Transformer-based odor recognition on E-nose platforms

Transformers have reshaped sequence modelling, yet their value for embedded electronic-nose (e-nose) systems remains largely anecdotal. We present a systematic evaluation of ConvTran, a lightweight self-attention network originally proposed for generic time-series analysis, in the specific context of portable gas sensing. Two open datasets acquired with an eight-sensor Bosch BME688 board underpin the study: CoffeePow-4 (3583 traces across three coffee powders plus ambient air) and the imbalanced Aroma-7 extension that adds three fragranced creams for a total of 4750 traces. ConvTran is benchmarked against five deep-learning baselines (FCN, ResNet, InceptionTime, LSTM-FCN, ALSTM-FCN) and the proprietary Bosch AI-Studio classifier using accuracy, macro-F1 and macro false-positive rate, complemented by on-device memory and latency profiling. On CoffeePow-4, ConvTran attains 96.42% macro-F1 with 1.71 ms inference time while occupying 115 kB of flash. On Aroma-7, it preserves 96.03% macro-F1 at 2 ms and 557 kB, remaining faster than ResNet or InceptionTime, whose accuracy drops by 1.3%. Fully convolutional networks accuracy exceed ConvTran by 0.6% on the four-class task but loses 17% when the number of classes increases, highlighting the advantage of global self-attention for chemically similar odors. These findings demonstrate that an off-the-shelf transformer can satisfy the requirements of embedded devices, delivering state-of-the-art accuracy across balanced and imbalanced odor sets.