A thrust allocation system is used to determine how the desired forces, computed by a high level control sytem, can be distributed among the thrusters. The main goal of the thrust allocation is to obtain the desired force, but other objectives can also be included. Such secondary goals can be to minimize fuel consumption, keep wear and tear of the thruster to a minimum and avoid overloading the power systems. The thrust allocation should also take forbidden sectors and actuator rate constraints into account. It is essential to safe operation that the allocation system provides a solution, and provides the solution in time. In this thesis MPC (Model predictive control) is suggested as a method to solve the control allocation problem for CyberRig I (a scaled model of a semi-submersible drilling unit). 3 MPC algorithms are simulated in matlab, and the most complete are chosen for on-line implementation. The algorithm is based on an extended thrust formulation, and allows for rotatable thrusters. The cost function penalizes change in thust magnitude and in the azimuth angle. Forbidden sector constraints and rate constrains, both for thrust magnitude and angle, are implemented. It is shown in simulations that the MPC algorithm performs well in comparison with an existing quasi-static method. Its main benefit over the quasi-static method is the ability to handle constraints. The cost of using MPC is increased computational efforts.

This report addresses the stabilization problem of a marine structure (i.e. cable/riser), connected to a surface vessel at one end and to a thruster unit at the other. Here, only motion in the lateral direction has been considered. Stabilization control laws are designed for position and velocity control of the robot system. The passivity of the control system is analyzed, and the closed loop system is shown to be asymptotically stable. Simulation results are presented.

This report investigates proportional-derivative (PD) controller and different versions of sliding mode controllers, including a 2-sliding mode controller invented by Arie Levant, applied to a pneumatic actuator on a truck clutch, The purpose of the clutch system is to develop a transmission system consisting of a normal clutch and transmission controlled automatically as an automatic gearbox, called automated manual transmission. The goal is to increase driver comfort and performance, as well as reduce fuel consumption. It is put an effort in implementing an accurate simulation model of the clutch system in Matlab Simulink. The model output includes clutch position, velocity and acceleration, actuator chamber pressure and temperature. The accuracy of the model developed is assumed to be accurate enough for control design. The only measurement available is position measurement, because more sensors increase cost. The measurement noise is low, which enables direct use of the position measurement for control. For the controllers and other parts of the control system that is dependent on velocity, acceleration or pressure estimates, the measurement has to be differentiated. Differentiation of noisy signals is problematic, therefore filters have to be used. In this report a first order low pass filter differentiator is compared to a robust differentiator, which is inspired by higher order sliding modes and developed by Arie Levant. The reason for comparing it with a very simple filter is the simplicity of the first order filter. It is easy to understand and tune. The performance of the robust differentiator is in this application not better than the first order filter. Therefore the first order filter is used in the tests. A simplified version of the simulation model is used in the design of the controllers. A PD controller with limited derivative action is tuned on the basis of a linearized version of the control model. PD controllers have turned out to perform well and is suitable for comparison, because they are independent of measurement filters, well known and have well established design methods. The PD controller is compared to different sliding mode controllers. The most promising sliding mode controller, which is a boundary layer controller with variable boundary layer width, is tested thoroughly on the simulation model. Different tests where the simulation model parameters are altered, are performed to investigate the robustness and performance properties of the controllers. The most promising sliding mode controller were supposed to be tested against the PD controller on a test truck at Kongsberg Automotive. New and faster valves were supposed to be tested in the truck. Unfortunately they did not arrived in time for the test and in addition, the driver circuit of the older and slower valves broke down under the test startup. Therefore the real tests could not be accomplished. A brief overview of the planned field tests is given. The PD controller and sliding mode controller are compared in a view focusing on robustness. It is found that the ideal sliding mode controller is highly robust, but not usable in practise for this application. The developed variable boundary layer sliding mode controller performs better than the PD controller on the simulation model. The reason may be that it is tailored to the reference trajectory used, as opposed to the PD controller which is tuned using Bode diagrams, gain and phase margins. Both controllers possess approximately the same robustness properties.

