Local Temporal-Spatial Structure Feature Extraction And Fault Detection Of Industry Processes

Timely process monitoring plays a critical role in maintaining the process safety and stability, as well as guaranteeing the production quality. In modern chemical process, numerous observations can be well collected and stored, which provide reliable basis for characterizing the process operating conditions. Multivariate statistical process monitoring (MSPM), which only depends on the process data for the feature extraction, process modeling and monitoring, has become one of the research hotspots in industrial process monitoring. MSPM-based methods typically employ dimensionality reduction to discover the underlying data properties and remove redundancy information, thus the performance of dimensionality reduction will directly influence the reliability of the monitoring performance. The conventional global based dimensionality reduction approaches, such as PCA, are mainly performed by the global data features. However, the detailed local neighborhood structure on the data manifold is failed to be discovered. From the perspective of dimensionality reduction, several effective monitoring methods to identify both spatial and temporal relationships among the process data are proposed based on the idea of local information based manifold learning method NPE. The case studies on the Tennessee Eastman process demonstrate the effectiveness of the proposed methods in fault detection.1. A new nonlinear dimensionality reduction method named kernel orthogonal neighborhood preserving embedding (KONPE) is proposed and applied for nonlinear fault detection, with the application of kernel-trick on the method orthogonal neighborhood preserving embedding (ONPE). As a local space structure based approach, neighborhood preserving embedding (NPE) aims at preserving the latent manifold of data hidden in the high dimensional observations. Imposing orthogonal constraints on NPE, ONPE can effectively improve its discriminating power and the ability of feature extraction. The developed KONPE explicitly considers the low-dimensional structure in data, thus the nonlinear modeling performance is well improved. Simulation results illustrate the superiority of KONPE in process monitoring in comparison with the widely used KPCA.2. A nonlocal space structure constrained feature extraction method named nonlocal structure constrained neighborhood preserving embedding (NSC-NPE)is developed, on the basis of the local information based manifold learning approach neighborhood preserving embedding (NPE). NPE mainly focus on preserving the local geometry structure of the process data, and does not give a constraint for the data points outside the neighborhood. As a result, the intrinsic data information may lose. Considering such deficiency, in the new method, the data relationship outside the neighbors is also given constraints. By utilizing the meaningful nonlocal variance information, NSC-NPE constructs a global information based dual-objective optimizations function for modeling the process data. The relationship among the data points in the neighborhood which represents the local structure and the relationships among the data points outside the neighbors which represents the nonlocal structure are both be considered in the objective function. As a result, the global geometrical structure of the data is totally exploited. Different from the global based model PCA, NSC-NPE deals with the global data relation by dealing with the data points in and outside the neighbors with different strategies, respectively. Thus, NSC-NPE can give more faithful representation of the data character and better monitoring performance.3. With the consideration of autocorrelation among data samples, a novel algorithm named time neighborhood preserving embedding (TNPE) is proposed by utilizing local information. Furthermore, by taking both the temporal and spatial information of the process data into consideration, the method named time and spatial neighborhood preserving embedding (TSNPE) is also given. In industry process, the process variables always have autocorrelation and the system has dynamic properties, such dynamic behavior varied according to specific process condition should be totally extracted. The local space geometry based algorithms may be invalid to deal with the samples with autocorrelation. From this point of view, the neighborhood space is constructed with respect to the time sequence adjacent points of each data point, and the dynamic relationship is represented as a linear combination of its nearest neighbors. The local dynamic character is preserved in the low dimensional space by keeping the reconstruction coefficients. A numerical example and the Tennessee Eastman process indicate the benefit of having process dynamic information included in modeling process. By taking the autocorrelation of the samples in the modeling process, the two methods may be easy to identify the intrinsic data geometry structure and preserve the spatial and temporal relationships effectively.4. Based on the local information based manifold learning algorithm, NPE, the basic theories and characteristics in applying the manifold learning on process monitoring are discussed. Research on process monitoring based on manifold learning has been well developed recent years, with much good results. However, the theoretical studies of these methods on this field is much few. Thus, the validity of manifold learning is discussed by analyzing the algorithm features, applied conditions, as well as the scope of application. Then, the discussions on the construction of the statistics upon the NPE model are given, which presents the differences between the local based method and the traditional method about the Hotelling’s T2and squared prediction error (SPE) statistics. Finally, the advantages as well as the disadvantages of the application in process monitoring by manifold learning are illustrated.