The parameter-feature-based transfer learning for mixed gas monitoring under small datasets

Gas monitoring is an important application in the field of gas sensor systems. Currently, gas monitoring methods primarily rely on processing large amounts of high-quality data using machine learning or deep learning algorithms. While this high quality and large quantity of data are sometimes hard to achieve due to the high costs associated with the data collection stage. To this end, a novel gas monitoring method developed for small datasets using parameter-feature-based transfer learning is proposed in this paper. First, a deep learning-based source model will be developed, which will use a large dataset that has undergone data augmentation for parameter training. Then, a parameter-feature transfer learning method will be used to transfer the model parameters to the target domain of the small dataset for mixed gas detection and recognition. Experiments were conducted on a convolutional neural network combined with a long and short-term memory network, and the results showed that the amount of data required was reduced by 96.5 % and 76.1 %, respectively, compared to the traditional experimental procedure. In addition, the developed model was able to achieve robust concentration prediction with accuracies of 96.3 % and 99.6 % in different gas monitoring tasks.