基于迁移学习算法的深部爆破振动速度预测

doi:10.16265/j.cnki.issn1003-3033.2023.06.0337

摘要/Abstract

摘要：

为了更好地预测深部矿山爆破振动速度,针对深部爆破振动速度预测中存在的样本量小、数据分布同浅部爆破不同的问题,将浅部地下矿山爆破数据中有用的知识迁移至深部矿山爆破振动速度预测模型中,提出一种逻辑回归迁移学习算法(LR-TrAdaboost),提升模型的样本容量及预测准确率;以某铜矿深部爆破振动速度预测为研究对象,结合该铜矿27条深部爆破数据以及梅山矿等5个地下金属矿204条浅部爆破数据,利用支持向量回归机(SVR)、回归迁移学习算法(TrAdaboost-R²)以及LR-TrAdaboost算法分别进行预测和对比。结果表明:3种算法的模型分数分别为0.24、0.38、0.81,均方根误差(RMSE)分别为0.152、0.107、0.06,LR-TrAdaboost算法预测误差相比SVR、TrAdaboost-R²分别降低了60.5%、43.9%;同时,LR-TrAdaboost在迭代次数为50时已经收敛,而TrAdaboost-R²在迭代次数100次后才收敛,收敛速度前者是后者的2倍;LR-TrAdaboost算法的预测性能更好。

关键词: 深部爆破振动速度, 逻辑回归迁移学习算法(LR-TrAdaboost), 支持向量回归机(SVR), 机器学习

Abstract:

In order to better predict the blasting vibration velocity of deep mines, aiming at the problems of small sample size and different data distribution in the prediction of blasting vibration velocity of deep mines, the useful knowledge in the blasting data of shallow underground mines was transferred to the prediction model of blasting vibration velocity of deep mines, and an LR-TrAdaboost (transfer learning) algorithm was proposed to improve the sample size and prediction accuracy of the model. Taking the prediction of deep blasting vibration velocity of a copper mine as the research object, combined with 27 deep blasting data of the copper mine and 204 shallow blasting data of five underground metal mines such as Meishan Mine, SVR, TrAdaboost-R² and LR-TrAdaboost algorithms were used for prediction and comparison respectively. The model scores of the three algorithms are 0.24, 0.38 and 0.81, and the root mean square error(RMSE) is 0.152, 0.107 and 0.06, respectively. Compared with SVR and TrAdaboost-R², the prediction error of the LR-TrAdaboost algorithm reduces by 60.5% and 43.9%, respectively. At the same time, LR-TrAdaboost converged when the number of iterations is 50, while TrAdaboost-R² converged after the number of iterations is 100, and the convergence rate is twice that of the latter. Research shows that the LR-TrAdaboost algorithm has better prediction performance.

Key words: vibration velocity of deep blasting, prediction error, support vector regression(SVR), machine learning

张西良, 焦灏恺, 李二宝. 基于迁移学习算法的深部爆破振动速度预测[J]. 中国安全科学学报, 2023, 33(6): 64-72.

ZHANG Xiliang, JIAO Haokai, LI Erbao. Prediction of vibration velocity of deep blasting based on transfer learning[J]. China Safety Science Journal, 2023, 33(6): 64-72.

图/表 11

表1

符号算法定义

序号	算法符号	具体定义
1	D_S	源域
2	$(x 1 s, x 2 s, …, x l s)$	源域样本集合,l为样本数量
3	$(y 1 s, y 2 s, …, y l s)$	源域样本标记集合
4	$(ω 1 s, ω 2 s, …, ω l s)$	源域样本权重集合
5	D_T	目标域
6	$(x 1 t, x 2 t, …, x l t)$	目标域样本集合,l为样本数量

表1

续表1

序号	算法符号	具体定义
7	$(y 1 t, y 2 t, …, y l t)$	目标域样本标记集合
8	$(ω 1 t, ω 2 t, …, ω l t)$	目标域样本权重集合
9	Test	目标域测试集
10	$μ$	样本排除参数
11	$D s$	源域训练集
12	$D t$	目标域训练集
13	$ρ$	权重校正参数
14	M	最大迭代次数

