Research On Identification Of Coal Mine Water Inrush Sources Based On Improved Model Of Extreme Learning Machine

The identification of coal mine water sources runs through early prediction and later treatment of water inrush prevention.In the early stage of source prevention,it is necessary to identify water sources accurately and quickly,so as to play an early warning role for water inrush prevention and control.The experimental methods and linear classification models of traditional water sources identification can no longer meet the rapid and accurate requirements.This thesis mainly focuses on the application of water sources identification in coal mine,and establishes the models suitable for practical application.With the research background of coal mine water sources identification of water sources,the thesis carries out the research of water sources identification model,focusing on water sources rapid identification,anti-interference and semi-supervision identification,multi-function integration identification and online identification.It mainly optimizes from several aspects of model,such as parameters optimization,structure design and training mode.The main research results of this thesis are summarized as follows:It studies the problem of accurate and rapid identification of water sources in coal mine.Laser induced fluorescence(LIF)technology is used as a new experimental method to obtain the fluorescence spectra of different aquifers in mine.Taking the fluorescence spectra of water sources as the research object,the spectral features are extracted to identify different water sources.It establishes nonlinear multivariate classification model by using the pattern recognition method in machine learning.It proposes the identification method of water source based on extreme learning machine(ELM).And it establishes the ELM rapid identification model of water sources,with determining the key parameters in model such as the number of hidden layer nodes and excitation function.As a new type of nonlinear feedforward neural network,ELM can meet the requirement of accuracy with infinite approximation capability,the requirement of detection speed with extreme running capability,the requirement of model operability in practical application with general learning performance.It studies the problem of anti-interference and semi-supervised learning about identification model of water sources in complex environment.When the basic ELM algorithm quickly trains the model and identifies the samples,it has performance fluctuation.Therefore,the performance of the model can be stabilized by optimizing the network parameters.With regularization optimizes the basic ELM model,so that the regularization ELM(RELM)model can deal with interference data and unlabeled data,and has the ability of anti-interference and semi-supervised learning.It proposes the identification model(L2-RELM)of water sources based on L2 norm regularization with equality constraint.L2-RELM model searches for the appropriate regularization parameter C through cross-validation,and C as penalty coefficient to minimize training error.The experiment shows that L2-RELM model can not only avoid overfitting and stabilize model performance,but also avoid the deviation of the external environment disturbance to identify water sources and enhance the anti-interference ability of the model.It proposes the identification model(GM-RELM)based on graph manifold.Based on the manifold assumption that the unidentified category samples is similar to the identified category samples,it constructs a Laplacian adjacency graph that reflects the similarity relationship among samples.By minimizing the energy function on graph,the classification function that satisfies the global consistency assumption is obtained.The experiment shows that GM-RELM uses a large number of unidentified category samples to assist the training of identification model,realizes semi-supervised classification,and improves the generalization ability of the network.It studies the problem of multi-function integration in water sources identification model.Mainly optimizing from the network structure design of model,it proposes the identification model(M-RELM)based on multi-hidden-layer integration.The data preprocessing and classification learning are integrated into a unified regularization model by setting up multi-hidden layers.In the preprocessing stage,it proposes to improve the auto encoder(AE)network with L2-RELM algorithm,so that performs the nonlinear feature extraction of water sources spectra layer by layer using the depth learning method.In the classification learning stage,the supervised classification based on L2-RELM and the semi-supervised classification based on GM-RELM are carried out according to whether the samples have category label or not.The integration of nonlinear feature extraction with AE-RELM and classification learning with L2-RELM or GM-RELM,forms the ML2-RELM or MGM-RELM model respectively and completes the multi-hidden-layer structure design of M-RELM model.M-RELM model completes the parameters transfer among multi-hidden layers and implements the dissemination of the learning function step by step.The experiment shows that M-RELM model optimizes the network structure design,expands multi-function learning ability of the model,adapts to the nonlinear identification of water sources in coal mine.It studies the flexibility and stability of online water sources identification.Mainly optimizing from the mode of training network,it proposes the online sequential regularization(OS-RELM)identification model.OS-RELM model changes the batch training mode with the sequential learning mode,so that the model has the ability of online identification of water sources.It optimizes the basic OS-ELM algorithm with L2 norm regularization,and abandons the restriction the samples block,so that it can deal with random size samples flexibly,and output the training results while learning.The experiment shows that OS-RELM model realizes the stability and flexibility of online identification of water sources.The thesis proposes some improvement measures for the traditional method of water sources identification and constructs a nonlinear multivariate classification model with ELM and improved algorithm.From the constraints of model such as the performance fluctuation,function singleness and constraint of training mode,the model is optimized by combining with the requirements of water inrush sources identification.The optimized model has the characteristics of anti-interference,function extensibility and online identification.Using the research method of “theory analysis,numerical simulation and experimental test”,it analyses and constructs the model from the actual application requirements of water sources identification in mine.On the standard dataset,it verifies the model's validity,further extends to the application of spectral data of water sources and assists the prevention and control of water inrush disaster in coal mine.