融合全监督学习的半监督矿石粒度预测算法

Semi-supervised Ore Granularity Prediction Algorithm Incorporating Fully Supervised Learning

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。

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.