(362z) Root Cause Identification Using Cross-Correlation Weighted Lag in Chemical Plants

Modern industrial plants consist of a large number of connected control loops that ensure better product quality and high plant productivity. Plant performance degrades over time due to various factors like aggressive tuning of controllers, presence of static non-linearities in control loops such as stiction etc. which would introduce faults/disturbances in the process. These oscillatory/non-oscillatory disturbances will propagate to other variables and would affect the overall performance of the plant. Detecting the root cause for a fault in a large industrial plant from routine operating data is very difficult. Most of the methods rely on causal maps or signed digraphs developed for the process using information provided in flow sheets or some data-based metrics. As accurate causal map identification from data is near to impossible, these techniques result in inefficient predictions. Those methods that rely on prior knowledge of the process are time consuming or may need information that is not readily available. Some techniques, such as the spectral envelope method detect the root cause directly from data and do not rely on causal maps. However, it is only restricted to oscillatory faults and does not work for non-oscillatory faults.

In this article, we propose a simple purely data-driven approach based on cross-correlation weighted lag metric (Ï„-metric) for identification of root cause in a multivariate process with connected control loops. The proposed algorithm can detect the root cause for both oscillatory and non-oscillatory faults and is computationally fast. The method does not require any prior knowledge unlike most techniques available in the literature. The proposed technique was tested rigorously using synthetic datasets generated in MATLAB Simulink for various processes with interconnected control loops. The Ï„-metric correctly identified the root cause for 86.31% of non-oscillatory faults and 89.96% of oscillatory faults showing an overall accuracy of 88.45%. The metric is also easy to calculate as it only requires the computation of variance and cross-correlation. The technique was tested on various industrial case studies as well and was found to perform better and faster than the existing techniques in literature.