No description available.

In thiswork, we develop a learning model predictive controller (LMPC) for energy-optimaltracking of periodic trajectories for a quadcopter. The main advantage of this controller isthat it is “reference-free”. Moreover, the controller is able to improve its performance overiterations by incorporating learning from the previous iterations. The proposed learningmodel predictive controller aims to learn the “best” energy-optimal trajectory over timeby learning a terminal constraint set and a terminal cost from the history data of previousiterations. We have shown howto recursively construct terminal constraint set and terminalcost as a convex hull and a convex piece-wise linear approximation of state and inputtrajectories of previous iterations, respectively. These steps allow us to formulate theonline planning problem for the controller as a convex optimization program, therebyavoiding the complex combinatorial optimization problems that alternative formulationsin the literature need to solve. The data-driven terminal constraint set and terminal costnot only ensure recursive feasibility and stability of LMPC but also guarantee convergenceto the neighbourhood of the optimal performance at steady state. Our LMPC formulationincludes linear time-varying system dynamics which is also learnt from stored state andinput trajectories of previous iterations.To show the performance of LMPC, a quadcopter trajectory learning problem in thevertical plane is simulated in MATLAB/SIMULINK. This particular trajectory learningproblem involves non-convex state constraints, which makes the resulting optimal controlproblem difficult to solve. A tangent cut method is implemented to approximate the nonconvexconstraints by convex ones, which allows the optimal control problem to be solvedby efficient convex optimization solvers. Simulation results illustrate the effectiveness ofthe proposed control strategy. / Vi utvecklar en lärande modell-prediktiv regulator för energi-optimalt följande av periodiskatrajektorier för en quadkopter. Den huvudsakliga fördelen med denna regulator äratt den är “referensfri”. Dessutom så klarar regulatorn att förbättra sin prestanda medtiden genom att inkorporera inlärning från föregående iterationer. Syftet med den föreslagnalärande modell-prediktiva regulatorn är att över en viss tid lära sig den “bästa”energioptimala trajektorian genom att lära sig den terminala bivillkorsmängden och denterminala kostnaden från historiskt data från tidigare iterationer. Vi har visat hur man kanrekursivt konstruera terminala bivillkorsmängder och terminala kostnader som konvexahöljen respektive konvexa styckvis linjära approximationer av tillstånds- och insignalstrajektoriernafrån tidigare iterationer. Dessa steg gör det möjligt att formulera onlineplaneringsproblemet för regulatorn som ett konvext optimeringsproblem och på så visundvika de komplexa kombinatoriska optimeringsproblemen som ofta krävs för alternativametoder som kan hittas andra publikationer. Den datadrivna terminala bivillkorsmängdenoch terminala kostnaden garanterar inte bara rekursiv tillåtenhet och stabilitet av LMPC,utan även konvergens till en omgivning av den optimala prestandan efter att ha uppnåttjämvikt. Vår LMPC-formulering innehåller linjär och tidsvarierande systemdynamik, somockså lärs från lagrade tillstånds- och insignalstrajektorier från tidigare iterationer.För att visa prestandan av LMPC så simuleras iMATLAB/SIMULINK ett problem ominlärning av quadkopter-trajektorier i det vertikala planet. Just det trajektorieinlärningsproblemetinnehåller icke-konvexa tillståndsbivillkor, vilket gör det resulterande optimeringsproblemetsvårt att lösa. En tangentsnitt-metod är implementerad för att approximera deicke-konvexa bivillkoren med hjälp av konvexa bivillkor, vilket möjliggör lösningen avdet optimala regleringsproblemet med effektiva lösare för konvexa optimeringsproblem.Simuleringsresultaten visar effektivitet av den föreslagna regleringsmetoden.

