Ilmari Juutilainen

Two-level clustering approach to training data instance selection: A case study for the steel industry

Nowadays, huge amounts of information from different industrial processes are stored into databases and companies can improve their production efficiency by mining some new knowledge from this information. However, when these databases becomes too large, it is not efficient to process all the available data with practical data mining applications. As a solution, different approaches for intelligent selection of training data for model fitting have to be developed. In this article, training instances are selected to fit predictive regression models developed for optimization of the steel manufacturing process settings beforehand, and the selection is approached from a clustering point of view. Because basic k-means clustering was found to consume too much time and memory for the purpose, a new algorithm was developed to divide the data coarsely, after which k-means clustering could be performed. The instances were selected using the cluster structure by weighting more the observations from scattered and separated clusters. The study shows that by using this kind of approach to data set selection, the prediction accuracy of the models will get even better. It was noticed that only a quarter of the data, selected with our approach, could be used to achieve results comparable with a reference case, while the procedure can be easily developed for an actual industrial environment.

Modeling the temperature of hot rolled steel plate with semi-supervised learning methods

The semi-supervised learning methods utilize both the labeled and unlabeled data to produce better learners than the usual methods using only the labeled data. In this study, semi-supervised learning is applied to the modeling of the rolling temperature of steel plate. Measurement of the rolling temperature in the extreme conditions of rolling mill is difficult and thus there is a large amount of missing response measurements. Previous research mainly focuses on semi-supervised classification. Application of semi-supervised learning to regression problems is largely understudied. Co-training is a semi-supervised method, which is promising in the semi-supervised regression setting. In this paper, we used COREG algorithm [10] to a data set collected from steel plate rolling. Our results show that COREG can effectively exploit unlabeled data and improves the prediction accuracy. The achieved prediction accuracy 16?C is a major improvement in comparison to the earlier approach in which temperature is predicted using physical-mathematical models. In addition, features that describe the rolling process and are applicable to input variables of learning methods are presented. The results can be utilized to develop statistical models for temperature prediction for other rolling processes as well.

Exceedance Probability Score: A Novel Measure for Comparing Probabilistic Predictions

Accurate prediction of exceedance probabilities is important in many applications. For example, in process planning and control, engineers should anticipate the risk that a product fails to meet its specification limits. Statistical comparison between candidate probability prediction methods is commonly performed using scoring rules, like the continuous ranked probability score (CRPS) and the logarithm score (LogS). In this work, a new scoring rule, the exceedance probability score, is proposed. The experiments in simulated and real industrial data show that the new scoring rule is useful in comparing and testing differences in the predictive accuracy of competitive probabilistic predictions in regression setting. The proposed scoring rule have some similarities with CRPS and LogS, but is more directly connected to the accuracy in the prediction of exceedance probabilities.

Quantile regression model for impact toughness estimation

The purpose of this study was to develop a product design model for estimating the impact toughness of low-alloy steel plates. The rejection probability in a Charpy-V test (CVT) is predicted with process variables and chemical composition. The proposed method is suitable for the whole production line of a steel plate mill, including all grades of steel in production. The quantile regression model was compared to the joint model of mean and dispersion and the constant variance model. The quantile regression model proved out to be the most effective method for modelling a highly complicated property at this extent. Next, the developed model will be implemented into a graphical simulation tool that is in daily use in the product planning department and already contains some other mechanical property models. The model will guide designers in predicting the related risk of rejection and in producing desired properties in the product at lower cost.

Modelling of Charpy V test rejection probability

Abstract The purpose of this study was to develop a product design model for estimating the impact toughness of low alloy steel plates. The rejection probability in a Charpy V test is predicted with process variables and chemical composition. Joint modelling of the mean and deviation was used in order to improve the results. The proposed method is suitable for the whole production line, including all grades of steel in production and it is not restricted to a few test temperatures. Using the proposed model the product design group could have recognised most of the rejections before production. Next, the developed model will be implemented into a graphical simulation tool that is in daily use in the product planning department and already contains some other mechanical property models. The model will guide designers in predicting the related risk of rejection and in producing desired properties in the product at lower cost.

Detection of Correct Moment to Model Update

When new data are obtained or simply when time goes by, the prediction accuracy of models in use may decrease. However, the question is when prediction accuracy has dropped to a level where the model can be considered out of date and in need of updating. This article describes a method that was developed for detecting the need for a model update. The method was applied in the steel industry, and the models whose need of updating is under study are two regression models, a property model and a deviation model, developed to facilitate planning of optimal process settings by predicting the yield strength of steel plates beforehand. To decide on the need for updating, information from similar past cases was utilized by introducing a limit called an exception limit for both models. The limits were used to indicate when a new observation was from an area of the model input space where the results of the models are exceptional. Moreover, an additional limit was formed to indicate when too many exceedings of the exception limit have occurred within a certain time scale. These two limits were then used to decide when to update the model.

A Method for Measuring Distance From a Training Data Set

A new method is proposed for measuring the distance between a training data set and a single, new observation. The novel distance measure reflects the expected squared prediction error when a quantitative response variable is predicted on the basis of the training data set using the distance weighted k-nearest-neighbor method. The simulation presented here shows that the distance measure correlates well with the true expected squared prediction error in practice. The distance measure can be applied, for example, in assessing the uncertainty of prediction.

Visualizing Predicted and Observed Densities Jointly with Beanplot

Beanplot is a graphical method for visualizing univariate distributions. Density forecasts have an important role to play in many applications. Although graphical methods are widely used for illustrating distributions, suitable graphical methods to help for the purposes of analysis and comparison of density forecasters do not exist. This article explains how density forecasts and related observed densities are visualized parallel using beanplots in different groups of data. The visualization method is illustrated with industrial and simulated data. The functionality extends the plotting function of R package beanplot and the developed functions are made available for R programming language.

Exceedance probability estimation for a quality test consisting of multiple measurements

The purpose of this study was to develop methods for exceedance probability estimation in the case of highly scattered measurement sets. The situation may occur when product quality is verified with several test samples, and thus, traditional point prediction based modelling methods are not sufficient. Density forecasting methods are needed when not only the mean but also the deviance and the distribution shape of the response depend on the explanatory variables. Furthermore, with probability predictors, the ranking methods for the model selection should be chosen carefully, when models trained with different methods are compared. In this article, the impact toughness of the steel products was modelled. The rejection probability in Charpy-V quality test was predicted with mean and deviation models, distribution shape model and quantile regression model. The proposed methods were employed in two steel manufacturing applications with different distributional properties.

Product design model for impact toughness estimation in steel plate manufacturing

The purpose of this study was to develop a product design model for impact toughness estimation of low-alloy steel plates. Based on these estimates, the rejection probability of steel plates can be approximated. The target variable was formulated from three Charpy-V measurements with a LIB transformation, because the mean of the measurements would have lost valuable information.The method is suitable for all steel grades in production and it is not restricted to a few test temperatures. There were differences between the performances of different product groups, but overall performance was promising. Next the developed model will be implemented into a graphical simulation tool that is in daily use in the product planning department and already contains some other mechanical property models. The model will guide designers in predicting the related risk of rejection and in producing desired properties in the product at lower cost.