Predicting Hidden Danger Quantity in Coal Mines Based on Gray Neural Network

Abstract

The hidden danger is the direct cause of coal mine accidents, and the number of hidden dangers in a certain area not only reflects the current safety situation, but also determines the development trend of safety production in this area to a large extent. By analyzing the formation and development law of the hidden dangers and hidden danger accident-induced mechanism in coal mines, it is concluded that there are some objective laws in the process of occurrence, development, weakening, and even stabilization of hidden dangers in a certain area. The development of the number of hidden dangers for a coal mine generally presents the law of similar normal distribution curve, with a certain degree of partial symmetry. Many years of hidden danger elimination in coal mines will accumulate large-scale hidden danger data. In this paper, by using the average value of hidden danger quantity in consecutive months to weaken the oscillation of hidden danger quantity sequence, and combining with gray model (1,1) and the neural network of extreme learning machine, and employing big data of hidden dangers available, a hidden danger quantity prediction model based on the gray neural network was established, and the experimental analysis and verification carried out. The results show that the model can achieve good prediction effect on the number of hidden dangers in a coal mine, which not only reflects the complex gray system behavior of hidden dangers of a coal mine, but also can predict dynamically. The safety management efficiency and emergency capacity of the coal mine enterprise will be greatly improved.

1. Introduction

Identification and control of hidden dangers is an effective way to prevent and control accidents in a coal mine [1,2]. The coal mines across China have established hidden danger identification system, and made hidden danger identification daily work. Such hidden dangers identified are enormous. According to statistics, the industrial and mining enterprises in 2013 alone had identified five million hidden dangers [3]. As coal mines intensify their efforts in work safety, the quantity of hidden dangers to be identified is ever-increasing, and big data has come into existence. However, after its inspection of the coal mines across China, the State Administration of Work Safety pointed out that identification and elimination was far below expectation, given that underreporting and false reporting were common, and the safety situation in the coal mines is still severe [4,5].

Research on the hidden dangers of coal mines presently focuses on statistical analysis on the quantity and features of the hidden dangers in different months or varieties. Regarding hidden dangers data extraction in the literature, [6] used the association rule and analyzed the rule of association of the responsible department for hidden dangers, hidden danger variety, hidden danger grade, and hidden danger scene. In [7], they applied 5W1H analysis approach and transformed loopholes into eight dimensions, including hidden danger time and space, and used the log-linear model to dig the relationship among the dimensions. In [8], they managed to predict the hidden danger quantity by the least square method and extreme learning machine. The purpose and essence of all these studies make use of big data to predict the features of hidden dangers to allow coal mine managers to make good preparations. Qualitative or quantitative research of those hidden dangers is of some applicational value, but prediction precision is not high and the operability is not strong. The hidden dangers largely determine the safety situation of a coal mine, and they play a decisive role in work safety.

Any accident is the result of the interaction of hidden dangers in time and in space. Hidden dangers cover four aspects: human, machine, environment, and management. The quantity of hidden dangers in a system shows a nonlinear relationship with influential factor, and is hard to be predicted by traditional classical mathematical models due to its randomness and volatility [9]. Hidden dangers are the result of the interaction of all risk factors in a system, and the hidden danger quantity features gray and uncertainty [10,11]. The gray prediction model is characterized by simple computation and only a few samples. Hence, prediction of hidden danger quantity can be considered as a gray system to be predicted. According to the characteristics of the gray system, it can be known that gray prediction is suitable for exponential growth. Therefore, prediction precision is yet to be improved. However, neural network is a nonlinear, self-adaptable system, and is fault-tolerant, but needs more samples and complex computation [12,13,14,15].

In order to further utilize the existing big data of hidden dangers and adopt scientific data processing algorithms to predict the number of hidden dangers in different areas of the coal mine and improve the shortcomings of the existing models, this paper combines the advantages of gray prediction and neural networks. We predict and then use the neural network to correct the residuals of the gray prediction model, which improves the accuracy and practicability of the model, and is of great significance for improving the level of coal mine safety production management.

