Problems in process plants can be loosely categorized into two main types: predefined or repeating problems, and undefined or non-repeating ones. While both types cause loss of income and unnecessary outlay, undefined problems have a far greater impact on revenue and productivity. Let’s look at how each type fits into the rubric of predictive analytics, and how Precognize’s SAM GUARD can predict and prevent both types of problems, generating clear return on investment. Our core competence is in identifying and solving the undefined, non-repeating problems; however SAM GUARD of course also has a solution for the defined problems, which are actually much easier to deal with.
What are Defined and Undefined Problems?
Predefined, repeating problems are when the problem is known. It’s something that has happened before and repeats fairly often. Repeating problems are common around process improvement, like aiming to maintain the golden batch.
Engineers can use Precognize’s SANDBOX to find out which tags are relevant for that problem, to spot the causes of the problem and find the parameters and monitor them if relevant. The SANDBOX can connect between seemingly unrelated parameters, identifying the root cause.
A bigger challenge lies in alerting plant engineers to non-repeating, undefined problems that arise seemingly out of nowhere, because there’s no historical data to prepare you for such events.
Undefined, Non-repeating Problems
Undefined problems are ones that do not repeat themselves, like most cases in the process industry. In these cases, the issue never occurs in the same place or in the same way twice, making it unexpected. Examples of undefined problems include equipment failures, small process deviations, and changes to the operating process because of equipment failure, malfunction, or abnormality.
In process plants, non-repeating problems occur more often than repeating problems, but they remain harder for most solutions to predict and analyze.
Here are several examples of non-repeating problems, and how SAM GUARD identified them and created new value for the plants:
Petrochemical Example: Partial Blockage in Heat Exchanger
At a petrochemical plant, SAM GUARD alerted due to substantial differences in temperature, combined with multiple parameters measured in the heat exchanger and other related parts of the plant. All of the measurements showed an accelerated fouling process, indicating a partial blockage in the heat exchanger which was the operational root cause.
As a result of this combined alert, the heat exchanger was cleaned, and the partial blockage was removed. It was determined that while the heat exchanger had been partially blocked it had led to higher energy costs and increased fluid velocity, causing potential erosion problems; if it had not been fixed in time, it would have eventually led to a plant shutdown and lost production.
Refinery Example: Preventing Major Pump Failure
At a refinery, SAM GUARD began sending alerts that indicated a potential pump failure. Indicators showed position tags of valve at 90%. The reformer tube temperature had dropped slowly to 750° from 785°; the low alarm threshold was set at 650° but SAM GUARD began sending alerts 100° prior to the set threshold. And the reformer draft opening dropped to 42% from 98%.
Thanks to the early alerts, which combined all indicators along with other aspects of the plant, measures were taken to avoid equipment failure and major emergency shutdown. By fixing this small problem ahead of time, a greater problem was eliminated that would have caused shutdown and lost production.
Solving Undefined Problems
Most predictive analytics solutions use machine learning (either supervised or unsupervised) to identify problems in process plants. Every plant has thousands, if not tens of thousands, of sensors, covering every asset. Only machine learning can make sense out of this onslaught of data; it is far too much for a human analyst to process.
However, supervised machine learning is dependent on historical examples, to “train” the model, and this is impossible for non-repeating problems. Without enough historical examples, supervised machine learning can’t be used, and by definition, non-repeating problems have few, if any, historical examples.
At the same time, unsupervised machine learning isn’t relevant, because the torrent of data is overwhelming. With so many data points and possible states, unsupervised machine learning cannot define any “normal” state to serve as a background against which to spot anomalies. From an unsupervised machine learning point of view, a standard petrochemical or chemical processing plant is made up entirely of anomalies. Also not very helpful.
Traditional predictive analytics solutions are of limited use in these situations, because the lack of defined tags and problems prevents the engineer from analyzing the problem, and the volume of data is too great for traditional solutions to spot patterns and come up with warnings ahead of time.
SAM GUARD Predicts Defined AND Undefined Problems
The good news is that there is a solution that can cope with undefined problems. Precognize SAM GUARD is unique in combining human domain knowledge with machine learning, enabling plant engineers to identify the causes and pinpoint solutions even for undefined, non-repeating problems. SAM GUARD draws on a plant domain model that takes into consideration the specific structure and behavior of the plant. This is rapidly created at the start of the engagement and it enhances the machine learning, guiding users to the right areas of the plant to identify potential issues that may never have occurred before.
It can be easier to predict and thus prevent the defined problems, but the undefined ones occur much more frequently and can be a massive headache to plant owners and operators.
Avoiding these unforeseen problems is the ultimate goal of every predictive analytics program in petrochemical and chemical plants and this adds much more value to the plant as a whole. With the help of SAM GUARD, plants can prevent these undefined, non-repeating problems.
In our upcoming webinar and a later blog post, we’ll talk about how to calculate ROI on both defined and undefined problems.