基于WOA-ELM算法的矿井突水水源快速判别模型

董东林,陈昱吟,倪林根,李 源,覃华清,韦仙宇

(中国矿业大学(北京) 地球科学与测绘工程学院,北京 100083)

摘 要:我国是世界上煤炭生产量和消耗量最大的国家,但由于水文地质条件复杂,在煤矿生产过程中煤层顶底板突水事故频发,常常造成严重的经济损失和人员伤亡,快速精准地判别水害来源是矿井突水水害防治的关键步骤。基于河北开滦赵各庄矿的67个水样实测数据,将六种离子的物质的量浓度作为输入项,突水水源类型为其输出项,应用鲸鱼优化算法(WOA)改进极限学习机(ELM)形成WOA-ELM判别模型实现突水水源判别。研究结果表明:以往的单一极限学习机具有稳定性差的缺点,采用鲸鱼算法对其权值和阈值进行迭代寻优,通过环形包围、气幕袭击、随机搜索3种方式的鲸鱼优化算法对最优参数进行搜索,收敛速度快、全局搜索能力强。根据座头鲸捕食行为建立的数学模型,由于猎物(突水)位置不确定,WOA算法首先假设当前的最佳候选解是目标猎物位置或最靠近猎物的位置,然后通过随机产生向量A和概率p来决定鲸鱼更新位置的方式。当|A|>1时随机搜索猎物;当|A|<1时,以0.5为分界点,p<0.5选择环形包围模式,p>0.5则通过螺旋运动来更新位置。依据最低适应度值得到最优个体的位置,最终将输出的42个最优权值和阈值赋给ELM模型,对待测样本进行判别。通过对比,WOA-ELM判别模型在矿井突水水源识别中的准确率高达95%以上,与单一ELM模型相比,准确率提升了15%左右。与支持向量机模型(SVW)、粒子群优化的极限学习机(PSO-ELM)模型以及灰狼优化算法改进的极限学习机(GWO-ELM)模型等相比,该模型具有更快的收敛速度与更高的精度,稳定性和泛化能力也均得到提升。

关键词:鲸鱼优化算法(WOA);极限学习机(ELM);矿井突水;水源判别;气幕袭击

为快速分析水害事故产生的原因,众多学者通过构建高效精准的水源判别模型来指导矿井安全生产。目前水源判别方法有很多,如水温水位法、水化学分析法、数理分析法、神经网络法、可拓识别法、GIS理论法、支持向量机法等[1-11]。然而,这些方法都有其自身的优势与局限。胡伟伟等[12]运用同位素法准确地判定了抚顺老虎台矿区的突水来源,该方法具有判别速度快、结果可靠的优点,但此方法成本较高,需考虑现实因素[13-14]。张好等[15]将PCA与Logistic模型相结合,在水源识别上取得了很好的效果,但Logistic模型具有分类精度不高、易产生过拟合等缺点,主要适用于线性拟合,难以适应绝大多数线性不可分的复杂庞大数据[16]。人工神经网络具有较强的自组织、自学习能力和较强的容错性,评价结果更客观,但是对训练样本要求高,训练样本的合适与否会严重影响到最终的评判结果[17-18]。灰色关联分析法计算简单、易于理解,能客观地找到各因素之间的密切程度,降低人为干扰因素[19],但在确定各含水层离子标准时需要大量的数据来进行模糊聚类分析,所得标准只适用于聚类分析样本采集地区[20]。对于可拓识别法[21-24],冯琳等[25]发现对一些在各含水层分布差异较小的数据来说,使用该方法易造成误判。因此,我们需从数据的特征、模型的精度与泛化能力、过度拟合现象、人为影响程度、经济条件等方面入手,以找到解决问题的最佳方法。笔者针对赵各庄矿突水水源,综合考虑各方面因素,选择鲸鱼算法对极限学习机的参数进行优化,旨在准确、快速地对矿井突水水源进行判别分类。

1 矿区概况

开滦赵各庄矿兴建于1909年,位于河北省唐山市古冶区内,具有典型的华北型石炭—二叠纪煤田特征。赵各庄矿井田的地质条件极为复杂,自寒武系至第四纪,发育有五大含水层系。经过调查发现,赵各庄矿的各种突水事故中,奥陶系含水层发生突水的频率最大。1992年7月,开滦矿务局统一矿区奥陶纪碳酸盐岩含水层及其以上含水层,将其定为7组,绘制出含、隔水层示意图(图1)[26-29]。其中,含水层IV和V均可再细分为A,B两个含水层段。根据布罗茨基水化学分类法:第四纪冲积层孔隙水的水化学类型以HCO3-Na·K型为主;煤系地层砂岩裂隙水的水化学类型以HCO3-Na型为主;奥灰水的水化学类型以HCO3-Ca·Mg型为主。

