(178b) Data-Driven Root Cause Analysis for Event Investigation Followed By Real-Time Monitoring and Prevention (extended version) | AIChE

(178b) Data-Driven Root Cause Analysis for Event Investigation Followed By Real-Time Monitoring and Prevention (extended version)


Noskov, M. - Presenter, Aspen Technology
Harmse, M. - Presenter, Aspen Technology, Inc.
Rao, A., Aspen Technology
Modern chemical process units have several thousand sensors installed that stream data into the Distributed Control Systems (DCS). The sensor data is then accumulated in historian databases spanning multiple years. Aspen Technology has recently introduced powerful new software that allows for causality analysis based on data patterns that appear when unique process events occur. Without the loss of generality, we focus on negative events (process upsets, shutdowns or equipment malfunctions), but the alternatives can be studied as well, e.g. an exceptionally good batch, or a faster than average plant startup. The starting point of investigation is a dataset that consists of a sufficiently long segment of continuous historical data record with good temporal resolution (e.g. minute data for one year or more). We represent the dataset as a collection of synchronized time series, for the available historian tags, each representing a process sensor. While it is not the only scenario, a negative event is represented as a particular pattern within a single tag. Such a tag might track some KPI parameter such as product purity or power consumption, etc. A fast parallel algorithm analyzes all tags within the dataset. The number of tags might be in thousands. The result of highly parallel analysis is a model that contains causal precursor patterns to precede the negative event. Our algorithm provides a natural view on sensors (tags) and their dynamics and thus leads to a streamlined analysis of precursors from a user perspective. Post training, our root cause analysis model is available for the real-time deployment. In real-time deployment, given an observation of any combination of precursors, we compute the probability of event to occur at some time in future up to a maximum time horizon. We can then observe the propagation of high probabilities of the event at some distant future (at time horizon) towards present time. Various warnings based on probability levels at different time spans (towards present time) can be created to monitor and prevent negative events.