Abstract: In order to improve the fault diagnostic accuracy of analog circuit, an improved l_p-norm multiple kernel extreme learning machine (ELM) diagnostic model is proposed based on the feature selection algorithm with one-dimensional ambiguity among fault features. In this model, the weighted classification error is incorporated into kernel ELM optimized objective function, and then a three-layer multiple kernel learning framework is constructed based on adaptive boosting (AdaBoost) algorithm, in which the training sample weights are adaptively adjusted so that the every layer in the model can focus on the different training samples. The proposed model provides an excellent strategy to improve the identifiability of classifier. Experimental results of two analog circuits show that the proposed model can achieve approximately consistent diagnostic performance under the constraints of different norms. For a fault with a single attri- bute, the diagnostic accuracy can be significantly improved, meanwhile the missing alarm and false alarm can be effectively balanced. For a fault with multiple attributes, the faults which are difficult to be identified can be accurately isolated into relevant ambiguity groups. Key

Key words: analogcircuit, l_p-normconstraint, multiplekernellearning, extremelearningmachine, ensemblelearning, ambiguitygroup, faultdiagnosis