图1 含、隔水层垂向示意
Fig.1 Vertical diagram of aquifer and waterproof layer

图1中编号为I,VI及VII的含水层为矿井间接充水含水层,主要特征如下:

(1)奥陶系岩溶承压含水层I。储水空间为古岩溶发育层和构造岩溶裂隙,富水性高、导水性强,具有较高承压水头。奥灰水的pH值为7.30~7.94。阳离子以Ca2+为主,其物质的量浓度为55.57~68.4 mmol/L,阴离子主要为其物质的量浓度为物质的量浓度为0.78~12.45 mmol/L,其物质的量浓度与水源的循环程度有关。

(2)A层以上顶板砂岩裂隙含水层VI。厚度较大,顶部为基岩风化带,砂岩经过强烈风化作用变成散砂状,黏土岩变成泥状,使裂隙弥合,含水性弱,相对阻滞了上部含水层向下补给。其水化学特征为pH值在7.04~8.60。阳离子以Na+为主,Ca2+,Mg2+的离子浓度相当,阴离子主要以为主。

(3)第四纪冲积层孔隙承压含水层VII。第四纪含水层以细、中砂为主,局部含砂、砾层,厚度不稳定,井田内由北向南、由西向东逐渐变厚,含水性差,对矿井安全生产基本没有影响。在井田范围内,冲积层与各时代的基岩含水层之间存在一定的水力联系。但井田内冲积层很薄,对煤层开采影响不大。其水化学特征为pH值在7.3~8.3,矿化度在165~353 mg/L。阳离子以Ca2+为主,阴离子以为主,水化学类型为HCO3-Ca型。

图1中编号为II,III,IV及V的含水层为矿井直接充水含水层,主要特征如下:

(1)14煤层~唐山灰岩砂岩裂隙承压含水层II。以粉砂岩为主,夹有泥岩、泥灰岩、灰岩,裂隙较发育。唐山石灰岩一般厚2.5 m左右,在井下钻探和开拓工程中揭露时,涌水量很小。

(2)12~14煤层砂岩裂隙承压含水层III。以中、粉砂岩为主,岩性致密坚硬,裂隙较发育,含水性不均匀。14煤层顶板为青灰色中砂岩,发育稳定,含水性相对较强。12煤层底板以灰色粉砂岩为主,含少量裂隙水,在裂隙发育或破碎带较常见出水,但涌水量不大。

(3)5~12煤层砂岩裂隙承压含水层IV。该层划分为2个含水层段(IVA,IVB),其中IVB段以中粉砂岩为主,含水性较强,在巷道揭露过程中均有不同程度的淋水现象。当煤层开采后,顶板垮落产生裂隙,导致大淋水或涌水,对7煤层的开采影响较大。IVA段以中、粗砂岩为主,岩性致密坚硬,裂隙较发育,含水性不均匀。

(4)5煤层顶板砂岩裂隙承压含水层V。该层划分为2个含水层段(VA,VB),以中、粉砂岩为主,夹粗砂岩,坚硬,裂隙发育,其含水性不仅和岩性有关,同时也受补给条件和构造断裂的控制,属含水性较强含水层。该层为煤层直接顶板,当5煤层开采后,顶板垮落,会有较大涌水出现。

这4个煤系砂岩裂隙含水层水的pH值为7.04~9.00,阳离子主要为Na+,其物质的量浓度一般在50 mmol/L以上,最大达98.85 mmol/L,Ca2+,Mg2+离子物质的量浓度相当,其物质的量浓度为20 mmol/L左右,具体各煤层顶底板水存在一定差异;阴离子主要为其物质的量浓度一般在70 mmol/L以上,最大可达93.86 mmol/L。Cl-离子物质的量浓度相对较小。

图2 ELM训练模型
Fig.2 ELM training model

2 WOA-ELM模型的基本原理

2.1 极限学习机的基本原理

极限学习机(Extreme Learning Machine,ELM)由黄广斌等[30]提出。该算法包括输入层、隐含层、输出层(图2),通过随机产生权值和阈值,然后利用最小二乘法解方程组得到输出权值。ELM不依赖于梯度下降,学习速度比BP(Back-Propagation)算法等传统前馈网络快,对于大量及非线性样本也具有很好的适用性[31-32]

设模型的输入层神经元个数为n,隐含层神经元个数为L,输出层神经元个数为m。给定N个不同的训练样本xjj=1,…,N,通过激活函数g(aibixj)可以将样本映射到另一个特征空间。对于通过映射得到的矩阵H有如下表达式:

=T

(1)

(2)

式中,H为神经网络的隐含层输出矩阵;T为网络的输出矩阵;β为隐含层节点与输出层节点的连接权值矩阵;ai为由输入层第1~n个节点与隐含层第i个节点的连接权值构成的矩阵,i=1,…,L,表示第i个隐含层神经元;bi为第i个隐含层节点的阈值。

在新的特征空间下,利用最小二乘法便可求解出最优输出权值

(3)

