Dražen Slišković

Adaptive soft sensor for online prediction and process monitoring based on a mixture of Gaussian process models

Abstract Linear models can be inappropriate when dealing with nonlinear and multimode processes, leading to a soft sensor with poor performance. Due to time-varying process behaviour it is necessary to derive and implement some kind of adaptation mechanism in order to keep the soft sensor performance at a desired level. Therefore, an adaptation mechanism for a soft sensor based on a mixture of Gaussian process regression models is proposed in this paper. A procedure for input variable selection based on mutual information is also presented. This procedure selects the most important input variables for output variable prediction, thus simplifying model development and adaptation. Apart from online prediction of the difficult-to-measure variable, this soft sensor can be used for adaptive process monitoring. The efficiency of the proposed method is benchmarked with the commonly applied recursive PLS and recursive PCA method on the Tennessee Eastman process and two real industrial examples.

Stream water temperature prediction based on Gaussian process regression

The prediction of stream water temperature presents an interesting topic since the water temperature has a significant ecological and economical role, such as in species distribution, fishery, industry and agriculture water exploitation. The prediction of stream water temperature is usually based on appropriate mathematical model and measurements of different atmospheric factors. In this paper, a probabilistic approach to daily mean water temperature prediction is proposed. The resulting model is a combination of two Gaussian process regression models where the first model describes the long-term component of water temperature and the other model describes the short-term variations in water temperature. The proposed approach is developed even further by modeling the short-term variations with multiple Gaussian process regression models instead with a single one. Apart from that, variable selection procedure based on mutual information is presented which is suitable for input variable selection when nonlinear models for stream water prediction are developed. The proposed approach is compared with traditional modeling approaches on the measurements obtained on the Drava river in Croatia. The presented methodology can be used as a basis of the predictive tools for water resource managers.

Estimation of difficult-to-measure process variables using neural networks - a comparison of simple MLP and RBF neural network properties

In this paper, two different artificial neural networks are tested and compared with regard to their application in the estimation of difficult-to-measure process variables. Two of the most commonly used neural networks, the MLP (multilayer perceptron) and RBF (radial basis function) neural networks, with simple structure and standard training methods are chosen as examples. Neural network training is based on available data from a database of process variables measured over a long time period. The database in this paper is obtained using a simulation model of a real process. Without going deeper into theoretical background, relative properties of these neural networks are given through the results obtained by testing the trained networks and analysis performed on these results.

Efficient postprocessing of edge maps for image segmentation based on greedy correction cost minimization

A highly efficient postprocessing technique which enables the result of edge detection to be used for image segmentation is proposed. The method starts from an edge map obtained by a standard edge detection tool, e.g., Canny edge detector, and corrects it to obtain an edge map in which every edge point belongs to a closed boundary of an image region. The correction of the original edge map assumes removing some of the existing edge points as well as inserting virtual edge points. The proposed edge map correction procedure consists of two stages: (1) edge linking, which closes the gaps in edge contours by inserting virtual edge elements, and (2) edge pruning, which rejects spurious contours thereby avoiding over-segmentation. The edge pruning procedure performs an iterative greedy minimization of a correction cost function, while keeping all contours of the edge map closed. The proposed approach is evaluated using a set of standard test images.

Adaptivna estimacija teško-mjerljivih procesnih veličina

There exist many problems regarding process control in the process industry since some of the important variables cannot be measured online. This problem can be significantly solved by estimating these difficult-to-measure process variables. In doing so, the estimator is in fact an appropriate mathematical model of the process which, based on information about easy-to-measure process variables, estimates the current value of the difficult- to-measure variable. Since processes are usually time-varying, the precision of the estimation based on the process model which is built on old data is decreasing over time. To avoid estimator accuracy degradation, model parameters should be continuously updated in order to track process behavior. There are a couple of methods available for updating model parameters depending on the type of process model. In this paper, PLSR process model is chosen as the basis of the difficult-to-measure process variable estimator while its parameters are updated in several ways—by the moving window method, recursive NIPALS algorithm, recursive kernel algorithm and Just-in-Time learning algorithm. Properties of these adaptive methods are explored on a simulated example. Additionally, the methods are analyzed in terms of computational load and memory requirements.

Application of PLS and LS-SVM in difficult-to-measure process variable estimation

Very often important process variables which are concerned with the final product quality cannot be measured by a sensor or the measurements are too expensive and often not reliable. In order to enable continuous monitoring of process variables and efficient process control, soft-sensors are usually used to estimate these difficult-to-measure process variables. Soft-sensor is based upon mathematical model of the process. Process model building is based on plant data, taken from the process database. In this paper two methods, namely, Partial Least Squares (PLS) and Least Squares Support Vector Machines (LS-SVM), are used for difficult-to-measure process variables estimation. The methods are used for modeling simulated fluid storage process and oil distillation process. Results are compared and discussed. Advantages and disadvantages of each approach are outlined with respect to this specific application area. Additionally, hybridization of these methods is proposed which exploits good properties of both methods.