变分推理的概率高斯/非高斯过程监测  

Variational inference-based probabilistic Gaussian/non-Gaussian for process monitoring

作　　者：任珈仪 任世锦[2] 潘剑寒 杨茂云[2] 李新玉[3] REN Jiayi;REN Shijin;PAN Jianhan;YANG Maoyun;LI Xinyu(College of Science,Nanjing University of Posts and Telecommunications,Nanjing 210023,China;School of Computer Science&Technology,Jiangsu Normal University,Xuzhou 221116,China;School of Mechanical and Electrical Engineering,China University of Mining&Technology,Xuzhou 221116,China)

机构地区：[1]南京邮电大学理学院,江苏南京210023 [2]江苏师范大学计算机学院,江苏徐州221116 [3]中国矿业大学机电工程学院,江苏徐州221116

出　　处：《电子设计工程》2021年第14期84-89,共6页Electronic Design Engineering

基　　金：国家自然科学基金资助项目(61703187)。

摘　　要：复杂工业过程存在明显的高斯/非高斯特性,过程对象受到外在以及本身状态因素影响导致过程数据带有随机不确定性。基于学生t-分布可以通过调整自由度参数逼近高斯分布和非高斯分布,提出了一种变分推理的概率高斯/非高斯成分抽取统一框架。该方法使用贝叶斯变分推理方法学习模型参数,同时从观测数据抽取高斯成分和非高斯成分。与传统方法不同,提出方法不仅可以自动调整非高斯成分数量,而且考虑了各个维度的噪声水平,具有较好的鲁棒性。在独立成分隐空间、高斯成分隐空间和残差空间构建了故障检测统计量。TE仿真平台重点研究故障3和故障9的检测效果,并与其他方法进行了对比,结果验证了所提算法的有效性。Complex industrial processes usually behavior implicit Gaussian and non-Gaussian charac-teristics,and the process components are often interfered by external situations and its healthy conditions,leading to the uncertainty of measured process data.Considering student t-distribution can approximate Gaussian distribution and non-Gaussian distribution by tuning the freedom parameter,a unified framework for probabilistic Gaussian/non-Gaussian component extraction is developed in this paper.Contrary to traditional approaches,the proposed method can extract Gaussian and non-Gaussian components from process data simultaneously,and model parameters can be achieved through Bayesian variational inference.The method can determine the latent dimensions of independent and Gaussian components automatically.Moreover,the noise levels associated with each variable can be modeled and the robustness of the method is enhanced.Three fault statistics corresponding to independent components,Gaussian components and residuals are constructed in latent spaces.Simulations on TE benchmark are conducted on detection of fault 3 and fault 9 to validate the efficiencies of the proposed method by comparing the results of the state-art-of existed methods.

关 键 词：学生t-分布 变分推理 故障诊断 概率高斯/非高斯成分 

分 类 号：TN967.1[电子电信—信号与信息处理]