Monitoring and diagnosis of process systems using kernel-based learning methods

[摘要] ENGLISH ABSTRACT: The development of advanced methods of process monitoring, diagnosis, and control hasbeen identified as a major 21st century challenge in control systems research and application.This is particularly the case for chemical and metallurgical operations owing tothe lack of expressive fundamental models as well as the nonlinear nature of most processsystems, which makes established linearization methods unsuitable. As a result, effortshave been directed in the search of alternative approaches that do not require fundamentalor analytical models. Data-based methods provide a very promising alternative in thisregard, given the huge volumes of data being collected in modern process operations aswell as advances in both theoretical and practical aspects of extracting information fromobservations.In this thesis, the use of kernel-based learning methods in fault detection and diagnosisof complex processes is considered. Kernel-based machine learning methods are a robustfamily of algorithms founded on insights from statistical learning theory. Instead of estimatinga decision function on the basis of minimizing the training error as other learningalgorithms, kernel methods use a criterion called large margin maximization to estimatea linear learning rule on data embedded in a suitable feature space. The embedding isimplicitly defined by the choice of a kernel function and corresponds to inducing a nonlinearlearning rule in the original measurement space. Large margin maximization corresponds todeveloping an algorithm with theoretical guarantees on how well it will perform on unseendata.In the first contribution, the characterization of time series data from process plants isinvestigated. Whereas complex processes are difficult to model from first principles, theycan be identified using historic process time series data and a suitable model structure.However, prior to fitting such a model, it is important to establish whether the time seriesdata justify the selected model structure. Singular spectrum analysis (SSA) has been usedfor time series identification. A nonlinear extension of SSA is proposed for classification oftime series. Using benchmark systems, the proposed extension is shown to perform betterthan linear SSA. Moreover, the method is shown to be useful forfiltering noise in time seriesdata and, therefore, has potential applications in other tasks such as data rectification andgross error detection.Multivariate statistical process monitoring methods are well-established techniques for efficient information extraction from multivariate data. Such information is usually compactand amenable to graphical representation in two or three dimensional plots. For processmonitoring purposes control limits are also plotted on these charts. These control limits are usually based on a hypothesized analytical distribution, typically the Gaussian normaldistribution. A robust approach for estimating con dence bounds using the reference datais proposed. The method is based on one-class classification methods. The usefulnessof using data to define a confidence bound in reducing fault detection errors is illustratedusing plant data.The use of both linear and nonlinear supervised feature extraction is also investigated.The advantages of supervised feature extraction using kernel methods are highlighted viaillustrative case studies. A general strategy for fault detection and diagnosis is proposedthat integrates feature extraction methods, fault identification, and different methods toestimate confidence bounds. For kernel-based approaches, the general framework allowsfor interpretation of the results in the input space instead of the feature space.An important step in process monitoring is identifying a variable responsible for a fault.Although all faults that can occur at any plant cannot be known beforehand, it is possible touse knowledge of previous faults or simulations to anticipate their recurrence. A frameworkfor fault diagnosis using one-class support vector machine (SVM) classification is proposed.Compared to other previously studied techniques, the one-class SVM approach is shown tohave generally better robustness and performance characteristics.Most methods for process monitoring make little use of data collected under normal operatingconditions, whereas most quality issues in process plants are known to occur whenthe process isin-control . In the final contribution, a methodology for continuous optimizationof process performance is proposed that combines support vector learning withdecision trees. The methodology is based on continuous search for quality improvementsby challenging the normal operating condition regions established via statistical control.Simulated and plant data are used to illustrate the approach.

[发布机构] Stellenbosch University