An Algorithm for Soft Sensor Development for a Class of Processes with Distinct Operating Conditions

Abstract

Soft sensors are increasingly being used to provide important information about production processes that is otherwise only available through off-line laboratory analysis. However, usually, they are developed for a specific application, for which thorough process analysis is performed to provide information for the appropriate selection of model type and model structure. Wide industrial application of soft sensors, however, requires a method for soft sensor development that has a high level of automatism and is applicable to a significant number of industrial processes. A class of processes that is very common in the industry are processes with distinct operating conditions. In this paper, an algorithm that is suitable for the development of soft sensors for this class of processes is presented. The algorithm possesses a high level of automatism, as it requires minimal user engagement regarding the structure of the model, which makes it suitable for implementation as a customary industrial solution. The algorithm is based on a radial basis function artificial neural network, and it enables the automatic selection of the model structure and the determination of model parameters, only based on the training data set. The testing of the presented algorithm is done on the cement production process, since it represents a process with distinct operating conditions. The results of the test show that, besides providing a high level of automatism in model development, the presented algorithm generates a soft sensor with high estimation performance.

Magnetic Characterization of MR Fluid by Means of Neural Networks

“…This type of artificial neural network uses radial basis functions as activation functions, and it consists of an input layer, a hidden layer and an output layer [29]. The hidden layer applies the radial basis function, typically a Gaussian function [30]. RBF networks are computational models used for their simple architecture, fast training speed and the opportunity to include real data in the system.…”