Development of KSVGRNN: A hybrid soft computing technique for estimation of boiler flue gas components

Abstract

In this research paper, a novel hybrid technique named KSVGRNN, which combines a multi-class support vector machine (SVM) and a generalized regression neural network (GRNN), has been developed for obtaining the composition of boiler flue gas mixtures. This hybridization was made by the support of K-means clustering algorithm and grid search technique. In the first phase, K-Means clustering technique has been utilized and the size of the training vectors has been reduced by employing a multiclass SVM. In the second, a GRNN has been trained for estimating the individual gas concentration in the flue gas mixture. The reduction of training vectors through SVM has been shown to improve the generalization capability of GRNN. Grid search has been utilized to obtain the optimal parameters of SVM. This hybrid technique has been validated by measuring its performance by processing volatile organic component (VOC) data acquired from quartz crystal microbalance (QCM) and SnO₂ semiconductor type sensors utilized by other researchers in this domain. Further studies have been carried out to assess the discriminating and estimation capability of the proposed hybrid technique for real-time flue gas data obtained from two different analyzers namely ORSAT^® and KANE^®. The outcome of these studies, observations and analysis clearly indicate the exceptional performance of the proposed hybrid model in classifying and estimating the flue gas components in the machine (Analyzer) independent manner.

Introduction

Gas sensing applications include analyzing a gas mixture for finding the components concentration from the sensor data using intelligent techniques like Artificial Neural Networks (ANN), Support Vector Machine (SVM), Fuzzy Logic (FL), to name a few. Control of excess air in Boilers improves boiler efficiency and also leads to air pollution control. Developing sensors to detect various gases for subsequent utilization of their properties in gas sensing areas has gained industrial importance due to their vital applications in industrial safety engineering. A lot of experimental works including a fundamental study on multi-mode quartz crystal gas sensors [1], chemically selective sorbent coating on QCR [2], and detection of Hydrogen in ambient air using a coated piezoelectric crystal [3] has been reported in the literature. In the past decade, a wide variety of pattern recognition techniques has been designed for developing gas recognition systems. Several computational models based on heuristics, Statistical parameters based designs, Optimization procedures have been developed for processing the digital data acquired from sensor responses for decision-making purposes leading to optimum resource utility [4,5].

Gas mixture composition finding consists of a few tool based computational stages such as identification, classification, prediction, estimation and optimization. This process requires periodical data observed from sensors. A good data acquisition system needs to be designed for processing the sizeable data acquired from sensors to classify them into appropriate classes. Once the classification task is over, a good estimator needs to be constructed so as to predict the gas components for quantification of each component with a practical level of accuracy.

Manual analysis of the data is time-consuming, error-prone and inconsistent. Moreover, the large sized data acquired from the gas sensors need a computational model which can capture the nonlinear relationships between sensor reading and the actual composition of gas mixtures. In this context, noteworthy contributions are found in the literature [6], [7], [8], [9], [10], [11], [12] which use Machine learning techniques [31] to address various issues such as feature selection, extraction, clustering classification prediction and estimation. Typically Neural Networks (NN) is employed to capture the nonlinearity in the prediction level. Various types of NN are used for applications ranging from classification to optimization. Since BFA involves prediction and estimation of gas components Generalized Regression Neural Network (GRNN) is found to be an appropriate tool.

The most important issue in machine learning with respect to processing sensor data is over-fitting which reduces the generalization capability of learned model. Since issues pertaining to generalization capability differ from classifier to classifier, designing an appropriate preprocessing technique as a classifier which augurs well with the GRNN is an important aspect of hybridization.

In this work, a hybrid technique has been developed by integrating the following: (i) K-Means clustering technique, (ii) Support Vector Machines based classifier (for multi- classification) and, (iii) a Generalized Regression Neural Network (GRNN). In the initial stage, the digital data acquired from the sensor responses are preprocessed for appropriate scaling of the attributes and similar data are grouped together using k-means clustering algorithm. In the next stage, an SVM with Radial Basis Function Kernel is implemented for classification. The support vectors, so generated in the higher dimensional space of the SVM are used to train the proposed GRNN. This technique is validated with volatile organic compounds data [33]. The quantitative discriminating capability of proposed technique has been analyzed in turn with data acquired from KANE [36] and ORSAT [34] analyzers, whose descriptions and specifications are furnished in [35].