A mixed gas concentration regression prediction method based on RESHA-ALW

Abstract

Previous research in gas regression predominantly concentrated on individual gas. However, in real world, gas typically coexists in mixed forms and the detection technology normally involves metal oxide sensors. This work introduces an efffcient model, denoted as RESHA-ALW, for mixed gas concentration prediction. A hybrid attention (HA) model is added into Residual Neural Network (RESNET) to reweight the features by generating a channel attention weight matrix and a locational attention weight matrix. Furthermore, an adaptive loss weighting (ALW) method is also applied to achieve the dynamic weighting of the different gases during the training process. The performance of the model is veriffed through the ffve-fold cross-validation experiments on the carbon monoxide (CO)-ethylene gas mixture dataset, with mean results of 0.25, 1.78, and 0.998 for ethylene (RMSE, SAMPE, and R2 ), as well as 7.97, 2.27, and 0.998 for CO, respectively, outperforming all the compared models. Then, the ALW method is applied to the comparison models, resulting in signiffcant increase in prediction performance of all the models, which conffrms the efffciency of the ALW method. Moreover, RESHA-ALW has demonstrated better performance in predicting the concentration of the mixture within a single task, as compared to RESHA’s performance when undertaking two separate tasks. This further validates the signiffcant application potential of our proposed model. Finally, we also conduct ablation experiments and compare the impact of choosing different activation functions on the model.

Tasks

Reproductions