MULTIMODE PROCESS FAULT DETECTION BASED ON PROBABILITY DENSITY SUPPORT VECTOR DATA DESCRIPTION

In view of the traditional gaussian kernel width parameters optimized based on the Euclidean distance is vulnerable to the influence of the modal data differences significant variance characteristics, a multimode process fault detection based on probability density support vector data description method is proposed. Gaussian mixture model was used for pattern recognition of multi-condition process data, and the probability density values were calculated respectively. The nearest neighbor and far neighbor samples under each condition were found by using the probability quantile, which were combined to form a new sample set to optimize the gaussian kernel width parameters. The optimal kernel width parameters were used to establish multiple sub-SVDD models. The method based on variable contribution graph was used for fault diagnosis. The proposed method was applied to numerical examples and Tennessee Eastman (TE) chemical process, and the experimental results were compared with traditional SVDD and DFN-SVDD to verify the effectiveness of the proposed method.