Application of Automatic Learning Methods for Modelling and Control of Industrial Processes

Process models are required to perform automatic planning and control of manufacturing processes. Reliable quantitative models of processes are rarely available, and the lack of process models is a major obstacle for application of automation and optimisation in many manufacturing environments. A reason for this lack of process models is that the physics of the processes are not fully understood, so the capability of analytically developed models to predict the state of a process is not sufficiently good. Instead, the development of the models must be based on experiments covering a comprehensive range of process conditions. In most cases the number of experiments needed to develop the models is large, making the development of the models very resource demanding. There is therefore a strong need to develop methods, which are much more resource effective than the presently available methods. In this project it is investigated if the use of methods from machine learning can make the development of stable, robust and reliable process models more resource effective. A number of methods from the area of machine learning is investigated e.g. decision trees, Bayesian network, decision graphs, artificial neural network, instance-based learning and genetic algorithms. By using these methods available knowledge about the process from experiments, analytical knowledge, experts, operators and sensors is attempted to be collected and used to make process models and control the process. There will be focus on the welding process to continue developing the knowledge and expertise there is in this field at Department of Production. The objective is to develop a system architecture and make an installation in the laboratory to exploit the advantages of the investigated methodologies and demonstrate the proposed system architecture. A system is developed to make and analyse welding experiments automatic to create reliable empirical data. The empirical data, analytical and expert knowledge is used to construct and to train a process planning model based on a Bayesian network. A process planning model based on a Bayesian network shows good welding results and in comparison to artificial neural network and regression is the Bayesian network model the most promising. The project was carried out as a PhD project financed by Department of Production. The PhD thesis was defended June 26^th 2007 and the PhD degree was appointed September 19^th 2007. Supervisor: Ole Madsen. (Morten Kristiansen, Department of Production, AAU)