RESUMO
The recognition of ignitable liquid (IL) residues in fire debris is a resource intensive but key part of an arson investigation. Due to the highly diverse and heavily loaded chemical matrix of fire debris samples, combined with the broad chemical composition of IL, the interpretation of the laboratory analysis results is a very challenging task for the forensic examiner. Fire debris samples are commonly analyzed using gas chromatography coupled to mass spectrometry (GC-MS). This method delivers both the total ion chromatogram (TIC) with the individually separated compounds and the underlying mass spectrum of each of the separated compounds. In this study, a completely new approach for the recognition of gasoline in fire debris samples is presented. First, the GC-MS data, including retention time, signal intensity, and mass spectrum is converted into a bitmap image. Five different data-to-image conversion approaches are tested, and their advantages and limitations are discussed. Subsequently, a convolutional neural network (CNN) is utilized to allocate the generated images to the classes "with gasoline" or "without gasoline". The applied approaches to generate a digital image and the pattern recognition of the CNN perform very well in the classification of unknown test samples. Depending on the data-to-image generation approach used, the rate of correct sample classification in the test dataset is between 95% and 98%. The machine learning approach in this study, as well as the complementary method presented in an accompanying article, are not only useful for the recognition of gasoline in fire debris but are equally applicable to any additional areas in which the interpretation of complex chromatographic and mass spectrometric is required.
RESUMO
The detection and identification of ignitable liquid (IL) residues in fire debris are two very challenging tasks in a fire investigation. To this day, the recognition of IL in fire debris includes the chemical analysis of the fire debris composition, followed by the examination and interpretation of the analysis result by a trained forensic examiner. Throughout the last decade, chemometrics and artificial intelligence have become increasingly important. In the present study, machine learning algorithms capable of recognizing gasoline residues in fire debris based on GC-MS data have been developed. Four methods, including random forest, gradient boosting, support vector machine, and naïve bayes are applied and used to classify fire debris samples into the two categories "with gasoline" or "without gasoline". A fifth method (logistic regression) did not converge due to well separated classes. A database comprising 360 measurements, including fire debris samples of real cases as well as fire debris samples spiked with known amounts of weathered gasoline (up to 99.6%), was available to train the machine learning algorithms (using 85% of the data) and to subsequently test the performance of the methods when classifying unknown samples (using 15% of the data). In general, the methods perform very well, as three of it succeeded to classify all test samples correctly without any false positive or false negative allocations. One (naïve bayes) was not trained enough to classify other (non-gasoline) IL correctly as "no gasoline". Furthermore, the random forest method reveals which chemical compounds are most relevant for the algorithm to classify the samples. In general, the presented approach is highly promising and could easily be extended or adapted to other types of IL. Similar to the neural network presented in the accompanying paper, such methods have the potential to serve as a fast screening technique for fire debris samples, thus supporting the forensic examiner by providing an additional independent opinion. Nonetheless, the definite identification of IL residues in fire debris always has to be accomplished by a forensic examiner.
