A minimalist Pd/Ag/SnO₂ array coupled with SE-enhanced multi-scale CNN for robust and accurate identification of CH₄/CO mixtures

Methane (CH₄) and carbon monoxide (CO) often coexist in hazardous environments such as coal mines and fire scenes, making the accurate and reliable discrimination of single gases and their mixtures critical for safety monitoring. In this work, a minimalist two-sensor Pd/Ag/SnO₂ array was fabricated on micro-hotplate substrates for gas identification. To effectively capture the dynamic features of gas interactions, transient responses were encoded into Markov Transition Field images and classified using an SE-enhanced multi-scale convolutional neural network (CNN). This framework supports the precise discrimination of seven gas classes, including pure CH₄, CO, and five binary mixtures with distinct concentration ratios. The proposed system achieved a peak classification accuracy of 100%, with even a single sensor response reaching 98.57%. Following periodic downsampling from 10 Hz to an effective sampling frequency of 1 Hz, the model achieved a classification accuracy of 89.3% with a 100-point (100 s) input window. These results demonstrate the potential of the proposed methodology for accurate gas identification using minimal hardware and truncated response windows.