Abstract: One of the central problems of many automation systems is Fault Detection and Diagnosis (FDD). The data from real systems are commonly high-dimension and hard-separated. Various mathematical techniques have been applied on these data. One of the problems in FDD, which so far has been still hard to solve, is how to deal with the data nearly reached the critical mass. Because the system is very unstable when it nearly reaches the critical condition, it may become very hard to collect data and make decision. An improved Support Vector Machine (SVM) classifier with a soft decision-making boundary is proposed in this paper. The boundary is constructed based on belief degrees of data and reflects the data distribution. A membership function of the critical condition is introduced to extract the critical state data. In order to deal with these critical state data, this paper introduces and discusses two different experimental strategies.