CN103745093A - Principal component analysis-extreme learning machine (PCA-ELM) based coal mine water inrush prediction method - Google Patents

Abstract

The invention relates to a PCA-ELM based coal mine water inrush prediction method and belongs to the field of coal mine water inrush prediction methods. The method comprises the steps of obtaining a plurality of data affecting coal mine water inrush in the coal mine normal exploitation running state; screening a plurality of factors affecting coal mine water inrush through the PCA to obtain a main controlling factor playing a decisive role in coal mine water inrush; partitioning sample data only containing the main controlling factor into a training set, a verification set and a testing set which are corresponding to model training, verification and testing respectively; building an ELM network model, and training the data through the ELM algorithm; verifying a coal mine water inrush prediction model through the verification data, beginning to rebuild a model from the PCA if the obtaining prediction result has no apparent advantages compared with other algorithms, and using the model as the prediction model for actually predicting the coal mine water inrush prediction condition if the prediction result is ideal.

Description

A kind of mine water inrush Forecasting Methodology based on PCA-EML

Technical field

The present invention relates to a kind of mine water inrush Forecasting Methodology, particularly a kind of mine water inrush Forecasting Methodology based on PCA-EML.

Background technology

Mine water inrush is one of major hidden danger in Safety of Coal Mine Production, correctly predicts in time gushing water, to ensureing having great importance smoothly of coal mining.

The factor that affects mine water inrush is a lot, and mostly these factors are ambiguity and fuzzy similarity between each influence factor, have complicated nonlinear relationship, are difficult to set up forecast model by classical mathematical theory.Based on this, experts and scholars have proposed the method for prediction water bursting in mine, there is the BP of employing algorithm to set up gushing water forecast neural network model based on water bursting in mine sample instance, mine water inrush Forecasting Methodology based on genetic neural network and water bursting in mine information being processed with support vector machine-rough set SVM-RS model, but the training speed of BP neural network is slow, easily be absorbed in local minimum point, SVM needs artificially to arrange kernel function in learning process, the parameters such as penalty coefficient, and need to consume a large amount of time carries out parameter adjustment, the mine water inrush Forecasting Methodology of extreme learning machine (ELM) algorithm in conjunction with principal component analysis (PCA) (PCA) therefore proposed.

ELM(extreme learning machine) algorithm is a kind of single hidden layer feedforward neural network learning algorithm, this algorithm does not need the biasing of the input weights memory hidden layer of adjusting network in training process, the hidden layer node number that network only need be set, just can produce unique optimum solution.Compared with traditional algorithm, this algorithm has that parameter is selected easily, pace of learning is fast and the advantage such as Generalization Capability is good.

PCA(principal component analysis (PCA)) method is the multivariate statistics technology of a kind of data compression and feature extraction, can effectively remove the correlativity between data, reduces the complexity of calculating.The influence factor of mine water inrush during as network input variable, comprises the information being relative to each other, certainly due to the not independence between network input variable in these influence factors, may cause information overlap, and then the complexity of increase network, reduce network performance, impact prediction precision.

Summary of the invention

The present invention seeks to provide the mine water inrush Forecasting Methodology based on PCA-EML that a kind of parameter is selected easily, pace of learning is fast and Generalization Capability is good.

To achieve these goals, the present invention is by the following technical solutions:

By principal component analysis (PCA), carry out the input parameter of optimization neural network, first utilize principal component analysis (PCA) to carry out pre-service to many factors data, eliminate the information overlap between legacy data, produce new separate training sample, retain as much as possible original information, then the input using the training sample of reconstruct as extreme learning machine, the structure complexity of reduction neural network, improves speed of convergence.

