Wojciech Rafajłowicz

Local Detection Of Defects From Image Sequences

Local Detection Of Defects From Image Sequences Our aim is to discuss three approaches to the detection of defects in continuous production processes, which are based on local methods of processing image sequences. These approaches are motivated by and applicable to images of hot metals or other surfaces, which are uniform at a macroscopic level, when defects are not present. The first of them is based on the estimation of fractal dimensions of image cross-sections. The second and third approaches are compositions of known techniques, which are selected and tuned to our goal. We discuss their advantages and disadvantages, since they provide different information on defects. The results of their testing on 12 industrial images are also summarized.

A modified filter SQP method as a tool for optimal control of nonlinear systems with spatio-temporal dynamics

A modified filter SQP method as a tool for optimal control of nonlinear systems with spatio-temporal dynamics Our aim is to adapt Fletchers filter approach to solve optimal control problems for systems described by nonlinear Partial Differential Equations (PDEs) with state constraints. To this end, we propose a number of modifications of the filter approach, which are well suited for our purposes. Then, we discuss possible ways of cooperation between the filter method and a PDE solver, and one of them is selected and tested.

Image-Driven Decision Making with Application to Control Gas Burners

Our aim is to propose a model-free approach to decision making that is based on the direct use of images. More, precisely, a content of each image is used – without further processing – in order to cluster them by the K-medoids method. Then, decisions are attached to each cluster by an expert. When a new image is acquired, it is firstly classified to one of the clusters and the corresponding decision is made. The approach is conceptually rather simple, but its success in on-line applications depends on the way of organizing learning and decision phases. We illustrate the approach by the example of a decision-making system for industrial gas burners.

Iterative Learning in Repetitive Optimal Control of Linear Dynamic Processes

Learning of stochastically independent decisions is a well developed theory, the main of its part being pattern recognition algorithms. Learning of dependent decisions for discrete time sequences, e.g., for patterns forming a Markov chain and decision support systems, is also developed, but many classes of problems still remain open. Learning sequences of decisions for systems with continuously running time is still under development. In this paper we provide an approach that is based on the idea of iterative learning for repetitive control systems. A new ingredient is that our system learns to find the optimal control that minimizes a quality criterion and attempts to find it even if there are uncertainties in the system parameters. Such approach requires to record and store full sequences of the system state, which can be done using a camera for monitoring of the system states. The theory is illustrated by an example of a laser cladding process.

Statistical Classifier with Ordered Decisions as an Image Based Controller with Application to Gas Burners

We consider a statistical decision problem as a tool for solving control problems with a camera in the loop. The first stage is features extraction from images. Its role is to process images in order to extract features relevant for the control problem. Then, they are fed as inputs to the Bayesian decision problem. At the second stage a loss function, which is a sum of squared deviations of decisions from true decisions is considered. Finally, an approximation of the optimal decision rule is proposed, using a learning sequence of decisions, which – together with feature extracting algorithms – form the control system. The proposed approach is illustrated by a system that is dedicated to control natural gas burners.

Fletcher's filter methodology as a soft selector in evolutionary algorithms for constrained optimization

Our aim is to propose a new approach to soft selection in evolutionary algorithms for optimization problems with constraints. It is based on the notion of a filter as introduced by Fletcher and his co-workers. The proposed approach occurred to be quite efficient.

(n+r)D systems and their control with reduced sensitivity to parametric uncertainties

Our aim is to propose an extension of nD systems by treating uncertain parameters of a system as additional independent variables. We recall known results on deriving equations for the sensitivity of the system state to parameter changes. Then, the problem of optimal control of linear systems extended by sensitivity equations with the quadratic criterion is stated. Its solution is relatively easy using the well known results for LQ optimal systems, but in our case the optimal controller is fed additionally by sensitivity signals. A numerical example indicates that the behavior of control system with reduced sensitivity is different than the behavior of classical systems, which is the price that we pay for parameter uncertainty.

Iterative learning in optimal control of linear dynamic processes

ABSTRACT We consider iterative learning control of linear repetitive processes in a setting, which is motivated by laser cladding processes that are controlled with the help of cameras and/or infra-red (IR) cameras. We develop a gradient-like learning procedure that is based on the Frechet derivative of a control quality criterion. Then, we prove its convergence. We also provide local bounds on parameter uncertainty for which the convergence of the learning process is still retained. The proposed approach is extensively tested using a very accurate model of the gantry robot. Finally, an example of the laser cladding process is discussed. Images from a camera and IR camera allow us to design a proper temperature profile, for which a laser power control signal is calculated by simulations of the the repetitive process.

Linear matrix-state systems and their use for image-driven control

Our aim is to consider a model-based approach to control of image-driven control systems, i.e., the systems in which images play the role of system states. As a tool for describing such systems we have selected n × n matrix state equations. For linear, time invariant systems we prove necessary and sufficient controllability and asymptotic stability conditions that can be efficiently verified by considering an accompanying system with n-dimensional states. Furthermore, the existing software can be used for their verification. The approach is illustrated by a semi-discretized heat equation with a control system, which is used to invoke large stresses in a ceramic plate.

Numerical Optimal Control of Integral-Algebraic Equations Using Differential Evolution with Fletcher’s Filter

Integral-algebraic equations are an interesting method of modeling real world problems with not too severe assumptions. We proposed a simple numerical method of using differential evolution. Constraints in optimal control problems are handled using a method based on the works of Fletcher and his co-workers’ filter.