[pt] Nesta dissertação foram desenvolvidas técnicas de controle por aprendizado acelerado e Neuro-Fuzzy, aplicadas em um sistema servo-hidráulico para ensaio de fadiga. Este sistema tem o propósito de fazer ensaios em materiais para prever a resistência à fadiga dos materiais. O trabalho envolveu quatro etapas principais: levantamento bibliográfico, desenvolvimento de um controle por aprendizado acelerado, desenvolvimento de um controle por aprendizado Neuro-Fuzzy, e implementação experimental dos modelos de controle por aprendizado proposto em uma máquina de ensaios de materiais. A implementação do controle por aprendizado acelerado foi feita a partir do modelo de controle desenvolvido por Alva, com o objetivo de acelerar o processo de aprendizagem. Esta metodologia consiste em fazer um controle do tipo bang-bang, restringindo a servo-válvula a trabalhar sempre em seus limites extremos de operação, i.e., procurando mantê-la sempre completamente aberta em uma ou outra direção. Para manter a servo-válvula trabalhando em seus limites de seu funcionamento, os instantes ótimos para as reversões são obtidos pelo algoritmo de aprendizado, e armazenados em tabelas específicas para cada tipo de carregamento. Estes pontos de reversão dependem de diversos fatores, como a amplitude e carga média da solicitação, e são influenciados pela dinâmica do sistema. Na metodologia proposta, a lei de aprendizado inclui um termo de momentum que permite acelerar a aprendizagem dos valores das tabelas constantemente durante a execução dos testes, melhorando a resposta a cada evento. O desenvolvimento de um controle por aprendizado Neuro-Fuzzy foi motivado pela necessidade de ter um agente com a capacidade de aprendizado e armazenamento dos pontos ótimos de reversão. Este modelo de controle também consiste na implementação de um controle do tipo bang-bang, trabalhando com a servo-válvula sempre nos seus limites extremos de operação. O instante de reversão é determinado pelo sistema Neuro-Fuzzy, o qual tem como entradas a gama (dobro da amplitude) e o valor mínimo do carregamento solicitado. O processo de aprendizado é feito pelas atualizações dos pesos do sistema Neuro-Fuzzy, baseado nos erros obtidos durante a execução dos testes, melhorando a resposta do sistema a cada evento. A validação experimental dos modelos propostos é feita em uma máquina servohidráulica de ensaios de fadiga. Para este fim, o algoritmo de controle proposto foi implementado em tempo real em um módulo de controle CompactRIO da National Instruments. Os testes efetuados demonstraram a eficiência da metodologia proposta. / [en] In this thesis, accelerated learning and Neuro-Fuzzy control techniques were developed and applied to a servo-hydraulic system used in fatigue tests. This work involved four main stages: literature review, development of an accelerated learning control, development of a Neuro-Fuzzy control, and implementation of the learning control models into a fatigue testing machine. The accelerated learning control was implemented based on a learning control developed in previous works, introducing a faster learning law. Both learning control methodologies consist on implementing a bang-bang control, forcing the servovalve to always work in its operational limits. As the servo-valve works in its operational limits, the reversion points to achieve every peak or valley in the desired history are obtained by the learning algorithm, and stored in a specific table for each combination of minimum and mean load. The servo-valve reversion points depend on a few factors, such as alternate and mean loading components, while they are as well influenced by the system dynamics. In the proposed accelerated methodology, the learning law includes one momentum term that allows to speed up the learning process of the table cell values during the execution of the tests. The developed Neuro-Fuzzy control also consists on a bang-bang control, making the servo-valve work in its operational limits. However, here the instant of each reversion is determined by the Neuro-Fuzzy system, which has the load range and minimum load required as inputs. The learning process is made by the update of the Neuro-Fuzzy system weights, based on the errors obtained during the execution of the test.The experimental validation of the proposed models was made using a servo-hydraulic testing machine. The control algorithm was implemented in real time in a C-RIO computational system. The tests demonstrated the efficiency of the proposed methodology.