2. Methodology

2.1. Principle of Hidden Danger Quantity Prediction

Hidden dangers represent unsafe human behavior, unsafe condition of the environment, unsafe condition of equipment, or defective management during work activity [16]. They evolved from hazards, which interact and eventually trigger accidents, as illustrated in Figure 1 [16,17]. If a workplace is considered as a system, each of its components is in continuous motion. The hidden dangers in the system are the middle status of motion, and accidents are the worst results of motion. Occurrence of accidents will produce new or secondary hidden dangers. The quantity of the hidden dangers in the system are the result of mutual restriction and interaction between the status and quantity at the previous moment and hidden danger identification and control capability. Hidden danger development has its objective regularity. A hidden danger cannot appear or disappear out of thin air. Its status value at any time is related to the status value at a previous time or previous timeframe. The majority of hidden dangers in a system do not appear alone, rather, they are induced by the available hidden dangers or the resulting accidents. Hence, the quantity of hidden dangers, on the one hand, reflects the level of hidden danger identification and control capability in the previous timeframe, and on the other hand, determines the development trend of hidden danger in future time. Although, due to the error of artificial examination, the hidden danger quantity identified every month shows a volatility. On the whole, the hidden danger quantity will increase, decrease, or stabilize along with the hidden danger control capability of the coal mine. Hidden dangers are hidden, latent, and inductive, so once appearing, they cannot be identified and eliminated in time. Plus, considering artificial reason and technical level, there is an error between the hidden dangers identified every month and actual ones in a coal mine. However, they are consistent on the whole, namely, hidden dangers quantity will increase, decrease, or stabilize along with the hidden danger identification, control capability, and geological condition. The development of the number of hidden dangers for a coal mine generally presents the law of similar normal distribution curve, with a certain degree of partial symmetry. Consequently, it is feasible to use the big data of historical hidden dangers to predict the hidden danger quantity in the coal mine.

From the formation and development mechanism of hidden dangers, it can be known that, hidden dangers appear random, uncertain, and sudden, but they are by nature regular and predictable. Due to their uncertainty, this paper at first employs the gray system theory to probe into the time law of hidden danger quantity, and to predict their change trend. Subsequently, a gray prediction model for hidden danger quantity is established. Then, as to their randomness, the paper adopts the extreme learning machine to correct the residual error of the gray prediction model. The prediction principle is illustrated in Figure 2.

2.2. GM (1,1) Model Predicting Hidden Danger Quantity

2.2.1. Mean GM (1,1) Prediction Model

The gray system was proposed by a Chinese scholar named Deng Julong in 1982. It is a mathematical tool that explores uncertain problem of not having much data and information. The approach functions to extract valuable information from some known information, and realizes correct description and effective monitoring of system behaviors and evolution law [18]. By raw data processing and gray model establishment, gray prediction extracts, reveals, and knows the evolution law of the system, and makes quantitative prediction of the future status of the system. GM (1,1) model is the underlying model of gray prediction theory, and also the most popular gray prediction model. The model is further divided into GM (1,1) even gray model (EGM), GM (1,1) original differential gray model (ODGM), GM (1,1) even differential gray model (EDGM), and GM (1,1) discrete gray model (DGM). Relative to other three models, EGM is more suitable for an oscillating sequence [19,20,21]. Because there exists a lot of influential factors behind hidden dangers in each coal mine area, and are affected by technical and managerial aspects in the capability of identifying and screening them, the sequence formed by the quantity of loopholes identified in a coal mine every month is an oscillating sequence. Therefore, this paper chooses GM (1,1) EGM. Its underlying principle is [22,23,24,25]:

Suppose the original sequence is:

$$X^{\left(0\right)}\left(t\right)=\left\{x^{\left(0\right)}\left(1\right),x^{\left(0\right)}\left(2\right),\cdots x^{\left(0\right)}\left(m\right)\right\}$$

(1)

m is the size of samples.