Concrete step is as follows:

(1) obtain colliery and normally exploit the numerous data that affect mine water inrush under running status;

(2) utilize principal component analysis (PCA) to screen the many factors that affects mine water inrush, obtain mine water inrush to rise the Dominated Factors of deciding factor;

The concrete steps of mine water inrush master control influence factor being screened by principal component analysis (PCA) are as follows:

1. get colliery and normally exploit the sampled data matrix under running status, and it is carried out to standardization obtain matrix Χ;

2. according to the data matrix Χ after standardization, calculate covariance matrix R;

3. according to covariance matrix R, obtain eigenwert, principal component contributor rate and accumulative total variance contribution ratio, determine major component number;

4. by major component number, determine load matrix W, W=XM, is the learning sample of the extreme learning machine model of foundation;

(3) sample data that only comprises Dominated Factors is divided into training set, checking collection and test set, corresponds respectively to training, checking and the test of model;

(4) set up extreme learning machine network model;

Set up extreme learning machine network model, and through extreme learning machine algorithm, these data are trained, its step is as follows:

1. given N training sample { x _i, t _i, i=1,2 ..., N; Initial hidden layer nodes is made as

excitation function g (x);

2. be input weights ω _iwith threshold value b _irandom assignment, wherein

3. according to the unified model formula of the SLFN of excitation function g (x) $\underset{i=1}{\overset{\tilde{N}}{\mathrm{\Σ}}}{\mathrm{\β}}_{i}g\left(x_j\right)=\underset{i=1}{\overset{\tilde{N}}{\mathrm{\Σ}}}{\mathrm{\β}}_{i}g(w_i\·x_j+b_i)=o_j,j=1,...,N,$ Be expressed in matrix as: H β=T,

$H({\mathrm{<figure-callout\; id="1"\; label="w"\; filenames="HDA0000446002590000011.png"\; state="\{\{state\}\}">w</figure-callout>}}_1,...,w_{\tilde{N}},{\mathrm{<figure-callout\; id="1"\; label="b"\; filenames="HDA0000446002590000011.png"\; state="\{\{state\}\}">b</figure-callout>}}_1,...,b_{\tilde{N}},x_1,...,x_{\tilde{N}})={\left|\begin{array}{ccc}g({\mathrm{<figure-callout\; id="1"\; label="w"\; filenames="HDA0000446002590000011.png"\; state="\{\{state\}\}">w</figure-callout>}}_1\&CenterDot;x_1+b_1)& ...& g(w_{\tilde{N}}\&CenterDot;x_1+b_{\tilde{N}})\\ .& & .\\ .& ...& .\\ .& & .\\ g({\mathrm{<figure-callout\; id="1"\; label="w"\; filenames="HDA0000446002590000011.png"\; state="\{\{state\}\}">w</figure-callout>}}_1\&CenterDot;x_N+b_1)& ...& g(w_{\tilde{N}}\&CenterDot;x_{\tilde{N}}+b_{\tilde{N}})\end{array}\right|}_{N\&times;\tilde{N}},$

$\mathrm{\&beta;}={\left|\begin{array}{c}{{\mathrm{\&beta;}}^{\mathrm{<figure-callout\; id="1"\; label="T"\; filenames="HDA0000446002590000011.png"\; state="\{\{state\}\}">T</figure-callout>}}}_1\\ .\\ .\\ .\\ {\mathrm{\&beta;}}_{\tilde{N}}^T\end{array}\right|}_{\tilde{N}\&times;m},T={\left|\begin{array}{c}{t}_1^T\\ .\\ .\\ .\\ t_N^T\end{array}\right|}_{N\&times;m}$ Calculate output matrix H; Wherein H is called the hidden layer output matrix of neural network; x _i=[x _i1, x _i2... x _in] ^t∈ R ⁿ, t _i=[t _i1, t _i2... t _im] ^t∈ R ^m, ω _ifor connecting the input weights of i hidden layer node; β _ifor connecting the output weights of i concealed nodes and output node; b _iit is the deviation of i concealed nodes; w _ix _jrepresent w _iand x _jinner product; Excitation function g (x) can be " Sigmiod ", " Sine " or RBF etc.

4. calculate output weights β: β=H ⁺y, wherein H ⁺for the Moore-Penrose generalized inverse of hidden layer output matrix H.

