An Improved KFCM Clustering Method Used for Multiple Fault Diagnosis of Analog Circuits

This study presents a new method for multiple fault diagnosis in moderate-sized analog circuits. Based on this method, a classifier is independently designed for each of the circuit components. Each of these classifiers only classifies the defect modes associated with the related component. The resultant effect is the much lower number of fault classes in each classifier than all of the circuit faults. Classifiers are designed based on a two-stage clustering method. Firstly, whole data are clustered by using Kmeans algorithm. Then, samples in each cluster are classified using a new version of KFCM algorithm. In this algorithm, initial cluster centers, as well as their number, are estimated by using an efficient method. Comparison with a neural network shows its very lower accuracy in classifying large number multiple faults. However, the proposed technique is capable of diagnosing multiple faults with acceptable accuracy (more than 91% for 15 single faults and more than 60% for most of the other multiple faults with the number greater than 200). Also this method has higher accuracy than traditional KFCM and YKFCM algorithms and this superiority rises along with increase in the number of faults.