Smart portable electronic nose system in combination with machine learning algorithms for the intelligent discrimination of fire debris

Abstract

The identification and discrimination of ignitable liquid residues (ILRs) in fire debris are crucial in forensic research for determining the intentionality of a fire. This study presents a new method using a portable sensor-based electronic nose (eNose) combined with machine learning (ML) algorithms for automated ILR classification. Six substrates (vinyl, nylon, linoleum, polyester, wood, and cotton) were burned with different ignitable liquids (gasoline, diesel, ethanol, and charcoal starter with kerosene), and samples were collected at intervals from 0 to 48 h after the fire had extinguished. Sensor responses from multiple sensors (SO2, H2S, CO, IRR, NO2, TBM, NH3, and ethanol) were collected over a duration of 140 s. The data were preprocessed using the first derivative and Savitsky-Golay filter, followed by low-level data fusion. A variable selection using the Boruta algorithm was applied, and both reduced and non-reduced matrices were used to train ML models. For detecting the presence of ILRs, random forest (RF) and support vector machine (SVM) models achieved 100 % accuracy. For discriminating between ILR types, the best performance was achieved by the RF model using the reduced matrix, correctly classifying 94.44 % of the samples. Only four sensors (SO2, H2S, CO, IRR) were necessary, indicating the potential for an optimized eNose design. This method offers advantages over traditional techniques, such as faster analysis, lower cost, and greater portability. Additionally, a web application was developed to allow users to automatically characterize fire debris using the algorithms.