(584b) Transfer Entropy-Based Dynamic Causal Maps Generation for Fault Diagnosis in Systems with Cycles

On-line fault detection and diagnosis in process industries has always been a challenge due to the large size and complexity of the plant resulting in enormous amount of potential interactions between variables. Causality represented in the form of a directed graph (digraph) is one of the most widely used tools for root cause analysis, once the presence of a fault is identified [1]. In conventional approaches of graph-based diagnosis, a digraph representing the actual desired process is developed and by tracing the path of variables/nodes showing deviation from desired behavior, root cause is identified. The performance of this approach depends on the accuracy of the causal map developed. Bauer et al. [2] proposed an algorithm using transfer entropy to detect direct causality and construct a causal map. This technique works well for systems that can be represented as directed acyclic graphs. In a chemical plant, however, feedback loops and recycling streams (e.g., refluxes) are common. Such cyclic effects make the problem of direct causality detection and root cause analysis harder.

In the proposed work, we use transfer entropy as a measure of causality between variables, as it is capable of capturing non-linear interactions between them. A weighted adjacency matrix derived from transfer entropy (based on process data) between each pair of variables is used to construct a weighted causal map. Onset of fault is accompanied by change in transfer entropy, which leads to change in the weighted causal graph identified. Tracking this change in the dynamic causal map serves as a useful metric for identification of faults, instead of relying on accurate estimation of the true map itself. In this work, we report on the performance of this approach on various simulated case studies, including the Tennessee Eastman process. The technique is found to satisfactorily locate the fault, either as a faulty cycle or a faulty node.

References

[1] V. Venkatasubramanian, R. Rengaswamy, and S. N. Kavuri, “A review of process fault detection and diagnosis part II: Qualitative models and search strategies,” Comput. Chem. Eng., vol. 27, no. 3, pp. 313–326, 2003.

[2] M. Bauer, J. W. Cox, M. H. Caveness, J. J. Downs, and N. F. Thornhill, “Finding the Direction of Disturbance Propagation in a Chemical Process Using Transfer Entropy,” IEEE Trans. Control Syst. Technol., vol. 15, no. 1, pp. 12–21, 2007.