式中,H+H的Moore-Penrose广义逆矩阵。

2.2 鲸鱼优化算法的基本原理

2.2.1 包围捕食阶段(Encircling prey)

该算法是根据座头鲸捕食行为建立的数学模型。由于猎物位置不确定,WOA算法首先假设当前的最佳候选解是目标猎物位置或最靠近猎物的位置,然后会不断更新其位置[33]。这种行为用下列方程表示:

D=|CX*(t)-X(t)|

(4)

X(t+1)=X*(t)-AD

(5)

其中,D为包围步长;X(t+1)为下一次迭代后解的位置向量;X*(t)为当前的最优解的位置向量;X(t)为当前解的位置向量;t为当前迭代次数;AC为随机系数向量;每次迭代过程中有更优解出现时就需要更新X*(t)。

AC由如下公式得出:

(6)

C=2r

(7)

其中,a为控制参数,在迭代过程中从2线性地下降至0;tmax为最大迭代次数,本例中取值为300;r为[0,1]的随机向量。

2.2.2 气幕袭击阶段(Bubble-net attacking method)

(1)收缩包围机制。新的个体位置可以定义在目前的鲸鱼个体和最佳鲸鱼个体间的任何一个位置[34],参见式(5)。

(2)螺旋式更新位置。计算鲸鱼和猎物之间的距离D′,构建方程:

X(t+1)=D′eblcos(2πl)+X*(t)

(8)

其中,D′=|X*(t)-X(t)|; b为常数,定义了对数螺线的形状,l为[-1,1]的随机数。

本模型以0.5为分界点,通过产生一个0~1的随机概率p来决定鲸鱼捕食方式。当|A|<1,概率p<0.5时,选择收缩包围机制;当|A|<1,概率p>0.5时,选择螺旋运动更新位置。

2.2.3 猎物搜索阶段(Searching prey)

当|A|>1时进入随机搜索阶段,随机地选择鲸鱼来更新位置,而不是根据已知的最优的鲸鱼个体。当达到最大迭代次数以后算法即终止[35-36]。随机搜索的数学模型如下:

D=|CXrand-X(t)|

(9)

X(t+1)=Xrand-AD

(10)

式中,Xrand为随机位置向量。A决定了座头鲸是否进入随机搜索状态。

3 WOA-ELM算法应用

3.1 数据选取

根据《赵各庄水文地质类型划分报告》,1959—2016年矿区突水类型以奥灰水害为主(发生次数占比44%),其次是老空水害(占比36%),砂岩水害和地表水害也时有发生。鉴于影响大的含水层水样数量多、影响小的含水层水样数量少的特点进行了训练与测试样本的筛选,选取老空水、奥灰水(来源于I含水层)、13煤层砂岩裂隙水(来源于III含水层)、12煤层砂岩裂隙水(来源于IV含水层的A段),共4种水样类型(67个水样)进行训练与测试,具体水样特征见表1,2。表1中,数字1为老空水,数字2为奥灰水,数字3为13煤层砂岩裂隙水,数字4为12煤层砂岩裂隙水,表2,3中的数字含义同表1。

3.2 模型建立过程

利用WOA优化ELM的流程如下:首先对原始数据进行归一化,然后设置ELM网络的节点个数和待优化参数的个数,初始化鲸鱼数量和种群适应度值,通过鲸鱼算法不停地迭代寻优,最后得到的适应度值最小时对应的个体即为最优解,将此最优参数赋值给ELM网络,对ELM网络进行训练后应用于水源判别中(图3)。

具体可分为以下4个步骤:

(1)数据归一化处理。

表1 赵各庄矿水样(训练集)
Table 1 Water sample data of Zhaogezhuang mine (training set)

各实测离子物质的量浓度/(mmol·L-1)Na+Ca2+Mg2+Cl-SO2-4HCO-3水样类型11.0566.3422.231.5081.2817.2114.3654.8340.662.6866.4930.83110.7054.6434.1811.3170.1417.5415.8648.4545.061.7981.8416.3718.9153.4337.468.4560.4731.0816.0554.1338.711.8176.5721.6215.6951.7041.831.6581.4116.85118.0240.940.821.7581.8316.2516.9544.4448.091.4485.5513.0012.5450.8545.592.4683.7613.7816.0949.6343.921.5881.7516.6717.6249.7542.401.6681.8816.4617.8648.5943.201.5482.3916.0714.7261.3533.833.2648.6048.1326.1458.4535.163.2249.5647.21210.4664.2824.9613.0615.7160.34211.1567.7921.0611.3619.9358.08211.8866.2021.9113.7725.3151.4828.8065.2724.1811.1614.7461.63212.4960.8524.9812.5723.6252.27210.6951.1737.619.5314.6369.0825.9754.0139.529.2213.6676.6438.1355.9235.519.3611.9778.1839.7451.2838.298.4513.9576.9938.3652.1439.129.4113.5776.9631.5467.1031.18.7614.5872.1439.8552.6136.578.4415.7175.23310.3151.0938.008.6514.6676.0839.4952.4937.58.7712.8877.7937.7155.3136.378.6713.1777.7939.4150.4338.878.4214.5976.60312.9050.3136.229.0412.3478.23315.1147.8136.858.4013.8177.4938.0860.9131.018.5615.4472.9448.4359.9731.558.9516.1671.8048.1658.0833.768.8714.9172.26412.6751.1735.7710.9912.7174.1147.7757.1334.889.0313.3274.1648.8057.5033.629.9415.3070.8144.6358.0137.3610.9617.0369.4743.7861.8834.325.7116.7173.4249.4059.5930.959.5415.3071.2342.9261.3035.789.5615.8571.0547.3360.3432.338.6116.4070.84411.9556.9930.879.0114.9373.8446.8062.7330.138.5716.8870.6145.5864.3629.948.816.4271.254

