Abstract: Support vector machines (SVM) are extensively exploited for small sample fault recognition. On the contrary, due to difficulties such as a priori knowledge acquisition and nonlinear recognition, fault recognition approaches based on Bayesian probability have found few applications. To deal with these problems, the latent function sigmoid compression Bayesian fault recognition (LS-BFR) is proposed. Based on Gaussian stochastic process and Bayesian probability, LS-BFR takes Gaussian regression as the latent function, and changes the regression output into probability by sigmoid compression. To improve the performance of LS-BFR for nonlinear fault recognition, kernel functions are introduced to perform latent Gaussian process regression, and a practical approach based on the Bayesian parameter estimation method is proposed to determine kernel function parameters. The performance of LS-BFR is validated by recognizing the faults of misalignment and unbalance simulated on a rotor testbed. Experiment results show that the LS-BFR approach based on latent function and S-compression can effectively perform small sample fault recognition and achieve better fault recognition accuracy than SVM.

Key words: Bayesian estimation, Fault recognition, Gaussian process regression, Latent function