融合全监督学习的半监督矿石粒度预测算法

doi:10.11872/j.issn.1005-2518.2024.03.040

摘要/Abstract

摘要：

针对选矿过程矿石粒度分析精度的提高依赖于有标签样本数量，以及传统全监督建模方法泛化性能较差的问题，提出了融合全监督学习的半监督矿石粒度预测算法。以运矿皮带上应用图像获取的矿石粒度数据作为研究对象，利用半监督学习获得无标签的图像识别矿石粒度样本伪标签，扩展数量有限的原始标签样本，以提高矿石粒度预测模型的性能。采用筛分法获取的矿石粒度数据集来验证融合全监督学习的半监督预测算法，结果表明，融合全监督学习的半监督预测算法的模型决定系数达到92.1%，均方根误差和平均绝对误差分别为0.023和0.02，相较于传统全监督建模方法，该模型的预测精度显著提高，为提高矿石粒度检测精度提供了有力的技术支撑。

关键词: 半监督学习, 粒度检测, 伪标签, 粒度分布, 机器学习, 矿石

Abstract:

Aiming at the problems that the improvement of the accuracy of ore particle size analysis in the ore dressing process depends on the number of labeled samples，and the application of the traditional fully supervised modeling method has poor generalization performance，a semi-supervised ore particle size prediction algorithm incorporating fully supervised learning was proposed.Taking the ore particle size data obtained by applying images on the ore transport belt as the research object，the ore particle size data was analyzed.Four kinds of ore particle size features namely，particle size，weighted arithmetic mean size，standard deviation and deviation coefficient was adopted as the input features.And three kinds of prediction models were established，namely，decision tree，GBDT and BP neural network.By stratified sampling of the original ore size labeled samples，a training set was constructed.Then use the semi-supervised learning to obtain the unlabeled image identification ore particle size samples pseudo-labels，screen out high-confidence pseudo-labeled samples，add the pseudo-labels judged by confidence to the original ore particle size label samples，expand the limited number of original labeled samples，and at the same time delete the corresponding samples in the unlabeled ore particle size samples.Finally，in order to improve the performance of the prediction mode，a new regression prediction model was constructed based on the expanded set of original labeled samples，.The ore particle size dataset obtained by sieving method was used to validate the semi-supervised prediction algorithm incorporating fully supervised learning.The results show that，compared with the traditional fully supervised modeling methods such as decision tree，ridge regression，Bayesian，etc.The model coefficient of determination of the semi-supervised prediction algorithm incorporating fully supervised learning reaches 92.1%，which is increased by 5%，5.4%，and 5.2%，respectively.The root-mean-square error is 0.023，which is reduced by 23.33%，23.33% and 20.69%，respectively，and the mean absolute error is 0.02，which is reduced by 23.08%，13.04% and 9.09%，respectively.The research shows that the prediction accuracy is significantly improved，which verifies the feasibility and reliability of the semi-supervised ore particle size prediction model incorporating fully supervised learning.It also provides a powerful technological support for the improvement of the accuracy of ore particle size detection，and further confirms the advantages of the semi-supervised learning，and provides a powerful technological support for the improvement of the accuracy of the semi-supervised learning.It further confirms the advantages of semi-supervised learning，provides new ideas and methods for the practical application of ore particle size prediction technology，and is expected to improve the production efficiency and quality control level in the process of ore processing and utilization.

Key words: semi-supervised learning, granularity detection, pseudolabeling, particle size distribution, machine learning, ore

中图分类号:

姜志宏, 陈澳. 融合全监督学习的半监督矿石粒度预测算法[J]. 黄金科学技术, 2024, 32(3): 539-547.

Zhihong JIANG, Ao CHEN. Semi-supervised Ore Granularity Prediction Algorithm Incorporating Fully Supervised Learning[J]. Gold Science and Technology, 2024, 32(3): 539-547.

图/表 9

表1

表2

表3

5种预测模型在实际矿石数据集上的决定系数"

模型	$R 2$
Tree	0.871
GBDT	0.899
Bayes	0.869
岭回归	0.867
BP神经网络	0.872

表3

图1

图2

表4

图3

表5

8种预测模型的评价指标对比"

模型	评价指标
模型	RMSE	MAE	$R 2$
BP神经网络	0.031	0.027	0.862
半监督学习模型	0.023	0.020	0.921
GBDT	0.030	0.024	0.871
RF	0.041	0.033	0.756
SVM	0.046	0.034	0.692
决策树	0.029	0.025	0.882
XGBOOST	0.030	0.025	0.872
贝叶斯	0.029	0.022	0.879

表5

表6

不同组合预测模型测试性能对比"

