An empirical study of the added value of the sequential learning of model parameters to industrial system health monitoring

Health monitoring provides opportunities to improve industrial system safety and to reduce system operation and maintenance cost due to more effective condition-based actions. Among the various methods for health monitoring, model-based methods exhibit advantages over data-driven methods in terms of explainability. This advantage is particularly promising for safety-critical systems, for example, nuclear power plants. However, the performance of model-based methods is heavily dependent on the knowledge of the parameters in the model for a system of interest. This knowledge may be lacking or be inaccurate initially, which poses challenges to applications of model-based health monitoring. Various methods for model parameter estimation, in particular, sequential learning, have been proposed in the literature. This research aims to investigate the added value of sequential learning of model parameters to industrial system health monitoring. Case studies based on solenoid-operated valve degradation are performed to illustrate such added values. Results based on synthetic data and experimental data demonstrate that, by considering the sequential learning scheme, the health monitoring accuracy is improved and in certain situations the uncertainty in the health monitoring result is reduced, compared to cases where the sequential learning scheme is not considered.