Abstract:

Process monitoring technology is an effective means to guarantee operation safety and product quality of modern industrial processes. Most of traditional process monitoring methods extract data features by dimensionality reduction and require process data obeying Gaussian distribution, linearity and other conditions. Therefore, traditional methods cannot obtain preferable detection results for faults occurred under complex operating conditions. A mixture of D-vine copulas model was proposed for fault detection. First, complex correlation among process variables were directly extracted without dimensionality reduction and a statistical model of process variables was established to accurately describing nonlinear and non-Gaussian processes. Then, model parameters were optimized by expectation maximization (EM) algorithm and maximum pseudo-likelihood estimation. Finally, a generalized Bayesian inference-based probability (GBIP) index was constructed for real-time monitoring by optimized model parameters as well as theories of the highest density region (HDR) and density quantile. Application of the proposed mixture model to a numerical example and the Tennessee Eastman (TE) benchmark process illustrated effectiveness and performance in fault detection.

Key words: process monitoring, nonlinear and non-Gaussian, dependence analysis, D-vine copulas

摘要：

过程监控技术是保证现代流程工业安全平稳运行及产品质量的有效手段。传统的过程监控方法大多采用维度约简方法提取数据特征，且要求过程数据必须服从高斯分布、线性等限制条件，对复杂工况条件下发生的故障难以取得较好的检测效果。因此，提出了混合D-vine copulas故障诊断模型，在不降维的情况下直接刻画数据中存在的复杂相关关系，构建过程变量的统计模型实现对存在非线性与非高斯性过程的精确描述。通过EM算法和伪极大似然估计优化混合模型参数，然后结合高密度区域（HDR）与密度分位数法等理论，构建广义贝叶斯概率（GBIP）指标实现对过程的实时监测。数值例子及在TE过程上的仿真结果说明了该混合模型的有效性及在故障检测中的良好性能。

关键词: 过程监控, 非线性非高斯, 相关性分析, D-vine copulas