基于多组分机械信号的湿式球磨机负荷混合集成智能识别模型

Mill load is one of the most difficulty-to-measure process parameters that relates to the global optimization of the grinding process. It is difficult to describe the fuzzy relation between the mechanical vibration & acoustic signals and the mill load. The recognition result of mill load status based on domain experts’ experience by hearing acoustic signal is subjective and one-sided. Moreover, it is difficult to obtain the labeled data in the industrial ball mill. In order to solve these problems, a hybrid ensemble intelligent identification model of wet ball mill based on multi-component mechanical signals is proposed. It includes:.1) The adaptive decomposition techniques are used to process the mechanical signal in time- and frequency-domain, and using a small number of labeled data to generate virtual samples, and then mixed with the real samples, data driven latent structure ensemble model is constructed to improve the stability and accuracy..2) Aim to a small number of labeled data and a large number of unlabeled data, deep neural network are used to extract the depth features selectivity. A fuzzy inference ensemble model based on data generation rules and expert complete rules is constructed to improve the generalization performance..3) The above constructed ensemble model based on different modeling mechanism are integrated as the main model of mill load reorganization. Then, error compensation model based on all features are constructed with the prediction error of main model as learning objectives. Thus, hybrid ensemble intelligent recognition model is obtained,, which is updated online based on evaluation of experts..4) Experimental verification and industrial application..This project is of great significance to realize the real time identification of mill load and optimization of the grinding process.

磨机负荷是与选矿全流程优化密切相关的难以检测参数。磨机的机械振动/振声等信号与磨机负荷间存在难以精确描述的模糊关系;专家依据经验“听音识别”磨机负荷存在主观性和片面性等问题;构建模型的有标签数据难以获得。针对这些问题，拟构建基于多组分机械信号的湿式球磨机负荷混合集成智能识别模型，包括：1)采用自适应分解技术对机械信号进行时频处理，利用少量有标签数据生成虚拟样本，与真实样本混合后构建数据驱动潜结构集成模型以提高稳定性；2)面对少量有标签和大量无标签数据，利用深度神经网络选择性提取深度特征，构建基于数据产生规则和专家完备规则的模糊推理集成模型以提高外推性；3)并联集成上述基于不同建模机理的异质模型作为主模型，与特征驱动的误差补偿模型组合后获得具有仿专家认知的混合集成智能识别模型，并依据专家评价在线更新模型；4)进行实验和工业应用研究。本项目对实现磨机负荷实时识别、磨矿过程优化控制意义重大。

磨矿过程是将破碎后的原矿通过球磨机研磨成粒度合格的矿浆，合格矿浆经选别过程产生精矿。磨矿过程是铁矿和各种有色金属矿选矿生产中的关键环节。磨机负荷大小直接关系到磨矿过程加工产品的质量、效率、能耗、物耗和安全运行。准确检测磨机负荷是实现磨矿过程优化运行的关键因素之一。针对磨矿过程的连续不间断运行特性和机械设备旋转封闭的特殊性导致获取完备训练样本的经济性差和周期性长等问题，提出了基于虚拟样本生成技术的多组分机械信号建模方法。针对经验模态分解算法及其改进版虽然能够对磨机筒体振动信号进行自适应分解，但存在模态混叠等问题，提出了基于变分模态分解（VMD）对磨机筒体振动信号分析方法。针对磨机筒体振动/振声多尺度频谱与磨机负荷参数间的模糊特性、多源多尺度频谱间的冗余性与互补性等问题，提出了基于多尺度频谱特征自适应提取与选择的选择性集成模糊推理软测量方法。针对如何融合球磨机系统研磨过程所产生的多模态机械信号构建磨机负荷参数预测(MLPF)模型的问题，提出了基于多模态特征子集选择性集成(SEN)建模的MLPF方法。针对潜结构映射模型和模糊推理模型在建模机理上存在差异性、运行专家只能对磨机负荷进行模糊认知和工业现场只能通过实验设计得到有限建模样本等问题，提出了基于数据和模糊推理的混合集成磨机负荷软测量方法。搭建了多通道实验磨机机械信号采集系统并进行了大量实验，基于针对实验磨机的分析结果和工业磨机的实际运行情况，搭建了多通道工业磨机机械信号采集系统并开展工业实验研究。本项目为磨机负荷实时检测和磨矿过程的运行优化提供有效支撑。