模型	评价指标
模型	RMSE	MAE	$R 2$
半监督+决策树	0.031	0.027	0.862
半监督+BP	0.023	0.020	0.921
半监督+GBDT	0.030	0.024	0.871
半监督+RF	0.041	0.033	0.756
半监督+XBGoost	0.046	0.034	0.692
半监督+贝叶斯	0.029	0.025	0.882
半监督+岭回归	0.030	0.025	0.872

表6

参考文献 0

	Chun F， Liu X M， Zhu X G，et al，2021.Numerical study on crushing law of iron ore under different impact velocity using CDEM［J］.IOP Conference Series：Earth and Environmental Science，861（4）：042069.DOI：10.1088/1755-1315/861/4/042069 . doi: 10.1088/1755-1315/861/4/042069
	Hu S D， Miao D Q，PedryczW，2022.Multi granularity based label propagation with active learning for semi-supervised classification［J］.Expert Systems with Applications，192：116276..
	Hu Yunqing， Qiu Qingying， Yu Xiu，et al，2020.Semi-supervised patent text classification method based on improved Tri-training algorithm［J］.Journal of Zhejiang University （Engineering Science），54（2）：331-339.
	Hu Z J， Yang Z Y， Hu X F，et al，2021.SimPLE：Similar pseudo label exploitation for semi-supervised classification［C］//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR），Nashville，TN，USA，15094-15103.DOI：10.1109/CVPR46437.2021.01485 . doi: 10.1109/CVPR46437.2021.01485
	Huang Faming， Pan Lihan， Yao Chi，et al，2021.Landslide susceptibility prediction modelling based on semi-supervised machine learning ［J］.Journal of Zhejiang University（Engineering Science），55（9）：1705-1713.
	Kang P， Kim D， Cho S，2016.Semi-supervised support vector regression based on self-training with label uncertainty：An application to virtual metrology in semiconductor manufacturing［J］.Expert Systems with Applications，51：85-106.DOI：10.1016/j.eswa.2015.12.027 . doi: 10.1016/j.eswa.2015.12.027
	Lan Fengchong， Pan Wei， Chen Jiqing，2023.Prediction of remaining useful life of real-world vehicle lithium-ion power battery based on Aseq2seq-PF［J］.Automotive Engineering，45（12）：2348-2356.
	Li Hongxiang， Wang Xiaoli， Yang Chunhua，et al，2021.Ore image segmentation method based on GAN-UNet［J］.Control Theory and Applications， 38（9）：1393-1398.
	Liu S S， Li S J，2022.A semi-supervised soft sensor method based on vine copula regression and tri-training algorithm for complex chemical processes［J］.Journal of Process Control，120：115-128.
	Lu Jiawang，2023.Study on Complex Particle Swarm Space Image Reconstruction and Screening Performance Evaluation in Screening Process［D］.Beijing：China University of Mining and Technology（Beijing）.
	Mao Gengxuan， Tu Yan， Cui Wenbo，et al，2022.Hyperspectral inve-rsion of soil heavy metal mass concentration based on semi-supervised regression［J］.Journal of Applied Sciences—Electronics and Information Engineering，40（6）：941-952.
	Shi Xudong， Xiong Weili，2020.Semi-supervised gaussian process regression modeling based on improved self-training algorithm［J］.Control Engineering of China，27（3）：451-455.
	Song Jian， Wang Wenlong， Li Dong，et al，2022.Injection molded part size prediction method based on stacking integrated learning［J］.Journal of South China University of Technology （Natural Science Edition），50（6）：19-26.
	Sun Xingwei， Yang Tongtong， Yang Heran，et al，2023.Research on thermal error of CNC machine tool feed system based on CNN-GRU combined neural network［J］.Chinese Journal of Scientific Instrumen，44（10）：219-226.
	Wang Jianming， Ye Yurong， Rao Chaomin，et al，2023.Prediction on composite interface bonding strength between ceramsite lightweight aggregate concrete and normal concrete based on GBDT algorithm［J］.Journal of Building Materials，26（2）：150-155，171.
	Wang Wei， Li Qing， Zhang Dezheng， al el，2023.A survey of ore image processing based on deep learning［J］.Chinese Journal of Engineering，45（4）：621-631.
	Xu Wen， Tang Jian， Xia Heng，et al，2022.Soft sensor of dioxin emission concentration based on Bagging semi-supervised deep forest regression ［J］.Chinese Journal of Scientific Instrument，43（6）：251-259.
	Xu Yongyang， Li Zixuan， Xie Zhong，et al，2020.Prediction of copper mineralization based on semi-supervised neural network ［J］.Earth Science，45（12）：4563-4573.
	Zhang G H， Fan Y B， Yang R S，et al，2023.Influence of ore size on the production of micro-sized ore particles by high-pressure gas rapid unloading［J］.Powder Technology，427（5）：118716.DOI：10.1016/j.powtec.2023.118716 . doi: 10.1016/j.powtec.2023.118716
	Zhuang Huimin，2022.Research on Semi-Supervised Learning Based ARP Attack Detection Method in SDIIoT［D］.Shanghai：Donghua University.
	胡云青，邱清盈，余秀，等，2020.基于改进三体训练法的半监督专利文本分类方法［J］.浙江大学学报（工学版），54（2）：331-339.
	黄发明，潘李含，姚池，等，2021.基于半监督机器学习的滑坡易发性预测建模［J］.浙江大学学报（工学版），55（9）：1705-1713.
	兰凤崇，潘威，陈吉清，2023.基于Aseq2seq-PF的实车锂离子动力电池剩余使用寿命预测［J］.汽车工程，45（12）：2348-2356.
	李鸿翔，王晓丽，阳春华，等，2021.基于GAN-UNet的矿石图像分割方法［J］.控制理论与应用，38（9）：1393-1398.
	卢佳旺，2023.筛分过程复杂粒群空间图像重构与筛分效果评价研究［D］.北京：中国矿业大学（北京）.
	毛耿旋，涂彦，崔文博，等，2022.基于半监督回归的高光谱土壤重金属质量浓度反演［J］.应用科学学报，40（6）：941-952.
	史旭东，熊伟丽，2020.基于改进自训练算法的半监督GPR软测量建模［J］.控制工程，27（3）：451-455.
	宋建，王文龙，李东，等，2022.基于Stacking集成学习的注塑件尺寸预测方法［J］.华南理工大学学报（自然科学版），50（6）：19-26.
	孙兴伟，杨铜铜，杨赫然，等，2023.基于CNN-GRU组合神经网络的数控机床进给系统热误差研究［J］.仪器仪表学报，44（10）：219-226.
	王建民，叶钰蓉，饶超敏，等，2023.基于GBDT算法的混凝土叠合面黏结强度预测分析［J］.建筑材料学报，26（2）：150-155，171.
	王伟，李擎，张德政，等，2023.基于深度学习的矿石图像处理研究综述［J］.工程科学学报，45（4）：621-631.
	徐雯，汤健，夏恒，等，2022.基于Bagging半监督深度森林回归的二噁英排放浓度软测量［J］.仪器仪表学报，43（6）：251-259.
	徐永洋，李孜轩，谢忠，等，2020.基于半监督神经网络的铜矿预测方法［J］.地球科学，45（12）：4563-4573.
	庄慧敏，2022.软件定义工业物联网下基于半监督学习的ARP攻击检测方法研究［D］.上海：东华大学.

