基于多尺度CNN特征及RAE-KELM的浮选加药状态识别  

Flotation Dosing State Recognition Based on Multiscale CNN Features and RAE-KELM

作　　者：张进[1] 廖一鹏[1] 陈诗媛 王卫星[1] Zhang Jin;Liao Yipeng;Chen Shiyuan;Wang Weixing(College of Physics and Information Engineering,Fuzhou University,Fuzhou,Fujian350108,China)

机构地区：[1]福州大学物理与信息工程学院,福建福州350108

出　　处：《激光与光电子学进展》2021年第12期401-410,共10页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(61471124,61601126);福建省自然科学基金(2019J01224)。

摘　　要：针对浮选加药状态在线检测困难、识别效率低和主观性强等问题,提出了一种基于多尺度卷积神经网络(CNN)特征及行列自编码核极限学习机(RAE-KELM)的浮选加药状态识别方法。首先,对浮选泡沫图像进行非下采样Shearlet多尺度分解,用CNN提取每个尺度图像的深度特征并进行多尺度特征融合;然后,构建RAEKELM,将用量子计算改进的细菌觅食算法用于RAE-KELM的参数优化;最后,通过自建数据集训练得到最优的RAE-KELM模型,实现了浮选加药状态的自适应识别。实验结果表明,本方法的识别准确率可达到98.88%;且本方法减少了人工干预,有利于提高生产效率。To address the problems associated with online detection,low recognition efficiency,and strong subjectivity of the flotation dosing state,this paper proposes a flotation dosing state recognition method based on multiscale convolutional neural network(CNN)features and ranks automatic encoder kernel extreme learning machine(RAE-KELM).First,the flotation foam image is subjected to non-subsampled Shearlet multiscale decomposition,and the CNN is used to extract the depth features of each scale image and perform multiscale feature fusion.Then,the RAE-KELM is constructed,and an improved bacterial foraging algorithm based on quantum computing is used to optimize the RAE-KELM parameters.Finally,the optimal RAE-KELM model is obtained through self-built dataset training to realize the adaptive recognition of the flotation dosing state.The experimental results demonstrate that the recognition accuracy of the method can reach 98.88%.Additionally,the method reduces manual interventions,which can improve production efficiency.

关 键 词：图像处理 卷积神经网络 非下采样Shearlet变换 行列自编码核极限学习机 量子细菌觅食算法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]