CN104156560A - Multi-level coal mine water inrush prediction method based on SaE-ELM (self-adaptive evolutionary extreme learning machine) - Google Patents

Abstract

A multi-level coal mine water inrush prediction method based on an SaE-ELM (self-adaptive evolutionary extreme learning machine) includes the steps of 1, researching a coal mine water inrush mechanism and selecting main controlling factors causing coal mine water inrush; 2, searching for a great deal of coal floor water inrush historical data as sample data, with each set of data including main controlling factors and maximum water inrush; 3, dividing the sample data into a training set and a test set, applied to training and testing of a model, respectively; 4, training the sample data with the SaE-ELM to establish a prediction model; 5, testing the prediction model with the data of the test set, comparing obtained prediction results with those obtained by other algorithms; if the prediction precision is high and the speed is high, using the prediction model for predicting whether a coal mine suffers water inrush or not and predicting the degree of water inrush.

Description

A kind of colliery many grades water-bursting predicting method based on SaE-ELM

Technical field

The present invention relates to a kind of colliery many grades water-bursting predicting method, particularly a kind of based on SaE-ELM ore deposit water-bursting predicting method.

Background technology

Mine water inrush is one of colliery five large disasters, predicts rapidly and accurately gushing water, is the guarantee of Safety of Coal Mine Production.Mine water inrush prediction relates to the factors such as hydrogeology, rock mechanics, mining conditions, has complicated nonlinear relationship between each factor, therefore by traditional mathematical theory, is difficult to set up model.

Experts and scholars have now proposed the method for multiple prediction mine water inrush, there is the genetic algorithm of employing to train BP neural network, set up coal seam gushing water Artificial Neural Network Prediction Model, although the method has improved training precision, but in view of the design feature of BP neural network, need to carry out parameter adjustment with a large amount of time; Separately have and set up the mine floor water-bursting predicting model that a kind of PCA combines with very fast learning machine ELM (Extreme Learning Machine), this model running speed and precision of prediction increase, but during ELM training pattern, all parameters of its network concealed layer are random generations, and training has certain randomness.In addition, China has divided gushing water type by the maximal value of gushing water amount (peak value), and existing most of method, only to whether gushing water is predicted, does not relate to the prediction of gushing water degree, is not easy to different gushing water situations to carry out pre-service.

ELM (extreme learning machine) is a kind of single hidden layer feedforward neural network, while utilizing ELM neural metwork training model, all parameters of hidden layer (input layer weights and hidden node deviation) are random generations, output layer weights are by calculative determination, therefore do not need to go to adjust parameter by the method for iteration, only need be the in the situation that of random given hidden layer parameter, method by least square is calculated output layer weights the training that just can complete neural network, therefore, it has in theory well solved the slow-paced problem of the neural neural network learning of feedforward.But, because all parameters of ELM hidden layer are random generations, make the model of training have larger training randomness, and SaE-ELM (Self-Adaptive Evolutionary Extreme Learning Machine) algorithm utilizes adaptive differential evolution algorithm optimization input weights and hidden layer deviation, the method has been avoided traditional E LM limitation, improve the degree of accuracy of generalization and prediction, retained the rapidity of ELM algorithm.

Summary of the invention

The present invention seeks to provide a kind of learning training and predetermined speed is fast, Generalization Capability is good and precision of prediction the is high colliery many grades water-bursting predicting method based on SaE-EML.

For achieving the above object, the present invention by the following technical solutions: study colliery mechanism of water inrush, choose and affect the Dominated Factors of mine water inrush and collect a large amount of mine water inrush historical datas as sample data, every group of packet is containing each Dominated Factors and maximum gushing water amount.Adopt single hidden layer feedforward neural network SaE-ELM to carry out model training, this algorithm is to utilize adaptive differential evolution algorithm optimization input weights and hidden layer deviation, method by least square is calculated output layer weights, compare with algorithm in the past, training and predetermined speed have been improved, and guaranteed higher prediction accuracy, concrete steps are as follows:

(1) research colliery mechanism of water inrush, chooses the Dominated Factors that affects gushing water;

(2) collect a large amount of mine water inrush historical datas as sample data, every group of packet is containing each Dominated Factors and maximum gushing water amount;

(3) according to the anti-regulation > > that harnesses the river in < < colliery, maximum gushing water amount is converted into gushing water type, according to data characteristics, all sample datas are divided into two component type and continuous types, and continuous data is normalized, and be divided into training set and test set;