Metrics

Viewed

Full text

Abstract

Cited

Shared

Discussed

样本数据类型	矿石粒级 /mm	图像识别矿石粒度分布/%	加权算术平均粒度 /mm	标准差	偏差系数	人工筛分矿石粒度分布/%
有标签样本数据	+10	0.322	3.245	1.586	0.244	0.290
	+5~10	0.315	2.434	1.190	0.183	0.285
	-5	0.363	0.811	0.397	0.061	0.425
	+10	0.501	4.019	1.267	0.158	0.474
	+5~10	0.356	3.014	0.950	0.118	0.312
	-5	0.143	1.005	0.317	0.039	0.214
	…
无标签样本数据	+10	0.359	3.421	1.548	0.226
	+5~10	0.330	2.566	1.161	0.170
	-5	0.311	0.855	0.387	0.057
	+10	0.325	3.286	1.582	0.241
	+5~10	0.318	2.465	1.187	0.181
	-5	0.357	0.822	0.396	0.060
	+10	0.366	3.468	1.504	0.221
	+5~10	0.338	2.601	1.148	0.166
	-5	0.296	0.867	0.383	0.055
	…

模型	最佳超参数	RMSE	MAE
Tree	max_depth = 4	0.030	0.026
RF	n_estimators = 20	0.040	0.027
GBDT	learning_rate = 0.01 max_depth = 5 n_estimators = 200	0.026	0.022
XGBoost	gamma = 0.0 max_depth = 4 min_child_weight = 4 n_estimators = 70	0.040	0.028
Bayes	alpha_1= 1e-08 alpha_2= 1e-06 lambda_1= 1e-06 lambda_2= 1e-08 n_iter= 100	0.029	0.022
多项式回归	Linearregression_fit _intercept= True polynomialfeatures_degree2	0.080	0.051
岭回归	Alpha = 0.3 gamma = 0.1 kernel = linear	0.030	0.023
SVM	C = 1.0 Gamma = 1.0 Kernel = linear	0.075	0.063
BP神经网络	Activation = relu Alpha = 0.0001 hidden_layer_sizes = （100，）	0.030	0.023