For $X^{\left(0\right)}\left(t\right)$, for X, do first-order accumulation to get first-order accumulation to generate a sequence of sequences (1-AGO) and obtain $X^{\left(1\right)}\left(t\right)$, and

$$X^{\left(1\right)}\left(t\right)={\displaystyle \sum }_{i=1}^t{x}^{\left(0\right)}\left(i\right)$$

(2)

Figure out the nearest neighbor mean Z⁽¹⁾(t) of $X^{\left(1\right)}\left(t\right)$, and

$$Z^{\left(1\right)}\left(t\right)=\frac{1}{2}\left(x^{\left(1\right)}\left(t\right)+x^{\left(1\right)}\left(t-1\right)\right)\hspace{1em}t=1,2,\cdots ,m-1$$

(3)

Then,

$$x^{\left(1\right)}\left(t\right)+a{z}^{\left(1\right)}\left(t\right)=b$$

(4)

The whitening differential equation of the above equation is

$$\frac{d{x}^{\left(1\right)}}{dt}+a{x}^{\left(1\right)}\left(t\right)=b$$

(5)

The vector $\widehat{a}={\left[a,b\right]}^T$ in the equation can be determined by least squares method, namely, $\widehat{a}={\left[B^{T}B\right]}^{-1}{B}^T$, and

$$Y=\left[\begin{array}{c}{x}^{\left(0\right)}\left(2\right)\\ x^{\left(0\right)}\left(3\right)\\ ⋮\\ x^{\left(0\right)}\left(n\right)\end{array}\right],B=\left[\begin{array}{ccc}-z^{\left(1\right)}\left(2\right)& 1& 1\\ -z^{\left(1\right)}\left(3\right)& 1& 1\\ ⋮& 1& 1\\ -z^{\left(1\right)}\left(n\right)& 1& 1\end{array}\right]$$

(6)

In the equation, a is known as development coefficient, which reflects the development status of a variable. b is known as gray actuating quantity.

Once each coefficient is figured out, gray GM (1,1) prediction model is determined

$${\widehat{x}}^{\left(1\right)}\left(t\right)=\left(x^{\left(0\right)}\left(t\right)-\frac{b}{a}\right)e^{-a\left(t-1\right)}+\frac{b}{a}$$

(7)

Then, inverse accumulation is made for the above equation, and the prediction value of any given input is worked out

$${\widehat{x}}^{\left(0\right)}=\left(1-e^a\right)\left(x^{\left(0\right)}\left(1\right)-\frac{b}{a}\right)e^{-a\left(k-1\right)}\hspace{1em}k=1,2,\cdots ,n$$

(8)

2.2.2. Buffer Operator Improvement GM (1,1)

The hidden dangers in a coal mine are artificially identified and entered. The quantity of each month shows a significant volatility and randomness according to the feature analysis of the hidden dangers and the difference of individuals’ hidden dangers identification capability. Though the data are scattered, they always possess an overall function according to the characteristics of the hidden dangers in the system of each area. In order to weaken the randomness of the sequence, the sequence, formed by the mean of the hidden danger quantity of each N consecutive months, is considered as a new gray sequence according to the identification characteristics. That is to say, the initial hidden danger quantity sequence in Equation (1) is generated the operator D by using the mean, that is, through the operation of formula (10) and a new sequence ${X^{\prime }}^{\left(0\right)}\left(t\right)$ is obtained.

$${X^{\prime }}^{\left(0\right)}\left(t\right)=\left\{{x^{\prime }}^{\left(0\right)}\left(1\right),{x^{\prime }}^{\left(0\right)}\left(2\right),\cdots {x^{\prime }}^{\left(0\right)}\left(n-4\right)\right\}$$

(9)

And

$${x^{\prime }}^{\left(0\right)}\left(t\right)=\frac{1}{4}{\displaystyle \sum }_{i-t}^{t+4}{x}^{\left(0\right)}\left(i\right)\hspace{1em}t=1,2,\dots ,n-4$$

(10)

Then use mean GM (1,1) model for prediction.

2.3. Residual Error Modified Model for Hidden Danger Prediction Based on PSO-RELM

2.3.1. ELM

Neural network has strong nonlinear fitting capability. It can better establish the quantitative relation of all factors in the system even if the qualitative relation of all the factors remains unknown [26,27,28,29]. Extreme learning machine in the neural network has been extensively used in the past years due to its simple structure, short training time, and high prediction precision. Hence, this paper adopts ELM to correct the residual error of gray GM (1,1) hidden danger quantity prediction.

