Automated classification of ignitable liquid profiles: Comparing machine learning and rule-based systems

In this work, two methods were developed for the automated classification of neat ignitable liquids and ignitable liquid residues. The performances of the two approaches were evaluated using laboratory-generated gas chromatography-mass spectrometry (GC-MS) data of various types of samples of ASTM E1618-19 classes. One method used machine learning with the Orange data mining software via image analysis for pattern recognition/comparison to known reference samples for the classification procedure. Five machine learning models, namely neural network, support vector machine, logistic regression, random forest, and k-nearest neighbors, were evaluated, and moderate classification accuracy (0.706-0.635), F1 score (0.703-0.626), precision (0.734-0.646), recall (0.706-0.635), and Matthews Correlation Coefficient (0.675-0.595) were achieved. The second method used a rule-based Python framework to automate the routine steps used by fire debris analysts for the interpretation of the GC-MS profile via identification of the major compounds (based on their mass spectra and retention times in the GC-MS chromatogram) in the sample, followed by classification using logic rules. The Python framework approach attained a higher classification accuracy (0.922), F1 score (0.924), precision (0.935), recall (0.922), and Matthews Correlation Coefficient (0.913). The two methods enabled the automated classification of ignitable liquid profiles and could serve as a preliminary assessment tool or an additional verification of the fire debris analysts' interpretations and conclusions of the analysis.