“Static, incomplete, superficial, wrong”. The traditional approach to risk analysis, as applied in the process industries, has been largely criticized in response to recent major accidents. Since it was first proposed, modifications and improvements have been made, and a formal accepted approach is included in several regulations and standards (as the recent development of guidelines for the ageing management in SEVESO installations). Quantitative Risk Assessment (QRA) is based on consolidated procedures. Nevertheless, the need of safety improvement asks for more advanced tools for hazard identification and risk evaluation. Besides considering technical aspects (e.g., malfunctions and process upsets), operational errors, organizational aspects, such as lack of attention and motivation to the safety culture, may lead to risk increment in terms of likelihood of undesired failures. Not all those aspects may be investigated with conventional QRA techniques, which have also the disadvantage of being intrinsically static and failing to capture risk variations during the lifecycle of a plant or production site. Despite their proved effectiveness, many hazards identification and risk assessment techniques lack the dynamic dimension, which is the ability to learn from new risk notions, experience, and early warnings. Now’s the time to go beyond the limits of conventional static methods for hazard identification and risk assessment; the risk assessment is, indeed, a very useful approach in support of this change but at the same time it is not exhaustive to capture also the possible “failure” in the interface/interaction among the several single components of a complex system beside their specific failures. This research work discusses a novel approach for dynamizing the risk assessment process, integrating measured process data, asset integrity and operative conditions. In the first part of the thesis, the inferential process and the application of Machine Learning to inference is discussed, and various applications of standard, and tailored, machine learning algorithms to industrial and environmental risks are detailed as case studies. The second part is focused on the resilience engineering. The resilience paradigm is discussed, as well as the concept of emerging properties of complex systems. it will be shown how real-time data analytics, through appropriate AI models, combined with the expert knowledge of process engineering, constitute the fundamental technological key to pursue the resilience of plants and processes. The third section integrates the aforementioned concepts within the wide framework of Systems Engineering. Accordingly, a dynamic and systemic model is presented, to address the significant shortcomings of the current risk analysis models. The Dynamic Asset-integrity and Risk Management System (DARMS) is designed starting from the Bow-tie technique, integrated with improved Machine Learning algorithms, to overcome the epistemic uncertainty in the prior probabilities and likelihoods of escalation factors and barriers. Subsequently, a Hidden Markov Model (HMM), based on Bayesian Inference, is developed to analyze real-time risk, and produce reliable predictions on the state of the whole system during the operations. The application of the proposed model is demonstrated on an Oil and Gas terminal under Seveso legislation. The results of the case study provide a better understanding of the advanced Data Driven modeling of accident scenarios. The proposed model will serve as a useful tool for the operational safety management of complex systems.

DARMS - Dynamic Asset-integrity and Risk Management System - How Machine Learning and Systems Engineering cooperate to enhance the resilience of complex systems

VAIRO, TOMASO
2022-05-16

Abstract

“Static, incomplete, superficial, wrong”. The traditional approach to risk analysis, as applied in the process industries, has been largely criticized in response to recent major accidents. Since it was first proposed, modifications and improvements have been made, and a formal accepted approach is included in several regulations and standards (as the recent development of guidelines for the ageing management in SEVESO installations). Quantitative Risk Assessment (QRA) is based on consolidated procedures. Nevertheless, the need of safety improvement asks for more advanced tools for hazard identification and risk evaluation. Besides considering technical aspects (e.g., malfunctions and process upsets), operational errors, organizational aspects, such as lack of attention and motivation to the safety culture, may lead to risk increment in terms of likelihood of undesired failures. Not all those aspects may be investigated with conventional QRA techniques, which have also the disadvantage of being intrinsically static and failing to capture risk variations during the lifecycle of a plant or production site. Despite their proved effectiveness, many hazards identification and risk assessment techniques lack the dynamic dimension, which is the ability to learn from new risk notions, experience, and early warnings. Now’s the time to go beyond the limits of conventional static methods for hazard identification and risk assessment; the risk assessment is, indeed, a very useful approach in support of this change but at the same time it is not exhaustive to capture also the possible “failure” in the interface/interaction among the several single components of a complex system beside their specific failures. This research work discusses a novel approach for dynamizing the risk assessment process, integrating measured process data, asset integrity and operative conditions. In the first part of the thesis, the inferential process and the application of Machine Learning to inference is discussed, and various applications of standard, and tailored, machine learning algorithms to industrial and environmental risks are detailed as case studies. The second part is focused on the resilience engineering. The resilience paradigm is discussed, as well as the concept of emerging properties of complex systems. it will be shown how real-time data analytics, through appropriate AI models, combined with the expert knowledge of process engineering, constitute the fundamental technological key to pursue the resilience of plants and processes. The third section integrates the aforementioned concepts within the wide framework of Systems Engineering. Accordingly, a dynamic and systemic model is presented, to address the significant shortcomings of the current risk analysis models. The Dynamic Asset-integrity and Risk Management System (DARMS) is designed starting from the Bow-tie technique, integrated with improved Machine Learning algorithms, to overcome the epistemic uncertainty in the prior probabilities and likelihoods of escalation factors and barriers. Subsequently, a Hidden Markov Model (HMM), based on Bayesian Inference, is developed to analyze real-time risk, and produce reliable predictions on the state of the whole system during the operations. The application of the proposed model is demonstrated on an Oil and Gas terminal under Seveso legislation. The results of the case study provide a better understanding of the advanced Data Driven modeling of accident scenarios. The proposed model will serve as a useful tool for the operational safety management of complex systems.
resilience engineering; risk assessment; complex systems; system safety; inference; machine learning; systems engineering
File in questo prodotto:
File Dimensione Formato  
phdunige_1779538.pdf

accesso aperto

Descrizione: PhD thesis
Tipologia: Tesi di dottorato
Dimensione 5.12 MB
Formato Adobe PDF
5.12 MB Adobe PDF Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: http://hdl.handle.net/11567/1080188
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact