%

C)

10,0% 20,0% 30,0% 40,0% 50,0% 60,0% 70,0% 80,0% 90,0% 100,0%

0% 5% 10% 15% 20% 25% 30% 35% 40% 45% 50% 55% 60% 65% 70% 75% 80% 85% 90% 95% 100%

Deadpan Actual machine respone Deadpan

(a) Existing calibration with over estimated set points.

A)

6,7% 13,3% 20,0% 26,7% 33,3% 40,0% 46,7% 53,3% 60,0% 66,7% 73,3% 80,0% 86,7% 93,3% 100,0%

0% 5% 10% 15% 20% 25% 30% 35% 40% 45% 50% 55% 60% 65% 70% 75% 80% 85% 90% 95% 100%

B)

10,0% 25,0% 40,0% 55,0% 70,0% 85,0% 100,0%

0% 5% 10% 15% 20% 25% 30% 35% 40% 45% 50% 55% 60% 65% 70% 75% 80% 85% 90% 95% 100%

C)

10,0% 20,0% 30,0% 40,0% 50,0% 60,0% 70,0% 80,0% 90,0% 100,0%

0% 5% 10% 15% 20% 25% 30% 35% 40% 45% 50% 55% 60% 65% 70% 75% 80% 85% 90% 95% 100%

(b) Existing calibration with under estimated set points.

A)

6,7% 13,3% 20,0% 26,7% 33,3% 40,0% 46,7% 53,3% 60,0% 66,7% 73,3% 80,0% 86,7% 93,3% 100,0%

0% 5% 10% 15% 20% 25% 30% 35% 40% 45% 50% 55% 60% 65% 70% 75% 80% 85% 90% 95% 100%

B)

10,0% 25,0% 40,0% 55,0% 70,0% 85,0% 100,0%

0% 5% 10% 15% 20% 25% 30% 35% 40% 45% 50% 55% 60% 65% 70% 75% 80% 85% 90% 95% 100%

C)

10,0% 20,0% 30,0% 40,0% 50,0% 60,0% 70,0% 80,0% 90,0% 100,0%

0% 5% 10% 15% 20% 25% 30% 35% 40% 45% 50% 55% 60% 65% 70% 75% 80% 85% 90% 95% 100%

(c) Existing calibration with optimally set points.

Figure 11: Existing calibration method.

In order to perform calibration, the use of a weight on the boom is required, which is of importance in calibrating some actions. This requirement is met by just equipping a bucket attachment on to the machine. For the start current (0%) set point for the booms lifting action, the process is done by setting a start current low enough to not cause any response from the machine. This while still being within the range of acceptable currents which the calibrated set point is required to be located within for the valve to be deemed fully operational. The current is then tested and the hydraulic pressure after the main valve is measured to determine if there is any actual response. If there is no response, the machine is left to settle for a few seconds before the current is raised and tested again. This whole process is then repeated until the desired pressure is met. For the end current (100%), the process differs, but still also depend on the approach of manual testing. Two currents are first measured, one high enough to be certain that the valve is fully open, and one lower than the first one but still higher than the previously set start current. A line between the points is then calculated and used to estimate the actual end current. If this current does not fall within the requirements, parts of this process will also be repeated. For setting the start current for lowering the boom, the process is similar to lifting, where it is gradually increased until a satisfying result is found. However, for the end current, the process is quite different, mainly due to two reasons. The first one is due to gravity, where a fully open valve could lead to speeds too high for the machine to operate ideally, thus requiring restrictions. The second reason is due to something called the float function and the design of the valves in the machine. The float function is a way that allows the operator of the machine to cause the boom and bucket to fall to the ground and lay there. This can then be used by the operator to scrape or even out the ground while operating the machine in reverse. To achieve this, the machine de-pressurises the cylinder by fully opening both cylinder ports to the tank while also connecting the pressure intake directly to the tank. This is a special state of the machine and is achieved with the special design of the main valves. By applying extra pressure to the valve, which is done by increasing the current above the end current thus trying to open the valve more than maximum, this state is activated. It is only possible to achieve for the lowering action of the boom, and therefore impose problems on the actions calibration process. The process that aims to find the end current uses a current that is high enough to ensure that the valve is fully open. This action can accidentally activate the float function of the machine, causing irregular and faulty data, while negatively impacting safety. Another method is therefore constructed in order to calibrate the end current for the boom lowering action. This action greatly depends on the need for a bucket or equally heavy attachment to be connected to the boom and focuses on measuring the speed of the action. The boom is fully raised with its attachment locked, before being lowered at a constant current. The speed of the

(18)

boom is measured and compared to an ideal speed. Depending on the result, the process might be repeated multiple times after increasing or decreasing the current. The tilting in and out set point calibration of the boom however follows the exact same process as the lifting calibration, where the start current is gradually increased and the end current is calculated and tested.

