A Quality Condition Monitoring Model For Multi-label Imbalanced Data

During industrial production,real-time monitoring of key quality characteristics is an important measure to ensure the final quality of products.In the intelligent production mode,the complexity of the manufacturing system continues to increase,and the post-mortem inspection technology relying solely on statistical process control can no longer meet the production needs.The rapid development of artificial intelligence technology and machine learning algorithms has provided new solutions for the quality control of industrial products.Make full use of the historical production data of the product,construct and predict the monitoring model,make real-time feedback adjustment to the abnormal status of the product,and realize the online monitoring of the quality status of the process-manufactured product.Compared with the discrete manufacturing mode,due to the dynamic customer demand and the complexity of the manufacturing process,the product quality data usually has the characteristics of high dimension,imbalance,and multi-objective in process manufacturing.In this paper,the problem of product quality status monitoring in the process industry is transformed into two typical machine learning problems,and the real-time production status of products is monitored through the mapping relationship between process data and product quality characteristics.The main improvements and research contents are as follows:(1)In view of the weak generalization ability of a single monitoring model,this paper considers stacking fusion of multiple models.Under the stacking framework,based on the correlation and prediction accuracy of each single classifier,a basic learning layer optimization method based on stacking learning is proposed.(2)The multi-objective characteristics of quality monitoring are considered.In this paper,the concept of generic attribute learning is introduced,a multi-label training process based on Lift algorithm is constructed,and combined with the Stacking algorithm,a multi-label monitoring model with dynamic adjustment capability is proposed.(3)From the perspective of feature selection,the disturbance of the unbalanced distribution of multiple marker sets to the monitoring model is studied.In this paper,the distance measurement process of RFML algorithm is re-weighted,and a feature selection method for multi-label industrial data is proposed.This paper uses the injection molding data of Foxconn enterprises,through exploratory research on the production process data,and after industrial feature extraction and fusion,a monitoring model of quality status is constructed.And the horizontal comparison with other combined algorithms verifies the effectiveness of the method in this paper,and provides decision-making help for the quality management of the process industry through the sensitivity analysis of the machine learning model.