Search

Systems for monitoring processes with multiple operating modes should be able to distinguish between changes in operating modes and developing faults. Process operators will make decisions regarding production and maintenance according to the information about faults, such as the occurrence of faults and the locations of faults, so that the process can run safely and efficiently. Whilst the development and application of kernel methods can improve the performance of monitoring systems, inappropriate usage of these methods can diminish the effectiveness of the methods. In this thesis the industrial considerations of operators are summarized and these considerations are incorporated into the development of kernel methods for process monitoring. The research in the thesis shows that kernel methods need to be designed and implemented properly for monitoring processes with multiple operating modes. The research in the thesis also aims to develop kernel methods that can generate useful results when applied to monitoring of processes with multiple operating modes. The thesis reports the following research outcomes: • A benchmark multimodal dataset from a pilot-scale experiment rig; • An investigation of the tuning of kernel methods and a tuning strategy for the radial basis function kernel; • A new kernel that can improve the monitoring performance when applied to multimodal data; • An on-line monitoring framework which can account for new operating modes in the process; • A way to define the contributions of process variables to a fault detection, in order to support fault diagnosis. The thesis delivers novel kernel methods for monitoring processes with multiple operating modes and gives guidelines for proper implementation of these methods. These outcomes extend the field of process monitoring. The research outcomes are relevant for industrial application because the practical considerations of end-users are incorporated in the development of the kernel methods. The thesis also contributes to the theory of process monitoring by proposing novel kernel methods for fault detection and diagnosis. The results in the thesis demonstrate that the new development of kernel methods can improve monitoring performance when applied to processes with multiple operating modes. Moreover, the monitoring results achieved by the new kernel methods can be interpreted and used by process operators. Open Access

Previously in her secondment at the University of Alberta, Nina created the continuous stirring tank heater case study. This case study can be used for testing techniques in various aspects of process systems engineering, such as system identification, fault detection and diagnosis, and controller design. With my research being about data-driven fault detection algorithms for process monitoring, this case study is of great interest to me. Although many previous works have already used this case study for testing the fault detection and diagnosis algorithms, there is potential for more practical faulty scenarios. Inspired by the discussions with Nina, I try to create two fault scenarios. The rst one develops gradually over time and the second one is to simulate a fault that may be physically induced in the experiment. The data from both healthy and faulty scenarios are collected Canonical Variate Analysis, which is a multivariate approach for modelling linear and dynamic processes, is applied to the data for fault detection. This test can be the starting point towards fault detection and prognosis for nonlinear and dynamic processes with multiple operating modes and transition periods.

Data-driven process monitoring has benefited from the development and application of kernel transformations, especially when various types of nonlinearity exist in the data. However, when dealing with the multimodality behavior that is frequently observed in the process operations, the most widely used radial basis function (RBF) kernel has limitations in describing process data collected from multiple normal operating modes. In this article, we highlight this limitation via a synthesized example. In order to account for the multimodality behavior and improve the fault detection performance accordingly, we propose a novel nonstationary discrete convolution kernel, which derives from the convolution kernel structure, as an alternative to the RBF kernel. By assuming the training samples to be the support of the discrete convolution, this new kernel can properly address these training samples from different operating modes with diverse properties and, therefore, can improve the data description and fault detection performance. Its performance is compared with RBF kernels under a standard kernel principal component analysis framework and with other methods proposed for multimode process monitoring via numerical examples. Moreover, a benchmark data set collected from a pilot-scale multiphase flow facility is used to demonstrate the advantages of the new kernel when applied to an experimental data set.

As data analytic techniques evolve and the accessibility of process measurements improves, data-driven process monitoring has enjoyed a quick development in both theoretical and application perspectives recently. Although abundant process measurements will facilitate data-driven process monitoring and lead to better monitoring indices, it becomes difficult to identify the underlying variables that are responsible for a fault directly with the monitoring indices as the scope of measured variables is getting broader. To restrain the scope and identify the source of fault, contribution plots are commonly used in fault diagnosis in order to quantify the influence of process variables in presence of fault. Nevertheless, as sophisticated monitoring techniques become more and more complicated, deriving corresponding contribution plots is challenging. The concept of deviation contribution plots is proposed to address this issue. By extending the original definition of contribution for linear processes, the deviation contribution is defined to quantify the contribution of deviations in originally measured variables to the deviation of monitoring indices. The ability of proposed deviation contribution plots to identify influential variables in monitoring algorithms based on nonlinear feature extractions is verified by both numerical simulation and the Tennessee Eastman Process benchmark case study.