续表1

图1

表2

表3

续表3

表4

表5

图2

图3

图4

参考文献 20

[1]	郑皓文, 赵根, 胡英国, 等. 基于ACOR-LSSVM算法的爆破振动速度预测[J]. 爆破, 2018, 35(3): 154-158.
	ZHENG Haowen, ZHAO Gen, HU Yingguo, et al. Prediction of blasting vibration velocity based on acor-lssvm algorithm[J]. Blasting, 2018, 35 (3):154-158.
[2]	施建俊, 李庆亚, 张琪, 等. 基于Matlab和BP神经网络的爆破振动预测系统[J]. 爆炸与冲击, 2017, 37(6): 1087-1092.
	SHI Jianjun, LI Qingya, ZHANG Qi, et al. Blasting vibration prediction system based on Matlab and BP neural network[J]. Explosion and Shock Waves, 2017, 37 (6): 1087-1092.
[3]	陈秋松, 张钦礼, 陈松, 等. 基于GRA-GEP的爆破峰值速度预测[J]. 中南大学学报:自然科学版, 2016, 47(7): 2441-2447.
	CHEN Qiusong, ZHANG Qinli, CHEN Song, et al. Prediction of blasting peak velocity based on gra-gep[J]. Journal of Central South University :Natural Science Edition, 2016, 47 (7): 2441-2447.
[4]	温廷新, 陈晓宇, 刘天宇, 等. 基于优化IGA-ELM模型的爆破振动特征参量预测研究[J]. 中国安全科学学报, 2017, 27(11): 37-42. doi: 10.16265/j.cnki.issn1003-3033.2017.11.007
	WEN Tingxin, CHEN Xiaoyu, LIU Tianyu, et al. Prediction of blasting vibration characteristic parameters based on optimized iga-elmmodel[J]. China Safety Science Journal, 2017, 27 (11): 37-42. doi: 10.16265/j.cnki.issn1003-3033.2017.11.007
[5]	庄福振, 罗平, 何清, 等. 迁移学习研究进展[J]. 软件学报, 2015, 26(1): 26-39.
	ZHUANG Fuzhen, LUO Ping, HE Qing, et al. Research progress of transfer learning[J]. Journal of Software, 2015, 26(1): 26-39.
[6]	张倩, 李明, 王雪松, 等. 一种面向多源领域的实例迁移学习[J]. 自动化学报, 2014, 40(6): 1176-1183.
	ZHANG Qian, LI Ming, WANG Xuesong, et al. A case study for multi-source domain transfer learning[J]. Acta Automatica Sinica, 2014, 40 (6): 1176-1183.
[7]	戴文渊. 基于实例和特征的迁移学习算法研究[D]. 上海: 上海交通大学, 2009.
	DAI Wenyuan. A case-and feature-based transfer learning algorithm[D]. Shanghai: Shanghai Jiaotong University, 2009.
[8]	高敬阳. 神经网络集成BOOSTING类算法研究[D]. 北京: 北京化工大学, 2012.
	GAO Jingyang. Study on BOOSTING class algorithm of neural network integration[D]. Beijing: Beijing University of Chemical Technology, 2012.
[9]	李号号. 基于实例的迁移学习技术研究及应用[D]. 武汉: 武汉大学, 2018.
	LI Haohao. Research and application of instance-based transfer learning technology[D]. Wuhan: Wuhan University, 2018.
[10]	何金虎. 基于迁移学习的软件缺陷预测模型研究[D]. 哈尔滨: 哈尔滨工业大学, 2019.
	HE Jinhu. Research on software defect prediction model based on transfer learning[D]. Harbin: Harbin Institute of Technology, 2019.
