论文部分内容阅读
对乙烯裂解炉建立实时监控模型具有重要的现实意义,而传统的多元统计过程监控方法都是假设过程处于单一工况下,而随着过程参数(进料负荷、产品组分等)的改变,工况也随之改变,传统方法便不再适用。本文针对工业过程中的多工况问题,提出了一种基于自适应模糊聚类的多模型过程监控方法,该方法可以减少监控方法对过程知识的依赖性,并且能够适应实际工业过程的非高斯性和非线性特征。首先对影响工况的过程变量利用自适应模糊聚类进行工况划分,然后对每种工况的建模数据分别利用最大方差展开(MVU)提取低维信息,再用支持向量数据描述(SVDD)建立多模型过程监控模型,最后再利用相应的统计指标进行过程监控。将上述方法应用在乙烯裂解炉上,并与基于高斯混合模型的多PCA方法(GMM-MPCA)进行了比较。仿真实验中,监控对裂解炉运行影响最大的33个变量,根据聚类有效性指标,将数据划分为5类时可以得到最佳的聚类效果。通过实验,将33维建模数据降到20维时误报率最小。仿真结果表明该方法在对非线性和非高斯性过程的监控上,能达到很好的效果,误报率和检测率均优于GMM-MPCA方法。
The establishment of real-time monitoring model for ethylene cracking furnace has important practical significance. However, the traditional monitoring methods of multivariate statistical process all assume that the process is under a single working condition. With the change of process parameters (feed load, product composition, etc.) Work conditions also change, the traditional method is no longer applicable. In this paper, a multi-model process monitoring method based on adaptive fuzzy clustering is proposed to solve the problem of multi-process conditions in industrial processes. This method can reduce the dependence of monitoring methods on process knowledge and can adapt to non-Gaussian Sexual and non-linear features. First of all, the process variables affecting the working conditions are classified by using adaptive fuzzy clustering. Then the MVV is extracted for the modeling data of each working condition, and then the support vector data description (SVDD ) To establish a multi-model process monitoring model, and finally use the corresponding statistical indicators for process monitoring. The above method was applied to an ethylene cracking furnace and compared with the multi-PCA method (GMM-MPCA) based on the Gaussian mixture model. In the simulation experiment, the 33 variables that have the most influence on the operation of the cracking furnace are monitored. According to the validity index of the cluster, the best clustering result can be obtained when the data is divided into five categories. Through experiments, the false alarm rate is the smallest when the 33-dimensional modeling data is reduced to 20 dimensions. The simulation results show that the method can achieve good results in monitoring nonlinear and non-Gaussian processes, and the false alarm rate and detection rate are better than the GMM-MPCA method.