Suppose there are N training samples $S_n=\left\{{\left(X_i,t_i\right)}_{i=1}^n\right\}$, where X_i = [x_i1, x_i1, ⋯ x_in]^T ∈ Rⁿ is the input value, and t_i = [t_i1, t_i1, ⋯ t_in]^T ∈ Rⁿ is the output value, then the regression model of the neuron function $f\left(x\right)$ of which the hidden node number is L can be expressed as [30]:

$${\displaystyle \sum }_{i=1}^L{\beta }_{i}f\left(W_i\cdot X_j+b_i\right)=o_j,j=1,2,\cdots ,N$$

(11)

where, f(x) is hidden activation function, $W_i={\left[w_{i1},w_{i1},\cdots w_{in}\right]}^T$ is input weight, ${\beta }_i$ is output weight, and $b_i$ is the biased value of the ith hidden node.

The solving objective of the neural network is to obtain the least error output, namely, to make

$${\displaystyle \sum }_{j-1}^L\parallel o_i-t_i\parallel =0$$

(12)

That is, there exists ${\beta }_i$,$W_i$ and $b_i$, which make

$${\displaystyle \sum }_{i=1}^L{\beta }_{i}f\left(W_i{X}_j+b_i\right)=t_j,j=1,2,\cdots ,N$$

(13)

Rewrite the above equation into a matrix form

$$H\beta =T$$

(14)

Where, $H$ is the output of hidden node, and

$$H={\left[\begin{array}{c}f\left(W_1{X}_1+b_1\right)f\left(W_2{X}_1+b_2\right)\dots f\left(W_L{X}_1+b_L\right)\\ f\left(W_1{X}_2+b_1\right)f\left(W_2{X}_2+b_2\right)\dots f\left(W_L{X}_2+b_{.L}\right)\\ ⋮\\ f\left(W_1{X}_2+b_1\right)f\left(W_2{X}_2+b_2\right)\dots f\left(W_L{X}_N+b_L\right)\end{array}\right]}_{N\times L}$$

(15)

$\beta$ is the weight of output layer, and

$$\beta ={\left[{\beta }_{1,}{\beta }_2,\dots ,{\beta }_L\right]}^T$$

(16)

T is expected output vector, and

$$T={\left[t_{1,}{t}_2,\dots ,t_N\right]}^T$$

(17)

In order to give the minimal error between input value and target value, the equivalent minimal loss function is

$$\mathrm{E}={{\displaystyle \sum }}_{j=1}^L{\left({{\displaystyle \sum }}_{i=1}^L{\beta }_{i}f\left(W_i{X}_j+b_i\right)-t_i\right)}^2$$

(18)

Among ELM neural network algorithm, once input weight and hidden biased value are determined at random, the solution of the hidden output matrix is exclusively determined. Then the question is translated to solve $=T$. According to the mathematical theory, obtain

$$\beta =H^{+}\mathrm{T}$$

(19)

Where H⁺ is the generalized inverse of matrix $H$, and

$$H^{+}={\left(H^{T}H\right)}^{-1}{H}^T$$

(20)

Moreover, we can prove that the norm of the solution obtained is minimal and exclusive. Hence, for given input value $x$, its prediction value can be obtained

$$t={{\displaystyle \sum }}_{i=1}^L{\beta }_{i}f\left(W_i{X}_x+b_i\right)$$

(21)

2.3.2. Regular Learning Machine RELM

Extreme learning machine (ELM) has the advantages of fast learning speed, etc. The solution obtained in Section 2.3.1 is the model established according to the least squares loss function, which only takes into account the least training error, namely, empirical risk. In statistics, the risk of the model consists of empirical risk and structural risk. Empirical risk reflects the training precision of the model, while structural risk reflects the generalization performance. The two combined determine the prediction precision of the model [31]. A sound model can not only obtain the training precision with better correctness, but satisfactory testing and prediction precision. That is to say, the model should avoid under-learning and over-learning. A study revealed that the smaller the weighted norm of feed forward neural network is, the more stable its network is, and the better its generalization performance is. As a result, the loss function of ELM neural network is translated into [32,33]. This method can overcome the above disadvantages, that is, the regular learning machine (RELM).

$$\{\begin{array}{l}E\left(H\right)=argmin\left({\epsilon }^2+\lambda {\beta }^2\right)\\ {\epsilon }_i={\displaystyle \sum }_{i=1}^L{\beta }_{i}f\left(W_i{X}_j+b_i\right)-T_i,i=1,2\dots N\end{array}$$

