CN107122860A - Bump danger classes Forecasting Methodology based on grid search and extreme learning machine - Google Patents

Abstract

The present invention proposes a kind of bump danger classes Forecasting Methodology based on grid search and extreme learning machine, this method：Using the influence factor data of known bump as extreme learning machine input, using bump danger classes as extreme learning machine output, optimize the hidden layer neuron number of extreme learning machine and the type combination of activation primitive using grid data service, corresponding extreme learning machine is set up according to each grid node, the predictablity rate of respective mesh node is determined using ten folding cross-validation methods to each model, selection predictablity rate highest node determines the hidden layer neuron number of extreme learning machine and the type of activation primitive, set up bump danger classes forecast model；By the influence factor data input bump danger classes forecast model of bump to be predicted, bump danger classes predicted value is obtained.This method is simple and easy to do, while ensure that model has good Generalization Capability.

Description

Prediction method of rock burst hazard level based on grid search and extreme learning machine

technical field

The invention belongs to the technical field of rock burst risk level prediction, and in particular relates to a method for predicting the rock burst risk level based on a grid search and an extreme learning machine.

Background technique

Rock burst is a common dynamic phenomenon. With the increase of coal mining depth in my country, the number of occurrences of rock burst is on the rise, and its damage is becoming more and more serious, causing a large number of casualties and property losses, which seriously threatens the safety of coal mines. Therefore, it is necessary to effectively predict the hazard level of rock burst.

At present, methods for predicting rock burst include drilling cuttings method using a single influencing factor, water content measurement and other methods. However, these methods only consider a single influencing factor, and there is a problem of low prediction accuracy. Recently, with the development of artificial intelligence, Many scholars use new technologies and methods to predict the risk level of rock burst, including artificial neural network method, GA-ELM method, Fisher discriminant analysis method, SVM model, etc. The above methods have achieved a lot of research results, but Due to the complex causes of rock burst and the characteristics of non-linearity and correlation of rock burst data, it is necessary to continue to explore new methods to predict the risk level of rock burst.

The extreme learning machine is a new learning algorithm with a single hidden layer structure. Compared with the traditional single hidden layer feedforward neural network, it has the advantages of fast learning speed, good generalization ability, and fewer adjustment parameters. The current application mainly uses optimization The algorithm optimizes the input weights and hidden layer deviations of extreme learning machines. For example, Zhu Zhijie et al. used genetic algorithms to optimize the input weights and hidden layer deviations of extreme learning machines to predict rock burst. However, the performance of extreme learning machines It is mainly affected by the number of neurons in the hidden layer and the activation function, and when the number of neurons in the hidden layer is large, there are more parameters to be optimized; in addition, Ding Hua et al. use genetic algorithm to optimize the optimal number of neurons in the hidden layer. The power of the shearer is predicted by using a progressive comparison to determine the activation function. However, the type of activation function has been fixed when the number of neurons in the hidden layer is optimized, and the weights of the input layer and the hidden layer The values and hidden layer thresholds are randomly generated, so it is difficult to guarantee the uniqueness of the running results; in addition, the over-fitting problem of the model is not fully considered during the training process of the parameters of the extreme learning machine, so the predictive performance of the model cannot be guaranteed.

Contents of the invention

Aiming at the deficiencies of the prior art, the present invention proposes a method for predicting the hazard level of rock burst based on grid search and extreme learning machine.

A method for predicting the hazard level of rock burst based on grid search and extreme learning machine, comprising the following steps:

Step 1: Obtain the known monitoring data of rock burst at different locations in the coal mine, the data of known influencing factors of rock burst Z=[z ₁ , z ₂ ,..., z _p ] ^T , The impact factor data Z′=[z′ ₁ , z′ ₂ ,…,z′ _p ] ^T to be predicted, and the known rockburst is monitored according to the magnitude and intensity classification standard Classify the data to obtain the rock burst hazard level corresponding to the rock burst influencing factor data, where z _i is the known influencing factor data of the i-th type of rock burst, and z′ _i is the i-th type of rock burst to be predicted Influencing factor data, i=1, 2, ..., p, p is the number of influencing factors of rock burst;

The rock burst influencing factors include: coal seam thickness, coal seam dip angle, buried depth, gas concentration and state parameters affecting rock burst;

The state parameters affecting rock burst include geological structure, coal seam dip angle change, coal seam thickness change, roof management, pressure relief state, and sound of coal guns.

