基于自适应多核潜结构映射选择性集成模型的磨机负荷参数预测
Mill load parameters forecasting based on adaptive multi-kernel projection to latent structure selective ensemble model

磨机负荷参数是影响选矿流程产品质量和产量的难以检测关键过程变量。磨机研磨产生的具有多尺度、高维和共线性等特性的多源机械信号频谱与磨机负荷参数间存在复杂的非线性映射关系。核潜结构映射(KPLS)算法适合构建基于频谱数据的磨机负荷参数预测(MLPF)模型。如何进行多核参数的自适应选择和多源信息的选择性融合是目前待解决的开放性问题。针对上述难点，本文提出一种面向MLPF的自适应多核潜结构映射选择性集成(SEN)模型。首先，基于经验模态分解(EEMD)和时频变换技术处理多源机械信号，得到基于不同时间尺度候选子信号的频谱数据；接着，采用KPLS和分支定界选择性集成(BBSEN)算法，构建基于多尺度频谱的候选子子模型和SEN子模型，其多核参数从依据先验知识确定的候选集合内自适应选择；最后，从候选子子模型和SEN子模型中优选获得不同时间尺度的候选子信号模型，并再次采用BBSEN算法优选集成子信号模型并加权组合，构建最终的MLPF模型。基于实验球磨机的实际运行数据仿真验证了所提方法的有效性。
Load parameters inside ball mill are difficulty-to-measure key process variables relative to production quality and quantity of the whole grinding process. There are complex nonlinear mapping relationships between mill load parameters (MLPs) and multi-source mechanical frequency spectral data with characteristics of multi-scale and multi-component and collinearity. Kernel project to latent structure (KPLS) algorithm is suitable to build mill load parameter forecasting (MLPF) model based on such frequency spectral data. How to make adaptive selection of multi-kernel parameter and how to make selective fusion of multi-source information are open issues. Aim to these problems, a new adaptive multi-kernel projection to latent structure selective ensemble (SEN) model for MLPF is proposed. At first, candidate sub-signals’ frequency spectral data with different time scales are obtained by using ensemble empirical model decomposition (EEMD) and time/frequency transformation techniques from multi-source mechanical signals. Then, candidate sub-sub-models and SEN-sub-models are constructed based on different frequency spectral data by using KPLS and branch & bound SEN (BBSEN) algorithms, whose multi-kernel parameters are selected from candidate kernel parameter set adaptively. Finally, the candidate sub-signal models are optimal selected from these candidate sub-sub-models and SEN-sub-models; BBSEN is used again to select ensemble sub-signal models from these candidate ones and to weight them. Therefore, the final MLPF model is constructed. Simulation results of a laboratory-scale ball mill show effectiveness of the proposed approach