(4) for N different training sample (x _i, t _i), x here _i=[x _i1, x _i2..., x _in] ∈ R ⁿ, t _i=[t _i1, t _i2... t _im] ∈ R ^m, hidden layer nodes is L, activation function is that the standard SLFNs model of g (x) is:

Σβ _jg(x _i)＝Σβ _jg(w _j,x _i,b _j)＝o _j,i＝1,...,N　

Wherein, w _j∈ R ⁿand b _j∈ R (j=1,2..., L) is j hidden layer parameter, w _jconnect j hidden layer node input weights, b _jit is the deviation of j concealed nodes.β _j∈ R ^mthe output weights between j concealed nodes and output node.W _jx _irepresent w _jand x _iinner product, g (x) is activation function, conventional have functions such as Sigmoid, Sine, Hardlim.

The N of a master pattern formula is expressed in matrix as: H β=T

Wherein:

$\begin{array}{c}H(w_1,\&CenterDot;\&CenterDot;\&CenterDot;,w_L,b_1,\&CenterDot;\&CenterDot;\&CenterDot;,b_L,x_1,\&CenterDot;\&CenterDot;\&CenterDot;,x_N)\\ =\left[\begin{array}{c}h\left(x_1\right)\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ h\left(x_N\right)\end{array}\right]={\left[\begin{array}{ccc}g(w_1\&CenterDot;x_I+b_I)& ...& g(w_L\&CenterDot;x_L+b_L)\\ \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \&CenterDot;\\ g(w_1\&CenterDot;x_N+b_I)& ...& g(w_L\&CenterDot;x_N+b_L)\end{array}\right]}_{N\&times;L}\end{array}$

$\mathrm{\&beta;}={\left[\begin{array}{c}{\mathrm{\&beta;}}_1^T\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ {\mathrm{\&beta;}}_L^T\end{array}\right]}_{L\&times;m};T={\left[\begin{array}{c}{t}_1^T\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ t_N^T\end{array}\right]}_{N\&times;m}$

In the most cases of real world applications, hidden layer nodes L is much smaller than number of training N, and this makes output matrix H is not a square formation, does not therefore have parameter b _j, w _j, β _j(j=1,2...L) sets up H β=T.By least square solution linear system H β=T, find β, obtain unique solution: h in formula ⁺it is the generalized inverse of the moore-penrose of H.

(5) produce at random NP initial population vector θ _k,G, k=1 wherein, 2 ... NP, every group of vector comprises all hidden layer parameter w _jwith b _j, the variation by repeating in each training and test process, intersect and select to operate and find optimum θ _k,G;

(6) adjust the value of hidden layer nodes L and the type of activation function g (x), speed and the degree of accuracy of the each training of record and test, choose and make model L value and g (x) type the most fast and accurately;

(7) calculate output weights β now;

(8) utilize test set data to test mine water inrush forecast model, predicting the outcome of obtaining compared with other algorithms, if precision of prediction is high, speed is fast, using it as can be to the colliery model that gushing water and gushing water degree are predicted.

Beneficial effect, owing to being the approximation problem of complicated nonlinear function between water-bursting predicting value and each gushing water, traditional Learning Algorithm exists training speed slowly, to be easily absorbed in the shortcomings such as selection sensitivity of local minimum point and learning rate.SaE-ELM algorithm arranges suitable hidden layer node number and activation function by continuous test, utilize adaptive differential evolution algorithm to random assignment input weights and hidden layer partially value be optimized, adopt least square method to calculate and export weights.To a large amount of historical data training, can access gushing water degree forecast model fast and accurately, the randomness that the method has avoided traditional E LM to bring because of random generation hidden layer parameter, has improved degree of accuracy and the speed of prediction.For this model prediction result, be different gushing water degree, different advanced warning grades can be set, be convenient to Mine Safe Production Management.

Accompanying drawing explanation

Fig. 1 is the colliery many grades water-bursting predicting process flow diagram based on SaE-ELM.