Step 2: Use the ^zscore standardization method to analyze the data Z ₌ [z ₁ , z ₂ , . [z′ ₁ , z′ ₂ , ......, z′ _p ] ^T is standardized, and the normalized known impact factor data of rock burst X=[x ₁ , x ₂ ,  … ..., x _p ] ^T and the data X′=[x′ ₁ , x′ ₂ ,……, x′ _p ] ^T of the impact factor data X′=[x′ 1 , x′ p ] T to be predicted;

Step 3: Take the standardized known impact factor data of rock burst X=[x ₁ , x ₂ , . . . , x _p ] ^T and its corresponding risk level of rock burst as a training sample set;

Step 4: Take the standardized and known impact factor data of rock burst in the training sample set as the input of the extreme learning machine, and use the corresponding rock burst hazard level in the training sample set as the output of the extreme learning machine, and use the grid search method Optimize the combination of the number of neurons in the hidden layer of the extreme learning machine and the type of activation function, establish a corresponding extreme learning machine according to each grid node, and use the ten-fold cross-validation method for each model to determine the prediction accuracy of the corresponding grid node , select the node with the highest prediction accuracy to determine the number of hidden layer neurons and the type of activation function of the extreme learning machine, and establish a prediction model for rock burst hazard level;

Step 4.1: Set the interval of the grid search method, set the interval of the number of neurons in the hidden layer according to the number of impact factors of rock burst, assign a value to the type of activation function, and set the number of rows of the grid search method as hidden The maximum value of the number of neurons in the layer, the number of columns of the grid search method is set to the maximum value assigned to the activation function type, and the search grid is established;

The assignment of the activation function type is an integer of 1 to 3, represented as a sigmoid function, a sin function, and a hardlim function, respectively.

Step 4.2: Use the number of rows where the nodes are located as the number of neurons in the hidden layer of the extreme learning machine, use the type of activation function corresponding to the number of columns where the nodes are located as the type of activation function of the extreme learning machine, and use the standardized neurons in the training sample set as The known influencing factors of rock burst are used as the input of the extreme learning machine, and the corresponding rock burst hazard level in the training sample set is used as the output of the extreme learning machine, and the extreme learning machine is established, and the current node is established by using the ten-fold cross-validation method. The prediction accuracy of the extreme learning machine;

Step 4.3: Determine whether the maximum number of nodes is currently searched, if so, perform step 4.4, otherwise, search for the next node, and return to step 4.2;

Step 4.4: Select the node corresponding to the model with the highest prediction accuracy in the extreme learning machine established based on all nodes as the search result, establish the extreme learning machine model according to the number of hidden layer neurons and activation function type corresponding to the node, and obtain the impact Prediction model of ground pressure hazard level;

Step 5: To predict the risk level of rock burst, the normalized data of influencing factors of rock burst to be predicted X′=[x′ ₁ , x′ ₂ ,  …, x′ _p ] ^T Input the prediction model of rock burst hazard level to obtain the predicted value of rock burst hazard level.

Beneficial effects of the present invention:

The present invention proposes a method for predicting the risk level of rock burst based on grid search and extreme learning machine. Since the mechanism of rock burst is complex and there are many influencing factors, the method of the present invention adopts the zscore method to standardize the data of the influencing factors and eliminates The influence of different dimensions on the model; the performance of the extreme learning machine is greatly affected by the number of hidden layer neurons and the type of activation function. The grid search method combined with ten-fold cross-validation is used to analyze the hidden layer neurons in the extreme learning machine. The number and type of activation function are combined and optimized. This method is simple and easy to implement, and at the same time ensures that the model has good generalization performance.

Description of drawings

Fig. 1 is a flow chart of the method for predicting the hazard level of rock burst based on grid search and extreme learning machine in a specific embodiment of the present invention.

detailed description

The specific embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings.

