Risk Asessment Of Coal Mine Gas Outburst Based Upon Small Sample

Coal industry which has extremely broad development prospects will remain significant in China's energy industry for quite a long time. However, current safety situation of mine production remains very serious because of frequent occurring of coal mine safety accidents. And there is very big disparity between western developed countries and China in coal mine safety management level. Gas accident is claimed as "First Killer" in coal mine safety production, and gas outburst is the most severe in coal mine disasters. Many scholars and engineering technical personnel who conduct research on mine gas outburst mechanism, evaluation and control technology have currently proposed dozens of hypothesis on coal mine gas outburst mechanism and made some progress. Nevertheless gas safety problems haven't been solved for the reason of nonlinearity, diversity and complexity in mine gas system. Mine gas outburst risk assessment must be strengthened by theory circle to provide coal enterprises with scientific and effective prevention means and control methods; therefore conducting research on mine gas outburst risk assessment research has important theoretical and practical significance.Theory and method of small sample evaluation have frequently been seen in national defense, aerospace, nuclear equipment and other high-tech fields. In these areas, expensive test cost makes the number of trials as little as possible which results in relatively limited sample data. Digging wider information under the circumstance of limited data is the problem need to be solved in small sample method. There is typically insufficient problem of negative-class data in mine gas security system, and negative-class samples, such as coal mine gas outburst, are infrequent events, so collecting sample data immediately on the accident scene is very difficult. Thus negative-class samples must be relatively small when compared with positive-class samples. Additionally, new-established coal mines lack historical data of gas outburst, therefore, small sample should be considered when evaluating with very limited samples. Moreover, sample acquisition will bring high risk and high costs which is due to extreme danger in mine gas safety data collection. According to above analysis, multiple-factors determine that scientific evaluation on mine gas outburst risk should base on small samples which hasn't been fully taken into consideration in traditional evaluation method and rarely involved in relative research. Apparently theory circle failed to propound effective methods to solve this problem even though evaluation based on small sample exists widely in reality.Mine gas outburst risk assessment method based on small samples proposed in this paper belongs to "Scenario Embedded Type" which possesses distinctive industry characteristics and focuses on combination of theory and practice. Research work of this paper starts from following aspects:(1) Theory and methods applied for small sample evaluation are compared in this paper. Similarities and differences among various evaluation methods and existing deficiencies of various methods in theory and application have been analyzed based on description of mathematical principle for various evaluation methods, subsequently situation and development direction of various evaluation theories and methods have been studied and indicated.(2) Mine gas outburst risk assessment index is chosen in accordance with gray relational analysis method in this paper. Grey system theory is suitable for solving the problem of small sample with poor information, and gray correlation analysis is one of the main contents of grey system theory. Various influencing factors of gas outburst are analyzed by applying grey relational analysis method. Correlation coefficient and correlation degree of every influencing factor are sorted by value of correlation degree in order to find main control factors which will be used as small sample evaluation indexes. Then measured data of mine for main controlling factors will be input into small sample evaluation model to assess risk.(3) Mine gas outburst risk assessment is established based on neural network which is data-driven "Black Box" model with a strong non-linear processing capacity and suitable for solving coal mine gas outburst system evaluation with property of complexity and nonlinearity. Impact of subjective factors can be effectively reduced by using neural network to assess risk of mine gas outburst. And established evaluation model can basically portray true complex nonlinear relationship between impact of mine gas outburst and actual level of coal mine gas outburst risk. So this research model has certain application value.(4) Normalizing method for training samples is improved in this paper. Improved standardization method has excellent properties of sequence keeping, difference ratio invariance, translation irrelevance, zooming irrelevance, interval stability and so on. Comprehensive evaluation is not feasible because of non-comparability of each training samples index. Improved normalizing method for training samples has been proposed to put forward specific improvement formula for training samples in neural network and support vector machine small sample evaluation method in this paper. As for neural network evaluation model, improved method avoids Stype activation function taking extreme value, margin with enough large connection weights and harsh conditions which enhances training speed. As for evaluation model of support vector machines, imbalanced data may cause dyscalculia for both inner product operation and exp operation used in kernel calculation. And improved method could preserve relationship between raw data and eliminate dimension influence and effect of variable mutation scale and numerical value, which could effectively reduce computational difficulty.(5) A new method which comprises adjusting step length grid search and K-CV cross-validation has been selected from kernel function of support vector machine parameter. In this method, values range of parameter a and Gauss RBF kernel function punish parameter C is chosen to set a large search step length to train each parameter assembly (σ,C) and calculate the accuracy of risk assessment. Accuracy of risk assessment is calculated by K-CV cross-validation method which can effectively avoid excessive learning and insufficient learning and guarantee that obtained result is also more persuasive. A area with good evaluation performance should be firstly located in the large area and then fractionized into grids, subsequently step length need be adjusted and grid search method will be applied to find the optimal parameters assembly which should be ordered by evaluation accuracy of K-CV cross-validation method to choose the parameters assembly with the highest evaluation accuracy as the optimal parameters for the model. Two parameters are considered to establish mine gas outburst risk evaluation model which has the best evaluation ability based on support vector machine.(6) C-SVM support vector machine model for gas outburst risk assessment which is good at solving problem of small sample is established in this paper. C-SVM support vector machine method can not only solve evaluation problem of complex nonlinear system, but also solve classification, function approximation and pattern recognition problems of small sample, which is the best research tools and methods for mine gas outburst risk assessment. Support vector machine model for gas outburst risk assessment established on basis of small samples prove to possess good generalization ability when researched by actual measured data in coal mine and can be used in actual mine gas outburst risk evaluation for conducting mine outburst prevention effectively and timely.(7) Empirical comparative study has been conducted on various mine gas outburst risk evaluation method in this paper and conclusion is drawn as follows: evaluation results of single index method and comprehensive evaluation index D, K method is not entirely consistent with actual results, because these two methods only emphasize impact of individual factors on mine gas outburst and aren't able to comprehensively and systematically characterize nonlinearity of mine gas outburst system with regard to multiple factors as well as grasp the essence of mine gas outburst risk assessment. Evaluation accuracy of C-SVM support vector machine is slightly higher than that of neural network which indicates support vector machine model shows higher generalization ability. Especially in the condition of small sample, C-SVM support vector machine still have high evaluation accuracy. However, generalization ability of neural network for establishing evaluation model apparently decreases under the circumstance of training small sample. It demonstrates:there is obvious advantage of support vector machine in resolving the small sample evaluation when compared with other models, and it is suitable for evaluating risk. Support vector machine which is based on the small sample evaluation can not only solve the problem of typical insufficiency of negative-class data, but also reduce risk of data gathering and cost of evaluation. So it is a small sample evaluation model or method of high practical value.In addition to above research results, some frontier issues worthy of further study in this field have also been pointed out in this paper.