Discrimination of gas mixtures remains a challenging task due to overlapping sensor responses and pronounced temporal evolution under real-time monitoring. To address this challenge, we investigate deep learning–based approaches for the analysis of CO–C2H4 gas mixtures using time-series signals acquired from 16-channel metal-oxide sensor (MOS) array under 12h real-time monitoring conditions. Convolutional, recurrent, and hybrid architectures, including CNN, GRU, LSTM, TCN, and their combinations, are evaluated for classification and regression multitasking. The results demonstrate that the GRU-based model consistently achieves the best overall performance, with a classification accuracy of 96.06% and a global regression R2 of 0.939. In contrast, hybrid architectures incorporating additional convolutional modules, such as CNN+GRU and cascaded TCN+GRU, do not provide further performance improvements, suggesting that the essential predictive information is primarily encoded in global temporal dynamics. Therefore, our study provides valuable insights for a robust and lightweight AI-driven readout system for the next-generation real-time monitoring device.