表2 赵各庄矿水样(测试集)
Table 2 Water sample data of Zhaogezhuang mine (test set)

各实测离子物质的量浓度/(mmol·L-1)Na+Ca2+Mg2+Cl-SO2-4HCO-3水样类型11.8149.1837.873.0871.8721.70112.3652.1134.832.9272.3022.78111.7750.8536.882.5178.2219.24110.3749.7339.472.1283.9113.9019.1650.2340.172.1382.7815.0916.9357.8639.664.4753.0141.8626.5259.3840.445.6151.8041.7025.5360.2938.494.6050.7444.6225.8764.6829.458.2520.9368.4826.2658.5339.735.1949.2443.23212.9648.2638.249.6513.8876.1739.6952.5037.169.4810.8679.58312.6347.1639.199.2014.7775.28314.4946.6237.709.4113.7476.91313.2849.6737.8510.0312.4377.0535.9261.8931.728.6115.9570.9948.3558.1133.548.7515.5471.9249.6060.3930.018.6516.3871.4348.4358.2130.858.7016.1271.2249.2757.2330.508.7715.9270.804

选取六种代表性的离子作为输入特征,由于采集的样本数据变化范围较大,且激活函数在(0,1)以外很平缓,区分度太小,故首先需要对输入数据进行处理,利用MATLAB软件中的mapminmax函数将其归一到[0,1]内。公式如下:

(11)

式中,x为样本某一特征的输入值;xmin为样本中输入值的最小值;xmax为样本中输入值的最大值;y为归一化后输出的值。

(2)设置模型参数。

① ELM网络参数:根据输入的数据特征,设置ELM模型隐含层的节点个数为6,故待优化参数Z根据下式计算得出:

Z=nL+m

(12)

在本次训练中,nLm均为6。

② 鲸鱼优化算法的参数:设置算法的下界、上界[-1,1],最大迭代次数为300,鲸鱼数量为30,并初始化每个鲸鱼的位置。

(3)迭代寻优。

根据训练过程中的误判率计算所有个体的适应度值后选取最优个体,然后根据最低适应度值更新最优个体的位置。在到达设置的迭代次数之前不断地进行寻优,直到达到误判率低于设定值或达到最大迭代次数。

图3 WOA-ELM模型流程
Fig.3 Flow chart of WOA-ELM model

(4)水源判别。

将输出的42个最优权值和阈值赋给ELM模型,对待测样本进行判别。

3.3 水源判别结果

利用MATLAB软件,将每次迭代的适应度值绘制成图,对训练样本和测试样本进行回判;同时将上述样本代入传统ELM模型中,绘制同样内容。结果如下。

图4 WOA-ELM算法迭代轨迹
Fig.4 Iteration trajectory of WOA-ELM algorithm

由图4可以看出在迭代次数不到20次的时候,WOA-ELM模型的适应度就已经降到最低值,之后的迭代适应度函数并没有变化,说明该模型的误判率已经达到最低。对于47个训练样本,由图5(a)和图6(a)可以看出,WOA-ELM模型的判别准确率高达95.74%,仅有2个样本被误判,2个样本均属于煤系地层砂岩裂隙水,含水层裂隙发育,构造条件复杂,需进一步研究水样的混合效应;而传统ELM方法只有85.11%。对于20个测试样本,由图5(b)和图6(b)可以看到,WOA-ELM模型准确率高达95%;而传统模型只有70%,准确率提高了25%,可见用WOA算法优化ELM模型十分成功,不但具有简便易行,迭代次数少的优点,还保证了极高的准确率。但由于采动效应、矿压与构造断裂等作用,12煤层至奥灰含水层间的裂隙发育程度不一,混合水效应为水源的识别带来一定程度上的困难。

