基于非参数统计检验的早期故障检测与诊断研究

Incipient fault detection and diagnosis is of great significance to the safety of equipment and industrial process. From the aspect of statistics, the influence of incipient fault to data can be divided into three categories, i.e., change in first order, second order and higher order statistics. To deal with the three kinds of incipient faults, this project proposes a fault detection and diagnosis framework based on nonparametric statistical test. The basis idea is to transform the single sample monitoring problem into distribution test between normal sample set and subset of test samples. The following aspects are considered: i) A fault monitoring strategy based on empirical likelihood is proposed to detect change in the first order statistic; ii)By projecting the normal data set and the test subset into a kernel space, the kernel mean discrepancy method uses the discrepancy between mean values of both sets in the kernel space as a test statistic; based on this statistic, a fault monitoring and diagnosis strategy to detect change in second order statistic is proposed. iii)To measure the difference between the probability density distributions of normal data set and test subset, a fault detection and diagnosis scheme is proposed using Kullback-Leibler divergence (KL divergence). The developed methods are then applied to metallurgical and chemical processes like blast furnace ironmaking process. The fault diagnosis methods based on distribution test do not rely on any distribution assumption; they are more sensitive to small and incipient process faults, and hence are suitable for early alarm of process faults. The research is helpful for improving the product quality and control performance of metallurgical and chemical processes; and is of great importance in both theory and practice.

早期故障的检测和诊断对设备和过程的安全具有重要意义，从统计角度看，早期故障在数据中的表现形式可分为：一阶统计量变化、二阶统计量变化和高阶统计量变化等。本项目针对以上故障类型，提出一套基于非参数统计检验的早期故障检测和诊断方法。首先将待测样本划分为固定长度的子集，将单样本检验问题转化为分布检验问题，具体包括：基于经验似然的故障检测方法，适用于检测一阶统计量的变化；对二阶统计量变化的早期故障，拟利用核均值偏差检验，将样本投影到核空间，比较其均值偏差，实现检测和诊断的目的；为比较正常样本与待测样本的概率分布，设计基于KL距离的早期故障检测与诊断方法。并将研究方法应用于高炉冶炼等冶金和化工过程。基于非参数统计检验的早期故障检测和诊断方法不依赖于任何分布假设，对微小、缓变的早期故障具有更高的灵敏度。项目的研究成果将有助于提高冶金和化工等生产过程的产品质量和控制性能，具有重要的理论和应用价值。

早期故障的检测和诊断对设备和过程的安全具有重要意义，从统计角度看，早期故障在数据中的表现形式可分为：一阶统计量变化、二阶统计量变化和高阶统计量变化等。本项目针对以上故障类型，提出一套基于非参数统计检验的早期故障检测和诊断方法。首先将待测样本划分为固定长度的子集，将单样本检验问题转化为分布检验问题，具体包括：基于经验似然的故障检测方法，适用于检测一阶统计量的变化；对二阶统计量变化的早期故障，拟利用核均值偏差检验，将样本投影到核空间，比较其均值偏差，实现检测和诊断的目的；为比较正常样本与待测样本的概率分布，设计基于KL 距离的早期故障检测与诊断方法。并将研究方法应用于高炉冶炼等冶金和化工过程。基于非参数统计检验的早期故障检测和诊断方法不依赖于任何分布假设，对微小、缓变的早期故障具有更高的灵敏度。经过项目组成员三年来的共同努力，基金项目的研究工作按照计划顺利进行，主要获得的成果包括：（1）发表、录用论文12篇，其中SCI期刊论文9篇，包括TOP期刊Automatica，AIChE Journal等杂志，EI期刊论文1篇，会议论文2篇；（3）参加国际会议1次，国内会议4次；（4）培养硕士生4人。