A hybrid Kalman filter and physics-informed neural network approach for leakage detection and localization in heat exchanger networks

Abstract

Leakage in heat exchanger network poses a critical latent threat to energy efficiency and process safety in industrial operations. This paper presents an integrated detection framework that synergistically combines Kalman Filter with physics-informed neural networks enable real-time detection and localization of leakage events. Kalman Filter is employed to preprocess noisy sensor data and accurately estimate key parameters, most notably the heat transfer coefficient, which is highly sensitive to leakage-induced deviations. These refined estimates serve as inputs for physics-informed neural networks, whose training is constrained by fundamental physical laws, enhancing fault detection accuracy. Validation via extensive simulations and experimental case studies demonstrates that the proposed framework reliably detects leakage flows as low as 1%, with an average inference time of only 0.76ms per sample. Compared with benchmark models, the proposed framework reduces prediction RMSE by 7%–15% and increases F1-score by 3%–5%, while maintaining millisecond-level responsiveness suitable for industrial real-time monitoring and precisely localizes the affected unit within complex heat exchanger network configurations. The integration of advanced state estimation and physics-constrained learning offers a robust strategy for improving the reliability, safety, and energy efficiency of industrial heat exchanger systems.