Plant InSites: Is AI Actually Ready for Process Troubleshooting?

Troubleshooting is built on experience, but much of that tribal knowledge is walking out the door. As industry veterans retire, chemical process operations are losing institutional memory. Training less-experienced workers can help, but it doesn’t replace hands-on know-how.

Over the past few decades, many of my clients have talked about using “expert systems” software to save institutional memory. These were early forms of artificial intelligence that fell out of favor because they were inflexible, required labor-intensive maintenance and coding. Artificial intelligence (AI) is the latest successor to the expert-system trend. Will AI succeed where past expert systems could not?

One thought about AI is that it’s finally an expert system that allows for capturing and saving institutional knowledge, a key capability for training less-experienced staff. There are many types of AI approaches with various models, and these approaches impact performance differently. Even so, they share some common features. They are all built upon data and observation. Many chemical process plants are collecting large volumes of plant data, storing it and mining it for conclusions about how the plant operates. This big data is the foundation for modern AI systems.

The largest AI models use billions of variables coupled with powerful statistical analysis to generate recommendations. One way to think about this is that big data represents plant experience. Then, the AI tool takes the big data and converts it into observations about what works and what might be wrong in a plant that’s not working. One challenge in building confidence in these models is that with enough variables, almost anything can be correlated. With a billion variables, one could curve-fit a plate of spaghetti — the equations may look impressive, but the relationships they describe are essentially meaningless.

In identifying problems, the challenge is that the plant is not operating in its normal range when a problem occurs. Plant operators change the plant operation or fix equipment. This means that only comparatively limited data is available for plant operations for any specific problem. For troubleshooting, the plant’s big data is always limited. The problem data is always at the edge of the information cloud. This creates a structural challenge. Troubleshooting the models is least reliable when they are most needed. Today, the models have use, but their recommendations should be treated with caution.

It’s unfortunate, but true for decades, that users tend to accept whatever results come from a computer. This applies to AI results even more than relatively straightforward calculations. “Hallucination” has become the polite term for “stuff the AI made up that is pure fantasy.” This is particularly risky when using AI to create legal filings. As I’m writing this column, one database is closing in on 1,400 identified cases with hallucinations. The real number is likely much higher than this.

Similar risks apply to AI systems when used for troubleshooting. Current AI solutions do not replace engineering analysis. Even so, plants are using them for identify critical issues in their operations. To minimize potential errors, these plants should take the following precautions:

1.  Ask for an explanation of every conclusion, and check each document that AI cites by going back to the original sources.
2. Verify every conclusion based on fundamentals of chemistry and physics (i.e. mass and energy balance are mandatory).
3. Is everything you think you see explained by the diagnosis given?
4. Do you see everything you should expect to see if the diagnosis given is true?
5. Log in as a different user, exclude the current explanation and ask for an alternate explanation. Repeat questions 1-4 and see which fits better.
6. Get the result reviewed by disinterested people with experience with the process and equipment involved.
7. Quantify the results of implementing a fix based on the AI results and that being wrong.

Review the results from all the steps here and decide if you are going to accept the AI-generated results or not. This is no guarantee the steps here will identify every problem. But they will find a lot of them and show prudent engineering judgment was being used.