APVC2021-A Gas Leakage Detection Method with Hybrid Acoustic Feature Selection and Stacking Ensemble Learning

Abstract

Abstract Model’s generalization and feature selection are always two challenging problems for gas leakage detection. This paper presents a method for gas leakage detection based on Hybrid-Feature-Selection-Stacking ensemble learning (HFS-Stacking), which fuses the Hybrid Feature Selection and Stacking ensemble learning. Firstly, a hybrid feature selection algorithm (HFS), which combines SFS-SVM, SFS-KNN, RFE-RF, RFE-XGB, and MIC algorithms, is proposed to select the optimal feature subset from multiple commonly used acoustic signal features; Then, SVM, KNN, random forest and XGBoost are designed as base learners in the stacking integration framework, which increase the generalization ability of the model. XGBoost is used as the meta-learner to output the classification results. The proposed gas leakage detection method gets the optimal subset of features and speeds of the subsequent model. Meanwhile, the constructed model can effectively improve the indicator of ,, and, and has a good generalization ability. The experimental results show that the HFS-Stacking algorithm using fewer features can effectively improve the training speed, accuracy, F1-score, AUC value and recall rate, and also has better robustness.