|Adm. – Grad.
|2011 – 2017
Causal Structure of Risk in Industrial Systems Using Dynamic Bayesian Networks
The thesis deals with detecting the causal structure of risk in industrial systems. We focus on the prioritization of risks in the form of correlated events sequences. To improve the established prioritization methods, the application of a new methodology using Dynamic Bayesian networks (DBN) is proposed. We are studying the development of a new user interface for industrial control systems and data acquisition, known as Supervisory Control and Data Acquisition (SCADA), to demonstrate the analysis method of risk causal structure. We perform a test based on a dataset of an actual SCADA system, obtained by UK authors from a semiconductor manufacturing plant. Our analysis use the R statistical software as a development platform, with classification algorithms implemented in the tool Tanagra. Our results show that: (1) the network of variables before and after the failure is represented by a limited and distinct number of factors; (2) the network of variables before and after the failure can be graphically represented dynamically in a user interface to assist in fault prevention and diagnosis; (3) variables related to the sequence of events at the time of failure can be used as a model to predict its occurrence (whose forecast quality is evaluated by the F1 measure), and find the main cause of it, thus making it possible to prioritize the requirements of the production system on the right variables to be monitored and manage in the event of a breakdown. The reliability of our fault forecasts is evaluated using the Train-Test, Cross-Validation and Bootstrap methods. These results have a significant value for industrial engineers, working as a team through a SCADA during the execution of the production system. Using this new, more intuitive interface, they will be able to more easily detect the probable cause of a system failure, and can intervene on the right factors with a higher confidence level.