Dynamic positioning of surface vessels involves control of vessels with changing dynamics, shifting conditions, for different operational tasks. A controller with fixed controller parameters cannot have an optimal performance for all these different cases, and autotuning of the controller would be very valuable. However, dynamic positioning is a complex task, and thus automatic tuning of a dynamic positioning controller is not less so. This thesis does not solve all problems that comes with autotuning of dynamic positioning systems, but it gives an overview of the problem and presents a novel performance index for station keeping. Furthermore, a hybrid controller that can function as a first step in solving the autotuning problem is suggested. The hybrid controller has a fixed controller structure and is a combination of a gain-scheduling controller and an adaptive controller. The adaptive controller is used in an idle (training/learning) mode to populate a look-up table with controller parameters, while the gain-scheduling controller work as a fast-changing dynamical controller, using the controller parameters stored in the look-up table. Each controller parameter set in the look-up table is optimized according to a vessel operational condition, which is defined as a function of environmental conditions (wind, waves, ocean current), vessel draught, and water depth. Optimization of the controller parameters for the different vessel operational conditions is carried out by two different autotuning methods; a genetic algorithm and a rule-based algorithm. Both of these autotuning methods are optimizing the controller gains in a nonlinear PID-controller. The performance index and the two autotuning methods are implemented in Matlab/Simulink, where simulation tests are performed for a 3 DOF mathematical model of a supply vessel. The test scenario includes two different vessel operational conditions, where the controller has been automatically tuned both for minimal position and heading deviation as well as weighting on the use of forces. A comparison of the two autotuning methods is also performed and finally a discussion of the behaviour and tuning of the suggested performance index is carried out.

A two-fluid model is used to derive a set of boundary conditions. The conditions are produced numerically, and try to imitate the behavior of output injection by using a linearized version of the model. In order to ensure that the model is hyperbolic, virtual mass terms are included in the momentum equations. An observer is presented, using OLGA, a fluid simulator, as its model. The boundary conditions derived are employed in the observer, and its convergence properties are shown to improve. A set of adaption laws for estimating parameters in a two-phase leak model is derived. Estimation of the leakage mass fraction is sacrifced in order to increase performance and stability. A model, also based on OLGA, is used to simulate a leak, and the observer prove to give good estimates of leak parameters as long as estimates of leakage mass fraction is available. Mass flow fraction seem to be a sufficient estimate. A wide range of scenarios are simulated, inspecting the weaknesses of the observer.

This report presents two optimization problems, where three cost-functions are suggested for each. The goal for the first optimization problem is to find a time variant look-ahead distance which improves the performance of the vessel in terms of the cross-track error, relative to constant look-ahead distances. The second optimization problem is an extension of the first, where a time variant surge velocity reference is also calculated. This results in smaller cross-track errors, at the cost of increased calculation time. It is assumed that the path and the desired surge velocity on the path is supplied by some external source. Existing kappa-exponential control laws are used to track the references resulting from the calculated optimal look-ahead distance and surge velocity reference. The predictions needed to solve the optimization problem are made from a model where the control laws are inserted into the dynamics to simplify the model. The optimization problems are solved for two different approaches. The first approach uses an LTV model for predictions and a QP-solver to solve the optimization problem. The second approach forms predictions of the states by numerical integration of the system dynamics and uses an optimization problem solver for general non-linear functions to solve the optimization problem. The latter approach generally results in longer calculation times but better accuracy, while the first approach yields convexity of the optimization problem. A passive observer is used to estimate the current such that it can be included in the predictions. Four of the six suggested cost-functions lead to significantly improved performance of the vessel in terms of the cross-track error. This is true both in the case of no disturbances and in the presence of a constant irrotational current. That is, the position of the vessel converges faster to the path, the vessel stays closer to the path and has less over-shoot in the cross-track error, for the optimal inputs than for constant inputs. The cost-functions includes weights for tuning where the tuning process is easy for some of the cost-functions. One of the cost-functions where only the optimal look-ahead distance is calculated, results in calculation times shorter than the time between samples, after a few time steps. This is a promising result since the application of MPC to fast-dynamic systems such as marine vessels is desired, but often problematic due to too long calculation times for solving the optimization problem. It takes more time to solve the optimization problem when the optimal surge velocity reference is found in addition to the optimal look-ahead distance.

Considerable progress has been achieved in recent years with respect to autonomous vehicles. A good example is the DARPA Grand Challenge, a competition for autonomous ground vehicles. None of the competing vehicles managed to complete the challenge in 2004, but returning in 2005, a total of five vehicles were successful. Effective collision avoidance is a requirement for autonomous navigation, and even though much progress has been done, it still remains an open problem. The focus of this thesis is on the development of a collision avoidance system for unmanned surface vehicles (USVs), which is compliant with the International Regulations for Avoiding Collisions at Sea (COLREGS). The system is based on a modified version of the Dynamic Window algorithm, taking both acceleration and lateral speeds into account for reactive collision avoidance. Path planning is provided by the Rapidly-Exploring Random Tree (RRT) algorithm, extended to use the A* algorithm as a guide, which significantly increases its efficiency. Extensive simulations have been performed in order to determine the value of the modifications done to the original algorithms, as well as the performance of the total control system. Full-scale experiments have also been carried out in an attempt to verify the results from the simulations. The collision avoidance system performed very well during the simulations, finding near-optimal paths through the environment and evading other vessels in a COLREGS-compliant fashion. In the full-scale experiments, important sensor data was erroneous, resulting in reduced avoidance margins. However, the collision avoidance system still kept the controlled vessel safe, showing significant robustness.

