Enhancing subtle features of gas detection signals using MFD_attention mechanism for improved detection in complex environments

Abstract

Hydrogen is widely used in industrial environments as a clean energy source. However, hydrogen leaks can easily lead to fires and explosions. Real-time monitoring of gas types and concentrations is essential to ensure industrial safety. When using semiconductor gas sensors to detect gas mixtures, the sensitivity of the sensors to different gases is often similar. As a result, subtle features of low concentration gases or gases with low sensitivity are easily masked or overlapped, and thus ignored in the algorithmic recognition process, greatly reducing the detection accuracy. To address this problem, this paper proposes a novel model that combines multi-scale filtering decomposition (MFD), kernel convolution, and an attention mechanism (AM) to extract subtle features from the data and enhance the useful information while suppressing redundant signals through adaptive weight assignment. We validated the model using commonly used regression algorithms: BP neural network and support vector machine. The results show that the total mean absolute error (MAE) root-mean-square error (RMSE) for predicting the concentration of hydrogen-methane mixtures showed a reduction of more than 50 %. In addition, to test the generalization ability of the model, we applied it to hydrogen-carbon monoxide mixtures and hydrogen-methane mixtures under different humidity conditions, and the results improved the prediction accuracy for both.