CONDITIONAL ENTROPY-BASED FEATURE SELECTION FOR FAULT DETECTION IN ANALOG CIRCUITS

Abstract

ABSTRACT: To detect parametric faults in analog circuits, a novel feature selection algorithm based on conditional entropy, which was integrated with a support vector machine (SVM)-based fault detection approach, was proposed in this study. In preventing the significant loss of the effective features, a sampling process was executed with a significantly higher frequency. The side effect of this process showed that raw observation vectors were of extremely high dimensions. To reduce computation overhead, the feature selection algorithm based on conditional entropy was put forward to compress raw observation vectors into new observation vectors. The conditional entropy was used to update the conditional probability of a fault based on new fault information, which eventually made the fault probability more clear. By applying the proposed feature selection algorithm, it can compress raw data more wisely, and maximize the information in choosing the dimensions to be included in the new observation vectors. Simulation results showed that the fault detection approach presented in the study could classify non-linear feature vector space of the analog circuits. and achieved a lower misclassification rate than other current methods (i.e., equidistant method and conditional probability-based method). Keywords:Fault detection, Conditional entropy, Support vector machine (SVM), Classification