(421c) Multi - Agent Based Process Supervisory System for Monitoring Large Scale Chemical Processes

Fault Detection and Identification in chemical process industries has been an active area of research for over three decades. Early and precise detection of process faults is essential to prevent off-spec products and also in many cases to prevent explosions and other serious accidents. The focus from academia has by and large been on developing a single monolithic monitoring strategy for process industries. But the sheer size and nature of chemical processes make the application of these monolithic strategies, which are computationally intensive, to real plants difficult. Moreover, modern sensors (Rosemount wireless sensors etc), instruments and equipments have some built-in diagnostic capabilities which are capable of determining faults within themselves, but current monolithic process monitoring strategies do no take advantage of such information. Also, a single method may not be capable of handling all the fault scenarios as each monitoring algorithm has its own inherent advantages and disadvantages. Improvement and robustness in monitoring chemical processes could be achieved by using multiple methods and by combining their results in a meaningful manner.

Faults in chemical process industries could occur at the process level, section level or more often at the equipment level. Attempting to monitor the overall process for identifying such instrument and equipment level failures may be futile as deviations in the process are often lagging indicators by which time the plant safety may be compromised. Hence multiple FDI agents capable of monitoring the plant at various sections and varying levels of granularity (tag level to unit level) are developed. For instance, a FDI agent is developed for monitoring the overall process, another for compressor units (equipment level), another agent which utilizes results from instrument's built-in diagnostic capabilities (sensor level), and so on. When multiple FDI agents are used they need to effectively interact with one another, hence an Ontology is developed as described in some of our previous work (Natarajan and Srinivasan, 2010).

A key issue in monitoring any complex system using multiple independent methods in parallel is that the individual methods may not always concur and are often in conflict with each other. A suitable decision-fusion/consolidator agent is used which performs the matching between the results from the various FDI agents by mapping their evidences to the process and fault ontologies and seeking coherence among the various evidences.

This multi agent system is implemented into the ENCORE multi agent architecture developed in our prior work (Natarajan and Srinivasan, 2010). These agents are compliant with the FIPA (Foundation of Intelligent Physical Agents) specifications, which enables inter-operability among agents. The efficacy of this system is demonstrated on the Tennessee Eastman case study and an industrial offshore oil and gas production platform.

References

Ng Y. S., and Srinivasan R., (2007). Multi-agent Framework for Fault Detection and Diagnosis in Transient Operations, Presented at the European Symposium on Computer Aided Process Engineering, Bucharest, Romania, Paper # T1-474.

Natarajan S., and Srinivasan R., (2010), A Distributed Intelligence System for Improving Fault Diagnostic Performance in Large Scale Chemical Processes, ESCAPE ? 20.