Practical Guidelines for Nonlinear Kalman Filters in Industrial Processes from Experimental Implementations

This study presents a practical implementation of nonlinear state estimators, namely Extended Kalman Filter and Unscented Kalman Filter, in industrial processes using the quadruple-tank system as a benchmark. The quadruple-tank system, known for its strong coupling and nonlinear dynamics, serves as a representative model for industrial applications. A physical prototype was constructed to assess the performance of the state estimators across various experimental scenarios, including both observable and non-observable configurations. The findings highlight that the Extended Kalman Filter consistently outperforms the Unscented Kalman Filter in estimation accuracy under non-observable setups, achieving mean absolute percentage errors (MAPE) as low as 12.4% with reduced sensor configurations. Practical recommendations are provided to optimize filter design, such as strategies for minimizing sensor usage, selecting configurations to maximize observability reducing number of sensors installed, and tuning covariance matrices to improve estimation robustness. The results also emphasize the critical role of the model's accuracy and its impact on the Unscented Kalman Filter, which proved more sensitive to modeling errors. This sensitivity underlines the importance of system-specific adjustments when applying nonlinear filters in real-world settings. These insights offer valuable guidance for the effective deployment of Kalman filters in cost-sensitive industrial environments, improving monitoring and control capabilities without compromising performance.