Ce travail vise à faire une pré-évaluation des paramètres de contrôle en vue de réduire la traînée sur véhicule réel. Deux mécanismes d’actionnement différents (Murata micro-blower et couteau d’air) ont été caractérisés et comparés en vue de déterminer leurs qualités ainsi que leurs limites. Les micro-blowers ont pour but d’exciter la couche limite en vue de perturber directement les structures tourbillonnaires formées dans la couche de cisaillement. Le couteau d’air étudié, à surface arrondie, pourrait être considéré comme un dispositif actif de réduction de la traînée à effet Coanda équivalent au dispositif passif de type boat-tail. Différentes stratégies de contrôle en boucles ouverte et fermée sont examinées, telles que le soufflage continu, le forçage périodique, le contrôle du mode glissant (SMC) et le contrôle par apprentissage (MLC). La SMC est un algorithme robuste en boucle fermée permettant de suivre, d’atteindre et de maintenir une consigne prédéfinie; cette approche présente l’intérêt d’avoir une capacité d’adaptation prenant en compte les perturbations extérieures inconnues. Le contrôle par apprentissage est un contrôle sans modèle qui permet de définir des lois de contrôle efficaces qualifiées et optimisées via une fonction coût/objectif spécifique au problème donné. Une solution hybride entre MLC et SMC peut également fournir un contrôle adaptatif exploitant les mécanismes d’actionnement non linéaires les plus adaptés au problème. L’ensemble de ces techniques de contrôle ont été testées sur diverses applications expérimentales allant d’une simple configuration académique de marche descendante jusqu’à des géométries présentant une structure d’écoulement représentatives de véhicules réels. Pour la configuration de marche descendante, l’objectif était de réduire expérimentalement la zone de recirculation via une rangée de micro-jets et de l’estimer par des capteurs de pression. Les contrôles d’écoulement ont été réalisés par forçage périodique ainsi que par MLC. On démontre dans ce cas que la MLC peut surpasser le contrôle par forçage périodique. Pour la configuration sur corps épais (corps d’Ahmed), l’objectif était de réduire et/ou de maintenir la traînée aérodynamique via un couteau d’air placé sur la partie supérieure du hayon arrière et évalué par le biais d’une balance aérodynamique. Le soufflage continu et le forçage périodique ont été utilisés dans ce cas comme stratégies de contrôle en boucle ouverte permettant ainsi de faire une comparaison avec les algorithmes SMC et MLC. La pré-évaluation des paramètres de contrôle a permis d’obtenir des informations importantes en vue d’une réduction de la traînée sur un véhicule réel. Dans ce cadre, les premiers essais de caractérisation sur véhicules réels ont été réalisés sur piste et un dispositif d’actionnement ainsi qu’un protocole expérimental sont également présentés en perspective à ce travail. / The present work aims to pre-evaluate flow control parameters to reduce the drag in a real vehicle. Two different actuation mechanisms (Murata’s micro-blower, and air-knives) are characterized and compared to define their advantages and limitations. Murata micro-blowers energized the boundary layer to directly perturb the vortex structures formed in the shear layer region. The air-knife has a rounded surface, adjacent to the slit exit, that could be considered as an active boat-tail (Coanda effect) for drag reduction. Different open-loop and closed-loop control strategies are examined, such as continuous blowing, periodic forcing, sliding mode control (SMC) and machine learning control (MLC). SMC is a robust closed-loop algorithm to track, reach and maintain a predefined set-point; this approach has on-line adaptivity in changing conditions. Machine learning control is a model-free control that learns an effective control law that is judged and optimized with respect to a problem-specific cost/objective function. A hybrid between MLC and SMC may provide adaptive control exploiting the best non-linear actuation mechanisms. Finally, all these parameters are brought together and tested in real experimental applications representative of the mean wake and shear-layer structures related to control of real cars. For the backward-facing step, the goal is to experimentally reduce the recirculation zone. The flow is manipulated by a row of micro-blowers and sensed by pressure sensors. Initial measurements were carried out varying the periodic forcing. MLC is used to improve performance optimizing a control law with respect to a cost function. MLC is shown to outperform periodic forcing. For the Ahmed body, the goal is to reduce the aerodynamic drag of the square-back Ahmed body. The flow is manipulated by an air-knife placed on the top trailing edge and sensed by a force balance. Continuous blowing and periodic forcing are used as open-loop strategies. SMC and MLC algorithms are applied and compared to the open-loop cases. The pre-evaluation of the flow control parameters yielded important information to reduce the drag of a car. The first real vehicle experiments were performed on a race track. The first actuator device concept and sensor mechanism are presented.

The demand for automated decision making, learning and inference in uncertain, risk sensitive and dynamically changing situations presents a challenge: to design computational approaches that promise to be widely deployable and flexible to adapt on the one hand, while offering reliable guarantees on safety on the other. The tension between these desiderata has created a gap that, in spite of intensive research and contributions made from a wide range of communities, remains to be filled. This represents an intriguing challenge that provided motivation for much of the work presented in this thesis. With these desiderata in mind, this thesis makes a number of contributions towards the development of algorithms for automated decision-making and inference under uncertainty. To facilitate inference over unobserved effects of actions, we develop machine learning approaches that are suitable for the construction of models over dynamical laws that provide uncertainty bounds around their predictions. As an example application for conservative decision-making, we apply our learning and inference methods to control in uncertain dynamical systems. Owing to the uncertainty bounds, we can derive performance guarantees of the resulting learning-based controllers. Furthermore, our simulations demonstrate that the resulting decision-making algorithms are effective in learning and controlling under uncertain dynamics and can outperform alternative methods. Another set of contributions is made in multi-agent decision-making which we cast in the general framework of optimisation with interaction constraints. The constraints necessitate coordination, for which we develop several methods. As a particularly challenging application domain, our exposition focusses on collision avoidance. Here we consider coordination both in discrete-time and continuous-time dynamical systems. In the continuous-time case, inference is required to ensure that decisions are made that avoid collisions with adjustably high certainty even when computation is inevitably finite. In both discrete-time and finite-time settings, we introduce conservative decision-making. That is, even with finite computation, a coordination outcome is guaranteed to satisfy collision-avoidance constraints with adjustably high confidence relative to the current uncertain model. Our methods are illustrated in simulations in the context of collision avoidance in graphs, multi-commodity flow problems, distributed stochastic model-predictive control, as well as in collision-prediction and avoidance in stochastic differential systems. Finally, we provide an example of how to combine some of our different methods into a multi-agent predictive controller that coordinates learning agents with uncertain beliefs over their dynamics. Utilising the guarantees established for our learning algorithms, the resulting mechanism can provide collision avoidance guarantees relative to the a posteriori epistemic beliefs over the agents' dynamics.

