Integrating Dynamic State Information to Enhance Gas Concentration Modeling Accuracy

Gas monitoring plays a crucial role in intelligent societies. Conventional models assume that sensors remain in equilibrium and rely on discretely collected training data. However, the sluggish diffusion processes inherent to gas sensors can lead to disparate responses even under identical states. In this study, the mechanism behind these discrepancies between different dynamic conditions is disclosed through a rigorous theoretical analysis. By adopting continuously collected data, the prediction scenario can be aligned with the training process, reducing the time required for data collection. Subsequently, the integration of dynamic state information into the modeling process becomes crucial to alleviate the dynamic bias. To address this issue, three modeling strategies are introduced: 1) constructing models based on distinct dynamic states; 2) utilizing shared feature extraction layers; and 3) embedding dynamic state labels. A comprehensive analysis of 12 machine learning algorithms across three datasets is conducted to assess their effectiveness in leveraging dynamic state information. Most algorithms exhibit a reduction in fitting errors of over 10% when dynamic information is considered in two external datasets. In a closed-loop demonstration using the authors’ dataset, the implementation of the three strategies yields reductions in fitting error of up to 43.64% convolutional neural network (CNN), 39.66% CNN, and 29.10% multilayer perception (MLP), respectively. This research demonstrates that integrating dynamic state information improves the accuracy of gas concentration modeling and holds significant implications for gas monitoring applications.