(183g) Iterative Fault Isolation for Integrated Chemical Systems Based on Approximate Linear Model Inversion

Chemical process integration by material and energy recycle is a common practice in the chemical industry, but fault detection and isolation have been challenging due to the complexity of the integrated system. In this study, we use the approximate inversion of strictly proper linear time-invariant subsystems to isolate the errors on an integrated plant.

The framework involves structuralized state space representation of the integrated system, a dynamic simulation of each subsystem (which may include process units as well as operating valves), an algorithm that simultaneously recovers the state and input variables of each subsystem based on output, as well as an iterative estimation of state and disturbance dynamics. For a strictly proper system where the feedthrough D matrix is absent, this improved algorithm presents opportunities to provide insights of the dynamics when the overall plant transfer function, G, is uncontrollable, or when the disturbance transfer function, G_d, has a large norm. In addition, observable variables are reconciled by an optimization formulation whose objective function is minimizing the norm of the error vector between two or more linked subsystems, and its constraints are determined by physical operations (e.g. valve percent opening, etc.) This methodology is fundamentally different from the data-driven fault detection methods, which neglects the process knowledge. Moreover, it improves the previous methods by creating an essentially “ideal” control of the simpler sub-models instead of the entire plant.

Finally, a numerical simulation of a chemical process network demonstrates the effectiveness of the methodology. We hope to create a real-time structure such that a fault can be detected as soon as it occurs. In future work, such a structure can have decentralized controllers to avoid the error from propagating downstream in the system and restoring the system back to normal operation, through either an active control action or a passive signal correction.