A prediction method of rock burst hazard level based on grid search and extreme learning machine, as shown in Figure 1, includes the following steps:

Step 1: Obtain the known monitoring data of rock burst at different locations in the coal mine, the data of known influencing factors of rock burst Z=[z ₁ , z ₂ ,..., z _p ] ^T , The impact factor data Z′=[z′ ₁ , z′ ₂ ,…,z′ _p ] ^T to be predicted, and the known rockburst is monitored according to the magnitude and intensity classification standard Classify the data to obtain the rock burst hazard level corresponding to the rock burst influencing factor data, where z _i is the known influencing factor data of the i-th type of rock burst, and z′ _i is the i-th type of rock burst to be predicted Influencing factor data, i=1, 2, ..., p, p is the number of influencing factors of rock burst.

In this embodiment, the influencing factors of rock burst include: coal seam thickness z ₁ , coal seam inclination z ₂ , buried depth z ₃ , gas concentration z ₄ and state parameters affecting rock burst.

State parameters affecting rock burst include geological structure z ₅ , coal seam dip angle change z ₆ , coal seam thickness change z ₇ , roof management z ₈ , pressure relief state z ₉ , and sound of coal cannon z ₁₀ .

Assign values to the state parameters that affect rock burst: set the corresponding assignments according to the different states of the state parameters that affect rock burst, and assign the state parameters that affect rock burst to integer values, as shown in Table 1:

Table 1 Assignment of state parameters affecting rock burst

In this embodiment, the rock burst monitoring data is classified according to the rock burst magnitude intensity classification standard, and the rock burst hazard level is obtained as 4 levels, which are level 1: micro impact, level 2: weak impact, and level 3: moderate impact , Level 4: Strong impact.

In this embodiment, the rock burst monitoring data and corresponding influencing factor data at different locations in the Yanshitai Coal Mine are obtained. The risk level of rock burst and the corresponding influencing factor data are shown in Table 2, wherein the first 26 sets of data The data of influencing factors is used as the data of influencing factors of known rockburst, the corresponding data of rockburst is used as the risk level of known rockburst, and the data of influencing factors in the last 10 sets of data are used as the data of influencing factors of rockburst to be predicted.

Table 2 Data of influencing factors of rock burst at different locations in Yanshitai Coal Mine and corresponding rock burst hazard levels

Step 2: Use the ^zscore standardization method to analyze the data Z ₌ [z ₁ , z ₂ , . [z′ ₁ , z′ ₂ , ......, z′ _p ] ^T is standardized, and the normalized known impact factor data of rock burst X=[x ₁ , x ₂ ,  … ..., x _p ] ^T and the standardized impact factor data X′=[x′ ₁ , x′ ₂ , . . . , x′ _p ] ^T to be predicted.

In this embodiment, the formula for standardizing the impact factor data of rock burst using the zscore standardization method is shown in formula (1):

Among them, x _ij is the jth value of the i-th type rock burst influencing factor data after normalization, μ _i is the mean value of the i-th type rock burst influencing factor data, and σ _i is the i-th type rock burst influencing factor data The standard deviation of , z _ii is the jth value of the collected i-th type rock burst influencing factor data, j=1, 2,..., N, N=36 is the total number of rock burst influencing factor data.

In this embodiment, the 36 groups of influencing factor data in Table 2 are standardized, wherein the first 26 groups of influencing factor data constitute X after normalization, and the last 10 groups of influencing factor data constitute X' after normalization.

Step 3: The standardized known impact factor data of rock burst ^X ₌ [x ₁ , x ₂ , .

Step 4: Take the standardized and known impact factor data of rock burst in the training sample set as the input of the extreme learning machine, and use the corresponding rock burst hazard level in the training sample set as the output of the extreme learning machine, and use the grid search method Optimize the combination of the number of neurons in the hidden layer of the extreme learning machine and the type of activation function, establish a corresponding extreme learning machine according to each grid node, and use the ten-fold cross-validation method for each model to determine the prediction accuracy of the corresponding grid node , select the node with the highest prediction accuracy to determine the number of hidden layer neurons and the type of activation function of the extreme learning machine, and establish a prediction model of rock burst hazard level.

Step 4.1: Set the interval of the grid search method, set the interval of the number of neurons in the hidden layer according to the number of impact factors of rock burst, assign a value to the type of activation function, and set the number of rows of the grid search method as hidden The maximum number of layer neurons, set the number of columns of the grid search method to the maximum value assigned by the activation function type, and establish a search grid.

