Research And Application On Early-warning System Of Coal Spontaneous Combustion Based On RS-SVM

Coal mine fire is one of the most major accidents, which impacts production safety seriously. It can disturb the schedule of production, or result in catastrophe. Coal spontaneous combustion is the main reason for coal mine fire. If the coal spontaneous combustion can be forecasted earlier, it will be able to reduce the economic losses and protect the safety of coal mine workers, which is a great significance to the development of coal mines.There are many methods to predict the coal spontaneous combustion. In this paper, the gas released from process of oxidation is used to analyze the risk of the coal spontaneous combustion. Coal mine will release a variety of gas in the process of oxidation, and there is a complex nonlinear relationship between the concentration of gas and the temperature of coal, which is hard to describe. Besides, some gas can not reflect the characteristic of coal spontaneous combustion. In order to solve the problem, a model based on Rough Set(RS) and Support Vector Machine(SVM) is proposed. Attribute reduction is made for the gas sample analysis and decision table to delete the redundant information and determine the index gas. And then the parameters for SVM are optimized with the decision table after reduction. At last, the index gas is checked from a special experiment and do the training and testing for RS-SVM with the experimental data.The experimental result shows that compared with the real SVM and the BP neural network, the RS-SVM has a higher accuracy rate and it can provide the basis for the decision-making for spontaneous combustion of coal.The main contents of this paper are as follows:1. The algorithm of Grid Searching, Particle Swarm Optimization and Genetic Algorithm are used to optimize the parameters of SVM to improve the classification performance.2. A real experience is made to simulate the process of self-heating of coal mine and collect and analyze the data to determine the index gas. The result is consistent with the result of attribute reduction The training and testing for RS-SVM is done with the experimental data.3. Different prediction algorithms are used to estimate the correct rate which provides the factual basis for RS-SVM model.