The ore dressing shaking table serves as a crucial apparatus for the separation and purification of strategic mineral resources,including tungsten,tin,tantalum-niobium,titanium,and rare earth elements.It is extensively utilized in ore dressing production.Nonetheless,the current level of automation in shaking tables is relatively low,with the adjustment of control parameters predominantly dependent on the expertise of experienced operators.These operators make necessary adjustments based on the position of the concentrate boundary within the bed sub-belt,a process that is labor-intensive and prone to causing fluctuations in beneficiation indices.Consequently,there is a pressing need to develop a self-adaptive adjustment system for shaking table control parameters,informed by the position of the concentrate boundary,to enhance both production efficiency and the level of automation.The initial step in achieving adaptive adjustment of shaking table control parameters is to construct a mapping relationship model between the control parameters and the coordinate values of the concentrate boundary.To achieve this objective,we have collected extensive data on the mapping relationships between various combinations of four control parameters,namely surface slope,lateral flushing,and others.In this study,we introduce the gated recurrent unit (GRU) to process time-sequence information within the shaking table concentrator production data.Additionally,we incorporate the squeeze-and-excitation (SE) attention mechanism to assign weights to different channels,thereby enhancing the model’s feature extraction capabilities and fitting accuracy.Consequently,we developed the CNN-GRU-Attention algorithm to perform a regressive analysis of the concentrator shaking table production data and established a “control parameter-concentrate boundary coordinate value”mapping relationship model.Comparative analysis demonstrates that the proposed algorithm outperforms the CNN-GRU,CNN-LSTM,and CNN-LSTM-Attention models.The SSA algorithm was employed to optimize three hyperparameters:the learning rate,the number of hidden neurons in the GRU layer,and the regularization coefficient for the CNN-GRU-Attention model.The optimal values identified were 0.0214,3,and 0.0007,respectively,significantly reducing the time required for parameter tuning and enhancing the efficiency of model training.This optimized SSA-CNN-GRU-Attention model was subsequently utilized to regress the boundary coordinates of the concentrate,yielding evaluation metrics of R²=0.98269,RMSE=0.79085,MAE=0.34362,and MAPE=0.0844%.Compared to the original model,the RMSE,MAE,and MAPE were reduced by 34.83%,51.11%,and 51.30%,respectively,thereby substantially improving the model’s trend-following capability and predictive accuracy.The“control parameter-concentrate boundary coordinate value”relationship model developed in this study satisfies the requirements for industrial beneficiation production using a shaking table and offers valuable insights for constructing an adaptive adjustment system for shaking table control parameters.
基于现代矿山智能化理念,将机器视觉、大数据技术和协作机器人等数字化技术与矿业深度融合实现矿业高质量发展是未来智能矿山建设的关键创新技术(郑玉荣等,2025)。近年来,借助机器视觉和深度学习等计算机技术来提升摇床的自动化和智能化水平取得了一些研究成果。对于摇床的自动化升级研究主要有2种形式:一是开发智能摇床巡检机器人,利用机器人在摇床选矿车间巡检采集床面矿物分带特征并完成精矿接矿板的截取位置调节等工作(杨文旺等,2020;刘惠中等,2023);二是通过开发摇床矿带特征提取算法和设计截取机构装置等方式对选矿摇床装备进行自动化改造升级(You et al.,2023;刘惠中等,2025)。上述研究为促进摇床自动化和智能化发展提供了支持,但是智能巡检机器人对车间环境要求较高,大部分摇床选矿车间环境无法满足巡检机器人的行走要求,因此难以推广应用。
卷积神经网络(CNN)是一种深度学习模型,以卷积操作为基础,能够有效捕获数据中的特征(Jiang et al.,2020)。CNN通过堆叠多个卷积层和池化层的层级结构,能够逐层提取从低级到高级的特征表示,这种层次化的特征提取机制使模型能够有效捕捉数据的多尺度特征信息,从而提升对复杂问题的建模能力(Lu et al.,2019;Jörges et al.,2020;许克应等,2022;Young et al.,2022)。CNN在处理回归分析问题上具有较高的稳健性和泛化能力,因此本文采用CNN构建摇床“控制参数—精矿边界坐标”模型。为应对摇床床面精矿带边界位置变化的时序性,以CNN为主体网络融合门控循环神经网络GRU提取时序信息。
在摇床选矿生产过程中,控制参数的调节对精矿带边界位置的影响作用具有延滞性,在控制参数调节后,精矿带边界位置随时间缓慢变化。为实现对摇床运行状态的实时监测,需要捕获“控制参数—精矿边界坐标”映射关系中蕴含的时序信息。门控循环单元(GRU)(Dey et al.,2017)和长短时记忆网络(LSTM)(Yu et al.,2019)在处理时序信息方面均表现良好,二者均引入门控机制,使网络能够更好地适应数据,且更容易训练。LSTM在反向传播过程中能够更有效地传播梯度,克服了循环神经网络(RNN)在处理长序列时容易出现梯度消失或梯度爆炸的问题。图3和图4分别为GRU和LSTM内部结构图。
式中:为更新门的输出,控制新的输入是否需要更新内部状态;为重置门的输出,控制是否忽略旧的隐藏状态;为当前时间步的候选隐藏状态,用于计算新的隐藏状态;为新的隐藏状态;为sigmoid激活函数;为与更新门相关的权重矩阵;为上一个时间步的隐藏状态;为当前时间步的输入;为与重置门相关的权重矩阵;为双曲正切激活函数;为与候选状态相关的权重矩阵; I 为单位矩阵;表示Hadamard乘积。
