TF-former: A concentration regression model for toxic and harmful mixed gases based on time-frequency analysis

Abstract

Accurate quantification of individual components in mixed gases remains challenging for metal oxide (MOX) electronic noses because cross-sensitivity can produce highly overlapping sensor responses. In this study, we propose TF-Former, a time–frequency Transformer framework for the simultaneous regression of carbon monoxide (CO) and ethylene (C2H4) concentrations from multichannel MOX sensor signals. For the public UCI dynamic gas-mixture dataset, electrical responses from 16 chemiresistive sensors were acquired at 100 Hz under a constant heater voltage of 5 V. The signals were downsampled to 1 Hz, standardized using statistics estimated from the training data only, and grouped by continuous constant-concentration exposure intervals to reduce information leakage across the training, validation, and test subsets. TF-Former jointly encodes the original time-domain sequence and single-level Haar discrete wavelet transform coefficients, and integrates the two representations through bidirectional cross-domain attention before concentration regression. Across five random seeds, TF-Former achieved R2 values of 0.9960±0.0019 for CO and 0.9931±0.0021 for C2H4 on the grouped UCI evaluation. On a second MOX-array dataset evaluated using held-out concentration combinations, TF-Former achieved R2 values of 0.9877±0.0101 for CO and 0.9886±0.0093 for C2H4. Additional Gaussian-noise perturbation experiments were conducted to assess prediction stability under corrupted inputs. These results demonstrate the potential of joint time-frequency representation learning for concentration estimation of known CO/C2H4 mixtures under the investigated conditions.