Research On Non-coal Mine Production Accident Safety Early Warning Technology Based On Support Vector Machine

At present,the safety production situation of China's non-coal mining is still severe,which seriously restricts the sustainable development of enterprises,therefore,it is of great practical significance and theoretical value to carry out the research of mining production safety accident risk early warning technology.Aiming at the difficulties in selecting and quantifying the indicators of production safety accidents in non-coal mining and the difficulties in constructing the early warning model of production safety accidents in mines,the thesis conducts theoretical analysis and numerical calculation research.A time series prediction model of mining accidents based on wavelet transform is proposed.Firstly,the training data is applied to the wavelet transform,and the time series is gradually multi-scale refined by the telescoping translation operation;then,the low-frequency signal series is predicted by the autoregressive sliding average(ARMA)model,and the high-frequency signal series is predicted by the fractal theory(FT);finally,the prediction results of each sub-series are superimposed to build the wavelet transform-based mining accident time series prediction model.The results show that the average relative error of the test sample prediction results is 15.13%,and the relative error fluctuates relatively smoothly.The gray correlation model was used to identify the top nine indicators with the highest correlation with the fatality rate of one million tons in non-coal mines,and improvements were proposed for three shortcomings of the model.The absolute difference of the data to be compared is replaced by the absolute difference of the variation,which solves the shortcoming of not being able to filter out the negatively correlated comparisons;finding a threshold value to "reward" or "punish" the correlation degree through parameter search,which solves the shortcoming of only rewarding and punishing when initializing.The shortcomings of the initialization process,which only rewarded and punished,were solved by introducing the weight value in the calculation of correlation degree,and solving the shortcomings of the default weight of each influence factor in the calculation of correlation degree.The results show that before the model improvement,the average value of the extreme difference of the 14 initial indexes correlation is 28%,and after the model improvement,the average value of the extreme difference is 38%,which expands the separation degree.The least squares support vector machine(LSSVM)prediction model was applied to the prediction of the million ton mortality rate of non-coal mines,and the kernel parameters of the LSSVM model were optimized by using genetic algorithm(GA),particle swarm algorithm(PSO)and grid search algorithm(GS)to reduce the subjectivity of kernel parameter selection.The computational results show that the PSO algorithm has the smallest mean absolute error and root mean square error for the prediction of the training samples,which are 0.0401 and 0.0154,respectively.the PSO algorithm is finally chosen as the parameter search method for this model,and the mean relative error of the support vector machine model based on the PSO optimization algorithm is 0.183251 for the 13 test samples.