[11]	徐阳, 张忠伟, 刘明. 利用信息交互最优权重改进神经网络的方法[J]. 吉林大学学报:信息科学版, 2019, 37(1): 107-112.
	XU Yang, ZHANG Zhongwei, LIU Ming. Method of improving neural network by optimal weight of information interaction[J]. Journal of Jilin University :Information Science Edition, 2019, 37 (1): 107-112.
[12]	YI Dong. The assessment of the outliers of logistic regression model and its clinical application[J]. Journal of Medical Colleges of PLA, 1995, 10(1): 61-62,66.
[13]	唐贤伦, 李洋, 李鹏, 等. 多智能体粒子群优化的SVR模型预测控制[J]. 控制与决策, 2014, 29(4): 593-598.
	TANG Xianlun, LI Yang, LI Peng, et al. Predictive control of SVR model for multi-agent particle swarm optimization[J]. Control and Decision, 2014, 29 (4): 593-598.
[14]	邵良杉, 赵琳琳. 爆破振动对民房破坏的鱼骨图-SVM预测模型[J]. 中国安全科学学报, 2014, 24(8): 56-61.
	SHAO Liangshan, ZHAO Linlin. Fishbone diagram and SVM prediction model of blasting vibration on the destruction of private houses[J]. China Safety Science Journal, 2014, 24(8): 56-61.
[15]	张倩, 李海港, 李明, 等. 基于多源动态TrAdaBoost的实例迁移学习方法[J]. 中国矿业大学学报, 2014, 43(4): 713-720.
	ZHANG Qian, LI Haigang, LI Ming, et al. Case transfer learning method based on multi-source dynamic TrAdaBoost[J]. Journal of China University of Mining and Technology, 2014, 43 (4): 713-720.
[16]	赵兴东. 谦比希矿深部开采隔离矿柱稳定性分析[J]. 岩石力学与工程学报, 2010, 29(增1): 2616-2622.
	ZHAO Xingdong. Stability analysis of isolated pillar in deep mining of qianbishi mine[J]. Chinese Journal of Rock Mechanics and Engineering, 2010, 29 (S1): 2616-2622.
[17]	龙明盛. 迁移学习问题与方法研究[D]. 北京: 清华大学, 2014.
	LONG Mingsheng. Problems and methods of transfer learning[D]. Beijing: Tsinghua University, 2014.
[18]	黄强. Web下的Python3编程环境分析[J]. 电脑编程技巧与维护, 2018(12): 21-22.
	HUANG Qiang. Analysis of Python3 programming environment under the Web[J]. Computer Programming Skills and Maintenance, 2018 (12): 21-22.
[19]	刘志凯. 基于Web的Python3编程环境[J]. 计算机系统应用, 2015, 24(7): 236-239.
	LIU Zhikai. Web-based Python3 programming environment[J]. Computer System & Applications, 2015, 24 (7): 236-239.
[20]	曹晓峰. 基于Vis/NIR高光谱和机器视觉技术的冬枣分级方法研究[D]. 咸阳: 西北农林科技大学, 2018.
	CAO Xiaofeng. Classification method of winter jujube based on Vis/NIR hyperspectroscopy and machine vision technology[D]. Xianyang: Northwest A&F University, 2018.