(5) utilize verification msg checking mine water inrush forecast model, if what obtain predicts the outcome and other algorithm comparisons, there is no clear superiority, from principal component analysis (PCA), start to re-establish model, comparatively desirable if predict the outcome, set it as forecast model and carry out actual prediction mine water inrush situation.

Beneficial effect, due to the mine water inrush Forecasting Methodology that has adopted principal component analysis (PCA) and extreme learning machine method to combine, principal component analysis (PCA) is predicted for mine water inrush with the extreme learning machine method method that combines, take mine water inrush historical data as sample, by principal component analysis (PCA), input data are compressed, reduced input variable dimension, remove mine water inrush is affected to the little factor, reduce to a certain extent the error that redundant data causes, improved model prediction precision; The mine water inrush forecast model of structure based on principal component analysis (PCA)-extreme learning machine, improved the independence between nerve network input parameter, reduce the complexity of network, improve network performance and precision of prediction, simultaneously, reduce the structure complexity of neural network, improve speed of convergence, met well the real-time of mine water inrush prediction.

Accompanying drawing explanation

Fig. 1 is the mine water inrush prediction process flow diagram based on PCA-ELM.

Embodiment

Embodiment 1: the input parameter that carrys out optimization neural network with principal component analysis (PCA), first utilize principal component analysis (PCA) to carry out pre-service to many factors data, eliminate the information overlap between legacy data, produce new separate training sample, retain as much as possible original information, then the input using the training sample of reconstruct as extreme learning machine, the structure complexity of reduction neural network, improves speed of convergence.

(1) obtain colliery and normally exploit the numerous data that affect mine water inrush under running status;

(2) utilize principal component analysis (PCA) to screen the many factors that affects mine water inrush, obtain mine water inrush to rise the Dominated Factors of deciding factor;

The concrete steps of mine water inrush master control influence factor being screened by principal component analysis (PCA) are as follows:

1. get colliery and normally exploit the sampled data matrix under running status, and it is carried out to standardization obtain matrix Χ;

2. according to the data matrix Χ after standardization, calculate covariance matrix R;

3. according to covariance matrix R, obtain eigenwert, principal component contributor rate and accumulative total variance contribution ratio, determine major component number;

4. by major component number, determine load matrix W, W=XM, is the learning sample of the extreme learning machine model of foundation.

(3) sample data that only comprises Dominated Factors is divided into training set, checking collection and test set, corresponds respectively to training, checking and the test of model;

(4) set up extreme learning machine network model;

Set up extreme learning machine network model, and through extreme learning machine algorithm, these data are trained, its step is as follows:

1. a given N training sample xi, ti}, i=1,2 ..., N; Initial hidden layer nodes is made as

excitation function g (x);

2. be input weights ω i and threshold value bi random assignment, wherein

3. according to formula $\underset{i=1}{\overset{\tilde{N}}{\mathrm{\&Sigma;}}}{\mathrm{\&beta;}}_{i}g\left(x_j\right)=\underset{i=1}{\overset{\tilde{N}}{\mathrm{\&Sigma;}}}{\mathrm{\&beta;}}_{i}g(w_i\&CenterDot;x_j+b_i)=o_j,j=1,...,N$ Calculate output matrix H;

4. calculate output weights β: β=H ⁺y, wherein, H ⁺for the Moore-Penrose generalized inverse of hidden layer output matrix H.

(5) utilize verification msg checking mine water inrush forecast model, if what obtain predicts the outcome and other algorithm comparisons, there is no clear superiority, from principal component analysis (PCA), start to re-establish model, comparatively desirable if predict the outcome, set it as forecast model and carry out actual prediction mine water inrush situation.