(22)

where, ${\left|\left|\epsilon \right|\right|}^2$ is empirical risk, ${\left|\left|\beta \right|\right|}^2$ is structural risk, and $\lambda$ is weight coefficient. The above problem is transformed into resolving the extremum by Lagrange.

$$L\left(\epsilon ,\beta ,\alpha \right)={\epsilon }^2+{\lambda }^2-\alpha \left(H\beta -T-\epsilon \right)$$

(23)

It is known that

$$\{\begin{array}{l}\frac{\partial L}{\partial \epsilon }=2{\epsilon }^T+\alpha \\ \frac{\partial L}{\partial \beta }=2\lambda {\beta }^T-\alpha H\\ \frac{\partial L}{\partial \alpha }=H\beta -T-\epsilon \end{array}$$

(24)

Resolve

$$\beta =\{\begin{array}{c}{\left(\lambda I+H^{T}H\right)}^{-1}{H}^{T}H H^{T}H is nonsingular matrix\\ H^T{\left(\lambda I+H{H}^T\right)}^{-1}T H{H}^T is nonsingular matrix\end{array}$$

(25)

2.3.3. PSO-RELM

The initial weight and hidden deviation of the extreme learning machine are determined at random, so the resulting model is just the optimal prediction model at given initial weight and hidden deviation. Therefore, this paper adopts particle swarm optimization (PSO) to optimize the initial weight and hidden deviation.

PSO was first proposed by Eberhart and Kennedy in 1995, which was intended to study the foraging behavior of bird swarm. Namely, in a specific scenario, a bird swarm is searching for food at random, each of the birds can be considered as a particle, neither of the birds knows where the food is, but they know how far their location is to the food. Then, all the birds will search toward the bird currently nearest to the food and also toward their position historically nearest to the food [34].

In PSO, each particle is the feasible solution of the problem. Each particle owns an adaptive value determined by the target function and the velocity value determining the flight direction and distance of the particle [35,36]. In D-dimensional space, suppose the number of particles is M, and suppose the global optimal position at t is

$$p_g^t={\left(p_{g1}^t,p_{g2}^t,\dots ,p_{gd}^t\right)}^T$$

(26)

Then, the historical optimal position of ith particle at t is

$$p_i^t={\left(p_{i1}^t,p_{i2}^t,\dots ,p_{id}^t\right)}^T$$

(27)

Then, the velocity of the ith particle at t+1 is

$$v_{id}{}^{t+1}=\omega v_{id}{}^t+c_1{r}_1\left(p_{id}{}^t-x_{id}{}^t\right)+c_2{r}_2\left(p_{gd}{}^t-x_{id}{}^t\right)$$

(28)

where, $\omega$ is inertial weight, $c_1,c_2$ are acceleration factor, normally 2, $r_1,r_2$ are random number between 0 and 1.

The position of the ith particle at t+1 is

$$x_{id}{}^{t+1}=x_{id}{}^t+v_{id}{}^{t+1}$$

(29)

This paper adopts PSO to optimize the initial weight and hidden deviation of ELM. The process is [37].

Step 1: A swarm is produced and maximal iterations are set. This paper sets the swarm at 20, and maximal iterations at 50. Each particle is composed of input weight and hidden deviation.

Step 2: For each individual particle, its output matrix is worked out by RELM algorithm. Then the root-mean-square error is obtained. The error is treated as fitness function. Each particle’s fitness value is computed. The individual and global optimal fitness particle’s parameter values are reserved.

Step 3: According to Equations (28) and (29), update the velocity and parameters of each particle.

Step 4: If maximal iterations are reached, computation ceases. The global optimal particle’s parameter values are considered as the parameters of the initial weight value and hidden deviation value of ELM. Otherwise, jump to Step 2.

3. Results and Analysis

This paper chooses the loophole quantity of a coal mine in seventeen consecutive months since January 2014 for the experiment. The original sequence is: X(0) =(777, 681, 715, 500, 762, 759, 627, 609, 392, 730, 690, 643, 550, 602, 604, 629, 611, 649, 493, 510, 466, 603, 581, 621). The first 22 numbers are training sample data, and the last 2 ones are testing sample data.