3.3 Problem Statement

Due to the reality of hydraulic systems suffering from hysteresis where the machine’s behaviour is subject to change over time and use, the method of calibration is used to conform machine behaviour and ensure quality.

Current

Speed

Figure 12: Behaviour of the machine speed with regards to current.

A way to automate or remove the need of this process is therefore of interest as an approach to reduce cost and increase manoeuvrability. The current solution, while able to produce an adequate controllability of the complete machine behaviour, fails to further improve the control within the actual machine response range. The machine behaviour is non-linear which in turn can cause unwanted behaviour during operation. This non-linear nature of a machine’s speed in regards to the current can bee seen in Figure 12, where the data is collected from the lifting action of one wheel loader. Some interesting details can be recognised from this picture, the first one is the non-linearity of the system, where the increased of current does not increase the speed in a linear fashion. Another interesting aspect is related to the physical limitation of the system where after a certain current, any further increase would not translate into an additional rise in speed. This is an important section of the curve since it represents the ideal location to set the end current. It is worth noting that the behaviour in Figure12does not start at zero current.

Having the automatic calibration performed and located on target hardware will introduce additional problems, mainly regarding the embedded system nature. The calibration method will need to exist on hardware alongside multiple software applications which all requires to collabo-rate and share computing resources. Another limitation is presented in the solutions that require designated hardware. Lastly the automatic calibration method will need to be effective. Auto-matic calibration should ideally perform both better and faster than existing manual methods for calibration in order to potentially replace current methods and increase their availability.

(19)

3.4 Research Questions

The thesis aims to answer the following research questions:

RQ1: What methods for automatic calibration of hydraulics are available and applicable to con-struction equipment vehicles?

RQ2: How do methods for automatic hydraulic calibration perform at controlling hydraulics in con-struction equipment vehicles?

(20)

4 Related Work

The thesis highly values the research written in areas closely related to the control and identification of hydraulic systems. The following subsections will present the current techniques in use for the control of a hydraulic system, where Section 4.1 describes an ANN based solution and Section

4.2 describes PID based techniques. Both approaches represent a great interest for this thesis, since the proposed methods are inspired from the solutions derived from such papers. Section4.3

explores approaches based on the Recursive Least Square method for identification of hydraulic systems, while Section4.4 describe other work in the field of hydraulic control.

4.1 Nodal Link Perceptron Network

For similarity reasons, this thesis is inspired by the work realised by Opdenbosch and Sadegh in their investigation of automatic calibration for hydraulics [14–17]. Their research has been conducted on a different set of valves from the ones this thesis focus on. The authors base most of their research on the use of Electro-Hydraulic Poppet Valves (EHPV) [14–17], where this thesis instead focus on another common kind of valve, namely spool valves. The process of automatic calibration is referred to by Opdenbosch [14] as the ability to learn a systems inverse input-output. The way they achieved this was by the use of neural networks, namely the Nodal Link Perceptron Network (NLPN) developed by Nader Sadegh [18,19]. The NLPN, previously known as the functional basis perceptron network [18], was designed as a control system for non-linear systems in mind, and can work in both discrete and continuous time. It has a design that allows the system to learn the control input as a function of the desired states of the plant. This only requires the network to learn the input-state map of the plan, which is restricted to the desired states, and not the dynamics of the entire plant. This design generally allows a neural network with fewer interconnections and reduced training [18]. Compared to older neural network methods such as the work of Narendra and Parthasarathy [20] where the learning of the inverse plant dynamics is done through an indirect offline approach, the NLPN uses direct learning to achieve this.

