Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Nowadays, global warming is a major scientific and social topic. One of the major contributors to global warming is the emission of methane. The Intergovernmental Panel on Climate Change estimates that 60% of all methane emissions are caused by human activity. Investments in pipelines or methane conversion plants are often not viable due the high capital costs. Hence, new technologies are needed to stimulate methane conversion for smaller sources. An interesting alternative is biological conversion to valueable chemicals with the aid of methanotrophic bacteria. It has the potential to become economically viable for small amounts of methane due to its low capital and operating costs. A promising candidate for industrial processes is Methylomicrobium buryatense 5GB1. It has a high growth rate, robust growth characteristics and can withstand a wide range of growth conditions. This thesis discusses the development of process control tools for such a methane bioconversion process. The main objective is to develop a controller which allows to regulate the biomass and lactate concentration of a methane bioconversion process using Methylomicrobium buryatense 5GB1. This can be divided into the following sub-objectives: (i) the implementation of a controllability study to examine which set of points can be reached (ii) the design and optimization of an integral feedback controller using the linear quadratic regulation problem and (iii) the design of a model predictive controller. A model of the methane bioconversion process is required in order to design process control tools. This is taken from the thesis ’Modelling and Observability Analysis of a Methane Bioconversion Process’ by Koen Michiels. The model is adapted accordingly and then linearized. A controllability study is performed for the linearized model, which concludes that the model is controllable. Next, an integral feedback controller is designed and optimized for the linear model and then applied to the non-linear model. The controller achieves reference tracking for a large range of setpoints and is able to deal with disturbances and measurement noise. Sensitivity analyses are performed to examine to influence of model uncertainties and a linear Kalman filter is implemented as well. Finally, a model predictive controller is designed for the linear model and then applied to the non-linear model. This controller achieves reference tracking faster than the integral feedback controller.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

The fermentation of sugars and starches to ethanol is an important business worldwide. Shrinking fossil fuel reserves and growing environmental awareness both push the world towards alternatives for crude oil, furthering the trend towards renewable ethanol. Due to the high production volumes, any improvement in the production process will have a big economic impact. This work handles the very high gravity wheat mash fermentation processes at the Cargill plants in Sas van Gent and Manchester. This process uses the well-known yeast Saccharomyces cerevisiae. The plant is not running optimally, but current monitoring with HPLC analysis is too expensive and not good enough to allow for insight in the process without extensive screening campaigns. Cell concentration and yeast health measurements are not present or unreliable. This work puts forward analysis of microscope images in brightfield as an alternative means of process monitoring. Using state-of-the-art image recognition technology, a better measure for estimating cell concentration and health was developed. Images were acquired with the Cellometer CBA microscope, segmented with U-Net and the segments were classified with a convolutional network. Thus, measures for cell concentration, yeast health and degree of clumping were constructed. Due to its high capacity for generalization, the model can easily be adapted for other processes and other cell morphologies.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Machine-learning to attain optimal control is becoming an attractive field within academia, necessitating a paralleled interest by the chemical processing industry to cement its importance within control theory. Given the growing interest in applying Gaussian processes (GPs) to describe the behavior of dynamical systems, it oers a plausible solution to the regression tasks required in control strategies. This attraction derives from its probabilistic, machine-learning feature that can be exploited for uncertainty-reduced control decisions. Further, chemical process industries are continually challenged in meeting production demands; while still adhering to the limitations imposed through safety, health and environmental regulations. The ability to perform at, or within set constraints, becomes imperative to reach both economic and customer satisfaction. Control theory and its derivatives are often seen as the bridging mechanism to achieve these requirements. Although PID controllers still perform the majority of these control activities; increasing complexity – catalyzed by the movement to ’intensify’ process operations– has necessitated advanced control procedures such as model predictive control (MPC). It is thus the purpose of this thesis to illustrate the feasibility of implementing a Gaussian process based model predictive controller (GPMPC) and assess the performance relative to current industrial controllers (including alternative MPC and PID schemes). Additionally, through the inherent inclusion of a quantified uncertainty with GP estimates; the over-/under-estimation for process uncertainties usually plaguing nonlinear and linear models, could be eliminated. The developed GPMPC algorithm was applied to an Arduino Temperature Control Lab; whereby through a series of experiments for dynamic setpoint tracking and disturbance rejection the controller performance could be assessed, according to the mean-squared error (MSE) between actual responses and defined setpoints. The investigation proved an appropriately trained GPMPC algorithm was superior to both an SSMPC and PID controller alternative, however, control action decisions were unsatisfactorily too erratic to allow implementation in an industrial setup where actuator wear becomes a consideration. An adjusted cost function including the estimation uncertainty, resultantly solved for control actions that were within known operating regions; preventing undesired controller aggression when faced with dynamic setpoint trajectories.