Strength of knowledge (SoK) judgments are a key component in characterising risks from an uncertainty-based risk perspective. A commonly applied assessment scheme for this purpose involves evaluating the SoK based on several criteria: the reasonableness of assumptions, the availability and reliability of data, the degree to which phenomena are understood, the level of expert agreement, and the extent to which the knowledge has been scrutinised. The applicability of this scheme has been widely discussed and debated in the risk research literature. However, less attention has been given to its actual use and to risk practitioners' evaluations. In Norway, many practitioners regularly apply the SoK scheme to address real-world risk problems, as various public and private sectors have incorporated it into their risk assessment standards and guidance documents. In this paper, we examine how 22 risk practitioners use the SoK scheme and assess its utility. The study aims to provide insights into the challenges and benefits of applying the SoK scheme in practical contexts, thereby enhancing its applicability for practitioners and informing its ongoing development within the research community. Key findings relate to the scheme’s practical value in structuring uncertainty assessments and informing risk management, alongside identified needs for improved guidance to support consistent judgment and effective communication.
Tool wear assessment is essential for machining quality, process stability, and maintenance decision making. Existing methods have improved wear classification and wear-value prediction, but many rely on single-modality inputs, shallow fusion, or black-box task-specific models, limiting their ability to explain diagnostic evidence and support practical inspection. To address these limitations, we propose an evidence-guided and stage-aware multimodal large language model framework based on Qwen3-VL-8B. The framework integrates tool images, sensor-derived representations, process context, and structured base-model evidence to jointly diagnose wear type, estimate wear value, and generate engineer-readable reports. These reports provide confidence estimates, evidence interpretation, risk assessment, inspection focus, and maintenance suggestions, thereby transforming multimodal wear prediction into explainable engineering decision support. Experiments on the Multimodal Automatic Tool Wear Inspection dataset show that the proposed method achieves 93.12% accuracy, 77.09% macro-F1 score, 16.64μm mean absolute error, 25.52μm root mean square error, and a coefficient of determination of 0.689. Compared with single-modality, simple fusion, conventional multimodal fusion, and directly prompted MLLM baselines, the proposed method improves wear-type discrimination, maintains competitive regression performance, and generates structured, evidence-grounded reports for practical decision support.
When affected by abrupt severe disruptions such as natural disasters and third-party construction damage, natural gas pipeline networks may suffer serious supply loss and require costly recovery. Existing studies insufficiently integrate supply resilience assessment with economic constraints, making it difficult to identify cost-effective resilience enhancement strategies. This study proposes a cost-resilience trade-off framework for natural gas pipeline networks under stochastic severe disruptions. Demand-fulfillment variation is used to quantify supply resilience, while component disruptions, recovery rules, degraded operation, transmission capacity constraints, and gas supply allocation are integrated into Monte Carlo simulation. Reinforcement costs, recovery costs, and economic losses are further incorporated into a single-objective optimization model that minimizes total cost under a minimum resilience requirement. The case study identifies case-specific turning values within marginal-benefit saturation ranges, namely a reinforcement level of approximately 80% and a recovery rate of approximately 2.664 day-1, beyond which additional reinforcement or recovery acceleration yields diminishing marginal benefits. Scenario-based optimization analyses show that optimal strategies vary with risk, demand, defect, and cost conditions, but the framework consistently reallocates resources between preventive reinforcement and post-disruption recovery to satisfy resilience requirements at minimum cost. The framework supports adaptive resilience investment, critical component protection, and recovery-resource configuration.
The stability of real-world networks—from biological ecosystems to technological infrastructures—depends strongly on the vulnerability of critical components. Identifying such vulnerable nodes remains challenging because simple structural heuristics often neglect complex system dynamics, whereas simulation-based vulnerability evaluations are computationally expensive at scale. We propose a simulation-driven temporal hypergraph learning framework that serves as a neural surrogate for expensive vulnerability simulations. The framework represents evolving networks as temporal hypergraphs to capture higher-order group interactions and temporal evolution, and uses perturbation-aware self-supervised learning to obtain robust node representations under limited labels. A heterogeneous knowledge distillation strategy is then used to transfer the ranking information produced by offline simulation oracles into a lightweight inference model. Experiments on real-world networks show that the proposed method achieves strong performance among learning-based baselines, while substantially reducing online inference cost. The results suggest that simulation-distilled neural surrogates provide a practical route toward scalable vulnerability assessment in large complex networks.
To ensure the safety and durability of a ship operating in ocean waves, it is essential to characterize not only the short-term but also the long-term statistical properties of wave-induced responses over the ship’s life. Long-term risk is typically evaluated through long-term prediction (LTP), which aggregates numerous short-term predictions (STPs) of extreme responses across sea states and operational conditions.This paper proposes an efficient design optimization methodology that explicitly incorporates long-term safety. The mean outcrossing rate (MOR) relative to a design threshold is adopted as the target reliability metric. To accelerate the LTP process, a Bayesian Active Learning (BAL) strategy is introduced, where a Gaussian Process Regression model of the MOR is adaptively updated. By treating the MOR as a log-normal process, the non-negativity of the MOR is ensured. Furthermore, the design optimization is executed within a single-loop framework, bypassing the conventional double-loop structure by simultaneously processing Bayesian Optimization and the BAL for LTP. The proposed framework is demonstrated through numerical examples involving the optimization of ship parameters under different nonlinear responses. Finally, a First Order Reliability Method (FORM)-based STP procedure and a multi-fidelity strategy combining low- and high-fidelity response models are introduced, thereby enabling high-fidelity nonlinear simulations while maintaining computational tractability.
Environmental contours (ECs) are widely used to represent multivariate extreme marine conditions for offshore structural design. However, their reliabilities strongly depend on the accuracy of joint probability distribution (JPD) modeling and the robustness of tail extrapolation. In this study, a novel semi-parametric JPD modeling framework is developed to accurately capture the dependence structures of wind–wave–current variables. This framework integrates the adaptive-bandwidth kernel density estimation (KDE)–Pareto marginal model with the nonparametric R-vine copula, and is termed the KPRC model. Furthermore, an expected quantile discrepancy (EQD) method is introduced to determine optimal thresholds for generalized Pareto distribution (GPD)-based peak-over-threshold (POT) analysis. Using measured wind–wave–current data from three National Oceanic and Atmospheric Administration (NOAA) buoy stations, the performance of the proposed KPRC model and EQD method is validated. Finally, ECs derived from the inverse first-order reliability method (IFORM), direct sampling method (DSM), and improved Direct-IFORM are comprehensively compared. Results show that the KPRC model achieves consistently high modeling accuracy across different stations and variable combinations. The EQD method improves the stability of GPD parameter estimation and enhances the cross-method consistency of resulting ECs. IFORM and improved Direct-IFORM yield similar outcomes, while DSM generally produces the most conservative contours. It is thus recommended for safety-oriented design; however, its strong sensitivity to POT threshold selection should be carefully considered to avoid excessive conservatism.