The way the NLPN work is further described Section6, but can be seen as an adaptive lookup table [14,17], and has been used to identify the inverse input-output model by Opdenbosch & Sadegh in various work, see [16,17,21]. In the work published in 2008 [21], Opdenbosch et al. used a control technique called NLPN-based Input Matching (NBIM). An NLPN is trained to learn the plants inverse input-state map online with the use of gradient descent technique. This is achieved by learning flow conductance coefficient of EHPV in a closed-loop architecture. In later work, Opdenbosch et al. [16] use the NLPN with an accompanying control system to control a hydraulic actuator. The included INCOVA logic control system, developed by Pfaff and Tabor [22], was introduced with the aim of equitably share the hydraulic flow among the components in the system. The hydraulic actuator was controlled with the use independent metering EHPV configured in a Wheatstone bridge and according to their experiments, this solution was able to produce good results using active learning. In a later paper, Opdenbosch et al. [17] combined the NBIM control technique with an additional algorithm in order to extended the functionality of the control system to enable fault detection. The tracking performance of the control system was tested and validated experimentally while the fault detection was done in computer simulation.

4.2 Proportional Integral Derivative Control

Research on Proportional Integral Derivative (PID) techniques are necessary for the work in this thesis, where three of the four implemented methods are based upon the PID controller. The focus of this section is in the field of hydraulics which uses a PID control approach. Cobo et al. [2] investigated an approach using closed loop velocity control. Their approach use a standard Proportional Integral (PI) and a dynamic valve transformation algorithm to remove jerking actions in a wheel loaders racking motion. The system was modelled and further evaluated with the help of a real world implementation for testing. The authors found that their solutions provided a linearising effect, similar to pressure compensated load sensing hydraulic systems.

Aly [23] presented an optimisation method based on Genetic Algorithm (GA) for a PID con-troller to be applied in a position control of a non-linear electro-hydraulic servo-system. The author

(21)

investigates the optimisation process behind the search of improved PID parameters which was realised by the application of a Genetic Algorithm approach. The entire procedure is tested with the use of Simulink and the results of these tests demonstrate that the optimised PID is able to improve the performance of the hydraulic-servo system.

Liu and Daley [24] adopt an optimal tuning strategy to improve the performance of a non-linear PID. By the use of the inverse model of the control dead-zone and a non-non-linear compensator, they were able to remove the occurrence of unexpected overshoot, big steady-state error and long settling times for closed-loop hydraulic systems. This solution was able to successfully improve the hydraulic control system in their experiments.

Liu investigated a self tuning approach for improving PI controllers used to operate induction motors [25]. By using a gradient descent optimization approach to the parameter space, the PI controller’s constant were able to be tuned online. This approach allows the system to better adapt to the highly non-linear behaviour of such system. Several other automated tuning procedures have been investigated in the literature [26–31].

Zheng et al. [32] demonstrated the performances and higher control precision obtained in the position servo control, compared to a conventional PID controller. This online adaptive tuning PID controller is realised by the use of fuzzy logic and its ability to solve the contradiction between response, frequency and overshoot for improving the control accuracy is investigate in this research. Another research in the field of self-tuning PID has been presented by C¸ etin and Akkaya [33]. In this article the authors analysed a Hybrid Fuzzy PID Controller with Coupled Rules (HFPIDCR) for position control of the hydraulic system. The evaluation, performed by simulations, is compared with the performance of a classic PID and a Fuzzy Logic Controller (FLC). The results of this study demonstrate that HFPIDCR is more effective than the other controllers compared, where its rise and settling time are shorter.

4.3 Recursive Least Square

The research on Recursive Least Square (RLS) is related to the work in this thesis by the fact that two implemented methods rely on this technique for training the networks. One other method also makes use of the population based version of this method to estimate the initial inverse curve. Ghazali et al. [34] investigated offline and online identification of a hydraulic system by using AutoRegressive with eXogenous input (ARX) models. The authors discovered that for offline identification, modelling the system with a third order model is optimal, due to additional increments of order providing no significant improvement. For online learning, the recursive least squares method was used, and the conclusion was that both online and offline methods were able to successfully model the system.

The work was later expanded upon by Ghazali et al. [35], where the authors adopt an RLS method in order to achieve recursive identification of an electro-hydraulic system represented by a discrete-time model in open-loop and closed-loop configurations. This techniques is used to estimate the unknown parameters of the system based on auto regression.

A study on self tuning control for Electro-Hydraulic System was also conducted by Ghazali et al. [36]. In this article a Recursive Least Square (RLS) method with covariance resetting technique is proposed to estimate parameters of the discrete-time model. The proposed identification method and self tuning controller are able to handle range of system dynamics without knowledge of the actual system. This could represent a great advantage in the development of systems models by reducing the engineering effort. The results denote that the proposed technique is capable of adapting to changes occurring in the model parameters. A comparison with a fixed controller has been adopted for achieving this.