3.1. Predicting the Hidden Danger Quantity by GM (1,1) with Weakened Buffer Operator

According to artificial hidden danger identification characteristics and hidden danger development law, the mean of hidden danger quantity of every four months is treated as a new sequence value and a buffer operator sequence is generated according to Equation (10). The fitted values and errors of the original sequence and buffer operator sequence are listed in

Table 1. The GM (1,1) fitted value of original sequence and buffer operator sequence.

No.	Original Sequence				Buffer Operator Sequence
	Actual Value	Fitted Value	Residual Error	Relative Error	Actual Value	Fitted Value	Residual Error	Relative Error
1	777	— —	— —	— —	— —	— —	— —	— —
2	681	680.68	−0.32	0.05%	— —	— —	— —	— —
3	715	673.19	−41.81	5.85%	— —	— —	— —	— —
4	500	665.78	−165.78	33.16%	668.25	— —	— —	— —
5	762	658.45	−103.55	14.20%	664.5	673.303	8.803	1.325%
6	759	651.20	−107.80	14.20%	684	665.669	−18.331	2.680%
7	627	644.03	−17.03	2.72%	662	658.122	−3.878	0.586%
8	609	636.94	27.94	4.59%	689.25	650.660	−38.590	5.599%
9	392	629.93	237.93	60.70%	596.75	643.283	46.533	7.798%
10	730	623.00	−107.00	14.66%	589.5	635.989	46.489	7.886%
11	690	616.14	−73.86	10.70%	605.25	628.779	23.529	3.887%
12	643	609.36	−33.64	5.23%	613.75	621.650	7.900	1.287%
13	550	602.65	52.65	9.57%	653.25	614.601	−38.649	5.916%
14	602	596.02	−5.98	0.99%	621.25	607.633	−13.617	2.192%
15	604	589.46	−14.54	2.41%	599.75	600.744	0.994	0.166%
16	629	582.97	−46.03	7.32%	596.25	593.933	−2.317	0.389%
17	611	576.55	−34.45	5.64%	611.5	587.199	−24.301	3.974%
18	649	570.21	−78.79	12.14%	623.25	580.541	−42.709	6.853%
19	493	563.93	70.93	14.39%	595.5	573.959	−21.54	3.617%
20	510	557.72	47.72	9.36%	565.75	567.451	1.701	0.301%
21	466	551.59	85.59	18.37%	529.5	561.018	31.518	5.952%
22	603	545.51	−57.49	9.53%	518	554.657	36.657	7.077%

. The prediction values and errors of the original sequence and buffer operator sequence are listed in

Table 2. GM (1,1) prediction values of original sequence and buffer operator sequence.

No.	Prediction Value of Original Sequence				Prediction Value of Buffer Operator Sequence
	Actual Value	Fitted Value	Residual Error	Relative Error	Actual Value	Fitted Value	Residual Error	Relative Error
1	581	539.51	−41.49	7.14%	540	548.37	8.37	1.55%
2	621	533.57	−87.43	14.08%	567.75	542.15	−25.6	4.51%

.

The main parameters of GM (1,1) prediction model established by the original sequence are: development coefficient a = 0.011; gray actuation quantity b = 693.052; average relative error is 12.150%.

The main parameters of GM (1,1) prediction model established by the buffer operator sequence are: development coefficient a = 0.011; gray actuation quantity b = 684.769; average relative error is 3.749%.

The computation results show that, compared with the original sequence, the precision of the fitted values of the GM (1,1) prediction model established by the buffer operator sequence is obviously improved, and the oscillation is markedly weakened.

3.2. Modification of Residual Error Based on PSO-ELM

The experimental environment is Matlab2016b. Hidden set function is Sig. Regularization parameter uses 0.01. At first, we adopted RELM to optimize the hidden layer node. Then, we used RELM to optimize the initial swarm and maximal iterations of particle swarm algorithm. Finally, we employed PSO-RELM for experimental analysis.

(1) Determination of particle swarm parameters

The larger the number of swarms is, the more the iterations it takes, the smaller the resulting error is, the more complex the training devoted, and the longer the time it takes. A good model should be simple as far as possible while the precision is met. RELM is used to test the error of the model with different swarm numbers. The trained mean square error of the model is considered as an adaptive function.

