Measured data abnormality in multimode batch processes directly influences model accuracy of data-driven multivariate statistical analysis and decreases performance of process monitoring and controlling. A dynamic hypersphere structure change (DHSC) derived method was proposed for detecting such data abnormality. First

mode dicing according to membership change was achieved by introducing sequence-constrained fuzzy C-means (SCFCM). Then

for each mode

support vector data description (SVDD) was used to build a static hypersphere for training data and a dynamic hypersphere for testing data. Finally

important support vectors were chosen as structures of hyperspheres and detection of abnormal data was achieved by identifying structure change of hyperspheres. Simulation experiment of penicillin fermentation process shows that the present method can achieve mode division of multimode batch processes and reduce influence of mode switch on detection accuracy of abnormal data. Using structure change of hypersphere to detect abnormal data can decrease false detection rate with high detection accuracy.