Abstract: As industrial processes become more complex, on-line monitoring of the processes are gaining importance for plant safety, maintenance, and product quality. To handle the nonlinear problem for process monitoring, a novel fault detection method was proposed by combining locally linear embedding (LLE)with support vector data description (SVDD).Firstly, LLE manifold learning algorithm was performed for nonlinear dimensionality reduction and thus the main feature of the collected data was extracted. Then, the mapping matrix from data space to feature space was calculated by using local linear regression which guaranteed the real-time property of the algorithm. Next, in order to avoid the influence of data noise on the traditional statistics, a fault detection model was established based on SVDD in the feature space, while a corresponding statistic and its control limit were determined at the same time. Finally, the feasibility and efficiency of the proposed method were shown through a numerical simulation and the Tennessee Eastman (TE)benchmark process.

Key words: locally linear embedding, support vector data description, fault detection

摘要： 随着工业过程日趋复杂，系统安全及产品质量的在线监控也变得日益重要。针对化工过程的非线性特点，提出了一种新的基于局部线性嵌入(locally linear embedding, LLE)流形学习算法和支持向量数据描述(support vector data description, SVDD)的故障检测方法。首先，使用LLE提取高维数据的低维子流形，进行维数约减，以保存更多原有系统的非线性特性，通过局部线性回归得到高维数据空间到低维特征空间的映射矩阵，保证了算法的实时性；然后，为了避免数据噪声的累加对传统统计量的影响，引入SVDD直接根据特征空间建立SVDD模型，构造统计量并确定其控制限；最后，通过数字仿真及Tennessee Eastman(TE)过程仿真研究验证了本文方法的有效性。

关键词: 局部线性嵌入算法, 支持向量数据描述, 故障检测