Multi-mode Process Monitoring Based On Gaussian Mixture Model

In recent years, with the urgent market demands of multi-species, multi-standard and high-quality products in modern society, industrial productions are depended more on the efficient processes of producing a variety of products and the security and the reliability of production processes have became the focus points.However, the changes of producing process programs or product types would lead to a variety of modes with different process characteristics in producing processes.Process monitoring of multi-mode is a complex issue because of many characteristics, such as, multi-variable, multi-operation, time-variant characteristics, and uncertainty of mode conversion. We should consider process monitoring not only in stable modes, but also in transitional modes between stable modes. Gaussian mixture model is used to describe a mixture density distribution. It can approximate any continuous distribution with enough members in the model.Gaussian mixture model is a semi-parametric density estimation method, which combines the advantages of parameter estimation method and the non-parametric estimation method. The process monitoring method based on Gaussian mixture model is proposed for multi-mode process.1. Mode Identification. Mode identification is judging mode type of data which does not have mode information. Whether in offline modeling step or online monitoring step, identifying corresponding mode information for data is the key. For offline mode identification, Gaussian mixture model is used to calculate the posterior probability of offline data. And the clustering algorithm is used to divide data into stable modes and transitional modes. For online mode identification, Gaussian mixture model is used to obtain the joint density. And the current process status is combined to judge mode type for online data.2. Offline Modeling. Gaussian mixture model is used for offline modeling. Different models are built separately to describe data characteristics for stable modes and the transitional modes. Model based on difference method is used to extract dynamic characteristics of transitional mode. This multi-model modeling method improves the accuracy of describing multi-mode process characteristics.3.Online Monitoring. A probability index based on Bayesian inference (BIP) is defined for process monitoring. BIP indicator is consisted of posterior probability and the probability based on Mahalanobis distance. Online monitoring is realized through the comparison of BIP indicators and the control limits. Finally,the simulation results show effectiveness of the proposed multi-mode process monitoring method.