Detection and Evaluation of Abnormal Events in Complex Industrial Processes

Detection and evaluation of abnormal events in industrial process systems is vital for safe and undisturbed operation. Failure to treat a process operating in an abnormal state can lead to a loss of system functions that are necessary for recovering the process to a nominal state. Advances in process monitoring technologies pose the issue of information overload, namely, the difficulty for an operator to monitor and understand the process information that is provided in real-time. In extreme cases, the operator accidentally disregards critical information, which leads to incorrect diagnosis. Disasters such as the Deepwater Horizon explosion and the crashing of Lion Air Flight 610 and Ethiopian Airlines Flight 320 share a common post-accident diagnostic report: though the primary cause of these incidents was not attributed to operator error, their outcomes would have been significantly mitigated if operators had been provided with relevant and accurate diagnostic information. To address the issue of information overload, this thesis proposes a statistical method for detecting and evaluating abnormal changes in the signal characteristics of process variables. The method is independent of process knowledge; it only requires a collection of samples for process variables gathered from past operations rather than a physical understanding of the system. The motive for this approach is based on the complexity of modern industrial process systems, as a detailed physical description of the influence of process inputs on process outputs for a
system comprising thousands of process variables may not be available. The thesis argues that the performance of the proposed method is associated with its ability to extract features from samples gathered from while the process was consistent with
nominal operating conditions. Existing methods provide satisfactorily performance.
of detecting abnormal events. The thesis argues that their performance of evaluating
abnormal events could be improved.
The scientific contributions of the research cover two topics, namely, the detection
of abnormal events in nonlinear, dynamic systems and the evaluation of abnormal
changes in process variables. A method is proposed for extracting features from
samples for process variables with an artificial neural network - a mathematical
model that describes a nonlinear function. Abnormal event detection is facilitated by comparing the features of new observations against those of samples gathered from
while the process was consistent with nominal operations. It is concluded that an
abnormal event has occurred if the disparity of this comparison exceeded a certain
threshold. Abnormal changes in process variables are evaluated by combining a
structural analysis of the artificial neural network with a contribution analysis of
process variables on the detected abnormal event. The novelty of the proposed
method is that it does not require prior instances of abnormal events to evaluate
abnormal changes in process variables