الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 89 |  |  |
| | | from | 89 | | |
Abstract
Fire ignition can cause catastrophic consequences on human beings/properties and the surrounding environment. Using fire has been one of the significant factors that helped to evolve the human civilization. However, out of control fire can cause serious humanitarian, economic and environmental damages that vary based on its type (causing factor), location, scale, etc. Schools are places that include valuable lives and properties that need to be protected against fire ignition. Efficient firefighting in schools can save many lives and resources. However, there are not enough efforts that have been made in the literature to develop early fire detection systems for schools.
In this thesis, a novel method of a data mining-based fire detection and prediction is proposed. It can be used to detect fire sources in different locations of Kuwaiti schools listed in the Education ministry. The proposed method framework consists of four main processes, namely, data acquisition, data preprocessing, feature analysis and selection, and classification/regression. The data acquisition is made with the help of the General Civil Defense Department of Kuwait. In the data preprocessing process, a set of operations are performed on the collected data including discretization and categorization. In the feature selection process, two feature selection techniques are used, namely Information Gain (IG) and Principle Component Analysis (PCA). The proposed system performs the fire detection process using five classification models, including Decision Tree (DT), Linear Regression (LR), Linear Discriminant (LD), Support Vector Machine (SVM), and Deep Belief Network (DBN). In addition, the proposed system performs the fire prediction process using two regression models, including Decision Tree (DT) and Linear Regression (LR).
The proposed fire detection and prediction system’s performance is evaluated under different scenarios using a number of appropriate performance metrics. According to the utilized dataset, the proposed method of feature selection approved its superiority to eliminate irrelevant and redundant features. The best detection performance has been achieved using DT and the LR classifiers without applying the feature selection step and by applying the feature selection step using the IG, with 100% detection accuracy. On the other side, the SVM and the LD classifiers have the best detection performance by applying the feature reduction step using the PCA with 100% detection accuracy. On the other side, the DT based regression has achieved the best fire prediction with 0.0 Root Mean Squared Error (RMSE) and 0.0 Mean Absolute Error (MAE).