基于图神经网络的最优装矿溜井选择模型

Optimal orepass selection model based on graph neural network

针对金属矿山井下选择装矿溜井时过度依赖人工经验,常导致决策不合理的问题,建立能够科学选择最优装矿溜井的预测模型,优化溜井选择决策以提升有轨运输效率.以安徽白象山铁矿–495水平为研究对象,采集该水平溜井、巷道信息及矿石采装、运输等数据.将数据预处理成能描述该水平路网结构的拓扑矩阵,与包含路段/溜井基础信息、路段/溜井轨迹信息、溜井时序料位信息的特征向量,用作模型训练与验证.结合拆分时序特征、优化池化输出、编码边特征的模型改进设计,构建并训练能兼顾溜井属性、路段属性、路网拓扑的时序图神经网络模型Time-series transformer graph convolutional network(T-TransGCN),将预测结果作为最优装矿溜井选择结果.验证结果显示:(1) T-TransGCN能兼顾节点属性与拓扑信息,相较时序人工神经网络Time-series multi-layer perceptron (T-MLP)与时序基准图模型Time-series graph convolutional network (TGCN),T-TransGCN稳定且有较强的拟合能力;(2)时序料位特征有助于T-TransGCN理解溜井近期动态,轨迹特征能动态反映不同溜井的重要程度,同时帮助模型理解相邻节点相似、岔路节点相似等信息.两类新特征的引入均能有效提升TTransGCN泛化能力;(3)引入边特征、优化T-TransGCN池化层输出、拆分时序料位特征,能进一步提升T-TransGCN拟合能力、泛化能力与稳定性.

Overdependence on manual experience frequently leads to the irrational selection of orepass.Therefore,a scheduling model needs to be established to make sound decisions on orepass selection,increase the efficiency of underground rail transport,and improve production efficiency in metal mines.In this study,the−495 level of the Baixiangshan iron mine in Anhui Province is used as a research object.Orepass information,tunnel information,and historical mining,loading,and transporting data are collected.The data are then preprocessed to obtain a 130-order matrix that can describe the rail transit topology.Several vectors containing road/orepass basic information,road/orepass trajectory information,and orepass chronological material-level information are used for model training and validation.Time-series transformer graph convolutional network,which is denoted as T-TransGCN,is a temporal graph neural network that integrates orepass features,road features,and rail topology information.T-TransGCN is proposed to determine the optimal orepass selection.It enhances performance through splitting temporal features,fine-tuning the pooling layer architecture,and embedding edge features.Validated results show that(1)the T-TransGCN model is better than the Time-series multi-layer perceptron(T-MLP)and the Time-series graph convolutional network(T-GCN).The label accuracy,F1 score,and Top-3 accuracy of T-TransGCN improve by 7.33%,17.00%,and 14.26%compared with those of T-MLP,which indicates that T-TransGCN can effectively integrate node attributes and topology information.Moreover,T-TransGCN has a relatively higher number of model parameters,more complex model structure,greater stability,and stronger fitting capability than T-GCN.(2)The addition of chronological material-level features to T-TransGCN increases its F1 score and Top-3 accuracy by 11.75%and 17.02%,while the addition of trajectory features improves them by 11.83%and 10.01%.Both new data preprocessing methods are effective in enhancing the generalization ability of T-TransGCN.The chronological material-level features help T-TransGCN understand the recent state of orepass,while the trajectory features reflect the importance of different orepasses dynamically.The trajectory features help the model understand structural information,such as the similarity of adjacent nodes or the similarity of forked nodes.(3)The addition of edge features further distinguishes orepass nodes from road nodes.The optimization of the outputs of the pooling layer helps avoid the distraction of unimportant information.When chronological features are split,the F1 score and Top-3 accuracy of T-TransGCN improve by 15.94%and 12.34%.This increment enhances the focus of the model on the chronological material-level information.The integration of the abovementioned model improvements further increases the fitting capability,generalization ability,and stability of T-TransGCN.