In this embodiment, the interval of the grid search method is set to 1, the number of neurons in the hidden layer is set to [1, 100] according to the number of factors affecting rock burst, and the type of activation function is assigned a value of 1 to 3. The integers are respectively represented as sigmoid function, sin function, and hardlim function. In this embodiment, the value of the sigmoid function is 1, the value of the sin function is 2, and the value of the hardlim function is 3.

In this embodiment, the number of rows of the grid search method is set to 100 rows, and the number of columns of the grid search method is set to 3 columns.

Step 4.2: Use the number of rows where the nodes are located as the number of neurons in the hidden layer of the extreme learning machine, use the type of activation function corresponding to the number of columns where the nodes are located as the type of activation function of the extreme learning machine, and use the standardized neurons in the training sample set as The known influencing factors of rock burst are used as the input of the extreme learning machine, and the corresponding rock burst hazard level in the training sample set is used as the output of the extreme learning machine, and the extreme learning machine is established, and the current node is established by using the ten-fold cross-validation method. The prediction accuracy of the extreme learning machine.

In this embodiment, the ten-fold cross-validation method is used to divide the standardized known impact factor data of rock burst in the training sample set into ten parts, and 9 parts of them are used as training data, 1 part as test data, and 1 part as limit data in turn. The input of the learning machine, after ten calculations, calculates the accuracy rate of the ten prediction results and the corresponding rock burst hazard level in the training sample set, and takes the prediction accuracy rate as the evaluation index of the corresponding grid node.

Step 4.3: Determine whether the maximum number of nodes is currently searched, if so, perform step 4.4, otherwise, search for the next node, and return to step 4.2.

In this embodiment, the maximum number of nodes is 97 rows and 1 column.

Step 4.4: Select the node corresponding to the model with the highest prediction accuracy in the extreme learning machine established based on all nodes as the search result, establish the extreme learning machine model according to the number of hidden layer neurons and activation function type corresponding to the node, and obtain the impact Prediction model of ground pressure hazard level.

In this embodiment, the prediction model of rock burst hazard level has a three-layer structure, as shown in formula (2):

Among them, M is the number of neurons in the hidden layer, v=1, 2, ..., M, ω _v is the connection weight between the input layer and the hidden layer, and β _v is the connection weight between the hidden layer and the output layer, b _v is the threshold value of neurons in the hidden layer, g(*) is the activation function of the optimized extreme learning machine, o _k is the prediction category of rock burst hazard level, x _k is the kth normalized input training sample set The impact factor data of rock burst, k=1, 2,...,N.

In this embodiment, the optimal node is 97 rows and 1 column, that is, the number of neurons in the hidden layer is 97, and the type of activation function is a sigmoid function, and the partial input layer and hidden layer of the extreme learning machine model are obtained. The weight and hidden layer threshold b are shown in Table 3:

Table 3 Weights of some input layers and hidden layers and hidden layer threshold b of the extreme learning machine model

The extreme learning machine model was established according to the corresponding parameters of the node, and the correct recognition rate of the model after ten-fold cross-validation was 0.84615.

In this embodiment, in order to compare with the proposed method, the naive Bayesian method and the AdaboostM1 method are used to establish the prediction model of rock burst hazard level. It is lower than 0.84615, indicating that the rock burst prediction model established by this method has better performance.

Step 5: To predict the risk level of rock burst, the normalized data of influencing factors of rock burst to be predicted X′=[x′ ₁ , x′ ₂ ,  …, x′ _p ] ^T Input the prediction model of rock burst hazard level to obtain the predicted value of rock burst hazard level.

Adopt method of the present invention, naive Bayesian method and AdaboostM1 method to predict corresponding rockburst risk level according to the data after standardization of last 10 groups of influence factor data in table 2, prediction result is as shown in table 4:

Table 4 Prediction results

It can be seen from the table that the prediction model established by the method in this paper can accurately predict the hazard levels of 9 groups of rockbursts in the data, and only misjudge the moderate rockburst of the 10th group of data as weak rockbursts, while the Naive Bayesian method Both the AdaboostM1 method and the AdaboostM1 method accurately predicted the grades of 8 groups. The Naive Bayesian method misjudged the medium rock burst in the second group and the seventh group as a strong rock burst, and the AdaboostM1 method respectively judged the micro rock burst in the fourth group The moderate rockburst in group 7 was misjudged as strong rockburst.

