基于图分割与协同训练的工业过程半监督软测量方法

Semi-supervised Soft Sensor Modelling Method Based on Graph Segmentation and Co-training

在工业过程中,质量检测数据与运行数据的比例失衡导致数据驱动的软测量性能受到限制。基于协同训练算法的半监督建模方法利用海量运行数据获得具有高置信度的伪标签以扩充有效建模数据,然而协同训练建模过程中的特征分割会影响建模中两模型的差异性进而影响最终的模型性能。针对该问题,本文提出一种基于图分割与协同训练偏最小二乘(GSCO-PLS)的半监督软测量算法。首先计算辅助变量之间的相关性指标作为权重获得以特征为顶点的有权无向图,将协同建模中的特征分割转化为平衡最小割图分块问题,以获得两个互相关性最小的特征子集。在此基础上通过协同训练利用运行数据更新初始模型实现半监督建模,通过对初始模型特征分割的优化实现对协同训练半监督软测量方法精度的提升。最后通过TE仿真过程的结果说明本文所提算法的有效性。

In industrial processes,the imbalance between quality data and operating data limits the performance of data-driven soft sensor models.Semi-supervised modeling methods based on co-training algorithms can utilize massive amounts of operating data to obtain pseudo-labels with high confidence,thereby augmenting the effective modeling data.However,the feature segmentation in the co-training modeling process affects the differences between the two models,thereby influencing the final model performance.To address this issue,this paper proposes a semi-supervised soft sensor algorithm based on graph segmentation and co-training partial least square(GSCO-PLS).Firstly,the correlation indices between auxiliary variables are computed as weights to construct a weighted undirected graph with features as vertices.The feature segmentation in co-training modeling is transformed into a balanced minimumcut graph partitioning problem,aiming to obtain two feature subsets with the least intercorrelation.On this basis,semi-supervised modeling is achieved through co-training,where the initial model is updated using operating data.The precision of the co-training semi-supervised soft sensor method is improved through optimization of the initial model's feature segmentation.Finally,the effectiveness of the proposed algorithm is demonstrated through simulations of the Tennessee Eastman process.