Control valve failure early warning based on a two-stage degradation model and deep neural networks

Control valves, serving as critical actuating components in process industries, are subjected to prolonged operation under harsh conditions including elevated temperatures, high pressures, and aggressive corrosive environments, rendering them susceptible to abrupt failures. Investigating early warning methodologies for control valve failures is of paramount importance for ensuring process safety and minimizing economic losses. To accurately characterize the degradation mechanism of control valves and achieve high-sensitivity failure forewarning, this paper proposes a warning framework driven by the synergy of physics-based modeling and data-driven intelligence. Grounded in the physical degradation mechanism, a two-stage coating–substrate degradation model is formulated. A constrained stochastic process is incorporated, and a Genetic Algorithm–Least Squares (GALS) hybrid strategy is employed for parameter calibration, enabling a unified representation of deterministic degradation trends and stochastic fluctuations. An enhanced DAMADICS simulation platform is utilized to generate full life-cycle degradation data. A dual-level feature extraction network, comprising a Long Short-Term Memory (LSTM) branch and a Multilayer Perceptron (MLP) branch, is designed to collaboratively capture static statistical features and dynamic temporal patterns. Integrated with a physics-constrained gradient anomaly detector, dual-source alarm probabilities are generated. The Dempster–Shafer (D-S) evidence theory is subsequently adopted for probability fusion to achieve graded failure warning. Experimental results demonstrate that the proposed framework attains an early warning lead time of 6052 steps with an accuracy of 97.2%, significantly outperforming conventional approaches. This work provides a reliable solution for predictive maintenance of control valves.