In the existing system identification algorithms,the commonly used Gaussian,student's t(St)and Laplace distributions all show symmetric statistical characteristics which makes them difficult to describe the asym-metric and skewed noise,therefore the performance of the corresponding algorithms may largely degrade with the skewed noise.To this end,this paper introduces the generalized hyperbolic skew student's t(GHSkewt)distribu-tion and proposes a robust identification algorithm for linear systems with the asymmetric and skewed noise.Firstly,the thick-tailed and skewed characteristics of the GHSkewt distribution are introduced detailedly and it is also proved that the standard student's t-distribution can be regarded as a special case of the GHSkewt distribu-tion;Secondly,the latent variables are introduced to mathematically decompose the GHSkewt distribution in order to facilitate the derivation and implementation of the algorithm;Finally,the system cost function with the latent variables is reconstructed under the expectation-maximization(EM)algorithm.The dynamic characteristics and noise distribution of the system are continuously learned from the contaminated data with iterative optimization,then the estimation of noise parameters and model parameters are realized.
关键词
鲁棒系统辨识/非对称偏斜噪声/广义双曲倾斜学生氏t分布/期望最大化算法
Key words
Robust system identification/asymmetric and skewed noise/generalized hyperbolic skew student's t(GHSkewt)distribution/expectation-maximization(EM)algorithm
维普
万方数据