Claims (3)

1. a kind of bump danger classes Forecasting Methodology based on grid search and extreme learning machine, it is characterised in that including Following steps：

Step 1：Obtain known bump Monitoring Data in coal mining mine at diverse location, known bump influence because Prime number is according to Z=[z₁, z₂..., z_p]^T, bump to be predicted influence factor data Z '=[z '₁, z '₂..., z ′_p]^T, and known bump Monitoring Data is classified according to bump earthquake magnitude intensity criteria for classification, obtain and impact ground The corresponding bump danger classes of influence factor data is pressed, wherein, z_iThe influence factor number of bump known to the i-th class According to z '_iFor the influence factor data of the i-th class bump to be predicted, i=1,2 ..., p, p be bump influence factor Number；

Step 2：Using influence factor data Z=[z of the zscore standardized methods to known bump₁, z₂..., z_p]^T With influence factor data Z '=[z ' of bump to be predicted₁, z '₂..., z '_p]^TIt is standardized, obtains standard The influence factor data X=[x of known bump after change₁, x₂..., x_p]^TWith the impact to be predicted after standardization Influence factor data X '=[x ' of pressure₁, x '₂..., x '_p]^T；

Step 3：By the influence factor data X=[x of the known bump after standardization₁, x₂..., x_p]^TAnd its it is corresponding Bump danger classes is used as training sample set；

Step 4：It regard the influence factor data of the known bump after the standardization that training sample is concentrated as extreme learning machine Input, training sample is concentrated to corresponding bump danger classes as the output of extreme learning machine, using grid search The hidden layer neuron number of method optimization extreme learning machine and the type combination of activation primitive, phase is set up according to each grid node Extreme learning machine is answered, the predictablity rate of respective mesh node is determined using ten folding cross-validation methods to each model, selection is pre- Survey accuracy rate highest node and determine the hidden layer neuron number of extreme learning machine and the type of activation primitive, set up impact ground Press danger classes forecast model；

Step 4.1：The interval of grid data service is set, hidden layer neuron number is set according to bump influence factor number Interval, the type to activation primitive carries out assignment, sets the line number of grid data service as the maximum of hidden layer neuron number, The columns of grid data service is set as the maximum of activation primitive type assignment, search grid is set up；

Step 4.2：Using the line number where node as extreme learning machine hidden layer neuron number, by the columns where node Corresponding activation primitive type as extreme learning machine activation primitive type, known to after the standardization that training sample is concentrated Training sample is concentrated corresponding bump danger etc. by the influence factor data of bump as the input of extreme learning machine Level sets up extreme learning machine as the output of extreme learning machine, using ten folding cross-validation methods, calculates the pole that present node is set up Limit the predictablity rate of learning machine；

Step 4.3：Whether judgement currently searches maximum node number, if so, performing step 4.4, otherwise, searches for next section Point, return to step 4.2；

Step 4.4：Choose the maximum corresponding node of model of predictablity rate in the extreme learning machine set up according to all nodes As search result, extreme learning machine mould is set up according to the corresponding hidden layer neuron number of the node and activation primitive type Type, obtains bump Hazard rank forecast model；

Step 5：Bump Hazard rank is predicted, by the influence factor number of the bump to be predicted after standardization According to X '=[x '₁, x '₂..., x '_p]^TBump danger classes forecast model is inputted, bump danger classes is obtained pre- Measured value.

2. the bump danger classes Forecasting Methodology according to claim 1 based on grid search and extreme learning machine, Characterized in that, the bump influence factor includes：Coal seam thickness, seam inclination, buried depth, gas density and influence impact The state parameter of ground pressure；

The state parameter of the influence bump includes geological structure situation, seam inclination change, Coal Seam Thickness Change, top plate Management, depressurization phase, sound coal report.

3. the bump danger classes Forecasting Methodology according to claim 1 based on grid search and extreme learning machine, Characterized in that, the integer for being entered as 1~3 of the activation primitive type, be expressed as sigmoid functions, sin functions, Hardlim functions.