Fast Gas Detection Method Based on Adaboost and Kernel-Based Broad Learning System

Abstract

Broad Learning System (BLS) is a feed-forward neural network model that performs well on classification and prediction tasks. However, BLS uses double random mapping to extract features, which will produce uncertainty. Furthermore, the performance of BLS is affected by the number of hidden nodes, which will lead to complex manual tuning of parameters. In this work, we propose a Kernel-Based BLS (KBLS) method based on Kernel mapping, in which the kernel matrix instead of the enhancement nodes is utilized to enrich the extraction of features, thus making the data more discriminative and avoiding the process of manual adjustment. To further improve the stability and robustness of the system, KBLS is used as weak learners to construct the AdaBoost model. The preprocessed data in the early response as the input of the model to classify ethylene, ethanol, methane, and CO. Furthermore, the compared experimental results of Ada-KBLS with other existing methods, including random forest (RF), support vector machine (SVM), gradient boosting decision tree (GBDT), and multilayer perceptron (MLP) show that the proposed Ada-KBLS model achieves the best performance among all the models in the most time windows, with a highest accuracy of 98%.

Keywords: Early response, ensemble learning method, electronic nose, kernel-based broad learning system (KBLS)