Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.799523
Title: Probabilistic modelling of temporal correlations in industrial alarm data
Author: Irimias, Robert
ISNI:       0000 0004 8505 2036
Awarding Body: University of Manchester
Current Institution: University of Manchester
Date of Award: 2020
Availability of Full Text:
Access from EThOS:
Access from Institution:
Abstract:
It is crucial that modern industrial plants maintain a safe environment in which processes can run efficiently. To achieve this, processes are being monitored via sensors and associated alarms. Data obtained through the monitoring system is displayed to human operators who must predict, identify and respond accordingly to faults of any nature and magnitude. However, operators can be overwhelmed leading to dangerous situations. It is thus desirable to create numerical analysis tools to facilitate their work. In the first part of this thesis, we explore the breadth of methods available in current literature in an attempt to give a unified overview of industrial data analysis. Next, given their direct relevance to the operators, the focus is on modelling discrete alarm data. A Bayesian network model is learnt from real historical alarm data and its limitations regarding the lack of representation of time dependencies are proven empirically. To overcome said limitations, a novel dynamic Bayesian Alarm network for representing alarm data is developed. It has a parsimonious, causally independent CPD and models time dependencies using geometric distributions. A new way of interpreting and analysing same-time/instant causation alarms is introduced in the model. Its applications in key alarm detection, clustering, flood identifications and are demonstrated on both real and synthetic data. The synthetic alarm data is obtained via a new heuristic random walk based method. Additional synthetic data is generated using the dynamic Bayesian alarm network and statistically compared to real data.
Supervisor: Shapiro, Jonathan Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.799523  DOI: Not available
Keywords: similarity measure ; prediction ; historical alarm data ; probabilistic model learning ; early detection ; Dynamic Bayesian Network ; alarm management ; Bayesian Network ; artificial data generation
Share: