选择性融合多尺度筒体振动频谱的磨机负荷参数建模

Modeling mill load parameters based on selective fusion of multi-scale shell vibration frequency spectra

中文关键词  多组分信号分解  信息融合  选择性集成建模  振动频谱  软测量

英文关键词  multi-component signal decomposition  information fusion  selective ensemble modeling  vibration frequency spectrum  soft sensor

中文摘要

      针对目前采用经验模态分解(empirical model decomposition, EMD)得到的系列子信号构建的磨机负荷参数 软测量模型泛化性能差、难以进行清晰物理解释, 以及EMD算法存在的模态混叠等问题, 本文提出了基于选择性融 合多尺度筒体振动频谱的建模方法. 首先采用EMD、集合EMD(ensemble EMD, EEMD)、希尔伯特振动分解(Hilbert vibration decomposition, HVD)共3种多组分信号自适应分解算法获得磨机筒体振动多尺度子信号的集合, 接着通过 相关性分析剔除虚假无关部分, 然后再将与原始信号相关性强的那部分多尺度子信号变换至频域, 进而更有利于构 建这些多尺度频谱与磨机负荷参数间的映射模型, 最后通过改进分支定界选择性集成(improved branch and bound based selective ensemble, IBBSEN)算法建立软测量模型, 实现对多源多尺度筒体振动频谱的最优选择性信息融合. 基于实验球磨机运行数据的仿真实验表明所提方法在模型可解释性和泛化性能上均优于之前研究所提出方法.

英文摘要

      Soft sensor models of mill load parameters based on a set of sub-signals obtained by empirical model decomposition (EMD) have many shortcomings, such as low modeling accuracy and difficult interpretation. Moreover, EMD cannot get rid of the mode mixing problem. Thus, we propose a new soft sensor approach based on selective fusion of multi-scale shell vibration frequency spectra. At first, three multi-component signal decomposition algorithms, such as EMD, ensemble EMD (EEMD) and Hilbert vibration decomposition (HVD), are used to obtain a set of shell vibration sub-signals with different scales. Then, the correlation analysis between these sub-signals and the original signal is made, and the false decomposed part is excluded. Those sub-signals that have strong correlation with the original signal are transformed into frequency domain, which is helpful to construct the mapping model between these multi-scale frequency spectrum and mill load parameters. Finally, a new improved branch-and-bound-based selective ensemble (IBBSEN) algorithm is used to construct soft sensor models. Thus, the optimized selective information fusion of the multi-source multiscale shell vibration frequency spectra is realized. Simulation results based on operating data from a laboratory-scale ball mill shows that the proposed method outperforms the existing soft sensor approaches.