A three layers multiple kernel learning model for analog circuit fault diagnosis

In order to improve the performance of multi-kernel diagnosis models, a novel three layers multiple kernel learning (MKL) algorithms for extreme learning machine (ELM) is presented by combining the radius-incorporated MK-ELM with AdaBoost strategy. Firstly, an improved MK-ELM model is established by replacing the radius of the minimum enclosing ball (MEB) with the trace of total data scattering matrix and replacing the classification error with the weighted classification error. Then, for the improved MK-ELM model, a three layer adaptive boosting framework is constructed by a three-stage iterative strategy based on Polynomial kernels, Laplace kernels and Gaussian kernels. In this way, proposed framework provides an excellent strategy to gradually improve classifier performance. The experiment results show that our algorithm has the overall best diagnosis performance compared with four state-of-the-art MKL models. Specifically, our algorithm can achieve comparable or even better classification performance than radius-incorporated MK-ELM, while incurring much less computational cost.