矿石粒级 /mm	图像识别矿石粒度分布/%	加权算术平均粒度/mm	标准差	偏差系数	模型预测矿石粒度分布/%
+5~10	0.33	2.566	1.161	0.170	0.318
-5	0.357	0.822	0.396	0.060	0.377
+10	0.366	3.468	1.504	0.221	0.285
-5	0.296	0.867	0.383	0.055	0.393
+10	0.311	3.167	1.606	0.254	0.297
-5	0.389	0.792	0.401	0.063	0.279
+10	0.305	3.156	1.600	0.254	0.302
+5~10	0.305	2.367	1.200	0.190	0.416
+5~10	0.321	2.515	1.176	0.175	0.350
-5	0.332	0.838	0.392	0.059	0.318
+10	0.366	3.468	1.530	0.221	0.386
+5~10	0.340	2.601	1.148	0.166	0.280
-5	0.360	0.803	0.388	0.060	0.324
+10	0.310	3.363	1.507	0.224	0.402
+5~10	0.380	2.522	1.130	0.168	0.282
-5	0.310	0.841	0.377	0.056	0.305
+10	0.310	3.313	1.535	0.232	0.415
-5	0.250	0.903	0.367	0.058	0.325
+5~10	0.380	2.606	1.114	0.160	0.396
+5~10	0.390	2.569	1.114	0.163	0.325
+10	0.300	3.150	1.596	0.253	0.388
-5	0.390	0.788	0.399	0.063	0.282
…

[1]	李地元, 杨博, 刘子达, 刘永平, 赵君杰. 基于集成树算法的岩石黏聚力和内摩擦角预测方法[J]. 黄金科学技术, 2024, 32(5): 847-859.
[2]	李振阳, 张宝岗, 熊信, 杨承业, 白玉奇. 基于PSO-XGBoost的露天矿山PPV预测模型研究[J]. 黄金科学技术, 2024, 32(4): 620-630.
[3]	陈桥, 姬龙雪, 董欣, 倪蓉, 李岩松, 佟琳琳, 杨洪英. 尼尔森选矿机富集机制对金矿分选效果的影响研究[J]. 黄金科学技术, 2024, 32(4): 685-693.
[4]	凡兴禹, 王雪林. 基于改进XGBoost算法的深部巷道松动圈智能预测研究[J]. 黄金科学技术, 2024, 32(1): 109-122.
[5]	吴荔, 匡文龙, 张志辉, 陈健龙, 张跃权, 刘兆阳, 黄英剑. 赣东北银山矿田铜矿石中伴生金赋存状态研究[J]. 黄金科学技术, 2023, 31(6): 888-899.
[6]	许方颖, 邹艳红, 易卓炜, 杨福强, 毛先成. 基于非均衡数据的ADASYN-CatBoost测井岩性智能识别——以胶西北招贤金矿床为例[J]. 黄金科学技术, 2023, 31(5): 721-735.
[7]	陈建宏,赵亚坤,杨珊,钟旭东. 基于蚁群—蚁周模型的大均化联合配矿及生产数据集成共享系统研究[J]. 黄金科学技术, 2023, 31(2): 292-301.
[8]	汤文聪,罗小燕. 基于FCM-WA联合算法的多种类矿石图像分割[J]. 黄金科学技术, 2023, 31(1): 153-162.
[9]	许可, 许德如. 江南造山带黄金洞金矿蚀变岩型金矿化形成机制研究[J]. 黄金科学技术, 2022, 30(2): 151-164.
[10]	韩梓晴,李孜军,徐圆圆. 基于偏序集的硫化矿石自燃倾向性评价[J]. 黄金科学技术, 2022, 30(1): 105-112.
[11]	胡建华,郭萌萌,周坦,张涛. 基于改进迁移学习算法的岩体质量评价模型[J]. 黄金科学技术, 2021, 29(6): 826-833.
[12]	郭彩莲,牛芳银,陈炳龙,王海军. 老挝巴勉石英脉型金矿石工艺矿物学研究[J]. 黄金科学技术, 2021, 29(6): 908-916.
[13]	张淦,陶干强,吴宇轩. 基于改流体放矿技术的放矿口尺寸试验研究[J]. 黄金科学技术, 2021, 29(3): 364-371.
[14]	陈鑫,王李管,李金玲. 矿石堆场品位模型构建及取料品位估算方法[J]. 黄金科学技术, 2021, 29(2): 287-295.
[15]	王牧帆,罗周全,于琦. 基于 Stacking 模型的采空区稳定性预测[J]. 黄金科学技术, 2020, 28(6): 894-901.