A Semi-supervised Learning And Numerical Simulation Forecast System In Coal Mine Floor Water Inrush

Frequent mine water hazards make inestimable casualties and property losses for enterprises and our country. Existing prediction and forecast techniques are not accurate enough, which is the main cause of frequent water inrush accidents in coal mine. Therefore, there is a necessity to study effective mine water inrush prediction method. Moreover, once the water hazard happened, due to the uncertainty and complexity of the underground, it brings great blindness and challenging to rescue effort. Numerical simulation of water inrush process can not only show the evolution trend and influence area of water disaster macroscopically, but also make an important guide role in rescue decision-making. In addition, because of the limited enterprise informatization level, plenty of data resource cannot use efficiently in the secondary relief efforts due to the lack of systematic management. Therefore, it is a necessity to build a coal mine floor water inrush forecast system.The dissertation mainly aims at studying the subject of mine underground water. Three research topics are included: prediction technique in mine water inrush, numerical simulation of mine water inrush process, and coal mine floor water inrush forecast system.In the view of mine water inrush prediction, there are two problems. On one hand, current application of water inrush prediction methods mainly based on subjective analysis of specific conditions by technicians, which has a great deal of uncertainty. To overcome this shortage, methods and theories in intelligent information processing are employed to study water inrush prediction. On the other hand, although there already are some water inrush prediction methods that based on intelligent machine learning, existing methods are mostly supervised learning. Supervised learning methods have the limitation that the trained model, which studied with small labeled data, always has poor generalization ability. However collecting large labeled samples in water inrush accident is technically difficult. Therefore, on the premise of small labeled sample learning, the dissertation study mine water inrush prediction problem in a semi-supervised way.In the view of numerical simulation for water inrush process, the theory of numerical simulation in fluid dynamics is combined to analyze the main drive of mine water out bursting. Also water flow rules and factors influencing water spread after mine water inrush are summarized from the perspective of fluid dynamics theory. The mine tunnel is modelling by UG 6.0, and Flow-3D is employed to numerically simulate the process of water inrush.In the aspect of implementation of informationalized management, a coal mine floor water inrush forecast system is built according to the reality demands. The system include information management subsystem, water inrush forecast subsystem and damage range analysis subsystem. Subsystems collaborate together to achieve informationalized and intelligence management in mine water inrush forecast.The main works are summarized as followed:(1) The formation of coal mine floor water inrush and water inrush channel are analyzed first. Key factors affecting coal seam floor water-irruption from the perspective of mine water inrush mechanism are then summarized. Main drive force of water inrush, the evolution of water inrush, and influence factors during spreading are expounded in dynamics way at last.(2) The gravitational force based semi-supervised algorithm is proposed to solve small labeled samples in coal seam floor water inrush prediction problem. The proposed method make full use of unlabeled samples during modelling, making up the shortages of poor generalization performance and unstable prediction results when small labeled samples are available. Compared with supervised methods, gravitational force based semi-supervised method has better performance in coal seam floor water inrush prediction.(3) The graph based semi-supervised ensemble algorithm is proposed to solve the unevenness prediction accuracy of existing prediction methods, which even do not work in actual application. The proposed method is a semi-supervised learning method, and also dealing with learning problem when small labeled samples being available. A graph is constructed in the algorithm, where unlabeled samples are vertices, and correlation between samples are edges. Results from variety of base predictors are ensemble, and attempted to be maximized. Experimental results in coal seam floor water inrush prediction show that graph based semi-supervised ensemble algorithm has satisfactory performance even while base predictors have poor prediction results.(4) In numerical simulation, theories of computational fluid dynamics are employed. Actual mine tunnel is modelling by UG 6.0, and Flow-3D software is used to numerically simulate according to rescue needs. To facilitate the simulation, the model of tunnel is divided into small grid, and boundaries are set. TrueVOF method is used to track the free surface, k-ε turbulence model is combined. By solving Navier-stokes equation, the mine water process numerical simulation is achieved.(5) A coal mine floor water inrush prediction system is built. Basic function modules are developed by using ArcGIS Engine component and c sharp language. The system can manage the water inrush data systematically. Water inrush prediction module is realized by combining with the Visual C#.net and MATLAB programming technology. While the water hazard influencing area analyze function is realized by combing data from numerical simulation and tunnel data stored in database.