In the same field, similar research as been conducted by Plummer and Vaughan [37], where the application of an indirect (self-tuning) adaptive controller to an electro-hydraulic positioning system is described, as well as RLS to estimate plant parameters. Results show its ability of achieving a rapid adaptation to changes in the plant characteristics, providing good responds under very significant and instantaneous changes. According to the authors, the comparison of using an equivalent fixed coefficient controller was unable at achieving the same results in the same situation. The performance of the described technique is achieved despite the presence of many non-linear characteristics.

(22)

Kadissi et al. [38] investigate a control strategy based on a robust non-linear strategy called backstepping for identification and real time control of an electrohydraulic servo system. Electro-hydraulic systems are known to be highly non-linear and for this reason the authors based their identification on the recursive least square technique. The experimental results are evaluated by comparison with a real-time classic PID controller.

4.4 Hydraulic Control

Another approach presented to modelling a hydraulic system was performed by Ling et al. [39]. The authors make use of an Adaptive Neuro-Fuzzy Inference System (ANFIS) in order to model the behaviour of an electro-hydraulic actuator. They evaluated the model by measuring the Root Mean Squared Error of the model compared to the system and managed to produced a best fit score of 98.86%. However worth noting is that the authors mention that the added complexity of the model produced by this approach makes it unfavourable to the simpler approach which produced a score of 97.82%.

Ghazali et al. [40] investigated ways to improve the quality of hydraulic models that have otherwise been simplified with constants. This was realised by investigating the impact of the hydraulic system receives by a change of system pressure and load effect. The authors found that the system pressure is able to provide a significant impact on the behaviour of the system. However, the authors were unable to successfully test the impact of the load on the system due to limitations in their workbench. This is important to the work presented in this thesis, since the same variables investigated are extremely influential in the hydraulics found in construction equipment machines.

Other ways of calibration, that do not fully rely on NLPN and RLS were also investigated, such as the work by Stephenson [41]. To diminish dead pan and improve the control of the electro hydraulic actuator, Stephenson developed a method for calibrating the control of the hydraulic valves [41]. By increasing current to the valve while measuring pressure of either the valve’s inlet or outlet, the method manages to find the valve’s “start current”. This allows the system to ignore the valves dead zone and thus removing the otherwise imposed dead pan in the controller.

(23)

5 Methodology

One of the most important qualities of science, according to Dodig-Crnkovic [42], is the continuous re-examination and self-correction process it is subject to. This highlights the recursive nature of the logic of science and the scientific method, which is based on a set of rules defining the way to pose questions and formulating successful hypotheses. This entire procedure is achieved through logical reasoning, observations and experiments, creating a very impartial process based on replicability, where everyone is able to replicate a test under the same conditions and determine if certain results are valid or not.

Considering that computer science has roots in multiple scientific fields, allows this process’ applicability. The mathematical roots gives it a more logical imprint, while other roots could give it a more empirical one. Embedded systems is one branch of computer science where this dualism between logical and empirical approach is applied.

For these reasons, this thesis will be based on an iterative process which will perform according to the schema depicted in Figure13, where the first step is represented by a study of the current and existing system solutions. This step intends to provide an understanding of the calibration process and allows for the extraction of metrics and methods for evaluation. The following four steps are part of an iterative process, where solutions will be constantly explored, discovered and evaluated. Exploration of existing methods, models and tools on the market is conducted to get an extensive view of possible problem solutions. Solutions that are unsuited for the task will be excluded and the exploration step will repeat.

A selection is done based on the results reported on the researched literature analysed during the exploration phase. If a method seem promising the selected method is implemented as a Matlab Simulink model in order to be evaluated. The evaluation step will give a direction to follow according to its results, where the implemented method could be discarded and the procedure will fall back into the “Select methods for further testing” step. On the other hand if it turn out to be suitable for a possible implementation in a real machine, it will move forward in progress. Before reaching the final comparison, all the implemented methods suited for a real machine application must be evaluated in order to provide metrics for the final step.

The objective of this iterative process is the ability of changing direction during each step of the research. Basing the progression on the results from the previous steps, allowing a more flexible approach capable of giving a good direction from the early stages. Another important benefit of this continuous approach is represented by the ease of fixing possible unsuitable approaches.