The desired properties of a multi-loop PID tuning procedure is to find some parameters that makes the plant robust, meet some desired performance requirements and guarantee failure tolerance. A detailed literature survey of the different multi-loop PID tuning procedures are presented. Properties of the Independent design methods, Detuning methods, Sequential closing methods, Iterative or trial and error methods, Optimization methods and Relay feedback approaches are described in detail and discussed. Most of the tuning procedures result in a too conservative design without integrity. It is shown how the integrity property may be achieved with a multi-loop Hinf optimal tuning method. How to deffine and solve such a Hinf optimal problem is presented. The desired properties of the multi-loop PID tuning procedure is obtained with this method. The method aim for its object to solve a Hinf optimization problem with Linear Matrix Inequality (LMI) constraints. The optimization problem is non-convex, so a Successive Semidefnite Programming (SSP) algorithm is used to find local solutions to the problem. Several initial points must be examen to aim for a global solution. The SSP algorithm is implemented in MATLAB, and applied at a distillation column example. The implemented algorithm does not converge to a solution. Hence no simulation results that back up the theoretical work is presented.

Most of the systems considering wave synchronization in the literature are concerned with moon pool operations. The objective of this thesis is to transition the existing knowledge from some of that work into marine cranes working over the side of the vessel. This aim is mainly motivated by the fact that cranes working over the side of the vessel are more versatile. They can be used for more comprehensive and differently shaped objects. In addition, vessels with cranes mounted this way are more flexible since the moon pool configuration consumes a considerable amount of deck space. The main contribution of this thesis is a consideration of the transition from cranes lifting through a moon pool to cranes lifting over the side of the vessel. This includes consideration of the roll and pitch motion of the vessel, as well as simulating the complete system with wave synchronization control. The system is also simulated without any motion compensation for comparison. The degrees of freedom (DOFs) available for motion compensation control in different crane configurations are discussed, and some comments are made on how the actuators may cooperate in order to achieve better motion compensation. It is found that nothing obstructs the use of control algorithms originally developed for moon pool cranes for cranes working over the side of a vessel. The behaviour of the systems is roughly equal in the two cases, and the physical aspects are the same regarding wave elevation, still water level, and wire behaviour. There is however, one aspect to consider when designing overboard cranes, and that is the response time needed for following the wave elevation with the payload. When the payload is lifted over the side of the vessel, the roll and pitch motions affect the movement of the crane tip. In the worst case scenario, a wave can have a velocity directed upward, while the crane tip has a velocity directed downward. If this is the case, a faster motion compensation system is needed than for the moon pool applications.

This paper describes development of an optimal 3D path planner with collision avoidance for a 9 DOF robot manipulator. The application of the robot manipulator will be on an unmanned oil platform where it will be used for inspection. Most of the time the robot manipulator will follow a pre-programmed collision-free path specified by an operator. Situations where it is desirable to move the end effector from the current position to a new position without specifying the path in advance might occur. To make this possible a 3D path planner with collision avoidance is needed. The path planner presented in this paper is based on the well known Probabilistic Roadmap method (PRM). One of the main challenges using the PRM is to make a roadmap covering the entire collision free Configuration space, Cfree, and connect it into one connected component. It is shown by empirical testing that using a combination of the Bridge Sampling technique and a simple Random sampling technique gives best Coverage of the Cfree space and highest Connectivity in the roadmap for the given environment. An algorithm that increases the Connectivity and sometimes provide Maximal Connection is also described. A backup procedure that can be executed on-line if a query fails is also presented. The backup procedure is slow, but it increases the chances of succeeding a query if the goal is in a difficult area. It is also investigated if the coverage and connectivity can be further improved by using the potential field planner when connecting the waypoints. Empirical testing showed that the improvements of Coverage and Connectivity were limited and the sampling and query time increased. The query time for a roadmap containing 400 nodes and one containing 1000 nodes was compared. It turned out that a large roadmap did not necessarily affect the query time negative because it made it easier to connect the start and goal nodes. Three existing path smoothing algorithms and a new algorithm, called Deterministic Shortcut, were implemented and tested. Empirical testing showed that the Deterministic Shortcut algorithm outperformed the others when it came to path smoothing versus time.