Author: Dzulkiflee Ismail

Publisher:

Published: 2012

Total Pages: 0

ISBN-13:

Pattern recognition is a term that can be used to cover various stages of the investigation of characterising data sets including contributing to problem formulation and data collection through to discrimination, assessment and interpretation of results. Chemometrics techniques and Artificial Neural Networks (ANNs) are pattern recognition techniques commonly used to visualise and gather useful information from multidimensional datasets i.e. datasets with n-samples with m- variables. Of the many chemometric techniques available, Principal Component Analysis (PCA) and Hierarchical Cluster Analysis (HCA) are the most commonly used in the evaluation of dataset(s) generated from the analysis of samples which have relevance to forensic science. By contrast, Artificial Neural Networks (ANNs) and in particular Self Organising Feature Maps (SOFM) and Multi Layer Perceptron (MLP) have had limited application in forensic science eventhough these pattern recognition techniques have been known for almost 30 years. This study focuses on the applicability of the Artificial Neural Networks (ANNs) to specific datasets of forensic science interest and compares these with 'conventional' PCA and HCA techniques. Datasets generated from the analysis of wax based products and lighter fuels were used. The wax based product data set contained information obtained from Thin Layer Chromatography (TLC), Microspectrophotometry (MSP), Ultra-Violet and Visible Spectroscopy (UV/Vis) and Gas Chromatography with Flame Ionisation Detector (GC-FID) analysis of a variety of products from multiple sources where discrimination by brand was the objective. The data provided for the lighter fuel samples was obtained from analysis of a number of brands, both unevaporated and evaporated by Gas Chromatography-Mass Spectroscopy (GC-MS) and the objective was to discriminate the samples by brand as well as link degraded samples from the same brand together. The wax based product analysis provided simple, straight forward data whilst the lighter fuel analysis provided a more complicated and challenging dataset to investigate in terms of facilitating sample discrimination and/or linkage. In all cases, the 'conventional' Principal Component Analysis (PCA) and Hierarchical Cluster Analysis (HCA) failed to provide any meaningful discrimination of the samples by product type regardless of the nature of the datasets. In contrast, the Artificial Neural Networks (ANNs) techniques provided full discrimination of the samples by product type even when the samples had undergone considerable ageing and weathering. This work has demonstrated the potential use of Self Organising Feature Maps (SOFM) and Multi Layer Perceptron (MLP) to datasets of forensic science relevance. The findings of this work provide avenues for further exploration of Artificial Neural Networks (ANNs) in forensic science.