Fast identification of flammable chemicals based on broad learning system

Abstract

Fast identification of flammable chemicals is essential for industrial production and laboratory safety. With the continuous advancement of sensor technology, data-driven methods have become a promising tool for gas identification. However, these methods face problems such as insufficient feature learning, unstable prediction of the single classifier, and overfitting caused by insufficient data. In this work, a kernel-based BLS (KBLS) method is proposed, in which the kernel matrix is used to calculate the sample distances and map the feature nodes to the kernel space to reduce the uncertainty. In addition, KBLS uses a pseudo-inverse method to solve the weights, which greatly avoids the risk of overfitting while improves computational efficiency. To avoid the errors caused by a single classifier for specific gas samples, KBLS is used as the weak learner and combined with the AdaBoost algorithm to form an Ada-KBLS classifier to achieve fast and accurate gas identification. In the Ada-KBLS model, the sample weights obtained by the previous weak learners are used to train the following weak learners. This method improves the classification performance by paying attention to difficult and misclassified samples and integrating the classification results of multiple weak learners. In addition, a dataset containing four flammable gases is used to verify the effectiveness of the Ada-KBLS model. The initial stage of all response data is divided into different time windows as the input of the model to test the fast gas identification ability of the method. The Ada-KBLS achieves an average classification accuracy of 98.4 % in the 4 s time window, the best among all models, and the training time is only 6.22 s. The result represents a 0.4 % improvement over the second-best model, KBLS, and a 4.5 % increase compared to the 93.9 % accuracy achieved by Random Forest (RF). In addition, the precision, recall, and F1-score of ethanol gas classification reach high values of 100 %. The experimental results demonstrate the robustness and effectiveness of the proposed method in handling the task of fast detection of flammable gases, thus promoting the application of BLS and ensemble learning in gas identification.