Semi-supervised Learning And Its Application To Coal Mine Gas Safety Information Processing

In recent years, coal mine safety production accidents happened frequently.Especially with the increasing mining depth, mining conditions became furthercomplicated. The coal mine gas accidents has become the constraints of sustainabledevelopment of the coal industry. The foundation and management of existingsecurity technology has been difficult to adapt to the urgent need for the efficientproduction of the current mine safety. Therefore, How to reduce and prevent theoccurrence of gas safety incidents, systematic investigation of the hidden mine gasaccident and carry out the analysis and prediction of the evolutionary mechanisms forcoal mine gas risk, get rid of the unsafe situation caused by gas incidents, is a topicworthy of further study. The thesis is based on the National Natural ScienceFoundation of China, and under the application background of mine specific projects,and study the Semi-Supervised Learning and its application to coal mine gas safetyinformation processing. The studies mainly include the following four aspects:1. According to the problem of inefficiencies of intelligent algorithm caused bythe high dimensionality of the coal mine gas safety data, this thesis presents a newway to define the marginal points as well as a discrimination criterion. Specifically, apenalty graph is designed on the marginal points and their nearest points fromdifferent class while an intrinsic graph on the non-marginal points and their farthestpoints from same class. Based on a new discrimination criterion, a non-parameterdimensionality reduction method was proposed. By this means, it can solve theproblem of parameters chosen in traditional graph based dimensionality reductionalgorithms. Experiment on several benchmark datasets demonstrated the effectivenessof our proposed method, and appreciate performance was get while achieving a lowercomplexity.2. With the construction of the information technology of coal mine, a largeamount of data continue to appear in the information technology platform. Usingunsupervised learning algorithm to process required experts spend a lot of energy todo the labeling work. According to this problem, this thesis proposes a method calledlocality preserving semi-supervised dimensionality reduction. It takes both thediscrimination information and geometry structure into account. Specifically, abetween-class graph is constructed on labeled samples and a nearest neighbor graphboth from the perspective of locality. A directly mapping can be achieved by solving a generalized eigenvalue problem. Effectiveness of the proposed method is achievedeven with little labeled samples.3. Gas time series prediction is an effective means for risk evaluation of the coalmine safety. Coal mine gas time series data having a high temporal correlation,however, the existing multivariate time series of semi-supervised regression algorithmonly consider the information of the spatial relationships and ignore the temporalinformation between samples. According to this problem, this thesis proposes asemi-supervised regression algorithm which took account of the temporal informationof samples. For time series an assumption of temporal smooth was proposed, andbased on this assumption, a regularization item that could reflect more underlineinformation of samples was constructed. Semi-supervised regression under manifoldregularization framework using the proposed regularization item was carried out. Theexperiments results show that, the proposed algorithm which uses temporal andspatial information of samples simultaneously could achieves better performance thanthe algorithms which only considers the spatial information of samples.4. From the point of view of the pattern classification, using coal mine gas datato classify the risk of coal mine safety could provide a basis for avoiding coal minegas explosion. The majority of existing semi-supervised classification methods onlybased on one of two assumptions, respectively clustering assumption and manifoldassumption. It is obviously the performance of algorithm will drop when theassumption is not suitable. According to this problem, this thesis proposes a methodcalled semi-supervised kernel based progressive SVM which combines both thecluster assumption and the manifold assumption. Specifically, a semi-supervisedkernel which reflect manifold information of the samples was constructed by warpingthe Reproducing Kernel Hilbert Space, and then the semi-supervised kernel was usedin SVM which was based on cluster assumption, finally, a progressive learningprocedure was used in the proposed method. Based on both two assumptions, theproposed algorithm could avoid the worse performance caused by one unsuitableassumption.This thesis mainly studied three aspects, respectively coal mine gas datadimensionality reduction, coal mine gas time series regression and coal mine gas dataclassification. The methods all use semi-supervised learning paradigm. Analysis andimprovement for algorithms were carried on based on the researches on relatedprevious work. New algorithms were proposed and experiments were carried on public datasets often used in the fields of face recognition and data mining. Throughcomparing the performance with related algorithms, the effectiveness of proposedalgorithms was verified. Finally, the proposed methods were applied in coal minesafety information processing. The application results show that the methods proposedby this thesis achieved a good performance in coal mine safety information processing,so it has a certain significance to improve coal mine production safety.