Claims (3)

1. the mine water inrush Forecasting Methodology based on PCA-EML, it is characterized in that: the input parameter that carrys out optimization neural network by principal component analysis (PCA), first utilize principal component analysis (PCA) to carry out pre-service to many factors data, eliminate the information overlap between legacy data, produce new separate training sample, retain as much as possible original information, then the input using the training sample of reconstruct as extreme learning machine, reduce the structure complexity of neural network, improve speed of convergence, concrete steps are as follows:

(1) obtain colliery and normally exploit the numerous data that affect mine water inrush under running status;

(2) utilize principal component analysis (PCA) to screen the many factors that affects mine water inrush, obtain mine water inrush to rise the Dominated Factors of deciding factor;

(3) sample data that only comprises Dominated Factors is divided into training set, checking collection and test set, corresponds respectively to training, checking and the test of model;

(4) set up extreme learning machine network model;

(5) utilize verification msg checking mine water inrush forecast model, if what obtain predicts the outcome and other algorithm comparisons, there is no clear superiority, from principal component analysis (PCA), start to re-establish model, comparatively desirable if predict the outcome, set it as forecast model and carry out actual prediction mine water inrush situation.

2. the mine water inrush Forecasting Methodology based on PCA-EML according to claim 1, is characterized in that: described screens mine water inrush master control influence factor by principal component analysis (PCA), and its step is as follows:

(1) get colliery and normally exploit the sampled data matrix under running status, and it is carried out to standardization obtain matrix Χ;

(2) according to the data matrix Χ after standardization, calculate covariance matrix R;

(3) according to covariance matrix R, obtain eigenwert, principal component contributor rate and accumulative total variance contribution ratio, determine major component number;

(4) by major component number, determine load matrix W, W=XM, is the ELM(extreme learning machine of foundation) learning sample of model.

3. the mine water inrush Forecasting Methodology based on PCA-EML according to claim 1, is characterized in that: set up extreme learning machine network model, and through extreme learning machine algorithm, these data are trained, its step is as follows:

(1) given N training sample { x _i, t _i, i=1,2 ..., N; Initial hidden layer nodes is made as

excitation function g (x);

(2) be input weights ω _iwith threshold value b _irandom assignment, wherein

(3) according to the unified model formula of the SLFN of excitation function g (x)

$\underset{i=1}{\overset{\tilde{N}}{\mathrm{\&Sigma;}}}{\mathrm{\&beta;}}_{i}g\left(x_j\right)=\underset{i=1}{\overset{\tilde{N}}{\mathrm{\&Sigma;}}}{\mathrm{\&beta;}}_{i}g(w_i\&CenterDot;x_j+b_i)=o_j,j=1,...,N,$

Be expressed in matrix as

：Hβ=T, $H(w_1,...,w_{\tilde{N}},b_1,...,b_{\tilde{N}},x_1,...,x_{\tilde{N}})={\left|\begin{array}{ccc}g(w_1\&CenterDot;x_1+b_1)& ...& g(w_{\tilde{N}}\&CenterDot;x_1+b_{\tilde{N}})\\ .& & .\\ .& ...& .\\ .& & .\\ g(w_1\&CenterDot;x_N+b_1)& ...& g(w_{\tilde{N}}\&CenterDot;x_{\tilde{N}}+b_{\tilde{N}})\end{array}\right|}_{N\&times;\tilde{N}},$

$\mathrm{\&beta;}={\left|\begin{array}{c}{{\mathrm{\&beta;}}^T}_1\\ .\\ .\\ .\\ {\mathrm{\&beta;}}_{\tilde{N}}^T\end{array}\right|}_{\tilde{N}\&times;m},T={\left|\begin{array}{c}{t}_1^T\\ .\\ .\\ .\\ t_N^T\end{array}\right|}_{N\&times;m}$ Calculate output matrix H; Wherein H is called the hidden layer output matrix of neural network; x _i=[x _i1, x _i2... x _in] ^t∈ R ⁿ, t _i=[t _i1, t _i2... t _im] ^t∈ R ^m, ω _ifor connecting the input weights of i hidden layer node; β _ifor connecting the output weights of i concealed nodes and output node; b _iit is the deviation of i concealed nodes; w _ix _jrepresent w _iand x _jinner product; Excitation function g (x) can be " Sigmiod ", " Sine " or RBF;

(4) calculate output weights β: β=H ⁺y, wherein H ⁺for the Moore-Penrose generalized inverse of hidden layer output matrix H.

Images