属性名	含义
炮孔间距H_s	同排炮孔间的距离
炮孔排距H_r	邻排炮孔间的距离
最小抵抗线W_d	药包中心或重心到最近自由面的最短距离
总药量S	一次爆破使用炸药总量
单段最大药量(最大单响药量)S_m	一次爆破中单个炮孔最大装药量(如未采用逐孔爆破方式, 则该样本特征值为最大单响药量)
振动测试质点与装药爆破点间的相对垂直距离H	振动测试质点与装药爆破点间的相对垂直距离
水平距离J	振动测试质点与爆破点之间的水平距离
矿石硬度系数U_L	矿石的坚硬程度的表征,用坚固性系数表示
围岩硬度系数U_K	围岩的坚硬程度的表征,用坚固性系数表示
节理频数G	矿岩每米长度上节理个数的平均值
爆破振动峰值速度v	特定质点受爆破影响的峰值振动合速度

H_s/m	H_r/m	W_d/m	S/kg	S_m/kg	H/m	J/m	U_L	U_K	G	v/(m·s^-1)
1.7	2.2	3.4	5 072	53	21.4	136	5	10	1.2	0.562
1.7	2.3	3.8	5 894	65	53.6	143	6	11	1.7	0.458
1.8	2	2.3	6 980	62	19.3	121	6	10	1	0.854
1.8	1.9	2.7	6 548	72	27.4	178	6	10	1.4	0.487
1.7	2.2	3	5 947	49	59.4	124	7	10	1.7	0.393
1.8	2	3.6	5 892	64	20.6	104	6	11	2.1	1.015
1.8	2.4	4.2	5 732	59	23.5	115	8	11	1.7	0.753
1.9	2.1	2.6	4 938	67	62.5	89	7	11	1.5	0.286
2	1.7	3.6	6 847	73	71.4	136	6	12	2	0.385
1.8	1.9	3.4	7 282	63	34.4	124	6	11	1.2	1.104
1.8	2	4.5	6 879	68	14.4	132	5	10	1.4	0.478
1.7	2.2	4.2	6 382	51	34.7	135	6	10	1.5	1.012

H_s/m	H_r/m	W_d/m	S/kg	S_m/kg	H/m	J/m	U_L	U_K	G	v/(m·s^-1)
1.8	2.3	4.4	5 837	44	52.8	95	5	10	1	0.865
1.9	2.3	3.4	4 973	49	37.6	103	5	11	2.2	0.747
1.9	2	3.7	6 847	57	56.6	114	7	11	2.7	0.479
1.9	2	2.3	6 437	53	37.8	102	6	10	1.3	0.854
1.9	2	3.8	6 271	68	35.4	94	7	12	1.3	0.794
2.1	2.3	5.1	7 382	64	21.7	88	7	11	1.2	0.673
2.1	2.3	4	5 893	58	34.3	113	5	12	2	0.378
1.8	2	3.9	6 347	68	47.8	107	6	12	1.3	0.741
1.8	1.8	5.5	5 983	63	74.6	121	6	11	1.2	0.513
1.8	2	3.7	5 372	54	12.4	89	5	11	1.2	0.383
1.8	2	3.2	5 894	51	43.7	93	6	11	1	0.497
1.8	2	2.9	6 389	58	53.8	99	5	11	1.4	0.854
2.5	2	3.6	7 894	67	27.8	117	7	10	1.5	0.583
1.6	2.1	4.1	6 839	74	31.2	103	6	12	1.3	0.568
1.7	2.2	5.2	5 932	73	38.6	116	5	11	1.1	0.649

H_s/m	H_r/m	W_d/m	S/kg	S_m/kg	H/m	J/m	U_L	U_K	G	v/(m·s^-1)
3.9	2.8	3.4	2 989	12	19.4	145.7	3	12	2.3	0.372
3.8	2,7	2.9	2 876	11.3	18.6	111.6	4	11	3.6	0.196
4	3	3.5	4 072	11	15.7	171.9	3	14	3.1	0.436
4	3	3.3	4 893	12	21.9	132.5	5	9	2.4	0.615
4	3	3.6	3 468	15	5.07	193.8	6	10	3.2	0.133
3.3	2.3	3.2	3 252	12.4	34.3	126.8	7	1	4.8	0.511
3	2	3.7	3 780	11.8	5.2	135.6	4	1	2.6	0.563
4	3	3	3 601	12.1	4,9	158.6	4	1	4.2	0.419
3	2	3.1	4 200	11.5	8.6	177.7	3	1	4.8	0.492
4	3	3.5	3 624	10.3	7.1	233.2	5	1	2.6	0.412
3.8	3	3.7	3 648	12.5	11.6	165.9	4	1	2.2	0.263
︙	︙	︙	︙	︙	︙	︙	︙	︙	︙	︙
4	2.4	3.5	4 500	11.8	31.8	139.8	4	1	5.8	0.505
3.7	3	3.2	2 380	9.2	16	117.1	8	6	6.4	0.236

样本序号	真实值 (v/m·s^-1)	SVR预测值 (v/m·s^-1)	Tradaboost-R² 预测值 (v/m·s^-1)	LR-Tradaboost 预测值 (v/m·s^-1)
1	0.513	0.839	0.548	0.537
2	0.383	0.274	0.416	0.394
3	0.497	0.531	0.464	0.507
4	0.854	1.025	0.873	0.869
5	0.583	0.624	0.621	0.563
6	0.568	0.497	0.584	0.574
7	0.649	0.726	0.675	0.692