No abstract

This article presents a novel methodology to model a type-1 singleton fuzzy logic system (T1 SFLS) for temperature prediction in a secondary metallurgical process that takes place inside a ladle furnace. The proposal generates approximations using the energy consumed and the time elapsed within the casting process as input data, without using other instruments. It is known that the temperature cannot be verified all the time in the ladle furnace because it is sealed when it is in operation, and when temperature is measured, there is an uncertainty level in the sensor reading that generates predictions of the temperature in the order of 2.5 % out of the real value. The three proposed methodologies for the T1 SFLS forecaster provide a more accurate approximation of the temperature with less than 1 % of uncertainty. The predicted temperature is used in decision making to generate the required chemical composition of the steel and to mark the appropriate times to aggregate the additives in the alloy and achieve the required chemical balance. Compared with the model used by the industry, the results obtained show that the use of the proposed fuzzy model gives the opportunity to increase the quality of the steel by improving the adjustment of the quantities of additives that are lost by oxidation.

This paper presents type-1 and type-2 radial basis function networks to evaluate quality features. The proposed methodology fuses the central composite design and the radial basis function neural networks in type-1 or interval type-2 model to generate a network that evaluates quality features in an industrial image processing. The advantages of this proposal include that training is not required to get an accurate result and that the generation of the fuzzy rule base using central composite design method and statistical indicators is simplified. Another advantage is the excellent results obtained with the proposal. Experimentation shows an error reduction of 90% when the interval type 2 Mandami Radial basis function neural network compared against its type-1 counterpart using the Gaussian membership functions onto a radial basis function network.

No abstract