Transfer Learning for Gas Emission Prediction:A Comparative Analysis of Six Machine Learning Methods and TabNet

Abstract

Abstract Gas emissions in the tunnel will be a significant hindrance to its safe construction and may result in major consequences such as injuries and financial losses. Due to the peculiar characteristics of gas emission, the sample size of data on gas emissions is typically tiny; therefore, it is crucial to prevent over-fitting and to raise the precision and stability of predictions. We propose in this study to use tunnel inflow data as the source data and gas emission data as the target data, and then use transfer learning to predict gas emissions. For transfer learning, several well-known and effective machine learning models are used: AAN (artificial neural network), ET (extra tree), GB (gradient boost), KNN (K-nearest neighbor), MLP (multilayer perception), SVM (support vector machine), and XGBOOST (extreme gradient boost). To ensure the correctness and effectiveness of the experiment, the Tabnet model without transfer learning is used as a comparison. The method's viability and effectiveness are then confirmed by comparison with three sets of actual measurement data and the Tabnet model without transfer learning. The research demonstrates that: The transfer learning method, which uses tunnel water gushing data as the source data and gas emission data as the target data, confirms the viability and effectiveness of the method through the prediction of three groups of measured data and the comparison with the accurate and efficient Tabnet model, Indications are promising that the approach can provide a novel framework for improving the accuracy with which gas emissions are predicted.