Development of Data-Driven Methods for Dynamic Risk Management in the Chemical Industry

Abstract

Large amounts of hazardous substances are handled and stored in chemical facilities, elevating the risk of accidental releases with potentially disastrous consequences. Over the past three decades, there has been a significant evolution in the domain of safety science, leading to the standardization and widespread implementation of Risk Managment (RM) frameworks designed to identify, quantify, evaluate, control, and manage the risk associated with industrial activities involving hazardous substances. However, canonical RM techniques suffer several limitations, such as their inherent staticity and inability to update the risk picture in evolving and degrading systems. To overcome these limitations, recent research has proposed to move toward a more dynamic and proactive approach to process safety, named Dynamic Risk Management (DRM), which aims at capturing risk variations in industrial facilities, taking into account the performance of the control system, safety barriers, inspection and maintenance activities, and the human factor. This paradigm shift raises the need for dynamic and inherently updatable tools to capture the intricate dynamics between risk-influencing factors. In this context, Machine Learning (ML) techniques emerge as valuable tools due to their inherent ability to make predictions under uncertainty and model complex nonlinear relationships between features. However, the potential of these techniques in the context of DRM is still scarcely explored. Therefore, this Ph.D. study seeks to contribute to the development of ML methods to enhance and support DRM. Specifically, this investigation formulates and presents practical ML-based methods to address critical tasks in DRM, namely consequence prediction, frequency evaluation, and monitoring of safety barriers. In addition, this study delves deep into the broader implications of adopting ML technologies, such as the intricate relationship between human expertise and AI, critically examining their respective contributions in the future of Risk Management. Different ML algorithms have been explored, including classification, clustering, regression, and Natural Language Processing. Diverse data sources have been utilized, such as alarm data, process data, and accident data. The contributions of this research include: - the assessment of recent advancements of ML in the domain of safety and reliability; - the development of classification models to predict the consequences of major accidents; - the investigation of ML models to aid Risk Based Inspection of hydrogen systems; - the use of regression models to predict the Time-To-Failure of tanks exposed to external fire; - the exploration of classification models and natural language processing algorithms to monitor and improve the performance of industrial alarm systems; - the integration between traditional risk assessment tools, data-driven models, and resilience analysis to evaluate safety barriers in environmental-critical facilities; - the analysis of the involvement between ML and human actors in Risk Management. Proposed methods have been tested on real-world case studies to demonstrate their efficacy. The results indicate that ML methods can be used to take advantage of the wealth of heterogeneous data made available by the widespread digitalization of industrial sectors in order to extract safety-relevant knowledge and provide critical support to DRM. Limitations and challenges have been acknowledged and discussed, including the challenges linked to imbalanced and cost-sensitive classification, the importance of data quality and sound preprocessing procedures, model interpretability, quantification of prediction uncertainty, and the challenges related to model selection and hyperparameters tuning. While the trajectory of progress suggests an increasing adoption of AI tools, domain knowledge and human expertise remain pivotal, ensuring effective oversight of intelligent systems, understanding the limitations of ML models, and contextualizing their predictions