(372v) Causality-Guided Observer Model Minimization of an Acetaminophen Chemical Production Process

In the pharmaceutical manufacturing industry, the effective placement of plant sensors plays a vital role in ensuring high efficiency and product quality. Acetaminophen, a widely used analgesic and antipyretic drug, requires a carefully monitored production process to meet stringent quality standards and maintain cost-effectiveness. However, the deployment of an extensive network of sensors can lead to increased complexity, maintenance costs, and data redundancy. This study addresses these challenges by applying a novel algorithm designed to optimize sensor deployment in the acetaminophen production process, with the goal of minimizing the number of sensors required while preserving monitorability.

The acetaminophen production process involves a series of complex chemical reactions and unit operations. Each stage of the process requires precise control and monitoring to ensure consistent product quality and yield. Traditional approaches to sensor placement often rely on heuristics or expert knowledge, which can result in suboptimal configurations and excessive sensor usage. To overcome these limitations, we propose a data-driven approach that leverages the inherent causality within the production process to guide sensor minimization.

At the core of our methodology is a liquid time-constant (LTC) modeling framework, which is particularly well-suited for capturing the dynamics of chemical processes. LTC models are inspired by the neural circuitry found in nature and have demonstrated remarkable ability to learn and represent complex nonlinear relationships between variables. By training an LTC model on historical process data, we can uncover the underlying causal structure and quantify the impact of various input variables on critical production outputs.

The first step in our approach involves constructing a comprehensive LTC model of the acetaminophen production process, incorporating all available sensor measurements as inputs. The model is trained using time-series data collected through a series of perturbation experiments in ASPEN Plus Dynamics. Once the LTC model is fully trained, we employ a novel causality-guided model minimization algorithm to identify the most informative sensors. The algorithm systematically quantifies the impact of each input variable on the model's predictive performance by calculating intervention coefficients. These coefficients measure the sensitivity of the model's outputs to perturbations in individual input features, providing a quantitative assessment of their importance.

Based on the intervention coefficients, the algorithm ranks the sensors in order of their significance to the process. Sensors with low intervention coefficients, indicating minimal impact on the model's predictions, are considered redundant and can be potentially eliminated. The algorithm then iteratively removes the least important sensors and retrains the LTC model, evaluating the model's performance at each step. This process continues until a desired balance between sensor minimization and model accuracy is achieved.

To validate the effectiveness of our approach, we test against additional system excitation time-series data from ASPEN Plus Dynamics. The results demonstrate that the optimized sensor network, obtained through the causality-guided model minimization algorithm, maintains and, in some cases, even enhances the accuracy and reliability of process monitoring compared to the original high-order sensor-intensive setup. Consequently, the minimized model achieves comparable or better performance in predicting critical quality attributes and detecting process anomalies, while requiring significantly fewer sensors.

In conclusion, this study presents a novel application of causality-guided model minimization for optimizing sensor deployment in the acetaminophen chemical production process. By leveraging the power of liquid time-constant modeling and intervention coefficient analysis, we demonstrate the potential for significant sensor reduction without compromising monitorability. The proposed methodology offers a data-driven, scalable solution for sensor optimization that can be adapted across various industries, promising substantial cost savings and improved operational efficiency in pharmaceutical manufacturing and beyond. Future research directions include extending the approach to multi-stage production processes, incorporating sensor reliability and maintenance considerations, and exploring the integration of the optimized sensor network with advanced process control and fault detection techniques.