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Nowadays, many modern industrial processes would not be operable without the use of automatic control which can be achieved by controllers. Controllers typically contain some control parameters that require tuning for their specific application. This can be performed using either conventional tuning methods or optimization methods with a predefined cost or objective function. However, some processes, especially continuous processes that operate for a longer period of time, are defined by some slowly changing dynamics such as fouling. In such processes, maintaining the control parameters constant during operation is not favorable and dynamic optimization of the control parameters would be more beneficial. A potentially suitable optimization method is Bayesian optimization using a Gaussian process surrogate model. This is a derivative-free and data-efficient optimization method. However, the applications of Bayesian optimization for dynamic optimization problems are limited in literature. In this thesis, a modified version of Bayesian optimization for dynamic optimization problems called adaptive Bayesian optimization is evaluated. First, a PI controller, which has two control parameters, is used to show that Bayesian optimization is very well suited for static controller tuning. Subsequently, adaptive Bayesian optimization is used to dynamically optimize the gain parameter of a PI controller and the experimental results are compared to the theoretical prediction for the optimal control parameter. Comparison of both methods showed close resemblances between both results indicating that adaptive Bayesian optimization is suited to track the optimal control parameter in time and capable of generating decent predictions for its location during the next iteration. Although the results of adaptive Bayesian optimization are promising, improvements can still be achieved. Here, the length scale expressing the rate of change along the time direction was fixed to a predefined value. Proper selection of this length scale once sufficient knowledge is available should enhance the performance of the optimization method. In addition, some intrinsic properties of Bayesian optimization limit its applicability. First of all, Bayesian optimization is data-driven implying that the method requires regular evaluations of the control parameter which can only be achieved if the process is disturbed regularly by a more or less similar disturbance to ensure that the evaluations are comparable. Second, the optimization method will always require a minimal amount of exploration of the input space which might not be feasible in some applications.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

In most recent decades, humanity has been becoming aware of its impact on the environment, and the potential dangerous consequences to the standard of living if its impact is not kept sufficiently in check. More often than not, greenhouse gas emissions enter the debate on global temperature increase. Whilst carbon dioxide — rightfully so — is mostly associated to global warming, methane is considered as the second most a harmful gas. Current industrial technology does not always allow to effectively and cost-efficiently mitigate methane release into the atmosphere. However, aerobic bioconversion of methane via a special class of bacteria, methanotrophs, may provide an economically viable solution to prevent some methane emissions. Of particular interest, Methylomicrobium buryatense is a species of methanotrophs, discovered around two decades ago. The microorganism is able to produce lactate as byproduct from methane consumption. Given that lactate has several usages, including use as precursor into bioplastics, Methylomicrobium buryatense has gathered interest for being used in industrial bioconversion. This strain is even more attractive for practical applications due to the relative ease of genetic modifications which could increase the lactate production. In order to better understand the metabolic behavior of Methylomicrobium buryatense, in silico models are utilized. In particular, dynamic enzyme-cost Flux Balance Analysis (deFBA), a type of constraint-based model, is applied in this thesis to simulate the metabolism of Methylomicrobium buryatense over time. In this thesis, an existing deFBA model of Methylomicrobium buryatense, is improved upon. This initial draft model has been constructed from earlier scientific publications. For the purpose of model improvement, some key model parameters are identified, and information obtained from literature review is used to adapt the model where necessary. Furthermore, it is attempted to utilize data of fed-batch lab experiments on Methylomicrobium buryatense to predict this experimental data from the deFBA model. The different experimental datasets are briefly compared as well. Experimental data show both qualitative and quantitative differences, such that different parameters are estimated depending on the considered data source. This thesis also contains an overview of the different methodologies and concepts used to the deFBA model improvement.