Table 3. Time consumed and ultimate errors of the model with different swarm numbers.

Swarm number	15	20	25	30	35
Ultimate error	0.0658	0.0717	0.0579	0.0608	0.0573
Time consumed (s)	3.028	3.905	4.207	4.862	6.067

and Figure 3 list the favorable errors and consumed time in case of different swarm numbers. (The errors are normalized errors.)

It is observed from Figure 3 and Table 3 that when the initial swarms are increased to a given number, the increased precision slows down, but the time consumed for the training of the model is largely linearly increased. Meanwhile, when the number of iterations reaches 20, the precision remains basically unchanged. Therefore, this paper chooses initial swarm number at 25, and maximal iteration number at 25.

(2) Optimization of hidden layer node parameters

Generally speaking, the more the hidden layer nodes, the more extract the fitted values, and the more complex the network. However, excessive hidden layer nodes may cause over-learning of the network, worsening its normalization capability.

Table 4. Errors and consumed time of the model with different hidden layer nodes.

Number of Hidden Layer Nodes	Mean Square Error of Training	Predicted Mean Square Error	Consumed Time
5	23.28	2.44	2.924s
6	23.22	2.35	2.928s
7	22.88	2.46	2.931s
8	22.43	2.68	2.938s
9	21.78	2.36	2.938s
10	21.48	2.546	3.233s
11	21.01	2.40	3.499s
12	21.02	2.52	3.067s

separately tests the training error and consumed time of different hidden layer nodes.

It is shown in Table 4 that with an increasing number of hidden layer nodes, the errors of the training generally becomes smaller and smaller, and time consumed longer and longer, suggesting that the more complex the network is, the longer the computation takes. However, the prediction precision does not improve continuously with the increasing of the number of hidden layer nodes. This indicates that more hidden layer nodes will not necessarily bring about improved precision of the model.

(3) Experiments

According to the analysis above, the number of initial swarms is set at 25, the number of maximal iterations is set at 25, and the data of the residual errors of every three months slide to generate the sample. The experimental environment is Matlab2016b. SIG is chosen as the hidden layer aggregation function. Normalization parameter is chosen as 0.01, and the number of hidden layer nodes is chosen as 7. After a lot of experiments, the mean of any ten prediction models is taken as the correction value of residual error.

Table 5. Fitted values of average hidden danger quantity of a coal mine.

No.	Actual Value	Fitted Value of GM	Residual Error	Optimized ELM	Fitted Value of Combination Model	Residual Error of Combination Model	Relative Error
1	668.25	— —	— —	— —	— —	— —	— —
2	664.5	673.303	8.803	— —	— —	— —	— —
3	684	665.669	−18.331	— —	— —	— —	— —
4	662	658.122	−3.878	— —	— —	— —	— —
5	689.25	650.660	−38.590	32.80	683.46	−5.79	0.84%
6	596.75	643.283	46.533	−27.01	616.27	19.52	3.27%
7	589.5	635.989	46.489	−44.16	591.83	2.33	0.39%
8	605.25	628.779	23.529	−20.96	607.82	2.57	0.42%
9	613.75	621.650	7.900	−9.71	611.94	−1.81	0.29%
10	653.25	614.601	−38.649	39.46	654.06	0.81	0.12%
11	621.25	607.633	−13.617	9.76	617.39	−3.86	0.62%
12	599.75	600.744	0.994	8.48	609.22	9.47	1.58%
13	596.25	593.933	−2.317	−9.00	584.93	−11.32	1.90%
14	611.5	587.199	−24.301	24.49	611.69	0.19	0.03%
15	623.25	580.541	−42.709	24.54	605.08	−18.17	2.92%
16	595.5	573.959	−21.54	18.22	592.18	−3.32	0.56%
17	565.75	567.451	1.701	−4.06	563.39	−2.36	0.42%
18	529.5	561.018	31.518	−13.80	547.22	17.72	3.35%
19	518	554.657	36.657	−34.95	519.71	1.71	0.33%

lists the fitted values of the combination model.

Table 6. Prediction value of average hidden danger quantity of a coal mine.

