基于多目标分层遗传算法的溢流粒度软测量

发布时间：2018-04-16 05:23

本文选题：溢流粒度 + 软测量　；参考：《大连理工大学》2015年硕士论文

【摘要】：磨矿过程的旋流器溢流粒度是判断磨矿分级作业生产状况及后续产品质量的重要指标。由于影响溢流粒度的因素很多且关系复杂,难以建立机理驱动粒度检测模型,因此,工业现场一般采用离线化验或在线检测的方法对溢流粒度进行检测。然而,离线化验方法满足不了实时性要求,在线检测方法因受噪声等因素影响测量精度不高。鉴于磨矿过程积累的大量历史数据,可以采用数据驱动软测量方法对溢流粒度进行估计,进而为磨矿过程的控制及决策提供参考信息。针对溢流粒度检测时存在的建模数据含噪声信号较高,辅助变量难以确定,对溢流粒度建立软测量模型既要求准确性又要求稳定性等问题,本文提出了一种基于多目标分层遗传算法的溢流粒度模糊建模方法,该方法将模糊模型分为四层：输入变量层、隶属度层、规则库层和系统集成层。输入变量层用于获取软测量模型的辅助变量,隶属度层用于获取隶属度函数类型及相关参数,对辅助变量进行模糊划分,规则库层用于确定模型的所有规则,系统集成层将前三层关联起来,代表一个完整的软测量模型。为达到各层共同进化的目的,本文设计了遗传算法各层编码策略,并构建了以平均绝对百分误差(Mean Absolute Percentage Error, MAPE)和均方根误差(Root Mean Square Error, RMSE)为标准的适应度函数来计算遗传算法每一层个体的适应度值。鉴于模糊模型训练过程中可能出现异常解,本文将L-M贝叶斯正则化方法融入训练过程。为验证本文方法的有效性,分别选取标准数据集和我国某选矿厂实际生产数据进行实验,并与已有多种方法进行对比实验,实验结果表明本文方法对含噪声磨矿数据进行软测量建模具有较好的准确性和稳定性。基于本文方法所实现的软件系统在实际应用中效果显著。
[Abstract]:The overflow granularity of hydrocyclone in grinding process is an important index to judge the production status of grinding classification operation and the quality of subsequent products.Because there are many factors affecting the overflow particle size and the relationship is complex, it is difficult to establish a mechanism-driven particle size detection model. Therefore, off-line or on-line detection methods are generally used to detect the overflow particle size in industrial field.However, the off-line test method can not meet the real-time requirements, and the measurement accuracy is not high due to noise and other factors.In view of the large amount of historical data accumulated in the grinding process, the method of data-driven soft sensing can be used to estimate the overflow particle size, thus providing reference information for the control and decision-making of the grinding process.In view of the high noise signal in the modeling data of overflow granularity detection, it is difficult to determine the auxiliary variables. The establishment of soft sensor model for overflow granularity requires both accuracy and stability.In this paper, a fuzzy modeling method of overflow granularity based on multi-objective hierarchical genetic algorithm is proposed. The fuzzy model is divided into four layers: input variable layer, membership layer, rule base layer and system integration layer.The input variable layer is used to obtain the auxiliary variables of the soft sensor model, the membership level is used to obtain the membership function types and related parameters, and the auxiliary variables are divided fuzzy, and the rule base layer is used to determine all the rules of the model.The system integration layer associates the first three layers and represents a complete soft sensor model.In order to achieve the goal of coevolution of each layer, the coding strategy of each layer of genetic algorithm is designed in this paper.The fitness function of mean Absolute Percentage error (MAPE) and Root Mean Square error (RMSE) is constructed to calculate the fitness of each layer of genetic algorithm.In this paper, L-M Bayesian regularization method is incorporated into the training process in view of the possible abnormal solutions in the process of fuzzy model training.In order to verify the validity of this method, the standard data set and the actual production data of a concentrator in our country are selected for experiments, and compared with many existing methods.The experimental results show that the proposed method has good accuracy and stability for soft sensor modeling of noisy grinding data.The software system based on this method is effective in practical application.
【学位授予单位】：大连理工大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TD921.4;TP18

【参考文献】