关键词: 特征选择；多标记学习；质量监测；不均衡

Abstract: For process manufacturing enterprises, the monitoring accuracy of product quality status directly affects the production and operating costs of the enterprise. Facing the multi-variable monitoring requirements and data imbalance in the process industry, previous studies mainly adopted partial modeling strategies or multi-output models, which had the problems of feature selection bias and low classification accuracy. In this regard, a quality status monitoring model is designed that combines SRFML feature selection and Lift learning strategy, and is aims to improve the monitoring effect of the model by sharing information between different targets. First, according to the ReliefF filtering mechanism, the idea of resampling is introduced to optimize the selection process of industrial features (SRFML); then, the selection result is used as the input of the Lift learning framework, and the unique association of each feature to be monitored is reshaped through the generic attribute learning method Attributes; finally, multiple SVM classifiers are used for training, and the quality status results of each target are obtained. The results show that the SRFML-Lift constructed in this paper has fully learned the key information of the original characteristics, and compared with other combination strategies, it has a better monitoring effect on the quality status and can be applied to the production management practice of the process industry.

Key words: feature selection; multi-label learning; quality monitoring; data imbalance