No.	Actual Mean	GM(1,1) Prediction Value	Correction Value of Optimized ELM	Prediction Value of Combination Model	Residual Error	Relative Error
1	540	548.37	−4.07	544.3	4.3	0.79%
2	567.75	542.15	20.14	562.29	−5.46	0.96%

lists the prediction values of the combination model.

Table 7. Prediction values of hidden danger quantity of a coal mine.

No.	Actual Value	Prediction Value	Residual Error	Relative Error
1	581	598.2	17.2	2.96%
2	621	599.16	−21.84	3.52%

lists the prediction values recovered as the hidden danger quantity of every month.

From Table 5, it can be known that GM (1,1) prediction model using buffer operator optimization largely fits the law of average hidden danger quantity of the mine for every four months. However, the stability of the hidden danger quantity is poor, and fitting precision is not sound, and some errors are even more than 7%. The precision is obviously improved after PSO-ELM model is used to correct the residual error. From Table 6 it is evident that the precision of the prediction values obtained by the model is high. As indicated in Table 7, when the mean of hidden danger quantity is translated to the hidden danger quantity of every month, the prediction precision somewhat decreases, but the error is still below 5%. Hence, using gray neural network to optimize the combination model is more suitable for the hidden danger quantity of a coal mine.

4. Discussion

On the one hand, the accurate prediction of the number of hidden dangers can reveal the overall safety status of the mine, on the other hand, it can guide the formulation and implementation of the hidden trouble detection plan, which is of great significance to the coal mine safety production control. For the forecast model and its prediction results in this article, we will discuss from the following aspects:

First of all, from the above experimental results, we can see that the relative error of the prediction results is 12.5% when mean GM (1,1) prediction model is adopted, the relative error of the prediction results is 3.749% when buffer operator improvement GM (1,1) is adopted, and the relative error of the prediction results is 3.24% when the buffer operator improvement GM (1,1)+PSO-RELM model is used. The relative error of the last two prediction models is less than 5%, which shows that the two models can better predict the number of coal mine hazards.

Secondly, hidden danger is the unsafe state of human, material, environment, and management, which exists objectively, but it is found out manually. Therefore, the number of hidden dangers found out every month in the coal mine is limited by management level and other factors, and there is a certain subjectivity, which brings great difficulties to the prediction. As shown in Table 7, taking the average number of hidden dangers for three consecutive months as the prediction quantity, the relative error of the prediction result is reduced to 0.875%, which shows that it can weaken the random influence of the number of hidden dangers caused by human factors to a certain extent.

Finally, the number of hidden dangers on the one hand is affected by the development law of the hidden danger itself, on the other hand, it is also affected by other external factors, such as the change of geological conditions, the transformation of coal mining technology [38,39,40], etc. The model adopted in this paper belongs to the scope of the time series prediction model. The time series prediction model only shows good results in the simulation of the sequence itself. In other words, the prediction model used in this study cannot reflect the impact of other factors on the system in time. Ignoring the changes of internal factors may lead to inaccurate prediction. For example, when the mine is shut down for rectification, the number of hidden dangers will sharply reduce and when an accident occurs in a coal mine, the number of hidden dangers will increase significantly, which needs to be overcome in the future research.

5. Conclusions

(1) Prediction of the number of hidden dangers in coal mines plays an important role in the management of coal mine enterprises to better understand the situation of production safety in each area of the coal mine. It can judge the authenticity of the number of hidden dangers and reduce the number of hidden dangers. According to the mechanism analysis of the formation and development of hidden dangers, this paper knows that its essence is regular and predictable.

(2) The combination model of gray system theory and artificial neural network was used to predict the number of hidden dangers in coal mines. The model is based on the mean GM (1,1) model, which requires fewer samples, and is simple to calculate. The number of hidden dangers of average consecutive months is used as a buffer operator to weaken the oscillation of the original sequence, which improves the accuracy of model prediction.

(3) The regularization parameter term is used to optimize the ELM to improve the generalization ability of the model. Besides, the PSO is used to optimize the initial weight and hidden layer deviation of the ELM, so as to obtain a better training model, make the model more stable, and have higher prediction accuracy.

(4) The gray neural network prediction model is adopted, which has low requirements for data samples. It combines the advantages of two prediction methods, that is, it reflects the complex gray system behavior of hidden dangers in coal mines, and can dynamically predict.