Embodiment

Embodiment 1: research colliery mechanism of water inrush, and to choose and affect the Dominated Factors of mine water inrush and collect a large amount of mine water inrush historical datas as sample data, every group of packet is containing each Dominated Factors and maximum gushing water amount.Data are made to normalized, and the data after standardization processing are divided into training sample and test sample book, utilize SaE-ELM to train, step is as follows:

(1) N different training sample (x _i, t _i), x here _i=[x _i1, x _i2..., x _in] ∈ R ⁿ, t _i=[t _i1, t _i2... t _im] ∈ R ^m, hidden layer nodes is L, activation function is g (x);

(2) produce at random NP initial population vector θ _k,G, k=1 wherein, 2 ... NP, every group of vector comprises all hidden layer parameter w _jwith b _j, the variation by repeating in each training and test process, intersect and select to operate and find optimum θ _k,G;

(3) adjust the value of hidden layer nodes L and the type of activation function g (x), speed and the degree of accuracy of the each training of record and test, choose and make model L value and g (x) type the most fast and accurately;

(4) calculate output weights β now;

(5) utilize test set data to test mine water inrush forecast model, predicting the outcome of obtaining compared with other algorithms, if precision of prediction is high, speed is fast, using it as can be to the colliery model that gushing water and gushing water degree are predicted.

Claims (2)

1. the colliery many grades water-bursting predicting method based on SaE-ELM, it is characterized in that: use SaE-ELM algorithm to set up forecast model, SaE-ELM is single hidden layer feedforward neural network, utilize adaptive differential evolution algorithm optimization input weights and hidden layer deviation, method by least square is calculated output layer weights, compares with algorithm in the past, has improved training and predetermined speed, and guaranteed higher prediction accuracy, concrete steps are as follows:

(1) research colliery mechanism of water inrush, chooses the Dominated Factors that affects gushing water;

(2) collect a large amount of mine water inrush historical datas as sample data, every group of packet is containing each Dominated Factors and maximum gushing water amount;

(3) according to the anti-regulation > > that harnesses the river in < < colliery, maximum gushing water amount is converted into gushing water type, according to data characteristics, all sample datas are divided into two component type and continuous types, and continuous data is normalized;

(4) sample data is divided into training set and test set, should be in training and the test of model;

(5) set up the forecast model based on SaE-ELM;

(6) test result of forecast model and test result based on BP, SVM scheduling algorithm are made comparisons, if precision of prediction is high, speed is fast, using it as can be to the colliery model that gushing water and gushing water degree are predicted.

2. the colliery many grades water-bursting predicting method based on SaE-EML according to claim 1, is characterized in that: utilize SaE-ELM algorithm to train sample data, set up the colliery many grades water-bursting predicting model based on SaE-ELM, its step is as follows:

(1) get history of coal-mines gushing water data, and it is normalized rear as training sample;

(2) for N different training sample (x _i, t _i), x here _i=[x _i1, x _i2..., x _in] ∈ R ⁿ, t _i=[t _i1, t _i2... t _im] ∈ R ^m, hidden layer nodes is L, activation function is that the standard SLFNs model of g (x) is:

Σβ _jg(x _i)＝Σβ _jg(w _j,x _i,b _j)＝o _j,i＝1,...,N

Wherein, w _j∈ R ⁿand b _j∈ R (j=1,2..., L) is j hidden layer parameter, w _jconnect j hidden layer node input weights, b _jit is the deviation of j concealed nodes.β _j∈ R ^mthe output weights between j concealed nodes and output node.W _jx _irepresent w _jand x _iinner product, g (x) is activation function, conventional have functions such as Sigmoid, Sine, Hardlim.

The N of a master pattern formula is expressed in matrix as: H β=T

Wherein:

$\begin{array}{c}H(w_1,\&CenterDot;\&CenterDot;\&CenterDot;,w_L,b_1,\&CenterDot;\&CenterDot;\&CenterDot;,b_L,x_1,\&CenterDot;\&CenterDot;\&CenterDot;,x_N)\\ =\left[\begin{array}{c}h\left(x_1\right)\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ h\left(x_N\right)\end{array}\right]={\left[\begin{array}{ccc}g(w_1\&CenterDot;x_I+b_I)& ...& g(w_L\&CenterDot;x_L+b_L)\\ \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \&CenterDot;\\ g(w_1\&CenterDot;x_N+b_I)& ...& g(w_L\&CenterDot;x_N+b_L)\end{array}\right]}_{N\&times;L}\end{array}$

$\mathrm{\&beta;}={\left[\begin{array}{c}{\mathrm{\&beta;}}_1^T\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ {\mathrm{\&beta;}}_L^T\end{array}\right]}_{L\&times;m};T={\left[\begin{array}{c}{t}_1^T\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ t_N^T\end{array}\right]}_{N\&times;m}$

In the most cases of real world applications, hidden layer nodes L is much smaller than number of training N, and this makes output matrix H is not a square formation, does not therefore have parameter b _j, w _j, β _j(j=1,2...L) sets up H β=T.By least square solution linear system H β=T, find β, obtain unique solution: h in formula ⁺it is the generalized inverse of the moore-penrose of H.