图5 训练集与测试集的判别结果(WOA-ELM)
Fig.5 Discriminant result of training set and test set

图6 训练集与测试集的判别结果(ELM)
Fig.6 Discriminant result of training set and test set

在多次进行MATLAB仿真训练之后发现,由于单一ELM的权值和阈值具有随机性,训练后的网络判别效果并不稳定,而WOA-ELM模型则对权值和阈值进行了优化,在不断迭代过程中找到最优参数,提高了单一ELM模型的稳定性和泛化能力。

4 模型对比

为了进一步的体现WOA-ELM相结合的优势,采用不同方法对水样进行对比判别,并将结果列出(表3)。对于单一ELM、Fisher判别法[37]以及支持向量机(SVM)[38],其水样判别的准确率都低于WOA-ELM模型。

表3 各种水源判别模型结果对比
Table 3 Comparison of results of various water source discriminant models

实际水样类型不同模型预测的水样类型单一ELMFisherSVMPSO-ELMGWO-ELMWOA-ELM12111111211111111111111111111111111242222224223222422222224244424233223333333333333333333333333333344233344444444444444444444444344444434444测试样本准确率/%708090859595

此外,还对ELM模型进行了与PSO[39]和GWO[40]2种优化算法的结合,结果表明WOA-ELM模型的准确率(95%)高于PSO-ELM模型的准确率(85%),但是和GWO-ELM模型相差无几。通过对表3以及3种结合模型的迭代曲线的对比观察发现(图7),尽管GWO和WOA两种优化模型的测试集判别准确率相同都为95%,但是WOA-ELM模型的最低误判率(即适应度的最低值)小于GWO-ELM模型,说明在大量数据情况下,WOA-ELM模型的准确率会更高;对于PSO-ELM模型来说尽管它的适应度曲线更低,但是它的实际判别准确度并不高于WOA-ELM模型,可能是由于过拟合导致的,同时经过多次仿真模拟发现,PSO-ELM模型同样存在不稳定性,多次运行的结果并不完全一致。

图7 不同优化算法迭代曲线
Fig.7 Iteration curves of different optimization algorithms

5 结 论

(1)通过对赵各庄矿历史突水事件及水文地质条件分析,认为老空水、奥灰水、13煤层砂岩裂隙水、12煤层砂岩裂隙水为该矿的主要突水水源。奥灰水阳离子以Ca2+为主,阴离子主要以为主,其物质的量浓度与水源的循环程度有关。煤系砂岩裂隙含水层水阳离子主要以Na+为主,Ca2+,Mg 2+离子物质的量浓度相当,阴离子主要以为主,Cl-离子物质的量浓度相对较小。

(2)运用鲸鱼优化算法(WOA)对传统极限学习机(ELM)的权值和阈值进行了优化,建立了WOA-ELM水源判别模型,以矿区的不同含水层的水样为样本,以6项常规水化学指标作为识别对象,成功的判别了该矿的突水水源,准确率高达95%。

(3)由于煤系含水层裂隙发育,构造条件复杂,其水样混合造成了识别的难度,需进一步研究水样的混合效应。

参考文献(References):

[1] QU Shen,WANG Guangcai,SHI Zheming,et al.Using stable isotopes(δD,δ18O,δ34S and 87Sr/86Sr) to identify sources of water in abandoned mines in the Fengfeng coal mining district,northern China[J].Hyrdogeology Journal,2018,26(5):1443-1453.

[2] WANG Dandan,SHI Longqing.Source identification of mine water inrush:A discussion on the application of hydrochemical method[J].Arabian Journal of Geosciences,2019,12(2):58.

[3] 石磊,徐楼英.基于水化学特征的聚类分析对矿井突水源判别[J].煤炭科学技术,2010,38(3):97-100,124.

SHI Lei,XU Louying.Prediction of mine water inrush sources based on cluster analysis of hydrogeochemical features[J].Coal Science and Technology,2010,38(3):97-100,124.

[4] WU Qiang,MU Wenping,XING Yuan,et al.Source discrimination of mine water inrush using multiple methods:A case study from the Beiyangzhuang Mine,Northern China[J].Bulletin of Engineering Geology and the Environment,2019,78(1):469-482.

[5] 周健,史秀志,王怀勇.矿井突水水源识别的距离判别分析模型[J].煤炭学报,2010,35(2):278-282.

ZHOU Jian,SHI Xiuzhi,WANG Huaiyong.Water-bursting source determination of mine based on distance discriminant analysis model[J].Journal of China Coal Society,2010,35(2):278-282.

[6] 张许良,张子戌,彭苏萍.数量化理论在矿井突(涌)水水源判别中的应用[J].中国矿业大学学报,2003(3):42-45.

ZHANG Xuliang,ZHANG Zixu,PENG Suping.Application of the second theory of quantification in identifying gushing water sources of coal mines[J].Journal of China University of Mining & Technology,2003(3):42-45.

[7] 王连国,宋扬.煤层底板突水组合人工神经网络预测[J].岩土工程学报,2001,23(4):502-505.

