Evolving and proactive risk modelling in underground working environments

Underground environments are critical to global resource extraction and urban development, yet they are inherently hazardous due to their dynamic and unpredictable nature. Rapid changes, including methane accumulation and environmental fluctuations, threaten worker safety, operational continuity, and infrastructure integrity. Traditional risk assessment methods, which rely on static models and expert judgment, struggle to capture real-time variations in gas concentrations, ventilation, and ground stability, leading to delayed hazard detection and inadequate mitigation. Fragmented sensor data integration and the opaque nature of advanced models further hinder trust and decision-making. These limitations contribute to severe consequences, such as methane explosions, ventilation failures, and operational disruptions, highlighting the urgent need for an integrated, real-time risk modelling system capable of dynamically monitoring, predicting, and mitigating hazards in underground working spaces. This thesis aims to address these shortcomings by presenting an environmental Digital Twin (DT) framework for evolving and proactive underground risk modelling that integrates real-time Internet of Things (IoT) sensor networks, advanced probabilistic modelling, and adaptive deep learning (DL) techniques to address the critical challenges of managing hazardous underground environments. By fusing heterogeneous sensor data with robust Bayesian Networks (BNs) and online learning algorithms, the research establishes a system capable of continuous, real-time monitoring and proactive risk prediction, thereby significantly enhancing operational safety and decision-making in environments prone to hazards such as methane explosions and inadequate ventilation. The first phase of this research establishes an IoT-Bayes fusion framework that integrates diverse sensor arrays deployed in complex underground settings. IoT sensors capture critical environmental parameters, such as gas concentrations, temperature, and ventilation conditions, in real time, while BNs model the intricate causal relationships among these variables. Detailed case studies in underground coal mines in Poland and China validate the framework's effectiveness in dynamically updating risk estimates and predicting potential hazardous events. This integration not only supports continuous safety monitoring but also offers an interpretable platform that merges quantitative data with expert knowledge. Building on this foundation, the thesis develops an innovative hybrid system that couples online Long Short-Term Memory (LSTM) networks with BNs to capture the temporal evolution of underground hazards. In this approach, the LSTM component continuously learns from incoming data and captures time-dependent relationships in sensor readings. Its predictive outputs are integrated within the BN to refine risk assessments and quantify uncertainties. A detailed case study from a Chinese coal mine demonstrates how this hybrid model achieves high predictive accuracy for methane concentrations and adjusts forecasts in response to emergent conditions, offering actionable, scenario-specific insights for preventive measures and emergency responses. The research further advances the field by constructing a comprehensive model of the underground environment using data-driven Dynamic Bayesian Networks (DBNs). This model is derived through a rigorous process involving static BN structure learning, parameter estimation, and subsequent dynamic extension. Continuously assimilating real-time data, the DBN functions as a digital replica of the physical system, mirroring current conditions and anticipating future changes. This virtual replica enables proactive decision-making by allowing operators to simulate various risk scenarios, assess the impact of mitigation strategies, and optimise safety protocols based on updated probabilistic predictions. To address data scarcity during the early phases of operations, the thesis incorporates online deep transfer learning and multi-sensor analysis. A baseline LSTM model, initially trained on open-source data from a Polish coal mine, is refined using supplementary data from a Chinese mine via an online transfer learning approach. This strategy significantly improves hazard prediction accuracy in data-sparse environments and ensures the monitoring system remains robust despite rapidly changing conditions. Overall, the fusion of DT technology, IoT integration, adaptive DL, and Bayesian inference sets a new benchmark for proactive risk management in underground settings. Collectively, the thesis makes a substantial theoretical and practical contribution by establishing a dynamic, integrated framework that unifies DT technology, IoT sensor integration, adaptive DL, and Bayesian inference. This comprehensive system overcomes the limitations of conventional risk modelling approaches by providing a real-time, data-driven solution that is both interpretable and scalable. The work sets a new benchmark for proactive underground risk modelling, paving the way for safer and more efficient operations in complex, data-sparse environments, and offering a robust foundation for future research in the integration of advanced digital technologies with risk assessment and management practices.