Algorithms comparison of feature extraction and multi-class classification for fault diagnosis of analog circuit

For the novelties or anomalies of faulty signals occur in a damage circuit and fault signals vary with different circuit damages. To ensure the accuracy and reliability of diagnosis, it is very important to extract the characteristic features of fault signals. Two feature extraction methods based on wavelet packet transform is proposed to treat transient signals: optimal wavelet packet transform (OWPT) and incomplete wavelet packet transform (IWPT). For the fault signals decomposed, the energy in each frequency band may be heightened or be reduced, so a novel ´energy-fault´ method is put forward to extract fault features. The problem of fault diagnosis of analog circuit is actually a pattern recognition problem. Nowadays, the binary tree support vector machines (BTSVMs) is usually used for multi-class classification, but the structure of the binary tree is closely related to the classification performance of binary tree support vector machines (BTSVMs). A new separability measure method based on the space distribution of pattern classes is applied to construct different binary trees. Three BTSVMs classifiers based on the separability measure are defined in this paper: inclined binary tree support vector machines (IBTSVMs), balanced binary tree support vector machines (BBTSVMs) and adaptive binary tree support vector machines (ABTSVMs). Simulation results show us that the OWPT method is prefect for soft fault diagnosis, the IWPT for hard fault diagnosis, and the BBTSVMs multi-classifier possesses better classification speed, the ABTSVMs multi-classifier better classification accuracy.