WANG Lianguo,SONG Yang.Prediction of water inrush combination of coal seam floor by artificial neural network[J].Chinese Journal of Geotechnical Engineering,2001,23(4):502-505.

[8] 董东林,武强,钱增江,等.榆神府矿区水环境评价模型[J].煤炭学报,2006,31(6):776-780.

DONG Donglin,WU Qiang,QIAN Zengjiang,et al.Water environment assessment model in Yushenfu mining area[J].Journal of China Coal Society,2006,31(6):776-780.

[9] 马雷,钱家忠,赵卫东.基于GIS和水质水温的矿井突水水源快速判别[J].煤田地质与勘探,2014,42(2):49-53.

MA Lei,QIAN Jiazhong,ZHAO Weidong.An approach for quickly identifying water-inrush source of mine based on GIS and groundwater chemistry and temperature[J].Coal Geology & Exploration,2014,42(2):49-53.

[10] 董东林,孙文洁,朱兆昌,等.基于GIS-BN技术的范各庄矿煤12底板突水态势评价[J].煤炭学报,2012,37(6):999-1004.

DONG Donglin,SUN Wenjie,ZHU Zhaochang,et al.Water-inrush assessment of coal 12 floor using a GIS-based bayesian network for Fangezhuang Coal Mine with collapse column[J].Journal of China Coal Society,2012,37(6):999-1004.

[11] 闫志刚,杜培军,郭达志.矿井涌水水源分析的支持向量机模型[J].煤炭学报,2007,32(8):842-847.

YAN Zhigang,DU Peijun,GUO Dazhi.SVM models for analyzing the headstreams of mine water inrush[J].Journal of China Coal Society,2007,32(8):842-847.

[12] 胡伟伟,马致远,曹海东,等.同位素与水文地球化学方法在矿井突水水源判别中的应用[J].地球科学与环境学报,2010,32(3):268-271.

HU Weiwei,MA Zhiyuan,CAO Haidong,et al.Application of isotope and hydro-geochemical methods in distinguishing mine bursting water source[J].Journal of Earth Sciences and Environment,2010,32(3):268-271.

[13] 张雁.水化学分析法判定矿井突水水源及示踪法验证[J].煤炭科学技术,2014,42(10):98-100,105.

ZHANG Yan.Hydro-chemical analysis method applied to determine mine water inrush sources and verification of tracing method[J].Coal Science and Technology,2014,42(10):98-100,105.

[14] DOU Huiping,MA Zhiyuan,CAO Haidong,et al.Application of isotopic and hydro-geochemical methods in identifying sources of mine inrushing water[J].Mining Science and Technology(China),2011,21(3):319-323.

[15] 张好,姚多喜,鲁海峰,等.基于主成分分析的多项Logistic回归模型的突水水源判别研究[J].高校地质学报,2017,23(2):366-372.

ZHANG Hao,YAO Duoxi,LU Haifeng,et al.Study on water source discrimination based on multinomial logistic regression model using principal component analysis[J].Geological Journal of Universities,2017,23(2):366-372.

[16] 武强,张波,赵文德,等.煤层底板突水评价的新型实用方法:基于GIS的ANN型、证据权型、Logistic回归型脆弱性指数法的比较[J].煤炭学报,2013,38(1):21-26.

WU Qiang,ZHANG Bo,ZHAO Wende,et al.A new practical methodology of coal seam floor water burst evaluation:The comparison study among ANN,the weight of evidence and the logistic regression vulnerable index method based on GIS[J].Journal of China Coal Society,2013,38(1):21-26.

[17] 吴岩,余智超.神经网络在矿井突水水源判别中的应用[J].工矿自动化,2011,37(10):60-62.

WU Yan,YU Zhichao.Application of neural net work in water source distinguishing of mine water inrush[J].Industry and Mine Automation,2011,37(10):60-62.

[18] 魏永强,梁化强,任印国,等.神经网络在判别煤矿突水水源中的应用[J].江苏地质,2004,28(1):36-38.

WEI Yongqiang,LIANG Huaqiang,REN Yinguo,et al.Application of neural network in discriminating water bursting source of the coal mine[J].Jiangsu Geology,2004,28(1):36-38.

[19] 岳梅.判断矿井突水水源灰色系统关联分析的应用[J].煤炭科学技术,2002,30(4):37-39.

YUE Mei.Application of grey system relational analysis to judging the source of water bursting in mines[J].Coal Science and Technology,2002,30(4):37-39.

[20] 杨永国,黄福臣.非线性方法在矿井突水水源判别中的应用研究[J].中国矿业大学学报,2007(3):283-286.

YANG Yongguo,HUANG Fuchen.Water source determination of mine inflow based on non-linear method[J].Journal of China University of Mining & Technology,2007(3):283-286.

[21] 门宝辉,梁川.城市环境质量综合评价物元模型及其应用[J].系统工程理论与实践,2003,23(3):134-139.

