Detection of imrosived explosives devices using electronic nose

The face of terrorism has changed with advancing technology. Modern communication, transportation facilities, high-powered weaponry, and other technological developments have significantly increased the destructive capacity of today's terrorist. On one hand, the means employed by terrorists have become more devastating compared to the past; on the other hand, their ability to effectively utilize these means has also increased in parallel. The fact that our country has been subjected to terrorist attacks for many years has been the main motivation for law enforcement agencies to enhance their abilities to prevent such attacks before they occur. Especially in locations where large crowds gather, such as airports, metro stations, train stations, bus terminals, parks, gardens, and shopping malls, preventive searches are conducted. However, these searches are mostly performed manually by hand and visual inspection, supported by metal detectors and X-ray devices. Law enforcement also employs specially trained bomb detection dogs to identify explosives; however, the possibility that dogs may make mistakes or mislead their handlers during searches is always present. In this thesis, an electronic nose was developed using MQ-brand gas sensors for the detection of explosive substances, and the odors of explosives were studied. Odor samples were collected from both individual and mixed quantities of ANFO (Ammonium Nitrate Fuel Oil), a type of homemade explosive, and various spices including mint, thyme, cumin, and coffee. The data were collected using a custom software developed in LabVIEW. For each sensor, 10 data points per second were recorded over a period of 60 seconds, resulting in a total of 426 measurements. The non-explosive class included 161 samples, consisting of 40 thyme, 40 mint, 40 cumin, and 41 coffee records. The explosive classes consisted of mixtures containing 10% ANFO + 90% spice (41 samples), 20% ANFO + 80% spice (41 samples), 30% ANFO + 70% spice (41 samples), 40% ANFO + 60% spice (42 samples), and 100% ANFO (100 samples). From the collected data, statistical features such as mean, standard deviation, sum, and median were extracted. These features were then divided into three sets: training (60%), validation (20%), and testing (20%), and were fed into classification models. The classification was performed using kNN-3, logistic regression, random forest, and extra trees algorithms. Based on the experimental results, the extra trees algorithm achieved the highest accuracy rate with 96.45%. Both random forest (93.61%) and kNN-3 (93.67%) also produced high accuracy rates, while the logistic regression model, with 85.65%, showed relatively lower performance compared to the other methods.