Causal discovery from process time-series data is challenging because industrial systems often combine fast contemporaneous interactions, time delays, feedback, self-dependence, and latent disturbances. Existing contemporaneous, temporal, and hybrid causal discovery methods can capture parts of these features, but their raw outputs often contain false edges and inferred causal directions. This work introduces a lagged directed partial ancestral graph (LDPAG) and proposes the Pruning–Sanitation–Recovery (PSR) algorithm, a modular framework that refines candidate mixed-delay causal graphs. PSR receives an initial graph from existing causal discovery methods and applies three refinement stages: (1) pruning removes spurious connections, (2) sanitation weakens unsupported endpoints, and (3) recovery reintroduces edge orientations. The framework is evaluated on a synthetic mixed-delay benchmark and a chemical process case study. In both settings, PSR improves skeleton recovery, increases the fraction of correct definite endpoints and source marks, and improves the recovery of latent confounders. These results support PSR as a refinement framework for producing more accurate causal graph representations of processes with multiple time delays, latent confounders, and feedback.
Data-driven industrial process monitoring heavily relies on the independent and identically distributed (IID) assumption, which is frequently violated in real-world applications by equipment aging and fluctuating conditions. This causes severe performance degradation in out-of-distribution (OOD) scenarios. To address this critical bottleneck, this paper comprehensively reviews OOD process monitoring, categorizing current methodologies into two dominant paradigms: adaptive mechanisms and robust modeling. We contrast adaptive strategies, which passively update models to accommodate data shifts, with robust approaches that proactively extract cross-domain invariant representations grounded in generalization theories. Finally, we summarize OOD benchmarks and highlight future directions.
Resilience of wastewater treatment plants (WWTPs) is critical for mitigating human impacts on aquatic ecosystems and enhancing environmental sustainability. However, WWTPs are typically designed on the basis of economic considerations, often neglecting system response to failures and disturbances. This work presents a systematic methodology for retrofitting WWTPs that explicitly accounts for both uncertainty in wastewater inflow and expected and unexpected events that may result in equipment failures. The approach integrates process synthesis via the P-graph framework for the generation of plausible retrofitting designs with a systematic method for resilience evaluation based on enumeration of possible failures. This method enables the identification of cost-effective retrofit configurations with built-in redundancy to enhance resilience. The framework is illustrated via a case study that considers extreme weather, such as uneven rainfall, as an event that increases the flow of water. The method results in the evaluation of 1,531 retrofit alternatives and identifies a diverse set of non-dominated solutions offering different compromises between cost and resilience. The identified retrofit alternatives exhibit trade-offs ranging from a resilience six times that of the original design at a cost 47 higher, to a 10% increase in resilience at an additional 3% cost. These results highlight that economically optimal designs may lack sufficient redundancy, while moderate increases in cost can significantly improve resilience. The proposed framework serves as a decision-support tool for systematic retrofitting, enhancing the information available to stakeholders and supporting more sustainable WWTPs by balancing economic performance, tolerance to failures, and practical implementation considerations.
To achieve the efficient and stable operation of a blast furnace (BF) in the steel industry while retaining the critical skills of expert human operators, we developed a novel human-machine interface (HMI) grounded in ecological interface design (EID) principles. The BF is a complex, large-scale chemical reactor to produce hot metal where internal temperatures can reach up to 2,000 °C. Due to the inherent difficulty of full process automation, the EID framework was adopted to support operators’ situation awareness and keep human expertise central to decision-making. The developed ecological interface (EI) enhances manual control using a transient model, allowing operators to explore the best control action by showing the predicted future states of controlled variables resulting from their hypothetical control actions. Furthermore, the developed EI provides operators with explicit information on the mass and energy balances linking control means and ends, which is intended to support situation awareness. To facilitate comprehensive training, an operation simulator, capable of running at 300 times real-time speed, was integrated with the developed EI. Human-in-the-loop experiments with 16 subjects showed that the EI significantly reduced the number of control actions in both scenarios and reduced the HMT control error and the duration of ΔP upper-bound violations in the complex scenario, whereas no statistically significant reductions were observed in coke rate or carbon intensity.
The accurate determination of pressure loss across filter media is of central importance in the design of filtration processes, as it directly affects energy efficiency and thus operating costs. While Darcy-based approaches reliably describe the flow of Newtonian fluids, no transferable models exist for non-Newtonian fluids flowing through multilayer sintered metal wire composite-cloths, despite their widespread use in industrial applications such as polymer melt filtration or heavy fuel oil filtration in marine engines. To address this gap, this study develops and systematically compares several modeling approaches for predicting the flow velocity of shear-thinning fluids through such composite-cloths, relying on quantities available in industrial practice, such as standardized air permeability measurements. The investigated approaches comprise a physics-based white box model derived from a modified Darcy law, a purely data-driven black box model based on gradient boosting, including a target-scaled variant, and hybrid models combining both paradigms in serial and parallel configuration. All models are evaluated across ten random data splits and compared using paired statistical tests. The hybrid structures outperform the purely data-driven model, with the parallel configuration achieving the lowest error at a mean MAPE of 5.59 ± 1.85 %. The reported performance reflects interpolation within the investigated parameter space of three cloth types, six pore sizes, and three silicone oils. Overall, the proposed hybrid modeling framework provides a viable method for predicting the flow velocity of non-Newtonian fluids in multilayer sintered metal wire composite-cloths.