Verdrängergesteuterte hydrostatische Antriebssysteme mit drehzahlvariablem Pumpenantrieb zeichnen sich durch ihre im Vergleich mit ventilgesteuerten Antrieben gute Energieeffizienz und die Möglichkeit der einfachen Stillsetzung aus, weisen jedoch durch die mangelnde Einspannung des Aktors geringere Eigenfrequenzen auf, was die Einstellung von Standardregelkreisen erschwert und meist eine geringere Dynamik und Positioniergenauigkeit zur Folge hat. Hydraulische Achsantriebe, die ohnehin über zahlreiche dominante Nichtlinearitäten verfügen und schwach gedämpft sind können in der Folge das ihnen innewohnende Potential nicht ausschöpfen. Mit einem Vergleich von Dynamik und Präzision verschiedener Antriebssysteme an Hauptantriebsachsen von Kunststoff-Spritzgießmaschinen mittlerer Baugröße wird zunächst das Leistungspotential analysiert. Auf dieser Basis werden Methoden zur Verbesserung der statischen und dynamischen Eigenschaften drehzahlvariabler verdrängergesteuerter Antriebe in Positions- und Druckregelung entwickelt, welche sich durch eine einfache Parametrierung und hohe Robustheit auszeichnen, da sie ohne einen geschlossenen Regelkreis funktionieren. Die dynamische inversionsbasierte Vorsteuerung ermöglicht dabei ein initial gutes Folgeverhalten, das durch die Anwendung einer iterativ lernenden Regelung in jedem Zyklus weiter verbessert wird. Um die Dynamik von Folgeregelungen mit weiteren Randbedingungen zu maximieren wird eine Methode entwickelt, mit der es möglich ist, eine Bewegungsvorgabe entlang der physikalischen Leistungsgrenzen des Antriebssystems zu berechnen und die wirkenden Begrenzungen aufzuzeigen. Die Erstellung von Bewegungsvorgaben sowie die Einstellung der lernenden Regelung sind dabei jeweils mit einem einzigen Parameter möglich. Die experimentelle Untersuchung und der Funktionsnachweis der entwickelten Methoden am Beispiel der Kunststoff-Spritzgießmaschine zeigt eine deutliche Steigerung der möglichen Dynamik verdrängergesteuerter Antriebssysteme, ein gutes Folgeverhalten sowie eine erhöhte Positioniergenauigkeit bei gleichzeitiger Unabhängigkeit von der Betriebstemperatur.:1. Einleitung und wissenschaftliche Problemstellung 7 2. Zielsetzung der Arbeit 11 3. Stand der Forschung und Technik 13 3.1 Architekturen hydraulischer Linearantriebe 13 3.2 Betriebsverhalten drehzahlvariabler verdrängergesteuerter Antriebe 16 3.3 Regelung hydraulischer Achsantriebe in Verdrängersteuerung 18 3.4 Vorsteuerungen und iterativ lernende Regelungen 21 4. Antriebstechnik in Kunststoff-Spritzgießmaschinen 25 5. Analyse der Leistungsfähigkeit von Antriebssystemen in SGM 27 5.1 Aufbau und Funktionsweise von Spritzgießmaschinen 28 5.2 Auswahl der Antriebssysteme 31 5.3 Analyse der Bewegungsdynamik 34 5.4 Analyse der Positioniergenauigkeit 37 5.5 Analyse der Druckregelung 39 5.6 Identifikation von Potentialen für die Leistungssteigerung 44 6. Trajektoriengenerierung entlang der Systemleistungsgrenzen 47 6.1 Kniehebel-Schließeinheit 48 6.2 Analyse statischer und dynamischer Restriktionen 50 6.3 Trajektorienentwurfsmethodik 59 7. Modellbasierte dynamische Vorsteuerung 67 7.1 Methodik 68 7.2 Mathematisch-physikalische Beschreibung 69 7.3 Inversionsbasiertes Steuergesetz 75 8. Modellbasierte lernende Vorsteuerung 77 8.1 Methodik 78 8.2 Entwurf modellbasierter normoptimaler iterativ lernender Regelungen 79 8.3 Stabilitätsnachweis 84 8.4 Generierung von Lernmodellen 86 9. Anwendung der Verfahren und Diskussion 91 9.1 Positionsregelung im geschlossenen hydrostatischen Kreis 93 9.2 Lastkraftregelung