Process Monitoring Using Independent Component Analysis Based On Gaussian Mixture Model

With the development of the automation and integration levels of industrial process, on the one hand, the correlation among the production devices and the coupling between variables have become higher and higher, which generates a great influence on the production process owing to the minor abnormal condition. Therefore, it is necessary for an enterprise to build a safe, stable, and reliable process monitoring system. On the other hand, due to the application of field bus technology and distributed control system (DCS), a lot of data have been collected and stored into the database of enterprises. Then, how to apply the contained information from data to the process monitoring is becoming a research hotspot.Independent component analysis is an effective method based on statistics theory, which does not have to assume that the process variables must obey the Gaussian distribution, and also makes use of the higher-order statistic information. ICA has become a dominant method in multi-variable statistic process monitoring. In this paper, an ICA based on Gaussian mixture model is proposed, which takes the characteristics of multi-situation in the process of real-world production and the downside of FastICA into consideration, adopts the entropy based on Gaussian mixture model to measure the non-Gaussian feature of a random variable. And, the differential evolution algorithm is then adopted to eliminate the excessive dependence on the initial value of the FastICA algorithm, the entropy of each independent component is sorted, and the SPE statistical variable is adopted to realize industrial process monitoring.To demonstrate the validity and reliability of the proposed method, this paper takes the typical process-semi-autogenous grinding in beneficiation-as background, in which the simulation experiments consists of the normal and the abnormal conditions. By using the real-world data coming from a plant, a series of comparative experiments are carried out. The experimental results indicate that the proposed method is reliable and stable for the industrial process monitoring.