Plant-wide processes monitoring and fault tracing based on causal graphical model

Plant-wide processes are usually characterized by large scale, multiple operating units and complicated interactions. Effective monitoring for such processes is imperative and challenging. Traditional data-driven methods have some limitations due to the neglect of internal relationships between operating units. This paper proposes a plant-wide monitoring and diagnostic framework based on the multi-variate statistical analysis and causal graphical inference. Initially, the optimized process decomposition is performed by combining the mechanistic knowledge and historical data from the perspective of improving the monitoring performance. Taking into account the fact that shared variables among the different subsystems lead to the information interaction rather than being independent as in the existing methods, the multi-variate causal model based on probability density estimation is established to identify the quantitative association of the process variables in a single subsystem. The complete model is structured by the link of shared variables. Finally, system anomalies are detected by changes in the probability density of the observed variables; the root cause is pinpointed by the causal inference. Experiments with the Tenessee Eastman (TE) process and Panamax bulk carriers demonstrate the applicability of the proposed methodology. A plant-wide monitoring and diagnostic framework based on the multi-variate statistical analysis and causal graphical inference is proposed. Optimized process decomposition is performed by combining the mechanistic knowledge and historical data from the perspective of improving the monitoring performance. Distributed fault diagnosis and tracing algorithm is implemented effectively in industrial process.image