Study existing calibration solution

Explore existing methods, models

and tools on the market

Select methods for further testing

Implementation of methods for

simulation

Implementation of methods for real

hardware (Proof of concept)

Evaluate simulation

Comparison of methods

Evaluate real world implementation

Comparison of simulated methods

(24)

The methods will be evaluated in simulation and real world tests performed on real machines. The methods will be evaluated on their ability to confirm to a predetermined set of actions with regards to the error between the desired and actual response. The methods are then to be compared to each other as well as a baseline which will be the existing method explained in Section3. An identical evaluation will be performed for the real machine testing, where a similar control signal will be used instead of lever movement in order to provide more accurate results.

The method will be evaluated looking at the absolute average error they produce for the pre-determined control sequence and by studying the machines behaviour.

5.1 Data Anonymisation

This thesis has been conducted with the collaboration of Volvo CE. For this reason the thesis makes use of proprietary non-public information. In order to protect this confidential information, an agreement was made to keep parts of the data anonymised.

(25)

6 Nodal Link Perceptron Networks

This section aim to describe and give a general idea of the method called Nodal Link Perceptron Networks (NLPN) constructed by Nader Sadegh [43]. This article presents an approach for tracking behaviour of a general class of non-linear systems using a Perceptron Neural Network (PNN). His work describes the functioning and properties of the NLPN functional approximator. It is essential to describe this approach since the main method of this thesis is NLPN based and can be used to approximate non-linear mappings according to Sadegh [18].

6.1 General Structure

An NLPN structure, depicted in Figure 14, is similar to a generic ANN previously described in Section2.4. This particular type of networks is well suited for approximating functions of the form f : Rn → Rm_{. The architecture is very simple with a single hidden layer. One distinctive feature}

of this network is that each neuron of the hidden layer has a particular activation function. All these functions together form a set of basis functions which can be chosen arbitrarily.

As seen in Figure 14, the input is a vector in Rn_{. Each components is then passed to the}

activation function which outputs a single scalar. Each element of the output is computed as a linear combination of these values. To give a more mathematical explanation, we have

W =h [W]₁ [W]₂ . . . [W]_m i is our RN ×m_{matrix of weights and}

Φ(x) = [ φ1(x) φ2(x) . . . φN(x) ] T

is the vector with our set of basis functions. We can express the function approximated by our NLPN as fW(x) = WTΦ(x) = N X i=1 WT iφi(x)

where x ∈ Rn _{is our input vector and f}

W(x) ∈ Rm is the output. Ф1 Ф2 ФN X2 Xn X f1(x) fm(x) [W1] [Wm] X1

Figure 14: NLPN structure based on the image by Opdenbosch [14].

For more details regarding the described theory consult Opdenbosch dissertation [14].

6.2 Simulink Structure

The Simulink implementation for the NLPN was provided by Patrick Opdenbosch as examples of 1 to 1, 2 to 1, and 4 to 2 dimensional networks. The network consisting of a 1-dimensional input

(26)

to a 1-dimensional output is illustrated in Figure15. This implementation allows the user to select three different types of basis functions used for the approximation, a Triangular, Gaussian and a Hyperbolic function. phi(k) e(k) W(k) RLS x1 sigma W f (x) PHI_V NLPN 1D to 1D uT Desired Approximation Constant

Figure 15: Simulink representation.

The NLPN block of this design takes three inputs: an input vector, a constant sigma and a matrix of weights. Sigma is used only in the case the Gaussian or the Hyperbolic basis functions are chosen. This parameter is not applied in the case a Triangular basis function is used. The output of this NLPN block consists of a scalar, representing the approximation of the desired function and a vector of basis function, which are then used to compute the weights. Weights are computed with a Recursive Least Square (RLS) block according to the error, computed as the difference between the desired output and the approximation generated by the output from the NLPN. Once adjusted, these weights are then passed back to the NLPN to improve the approximation. This follows the iterative process that characterise a NN, where the predicted output is continuously adjusted according to the learning process, represented by the RLS in this specific design.

6.3 Motivation

This particular method has been chosen for its flexibility. In fact the main advantage of using basis functions, which are the main component of an NLPN, is the ability to arbitrarily select a precision for the approximation. In addition to this, a good selection of basis functions are capable of approximating even non-linear functions, which is a advantage for the scope of this thesis. In addition to the aforementioned advantage, this design is simple to be extended to higher dimensions, where multiple factors, such as pressure and load of the system, can be taken in consideration. The NLPN is not consider a deep network, and for this reason the computation required for the training is less than other NN based approaches.

(27)

7 Simulations and Models

The solution implementations were realised in the Matlab Simulink environment which is a common tool used for simulation of systems [44_{]. Additionally, the Matlab Simulink environment}

allows for code generation and is currently used at Volvo CE for the implementation of their software along with its ability to communicate with other third party tools such as the rapid prototyping tool called Speedgoat1.

