在线KPLS建模方法及在磨机负荷参数集成建模中的应用

针对过程非线性、基于历史数据构建的离线模型泛化性差以及基于滑动窗口 和每样本递推更新的在线建模方法难以均衡建模精度和建模速度等问题, 提出了一种在线 核偏最小二乘(On-line kernel partial least squares, OLKPLS)建模方法. 该方法依据新样本与建模样本间的近似线性依靠(Approximate linear dependence, ALD)值和代表工业过程特性漂移幅度的 阈值, 选择有价值样本更新KPLS模型, 并采用合成数据和Benchmark平台数据对该方法进 行了仿真验证. 针对基于离线历史数据建立的融合多传感器信息的磨机负荷参数集成模型难以适应磨 矿过程时变特性的问题, 提出了基于OLKPLS和在线自适应加权融合算法的在线集成建模方 法, 并通过实验球磨机的实际运行数据仿真验证了方法的有效性.

On-line KPLS Algorithm with Application to Ensemble Modeling Parameters of Mill Load

Industrial processes have the characters of strong nonlinearity, and soft sensor models constructed based on historical data have low generalization. Normally used moving window and sample recursive updating on-line modeling approach can not obtain a trade off between the modeling accuracy and the modeling speed. To aim at these problems, a novel on-line kernel partial least squares (OLKPLS) algorithm is proposed in this paper. The approximate linear dependence (ALD) value between the new sample and the modeling samples is calculated. According to the threshold value that represents the concept drift amplitude of the industrial process, only new samples satisfy the ALD condition to update the KPLS model. Synthetic data and benchmark data are used to validate this approach. To aim at the multi-sensor information fusion ensemble model of mill load parameters based on off-line historical data, an on-line ensemble modeling approach based on OLKPLS and on-line adaptive weighting fusion algorithm are proposed, which are validated using the simulation results based on operating data of the laboratory ball mill.