油水界面测量过程中自适应阈值聚类优化算法

Clustering optimization algorithm with adaptive threshold for oil-water interface detection process

精确检测原油储罐内油水界面及液位高度是保证净油外输含水率控制精度及联合站盘库系统计量精度的前提条件,是石油化工过程系统工程中的重要环节。鉴于油水界面测量过程中传统分类统计算法和经典K-means聚类算法存在依赖人工选取典型值和初始聚类质心、计算结果不确定性以及精度难于保证等问题,本文提出了一种改进的K-means自适应阈值聚类优化算法。该算法能自动获取最优初始阈值,并改进了油水界面测量传统分类统计算法和经典K-means聚类算法的思想,可实现最优数据分类。首先采用自适应阈值查找算法自动查找油水界面最优初始阈值,其次采用改进K-means聚类优化算法对油水界面数据进行最优划分,最后根据最优化聚类结果计算油水界面及液位高度。实验结果表明:相对于油水界面测量的传统分类统计算法和经典K-means聚类算法,该算法无需人工选值、能够完全保证计算结果的准确性,且比经典K-means聚类算法和其他改进K-means聚类算法所需的迭代次数更少、运行时间更短。

Accurately detecting the oil-water interface and the liquid level in a crude oil tank is very important for petrochemical process system engineering related to the accuracy of the net oil moisture control and that of combination station check system metering. In order to solve the problems that the conventional classification algorithm needs to select typical values manually and the traditional K-means clustering algorithm also needs to select initial clustering centers manually, and the results might not be accurate and stable, this paper proposed an adaptive threshold algorithm for oil-water interface detection, which is based on improved K-means clustering optimization algorithm. The proposed algorithm can calculate the best typical values automatically, and improves the conventional classification algorithm and the traditional K-means clustering algorithm. The proposed algorithm works as follows: first, find the typical values automatically by using an adaptive threshold searching algorithm, then search the optimal points by using the improved K-means clustering optimization algorithm, and finally calculate the oil-water interface and the liquid level according to the optimal partition. Experimental results showed that, compared with the conventional classification algorithm and the traditional K-means clustering calculation algorithm, the proposed adaptive threshold clustering algorithm does not need artificial selection, and can guarantee the results to be correct. Moreover, the number of iterations and running time needed by our proposed algorithm are smaller than those by the traditional and other improved K-means clustering calculation algorithms.