Diagnosis of process faults in chemical systems using a local partial least squares approach

Abstract

This article discusses the application of partial least squares (PLS) for monitoring complex chemical systems. In relation to existing work, this article proposes the integration of the statistical local approach into the PLS framework to monitor changes in the underlying model rather than analyzing the recorded input/output data directly. As discussed in the literature, monitoring changes in model parameters addresses the problems of nonstationary behavior and presents an analogy to model-based approaches. The benefits of the proposed technique are that (i) a detailed mechanistic plant model is not required, (ii) nonstationary process behavior does not produce false alarms, (iii) parameter changes can be non-Gaussian, (iv) Gaussian monitoring statistics can be established to simplify the monitoring task, and (v) fault magnitude and signatures can be estimated. This is demonstrated by a simulation example and the analysis of recorded data from two chemical processes. 2008 American Institute of Chemical Engineers AIChE J, 54: 2581-2596, 2008 Keywords: condition monitoring, fault detection, fault diagnosis, partial least squares, local approach IntroductionThe condition monitoring of industrial processes is standard practice in many industries, particularly the chemical industry. Based on the large body of work in the area of fault detection and diagnosis (FDD), the goals of condition monitoring are threefold: (i) to detect anomalous behavior in a process from the measured process variables, (ii) to locate and identify the source of such behavior, and (iii) to identify the magnitude of the fault. An ideal condition monitoring procedure can prevent a major malfunction in processing units by identifying incipient faults before they actually cause significant damage. Detailed discussions of condition monitoring and fault detection aspects can be found in many survey papers, for example, Refs. 1-4, or research texts. 5-7Techniques for process monitoring and FDD are based on a variety of paradigms including signal-based techniques [8][9][10][11] that are mainly applied to mechanical systems, model-based techniques 4,12-14 which address a wider spectrum of applications, rule-based techniques, [15][16][17][18] With advances in instrumentation and computer technology, the number of recorded process variables has increased to the point that some or many of the measured variables can be highly correlated. 27 In such cases, the creation of condition monitoring models is hindered by multicollinearity and suitably derived methods must be used. The most popular of these methods are statistical-based techniques, such as principal component analysis (PCA), partial least squares (PLS), and their extensions.In a previous paper, 28 a statistical theory was noted which can be readily used for the early detection of incipient faults, known as the local approach. The theory can be used to transform the problem of detecting a fault in a stochastic process to that of monitoring the mean of a Gaussian vector that is constructed from the...

Fault detection in non-Gaussian vibration systems using dynamic statistical-based approaches

“…The underlying data model, describing the correlated vibration signals, is of Kalman innovation form and represents a state space model. The statistical local approach has been proposed as a method for detecting abrupt changes [28], and overcomes the effects of non-Gaussian process data [30]. This technique has been applied in many areas, such as bridge damage detection and mechanical vibration monitoring [29,31,32].…”

“…[33] proposed the integration of statistical local in an MSPC framework to detect incipient changes in the variable covariance structure. This second approach incorporates the statistical local approach into the dynamic monitoring framework which is based on SMI and therefore represents a dynamic extension of the work in [30,33].…”

Kernel PLS based prediction model construction and simulation on theoretical cases

“…Besides KPLS, several other variants of PLS have also been widely studied and implemented in industrial process in recent years, such as recursive PLS [8], multiblock PLS [11,13], local PLS [10] and dynamic PLS [1,20,31], etc. To achieve better fault identification results, total kernel PLS (T-KPLS) has also been introduced [16].…”

Kernel Generalization of PPCA for Nonlinear Probabilistic Monitoring

“…[1][2][3][4][5][6][7] By projecting the data into a lower-dimensional space, these MSPC methods can accurately characterize the operation state of the monitored process systems. Different statistics such as T 2 and SPE have been constructed for monitoring purpose.…”