基于长程时空特征与多尺度外观特征的锌精选工况识别  

Working Condition Recognition of Zinc Concentration Based on the Long-term Spatiotemporal Feature and Multi-scale Appearance Feature

作　　者：林振烈 张虎 袁鹤[3] 唐朝晖 LIN Zhenlie;ZHANG Hu;YUAN He;TANG Zhaohui(Fankou Lead-Zinc Mine,Shenzhen Zhongjin Lingnan Non-ferrous Metal Company Limited,Shaoguan 510050,China;School of computer science and Engineering,Changsha University,Changsha 410022,China;School of Automation,Central South University,Changsha 410083,China)

机构地区：[1]深圳市中金岭南有色金属股份有限公司凡口铅锌矿,韶关510050 [2]长沙学院计算机科学与工程学院,长沙410022 [3]中南大学自动化学院,长沙410083

出　　处：《有色金属工程》2023年第2期79-89,共11页Nonferrous Metals Engineering

基　　金：国家自然科学基金面上项目(62171476,61771492);国家自然科学基金广东联合重点基金项目(U1701261)。

摘　　要：锌精选作为锌浮选的最后一道流程,其工况直接决定锌浮选最终产品质量。现有基于卷积网络的浮选工况识别方法具备挖掘隐藏特征的能力,取得了良好效果,但仍存在表征能力有限、模型参数大等问题。为此,提出了基于长程时空特征与外观特征的锌精选工况识别模型。首先,提出基于分离三维卷积网络(Separable 3D Convolutional Neural Network,S3D CNN)与注意力机制的泡沫视频相邻帧间短程时空特征提取方法,获得特征聚焦的泡沫视频相邻帧间短程时序信息。然后,在短程时空特征的基础上采用双向卷积长短时记忆网络(Bi-directional Convolutional Long Short-Term Memory,BiConvLSTM)提取泡沫视频帧间的长程时空特征,获取泡沫视频帧间的长程动态时序信息。最后,采用基于残差网络和迁移学习的二维卷积网络提取泡沫图像的多尺度外观特征,并融合长程时空特征,对锌精选工况进行识别。实验结果表明,与现有卷积网络方法相比,所提模型在工况识别精度和模型参数上性能更佳。As the last process of zinc flotation,the working condition of zinc concentration directly determines the final product quality of zinc flotation.The existing flotation condition recognition method based on convolution neural network can mine hidden features and has achieved good results,but there are still some problems such as limited representation ability and large model parameters.Therefore,a working condition recognition model of zinc concentration based on long-term spatiotemporal feature and multi-scale appearance feature is proposed in this article.First,short-term spatiotemporal features between adjacent frames of froth video are extracted based on the separate three-dimensional convolutional network(S3D)and attention mechanism,and they can obtain the short-term temporal information between adjacent frames of froth video with feature focus.Then,based on the short-term spatiotemporal features,a bi-directional convolutional long short-term memory(BiConvLSTM)is used to extract the long-term spatiotemporal features of froth video,which can obtain the long-term dynamic temporal information of froth video.After that,a two-dimensional convolution neural network based on ResNet and transfer learning is used to extract the multi-scale appearance features,and the multi-scale appearance features and long-range space-time features are integrated to identify the zinc concentration conditions.The experimental results show that compared with the existing convolution neural network,the proposed model has better performance both in the accuracy and model parameters.

关 键 词：锌精选 工况识别 卷积网络 泡沫视频 长程时空特征 多尺度外观特征 

分 类 号：TF8[冶金工程—有色金属冶金] TF3[自动化与计算机技术—检测技术与自动化装置] TP23[自动化与计算机技术—控制科学与工程] TP29