Improved KECA for process monitoring using ensemble learning and Bayesian inference

Kernel entropy component analysis (KECA) has been introduced into process monitoring recently. Generally, there exist different kinds of faults in actually industrial process and the fault information is unknown. These faults contain different information, that a single kernel function with explicit width parameter is not effective enough to detect different types of faults. The width parameter, meanwhile, is usually determined by experience in kernel function. To address theses issues, we incorporate ensemble learning approach with Bayesian inference into KECA approach. This proposed approach overcomes the sightless of the parameter selection and is suitable for detecting different kinds of faults. We test it in the complicated Tennesse Eastman (TE) benchmark process, the result shows that the monitoring performance is markedly improved.