CNN-GRU-Attention算法流程如图6所示,其中模型的超参数设置非常关键,直接控制着模型性能的优劣。对超参数手动调优需要根据经验不断尝试,不仅工作繁琐、耗时较长,而且通常无法达到模型的最优效果。为提高模型质量和模型的训练效率,本文利用麻雀搜索算法(SSA)(Xue et al.,2020)对CNN-GRU-Attention模型的超参数进行寻优。
式中:为在第t+1次迭代时第i个麻雀在j维中的位置,其中i=1,2,…,n;j=1,2,…,d;为随机数;和分别为预警阈值和安全值,其中,;Q为服从高斯分布的随机数; L 为元素均为1的d维行矩阵;<ST时发现者安全的执行广泛搜索;≥ST时觅食环境存在危险,麻雀种群集体位置迁移,位置表示为
式中:是最新迭代次数时发现者所处的局部最优位置;为当前迭代全局最差位置。,其中是一个 d 维行矩阵,其元素随机赋值为1或-1。当,第i个加入者能源储备较低,需搜寻其他位置觅食。麻雀种群中的加入者依据式(8)和式(9)的位置状态进行搜索。在觅食过程中,麻雀种群在初始位置随机产生10%~20%的警戒者,其位置状态表示为
为验证CNN-GRU-Attention模型对不同控制参数条件下的精矿边界坐标值拟合的精准性,分别与CNN-GRU模型(王世杰等,2023)、CNN-LSTM模型(荣光旭等,2023)和CNN-LSTM-Attention模型(Zhang et al.,2019)的性能进行对比分析。利用遗传算法(GA)、鲸鱼优化算法(WOA)和麻雀搜索优化算法(SSA)对CNN-GRU-Attention模型的超参数进行寻优,各模型分别重复30次训练—预测过程,预测结果的RMSE均值及其标准差见表1。节选各模型对精矿边界坐标预测结果如表2所示,其中未进行超参数优化时4种模型拟合效果如图10所示。
DeyR, SalemF M,2017.Gate-variants of gated recurrent unit(GRU) neural networks[C]//2017 IEEE 60th International Midwest Symposium on Circuits and Systems(MWSCAS).Boston:IEEE.
[2]
JiangY Y, YangF, ZhuH L,et al,2020.Nonlinear CNN:Improving CNNs with quadratic convolutions[J].Neural Computing and Applications,32(12):8507-8516.
[3]
JörgesC, BerbenbrinkD, GottschalkH,et al,2023.Spatial ocean wave height prediction with CNN mixed-data deep neural networks using random field simulated bathymetry[J].Ocean Engineering,271:113699.
[4]
LiXindong, HuangWanfu, XiaoYuanhua,et al,2012.Beneficiation process for tantalum and niobium recovery from primary slime of tantalum-niobium ore[J].Rare Metals and Cemented Carbides,40(3):5-7.
LiuHuizhong, RuiZuowei, ZhuHejun,et al,2025.Recognition on ore zone separation points target detection and identification in mineral processing shaking table based on improved YOLOv5 algorithm[J].Nonferrous Metals Science and Engineering,16(1):115-124.
[7]
LuL Q, YiY H, HuangF L,et al,2019.Integrating local CNN and global CNN for script identification in natural scene images[J].IEEE Access,7:52669-52679.
[8]
RongGuangxu, LiZongyang,2023.Application of CNN-LSTM model in slope reliability analysis[J].Gold Science and Technology,31(4):613-623.
[9]
WangKai, LiKangnan, DuFeng,et al,2023.Prediction of coal-gas compound dynamic disaster based on convolutional neural network[J].Journal of Mining Science and Technology,8(5):613-622.
[10]
WangShijie, XiaoMeng, HuaGuowei,et al,2023.Research on uplift pressure prediction of concrete dam based on CNN-GRU[J].Journal of Safety Science and Technology,19(7):99-105.
[11]
XuKeying, ShuPing, BaoHua,2022.Visual tracking combining attention and feature fusion network modulation[J].Laser and Optoelectronics Progress,59(12):160-171.
[12]
XueJ K, ShenB,2020.A novel swarm intelligence optimization approach:Sparrow search algorithem[J].Systems Science and Control Engineering,8(1):22-34.
[13]
YangJunyan, QiShen, LiuHai,et al,2018.Progress of research related to the comprehensive recovery and utilization of tungsten smelting slag[J].Chinese Journal of Engineering,40(12):1468-1475.
[14]
YangWenwang, HeQinglang, LanXixiong,et al,2020.Development and application of intelligent inspection robot for table concentrator[J].Nonferrous Metals(Mineral Processing Section),(5):102-106.
[15]
YouK S, WenC Y, LiuH Z,2023.Research on intelligent implementation of the beneficiation process of shaking table[J].Minerals Engineering,199:108108.
[16]
YoungS I, ZheW, TaubmanD,et al,2022.Transform quantization for CNN compression[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,44(9):5700-5714.
[17]
YuY, SiX S, HuC H,et al,2019.A review of recurrent neural networks:LSTM cells and network architectures[J].Neural Computation,31(7):1235-1270.
[18]
ZhangY S, ZhengJ, JiangY R,et al,2019.A text sentiment classification modeling method based on coordinated CNN-LSTM-Attention model[J].Chinese Journal of Electronics,28(1):120-126.
[19]
ZhengYurong, JinJunbao, WuXinnian,et al,2025.Research progress and application of key technologies in smart mine[J].Gold Science and Technology,33(2):404-416.
[20]
ZhiWenke, WangFei, ChenXiaoyi,et al,2024.Research progress of scandium resources and refining technology of scandium[J].Journal of the Chinese Rare Earths Society,42(1):13-29.