Advances in Automation and Artificial Intelligence continue to enhance the autonomy of process plants in handling various operational scenarios. However, certain tasks, such as fault handling, remain challenging, as they rely heavily on human expertise. This highlights the need for systematic, knowledge-based methods. To address this gap, we propose a methodological framework that integrates Large Language Model (LLM) agents with a Digital Twin environment. The LLM agents continuously interpret system states and initiate control actions, including responses to unexpected faults, with the goal of returning the system to normal operation. In this context, the Digital Twin acts both as a structured repository of plant-specific engineering knowledge for agent prompting and as a simulation platform for the systematic validation and verification of the generated corrective control actions. The evaluation using a mixing module of a process plant demonstrates that the proposed framework is capable not only of autonomously controlling the mixing module, but also of generating effective corrective actions to mitigate a pipe clogging with only a few reprompts.

Process plants are becoming increasingly autonomous to compensate for the lack of experienced plant operators. However, certain activities still need to be carried out by humans. To move from a low to a high level of autonomy, an evaluation of operator and plant behavior is necessary. This provides insights into suitable automation measures. Our contribution presents two methods for the creation, integration and evaluation of operator behavior in plant simulations. This will enable the identification and learning of suitable control actions for different levels of plant autonomy at an early stage in plant engineering.

Real-time optimization plays a key role in improving energy efficiency and the operational effectiveness of industrial systems. To deal with unknown process characteristics and safety constraints, a novel safe adaptive real-time exploration and optimization (ARTEO) algorithm is proposed recently for safety-critical industrial systems. ARTEO utilizes the Gaussian process (GP) regression to model unknown plant characteristics and enforces safety constraints using confidence intervals provided by the GP models. Due to changing process characteristics, the GP models need to be updated online by incorporating new observations and the computational complexity of model adaptation increases with a growing dataset. This work proposes an alternative ARTEO implementation by using ensemble learning, namely Ensemble-ARTEO. The Ensemble-ARTEO learns unknown plant characteristics through an ensemble of parametric regression models and calculates uncertainty by the variance of ensemble predictions. The predictive uncertainty is integrated into the optimization objective to further drive exploration. The ensemble members are updated efficiently online to capture the changing process characteristics. We demonstrate the effectiveness of our proposed Ensemble-ARTEO approach in an industrial refrigeration process. Experimental results show that our method enables tracking the desired cooling demand while satisfying the safety constraints.

Real-time optimization strategies aim to improve the operating performance of a process using a model of its input-output behaviour. This is challenging when the true system characteristics are not fully known and there are safe operating limits that must be respected. In this work, we evaluate the performance of an adaptive, real-time, exploration and optimization algorithm on a simulated refrigeration plant, and show how incorporating prior knowledge based on engineering principles can improve its performance, especially during the early stages of learning when there are few observed data. The results indicate that exceedances of the safe operating limit are avoided and the improved models still learn the true system characteristic, albeit more slowly than the standard models fitted without prior knowledge.

We present the development of a semi-supervised regression method using variational autoencoders (VAE) for soft sensing of process quality variables. Recently, use of VAEs was proposed for regression applications based on variational inference. In this work, We extend this approach of supervised VAEs for regression to make it learn from both labelled and unlabelled data leading to a semi-supervised VAE for regression (SSVAER) formulation. The probabilistic regressor resulting from the variational approach makes it possible to estimate the variance of the predictions simultaneously, which provides a means for online uncertainty quantification for soft sensors. We provide an extensive comparative study of SSVAER with previously proposed semi-supervised learning methods on two soft sensing benchmark problems using fixed-size datasets, where we vary the percentage of labelled data available for training. In these experiments, SSVAER achieves the lowest test errors in 11 of the 20 studied cases, compared to other methods where the second best method gets 4 lowest test errors out of the 20.