Affiliations: [a] School of Engineering, Sun Yat-Sen University, Guangzhou, Guangdong, China | [b] Guangdong Provincial Key Laboratory of Fire Science and Technology, Guangzhou, Guangdong, China
Correspondence:
[*]
Corresponding author: Shanjun Mo, Guangdong Provincial Key Laboratory of Fire Science and Technology, Guangzhou 510006, Guangdong, China. E-mail:[email protected]
Abstract: Various methods have been explored to differentiate these electric molten
marks including primary molten marks (PMMs, the cause of fires) and
secondary molten marks (SMMs, were caused by the fires). However, only a
small number of specimens were tested by related methods and none of these
methods are reproducible. In order to build an objective and developmental
method for analyzing electric molten marks, many molten marks were made
under different current and heat treatment temperature and time. And some
descriptors were extracted from metallographic images of primary molten
marks (PMMs) and secondary molten marks (SMMs) with help of digital image
processing method. For purpose of identifying the type of electric molten
mark, some judgment models were established with the aid of support vector
machine (SVM). There were 380 molten marks from the experiment, 288 of them
were PMM and 92 of them were SMM. All of them were used in training model, 5
PMMs and 5 SMMs were used to test these models. Then, these judgment models
were selected and optimized by different methods. Thereby, an optimal model
was found, the classification accuracy of this model was above 66.2%.
Finally, this model was tested by actual fire for the forensic application,
the accuracy was 100%.
Keywords: Fire investigation, molten mark, metallographic image, support vector machine (SVM), genetic algorithm (GA), particle swarm optimization (PSO)