MEN Baohui,LIANG Chuan.Establishment and application of matter element model for evaluating urban environmental quality[J].Systems Engineering Theory & Practice,2003,23(3):134-139.

[22] 刘立民,刘汉龙,连传杰.建筑物采动损坏等级评定的物元模型及其应用[J].煤炭学报,2004,29(1):17-21.

LIU Limin,LIU Hanlong,LIAN Chuanjie.Matter-element model for buildings’ damage grade evaluation and it’s application caused by mining subsidence[J].Journal of China Coal Society,2004,29(1):17-21.

[23] HE Yongxiu,DAI Aiying,YANG Weihong,et al.Risk assessment of urban network planning in China based on the matter-element model and extension analysis[J].International Journal of Electrical Power and Energy Systems,2010,33(3):775-782.

[24] 张瑞钢,钱家忠,马雷,等.可拓识别方法在矿井突水水源判别中的应用[J].煤炭学报,2009,34(1):33-38.

ZHANG Ruigang,QIAN Jiazhong,MA Lei,et al.Application of extension identification method in mine water-bursting source discrimination[J].Journal of China Coal Society,2009,34(1):33-38.

[25] 冯琳,李凤莲,张雪英,等.改进可拓识别方法及其在突水水源判别中的应用[J].工矿自动化,2015,41(2):61-65.

FENG Lin,LI Fenglian,ZHANG Xueying,et al.Improved extension identification method and its application to water bursting source discrimination[J].Industry and Mine Aut-omation,2015,41(2):61-65.

[26] 武强,杨柳,朱斌,等.“脆弱性指数法”在赵各庄矿底板突水评价中的应用[J].中国煤炭地质,2009,21(6):40-44.

WU Qiang,YANG Liu,ZHU Bin,et al.Application of vulnerability index method in coal floor water bursting evaluation in Zhaogezhuang coal mine[J].Coal Geology of China,2009,21(6):40-44.

[27] 李建民,朱斌,武强.赵各庄矿大倾角煤层综放开采突水条件的力学机理分析[J].煤炭学报,2007,32(5):453-457.

LI Jianmin,ZHU Bin,WU Qiang.Mechanical mechanism analysis of water inrush condition on mechanized top-coal caving in steep seam in Zhaogezhuang coal mine[J].Journal of China Coal Society,2007,32(5):453-457.

[28] 才向军,韩瑞刚,孟璐,等.赵各庄矿闭坑地下水安全警戒水位控制研究[J].煤炭工程,2020,52(9):116-121.

CAI Xiangjun,HAN Ruigang,MENG Lu,et al.Safe and warning water level control of closed pit groundwater in Zhaogezhuang Mine[J].Coal Engineering,2020,52(9):116-121.

[29] 牛培峰,吴志良,马云鹏,等.基于鲸鱼优化算法的汽轮机热耗率模型预测[J].化工学报,2017,68(3):1049-1057.

NIU Peifeng,WU Zhiliang,MA Yunpeng,et al.Prediction of steam turbine heat consumption rate based on whale optimization algorithm[J].CIESC Journal,2017,68(3):1049-1057.

[30] HUANG Guangbin,ZHU Qinyu,SIEW CheeKheong.Extreme lea-rning machine:Theory and applications[J].Neurocomputing,2006,70(1-3):489-501.

[31] 褚鼎立,陈红,王旭光.基于自适应权重和模拟退火的鲸鱼优化算法[J].电子学报,2019,47(5):992-999.

CHU Dingli,CHEN Hong,WANG Xuguang.Whale optimization algorithm based on adaptive weight and smulated annealing[J].Acta Electronica Sinica,2019,47(5):992-999.

[32] HUANG Guangbin,ZHOU Hongming,DING Xiaojian,et al.Extreme learning machine for regression and multiclass classification[J].A Publication of The IEEE Systems,Man,and Cybernetics Society,2012,42(2):513-529.

[33] 边凯.赵各庄矿深部煤层底板突水危险性与断裂滞后突水评价[D].北京:中国矿业大学(北京),2015.

BIAN Kai.Risk assessment of water inrush from deep coal floor and lagging water inrush from fractures in Zhaogezhuang mine[D].Beijing:China University of Mining and Technology(Beijing),2015.

[34] 徐睿,梁循,齐金山,等.极限学习机前沿进展与趋势[J].计算机学报,2019,42(7):1640-1670.

XU Rui,LIANG Xun,QI Jinshan,et al.Advances and trends in extreme learning machine[J].Chinese Journal of Computers,2019,42(7):1640-1670.

[35] 王珂珂,牛东晓,甄皓,等.基于WOA-ELM模型的中国碳排放预测研究[J].生态经济,2020,36(8):20-27.

WANG Keke,NIU Dongxiao,ZHEN Hao,et al.Forecast of carbon emissions in China based on WOA-ELM model[J].Ecological Economy,2020,36(8):20-27.

