Author

Savo Kordic

Date of Award

2011

Document Type

Thesis - ECU Access Only

Publisher

Edith Cowan University

Degree Name

Doctor of Philosophy (Computer Science)

School

School of Computer and Security Science

Faculty

Faculty of Computing, Health and Science

First Supervisor

Associate Professor Chiou-Peng Lam

Second Supervisor

Professor Huaizhong Li

Third Supervisor

Dr Jitian Xiao

Abstract

The safety and profitability of chemical plants depend on the performance of alarm systems. However, a variety of contributing factors such as over-alarming, a lack of configuration management practices and a reduction in staffing levels due to budget constraints, has often led to the degradation of these systems. Consequently, in many emergency situations excessive numbers of inappropriate alarms are generated, making the alarm system very difficult to use in the situations where it is most urgently needed. Plant personnel carry out periodic alarm rationalisation exercises to address this situation. However, these are manual processes which are very labour intensive and costly. Nowadays, large industrial processes such as chemical plants and petroleum refineries have databases with the ability to store terabytes of data. While it is possible to manually extract the information required for alarm rationalization, the extensive quantity and complexity of data has made the analysis and decomposition a very laborious task. This research presents a novel approach to support alarm analysis by extracting relationships of alarm points from alarm databases in a cost-effective way. To reduce the search space, this research proposes a method of context-based data extraction associated with co-occurrences of alarm tags, which takes advantage of local event-based segmentation. It also introduces filtering strategies that incorporates domain specific concepts to remove spurious data points, before the standard data mining algorithm is applied to discover frequent episodes from alarm sequences. Unlike previous work where mining thresholds were arbitrarily chosen first, prior to data mining, this project also proposed a data-driven approach for deriving “adaptive” thresholds that are more relevant to the context of the analysis and which are then used for guiding the data mining process. Since frequency distribution of typical chemical plant alarm data is skewed (J-shaped) owing to the presence of nuisance alarms, a relatively high support value used in data mining may result in a situation where interesting patterns with a low number of occurrences may be missed. On the other hand, if the support threshold is too low then the collection of all generated patterns and rules could be too large for the user to comprehend.Evaluation of the proposed approach involved employing the developed techniques/tools for diagnosing alarm problems in historical alarm data obtained from a Vinyl Acetate process model simulation, and also in data from a real plant. The first challenge was to obtain accurate information about relationships between 27 alarm tags in the simulated datasets, and then the final stage of evaluation used real plant data with more than one hundred individual alarm tags. The proposed approach was designed as an interactive exploration tool for the purpose of analysing data to find groupings of co-occurring alarm tags and improving presentation of alarm information. Experimental results showed that the proposed approach found patterns from the simulated data sets which can be validated against the Vinyl Acetate Model. Analysis of the real plant data also extracted some interesting patterns with low frequency of occurrences. Since the developed approach can carry out the main bulk of the tedious tasks of analysis, the proposed approach is very cost effective as the cost of computer time is very cheap compared to that of a process engineer.

Share

 
COinS