(3) produce at random NP initial population vector θ _k,G, k=1 wherein, 2 ... N, every group of vector comprises all hidden layer parameter w _jwith b _j:

${\mathrm{\&theta;}}_{k,G}=[w_{1,(k,G)}^T,\&CenterDot;\&CenterDot;\&CenterDot;,w_{L,(k,G)}^T,b_{1,(k,G)}^T,\&CenterDot;\&CenterDot;\&CenterDot;,b_{L,(k,G)}^T]$

Wherein G represents current population, k=1, and 2 ..., NP.

Then, by mutation strategy, generate the individual v of variation _k,G, four kinds of conventional mutation strategies are as follows:

v _k,G＝θ _r1,G+F·(θ _r2,G-θ _r3,G)

v _k,G＝θ _r1,G+F(θ _best,G-θ _r1,G)

+F(θ _r2,G-θ _r3,G)+F(θ _r4,G-θ _r5,G)

v _k,G＝θ _r1,G+F(θ _r2,G-θ _r3,G)+F(θ _r4,G-θ _r5,G)

v _k,G＝θ _r1,G+k·(θ _r1,G-v _k,G)+F(θ _r2,G-θ _r3,G)

Wherein, k random value between 0～1; Zoom factor F is that average is that 0.5 equation is 0.3 Gaussian distribution; R1, r2, r3, r4, r5 are mutually different random integers between 1～NP.

After mutation operation completes, utilize following formula to complete θ _k,Gwith the individual v of variation _k,Gbetween discretize interlace operation generate the individual u of test _k,G.

$u_{k,G}\left(j\right)=\left\{\begin{array}{cc}{v}_{k,G}\left(j\right),& \mathrm{if}({\mathrm{rand}}_j\&le;\mathrm{CR})\mathrm{or}(j=j_{\mathrm{rand}})\\ {\mathrm{\&theta;}}_{k,G}\left(j\right),& \mathrm{otherwise}\end{array}\right.$

In formula, the factor CR that intersects is that average is that 0.5 equation is 0.1 Gaussian distribution; Rand _jit is the random number between 0～1; J and j _randthe random integers between 1～L, in order to avoid testing individual U _k,Gcopy v completely _k.Gsituation.

Utilize following formula to obtain root-mean-square error (RMSE) to each individuality in population, using this error as adaptive value, obtain optimum population θ of new generation _{k, G+1}complete and select operation.

${\mathrm{RMSE}}_{k,G}=\sqrt{\frac{\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\left|\right|\underset{j=1}{\overset{L}{\mathrm{\&Sigma;}}}{\mathrm{\&beta;}}_{j}g(w_{j,(k,G)},b_{j,(k,G)},x_i)-t_i\left|\right|}{m\&times;N}}$

${\mathrm{\&theta;}}_{k,G+1}=\left\{\begin{array}{cc}{u}_{k,G+1}& \mathrm{if}{\mathrm{RMSE}}_{{\mathrm{\&theta;}}_{k,G}}-\mathrm{RMS}{E}_{u_{k,G+1}}>\mathrm{\&epsiv;}\&CenterDot;{\mathrm{RMSE}}_{{\mathrm{\&theta;}}_{k,G}},\\ u_{k,G+1}& \mathrm{if}|{\mathrm{RMSE}}_{{\mathrm{\&theta;}}_{k,G}}-{\mathrm{RMSE}}_{u_{k,G+1}}|<\mathrm{\&epsiv;}\&CenterDot;{\mathrm{RMSE}}_{{\mathrm{\&theta;}}_{k,G}}\\ & \mathrm{and}\left|\right|{\mathrm{\&beta;}}_{u_{k,G+1}}\left|\right|<\left|\right|{\mathrm{\&beta;}}_{{\mathrm{\&theta;}}_{k,G}}\left|\right|,\\ {\mathrm{\&theta;}}_{k,G}& \mathrm{else}.\end{array}\right.$

Repetitive cycling sudden change, intersection and selection operation, draw optimum θ until reach maximum iteration time _k,G.

(4) adjust the value of hidden layer nodes L and the type of activation function g (x), speed and the degree of accuracy of the each training of record and test, choose and make model L value and g (x) type the most fast and accurately;

(5) calculate output weights β now.