A novel method for H₂S concentration prediction under small sample based on ECA-1DCNN-XGBR

Although gas concentration prediction based on deep learning has made significant progress, the accuracy is typically achieved on the basis of a large number of training samples, making it challenging to meet the requirements of real industrial scenarios. Moreover, traditional neural networks often face issues such as insufficient feature extraction or overfitting in the condition of small sample. In this work, a novel detection method that combines one-dimensional convolutional neural network (1DCNN) featuring efficient channel attention (ECA) mechanism with extreme gradient boosting regressor (XGBR) is proposed to address the aforementioned issue, and simultaneously, a high-quality dataset of H ₂ S with small sample has also been collected through an automated gas data acquisition system fully operated by a computerized environment. Due to the special ensemble structure and regularization terms, XGBR can resist overfitting under small sample condition. Furthermore, the deep feature extraction capabilities of neural networks, coupled with the characteristic of attention mechanism to focus on key features, empower ECA-1DCNN to efficiently extract features. The experimental results demonstrated that the R ² of ECA-1DCNN-XGBR reached 0.9999, and a RMSE of 0.584 and an MAE of 0.374 were simultaneously achieved. Meanwhile, compared with traditional machine learning and deep learning models, the proposed method performed best in regression prediction tasks. These results indicate proposed method performs excellently in the prediction of H ₂ S gas concentrations under small sample, with high accuracy and reliability.