Simulink is a graphical block diagram based modelling technique where the user constructs their applications by the interconnection of various blocks, allowing for a simulation model made up from linked blocks, thus its name [44_{]. The Simulink environment is an addition to the original}

Matlab program, which allows for easy communication between the two environments.

A PID based solution was initial considered due to its simple design and record of being used in practise [10]. The initial PID strategy was however dropped after results from initial testing aligned to the theory of the PID controller being unsuitable for such a non-linear system. The early testing was done with the PID where the input was the error calculated from the reference speed taken from the Speed Control block and the plant. The non-linearity present in the hydraulic systems, especially the large deadpan in the valves start current, presented itself quite problematic to handle for a PID, causing highly oscillating and unstable results. The PID being an approach mainly used for linear problems, was as expected to be unable to alone handle such systems. This method was then evolved into the solution explained in Section7.4.

7.1 Speed Control Block

The speed control is an important component within the control strategy and is reused within almost all implementations presented. Its main purpose is converting the percentage received from the lever signal into a speed. It consists of a different construction depending on the action of the system it will control. The lift action Speed Control block has two inputs, Max Speed and Command. Since the machine maximum speed can vary depending on the hardware characteristics, it is important for a control strategy that is able to satisfy both slower machines incapability to meet the target speed, as well as not restricting faster machines from performing at optimal performance. The maximum speed of the machine is found by the Max Speed block seen in Figure 17. An initial maximum speed within a general range of allowed maximum speeds for the existing machine is set inside the Max Speed block. The approach chosen is then to periodically lower the maximum speed slightly within the block. This design depends on the fact that oscillations in the system are present to further increase the maximum speed available to the plant. The block updates the value by using the maximum of the previously set maximum speed and the result of a moving average block. This is calculated as an average of the last predetermined set of samples.

The control strategy for the lowering action differs slightly, whereas the Speed Control block does not take in consideration the maximum speed of the machine. Since lowering at the maximum operating speed can become uncontrollable due to the effects of the GET load and the effect of gravity, a predetermined speed limit for operating the lowering action is set. This removes the need for the Max Speed block and simplifies the design.

7.2 Simulated Plant

The plant model used in the simulation is constructed based on partial real data, expertise knowl-edge and heuristics. The data provided is similar to the data found in Figure12, where it contains a specified speeds with regards to a current. The data was however not complete, as there were areas in the input space that were missing. The reason behind this lack of data derives from the fact that retrieving it for the entire input space is time consuming for technicians and experts. For this reason it was decided that a fitted curve through this space would provide a representation within sufficient accuracy. Any inaccuracy could be argued to be a result of the individual machine behaviour. The regression analysis method used to fit a function through the data set representing the current to speed curve was the “Piecewise Cubic Hermite Interpolating Polynomial” or PCHIP. The curve was interpolated by the use of the pchip(x,y) function available in Matlab [45]. The

Methods for Automatic Hydraulics Calibration in Construction Equipment

School of Innovation Design and Engineering

V¨

aster˚

as, Sweden

Thesis for the Degree of Master of Science (120 credits) in

Computer Science with Specialization in Embedded Systems

30.0 credits

Methods for Automatic Hydraulics

Calibration in Construction Equipment

Peter Charbachi

pci13001@student.mdh.se

Filippo Ferrario

ffo16001@student.mdh.se

Examiner: Daniel Sundmark

M¨

alardalen University, V¨

aster˚

as, Sweden

Supervisors: Alessandro Papadopoulos & Saad Mubeen

M¨

alardalen University, V¨

aster˚

as, Sweden

Company Supervisor: Lars-Erik Larsson

Volvo Construction Equipment,

Eskilstuna, Sweden

Table of Contents

List of Figures

1

Introduction

1.1

Thesis Goals

1.2

Thesis Outline

2

Background

2.1

Wheel Loader Machines

2.2

Hydraulics

2.3

Calibration

2.4

Artificial Neural Network

∑

2.5

Proportional Integral Derivative

∑

∑

System

­1

r

e

Controller

u

y

e(τ)dτ

K

∫

K

de

dt

e

K

2.6

Regression Analysis

3

Problem Formulation

3.1

Calibration Principles

3.2

Existing Method of Calibration

3.3

Problem Statement

3.4

Research Questions

4

Related Work

4.1

1

_∫

_dt