稀疏贝叶斯相关向量机的模拟电路故障诊断

Analogous Circuit Fault Diagnosis on Sparse Bayesian Relevant Vector Machine

模拟电路故障诊断受制于传统的机器学习方法需要人为设定参数,分类效果依赖于参数设定是否成功,无法进行在线诊断。为此,提出一种基于稀疏贝叶斯相关向量机理论的模拟电路故障诊断模型,改进权值更新算法,设定阈值提前剔除非相关权值,减少算法运行时间,加快权值更新速度。在贝叶斯框架下对分类函数的权重进行推断,并得到各分类的后验概率,从而判断分类结果的置信度,辅助诊断决策。仿真结果表明,与支持向量机相比,该模型在精度相当的情况下,需要的相关向量更少,更具稀疏性和泛化性,分类时效性更高,适合模拟电路的在线检测。

Analogous circuit fault diagnosis is influenced by parameter selection of classical machine learning approach,the result of classification relies on parameter whether suitable or not,that is unable to carry on diagnosis online.This paper proposes an analogous circuit fault diagnosis model based on Relevant Vector Machine（RVM） from the sparse Bayesian theory,and improves the weight renewal algorithm.The hypothesis threshold value picks out unrelated weights before they approach infinity,this can reduce the algorithm running time and speed up the weight refresh.RVM can infer the discriminant function under the Bayesian framework.Moreover,it can obtain posterior probability of each classification,thus can judge the degree of classification result confidence,assist diagnosis decision-making.The result indicates that RVM need less relevance vectors than support vector machine with comparative default accuracy,sparser and generalizing.It suits to online fault detection.