im offenen Kreis 108 9.3 Ablösende Regelung: Geschwindigkeit - Last 123 10. Zusammenfassung 127 11. Literatur 133 12. Anhang 145 / Displacement-controlled hydrostatic drive systems with variable-speed pump are characterized by their good energy efficiency and the possibility of simple shutdown compared with valve-controlled drives, but they have lower natural frequencies, which makes the application of standard closed-loop control more difficult and usually results in lower dynamics and positioning accuracy. As a result hydraulic drives, which already have numerous dominant nonlinearities and are weakly damped, cannot exploit their full potential. The work starts with a comparison and an analysis of the dynamics and precision of different drive systems on main drive axes of medium-size plastic injection molding machines. On this basis, methods are developed for improving the static and dynamic properties of variable-speed displacement-controlled drives in position and pressure control. These methods are characterized by simple parameterization and high robustness without relying on a closed-loop control. In this context, the dynamic inversion-based feedforward control allows for a good tracking performance, which is further improved by applying a cycle-wise iterative learning control. In order to fulfill the dynamics of follow-up control with position boundary conditions, a method is developed which allows for calculating a motion specification along the physical performance limits of the drive system and to show the existing limitations. The creation of motion presets as well as the setting-up of a learning controller may be done with one single parameter. Experimental investigation of the developed methods using the example of the plastic injection molding machine shows a significant increase in dynamics of displacement-controlled drive systems, good follow-up behavior, and increased positioning accuracy while remaining independent of the operating temperature.:1. Einleitung und wissenschaftliche Problemstellung 7 2. Zielsetzung der Arbeit 11 3. Stand der Forschung und Technik 13 3.1 Architekturen hydraulischer Linearantriebe 13 3.2 Betriebsverhalten drehzahlvariabler verdrängergesteuerter Antriebe 16 3.3 Regelung hydraulischer Achsantriebe in Verdrängersteuerung 18 3.4 Vorsteuerungen und iterativ lernende Regelungen 21 4. Antriebstechnik in Kunststoff-Spritzgießmaschinen 25 5. Analyse der Leistungsfähigkeit von Antriebssystemen in SGM 27 5.1 Aufbau und Funktionsweise von Spritzgießmaschinen 28 5.2 Auswahl der Antriebssysteme 31 5.3 Analyse der Bewegungsdynamik 34 5.4 Analyse der Positioniergenauigkeit 37 5.5 Analyse der Druckregelung 39 5.6 Identifikation von Potentialen für die Leistungssteigerung 44 6. Trajektoriengenerierung entlang der Systemleistungsgrenzen 47 6.1 Kniehebel-Schließeinheit 48 6.2 Analyse statischer und dynamischer Restriktionen 50 6.3 Trajektorienentwurfsmethodik 59 7. Modellbasierte dynamische Vorsteuerung 67 7.1 Methodik 68 7.2 Mathematisch-physikalische Beschreibung 69 7.3 Inversionsbasiertes Steuergesetz 75 8. Modellbasierte lernende Vorsteuerung 77 8.1 Methodik 78 8.2 Entwurf modellbasierter normoptimaler iterativ lernender Regelungen 79 8.3 Stabilitätsnachweis 84 8.4 Generierung von Lernmodellen 86 9. Anwendung der Verfahren und Diskussion 91 9.1 Positionsregelung im geschlossenen hydrostatischen Kreis 93 9.2 Lastkraftregelung im offenen Kreis 108 9.3 Ablösende Regelung: Geschwindigkeit - Last 123 10. Zusammenfassung 127 11. Literatur 133 12. Anhang 145