Optimisation of electronic nose performance with multi-attention and domain adaptation for fire forensics

Abstract

The identification of ignitable liquids at fire scenes is a crucial component of forensic investigations. However, conventional analytical methods such as gas chromatography-mass spectrometry, often require costly equipment, specialised expertise, and extended analysis times, which limits their effectiveness for rapid on-site detection. As such, electronic nose (e-nose) technology presents a cost-effective and portable alternative. Nevertheless, inconsistencies in sensor responses across different platforms pose challenges to model transferability which requires independent data collection and model training for each device. This study introduces a multiple attention adversarial transfer learning (MAATL) network aimed at addressing cross-platform variability in ignitable liquid detection using e-noses. The MAATL framework incorporates a multiple attention mechanism to optimise sensor signals, a multi-scale one-dimensional convolutional network for feature extraction, and adversarial learning techniques to enhance domain adaptation. Experimental validation involving five e-nose platforms and four classes of ignitable liquids, namely gasoline, diesel, alcohol, and diluent. The results demonstrated an average classification accuracy of 87% with peak accuracy reaching as high as 97.3%. Furthermore, the Fréchet inception distance (FID) metric indicated significant distribution differences between e-nose platforms, with values ranging from 12.8 to 35.6. Overall, these findings suggest that the proposed method enhances the reliability and scalability of e-nose-based ignitable liquid detection, thereby contributing to more efficient forensic investigations and expanding potential applications in chemical sensing.