[36] SAEED Samadianfard,SAJJAD Hashemi,KATAYOUN Kargar,et al.Wind speed prediction using a hybrid model of the multi-layer perceptron and whale optimization algorithm[J].Energy Reports,2020,6:1147-1159.

[37] 陈红江,李夕兵,刘爱华,等.用Fisher判别法确定矿井突水水源[J].中南大学学报(自然科学版),2009,40(4):1114-1120.

CHEN Hongjiang,LI Xibing,LIU Aihua,et al.Identifying of minewater inrush sources by Fisher discriminant analysis method[J].Journal of Central South University,2009,40(4):1114-1120.

[38] 钱家忠,潘婧,赵卫东,等.基于SVM的潘三矿B8组与C13组煤开采中突水水源判别模型[J].系统工程理论与实践,2011,31(12):2425-2430.

QIAN Jiazhong,PAN Jing,ZHAO Weidong,et al.The discrimination model based on SVM of inrushed water source in the coal mining on the level of B8 and C13[J].Systems Engineering Theory & Practice,2011,31(12):2425-2430.

[39] 杨维,李歧强.粒子群优化算法综述[J].中国工程科学,2004,6(5):87-94.

YANG Wei,LI Qiqiang.Survey on particle swarm optimization algorithm[J].Strategic Study of CAE,2004,6(5):87-94.

[40] SEYEDALI Mirjalili,MIRJALILI Seyed Mohammad,LEWIS Andrew.Grey wolf optimizer[J].Advances in Engineering Software,2014,69(3):46-61.

Fast discriminant model of mine water inrush source based on WOA-ELM algorithm

DONG Donglin,CHEN Yuyin,NI Lingen,LI Yuan,QIN Huaqing,WEI Xianyu

(School of Geosciences & Surveying Engineering,China University of Mining and Technology(Beijing),Beijing 100083,China)

Abstract:China is the country with the largest coal production and consumption in the world.However,due to the complex hydrogeological conditions,water inrush accidents occur frequently in the coal seam roof and floor during coal mining,causing serious economic losses and casualties.Identifying the source of water hazards rapidly and accurately is the key step in mine water inrush prevention and control.Based on the measured data of 67 water samples from Zhaogezhuang mine in Kailuan,the ion concentrations of six ions,including taken as input items,and the type of water inrush source was taken as output item.The WOA-ELM discriminant model was developed by using the Whale Optimization Algorithm (WOA) to improve ELM for achieving the water inrush source discrimination.The results show that the single extreme learning machine has the disadvantage of poor stability,and the whale algorithm is used to iteratively optimize its weight and threshold values.The whale optimization algorithm using three methods of ring encirclement,air curtain attack,and random search is used to search for optimal parameters,with fast convergence speed and strong global search ability.A mathematical model is established based on the predation behavior of humpback whales.Due to the uncertainty of prey position (water inrush),the WOA algorithm first assumes that the current best candidate solution is the target prey position or the closest position to the prey,and then randomly generates the vector A and the probability p to determine the way that the whales update the position.When |A|>1,random search prey is chosen.When the |A|<1,the probability of 0.5 is taken as the cut-off point.When p is less than 0.5,the enveloping mode is selected.Otherwise,the position is updated by spiral motion.According to the lowest fitness value,the position of the optimal individual is obtained.Finally,the output 42 optimal weights and thresholds are assigned to the ELM model to identify the samples to be tested.By comparison,the accuracy of the WOA-ELM discriminant model in identifying the mine water inrush source is up to 95%,about 15% higher than that of the single ELM model.Compared with the Support Vector Machine (SVM) model,the Particle Swarm Optimization-Extreme Learning Machine (PSO-ELM) model and the Gray Wolf Optimization-Extreme Learning Machine (GWO-ELM) model,the WOA-ELM model has a faster convergence speed and higher precision.The stability and generalization ability of the model are also improved.

Key words:whale optimization algorithm;extreme learning machine;bubble-net attacking;mine water inrush;water source discrimination

中图分类号:TD745

文献标志码:A

文章编号:0253-9993(2021)03-0984-10

移动阅读

收稿日期:2020-11-24

修回日期:2021-01-22

责任编辑:黄小雨

DOI:10.13225/j.cnki.jccs.2020.1827

基金项目:国家自然科学基金资助项目(41972255);国家重点研发计划资助项目(2017YFC0804104);国家自然科学基金联合基金资助项目(U1710258)

作者简介:董东林(1969—),男,陕西乾县人,教授,博士。E-mail:ddl@cumtb.edu.cn

引用格式:董东林,陈昱吟,倪林根,等.基于WOA-ELM算法的矿井突水水源快速判别模型[J].煤炭学报,2021,46(3):984-993.

DONG Donglin,CHEN Yuyin,NI Lingen,et al.Fast discriminant model of mine water inrush source based on WOA-ELM algorithm[J].Journal of China Coal Society,2021,46(3):984-993.