WMDCNN: A deep learning method for gas sensor arrays utilizing feature correlation and CACLoss

Existing gas recognition methods rely on two primary methodologies: traditional machine learning, which is resource-efficient but struggles with complex cross-interference, and deep learning, which effectively handles complexity at the expense of high computational costs. This study proposes a novel deep learning framework for gas sensor arrays based on feature correlation analysis. First, highly discriminative heuristic feature subsets are constructed by prioritizing feature extraction methods that yield the highest Pearson correlation coefficients. This step is validated using a robust base discriminator. Second, a wise multi-dimensional convolutional neural network (WMDCNN) is introduced to process these purified feature vectors. A novel class anchor clustering loss (CACLoss) is integrated to regularize the latent space, effectively enforcing dense intra-class clustering and maximizing inter-class margins to resolve boundary blurring in complex gas mixtures. Last, the proposed framework is rigorously evaluated on proprietary and public datasets. Experimental results demonstrate that the WMDCNN achieves a superior mixed-gas classification accuracy of up to 97.51% and 97.89% on both datasets. Furthermore, in concentration prediction tasks, the root mean square error (RMSE) of the proposed model is reduced by 37.85% to 59.66% compared with that of the baseline model. These results confirm that the WMDCNN framework offers a highly accurate, computationally efficient, and drift-tolerant solution for mixed gas recognition.