Process working condition recognition based on the fusion of morphological and pixel set features of froth for froth flotation
Process working condition recognition based on the fusion of morphological and pixel set features of froth for froth flotation
复制标题
DOI：
10.1016/j.mineng.2018.08.017
复制
发表时间：
2018-11
影响因子：
4.8
通讯作者：
Xiaoli Wang;Chen Song;Chunhua Yang;Yongfang Xie
中科院分区：
文献类型：
--
作者：
Xiaoli Wang;Chen Song;Chunhua Yang;Yongfang Xie
工艺条件识别是提高泡沫工艺性能的有效途径。以往基于机器视觉的浮选过程状态识别算法中，所使用的特征包括灰度值、气泡尺寸分布、负荷等，本质上是灰度图像的统计结果，在提取过程中存在局部气泡结构信息的丢失。同时，大量的图像数据没有得到充分利用。因此，本文利用深度神经网络提取像素集特征，提出一种利用大量图像数据的基于气泡图像形态学和像素集特征的两步条件识别方法。首先，泡沫图像被分割成单个气泡图像。接下来，提取单个气泡图像的形态特征向量，并通过K均值聚类为单个气泡图像分配分类标签。对大量历史图像进行分析和标记，训练卷积神经网络（CNN），提取每个气泡图像的像素集特征。然后使用加权均值漂移算法融合气泡图像的形态特征向量和像素集特征以进行气泡图像聚类。计算泡沫图像中各种类型的气泡的频率，以形成泡沫图像的气泡频率集合。然后提出了一种基于一段时间内图像序列的两步工况识别策略。该策略通过两个主要步骤将所有泡沫图像的气泡频率集以及从相应时间段的泡沫图像中分割出的气泡图像的气泡频率集与典型浮选工况的图像的气泡频率集进行匹配，以确定当前的工况。使用工业数据的测试结果证明了该方法的高精度和计算速度。
Process condition recognition is an effective way to improve the froth process performance. In previous condition recognition algorithms based on machine vision in flotation process, the used features including gray value, bubble size distribution, load, etc., are essentially statistical results of gray level images and there is local bubble structure information loss in their extraction procedure. Meanwhile, the large number of image data are not adequately utilized. Thus, in this paper, deep neural network are used to extract the pix set features, and a two-step condition recognition method based on bubble image morphology and pixel set features is proposed, which utilizes the large number of image data. First, froth images are segmented into single bubble images. Next, the morphological feature vector of a single bubble image is extracted and classification labels are assigned to the single bubble images via K-means clustering. A large quantity of historical images is analyzed and labeled to train a convolutional neural network (CNN) by which the pixel set features of each bubble image are extracted. The morphological feature vector and pixel set features of the bubble images are then fused for bubble image clustering using the weighted mean-shift algorithm. The frequencies of various types of bubbles in a froth image are calculated to form a bubble frequency set for the froth image. A two-step working condition recognition strategy based on image sequence over a time period is then proposed. In this strategy, the bubble frequency sets of all froth images and those of the bubble images segmented from the froth images over the corresponding time period are matched with those of the images of typical flotation conditions in two main steps to determine the current working condition. Test results using industrial data demonstrate the high accuracy and calculation speed of the proposed method.