When you hear AI or ML software can address your factory optimization initiatives, you may ask yourself: What can software do that engineers can’t? If I don’t have enough bandwidth, why don’t I just hire more people?
There are two reasons why this approach won’t yield value (besides the difficulties of hiring in the current economy). First is complexity—many industrial use cases are too complex for engineers to solve easily, or at all. Second is volume. A typical organization has far too many use cases to solve without increasing headcount massively, which is simply not practical for most industrial organizations.
To better understand the complexity of use cases, let’s consider all the problems in a plant that would benefit from optimization. These can be sorted into three tiers: simple data analytics problems; problems involving complex correlations across multiple assets or stages; and bigger problems requiring a black belt or higher. The more complex these problems become, the more potential value a solution can generate; however, tackling a complex, high-value problem also takes a lot of time and effort, which is why these typically get left on the back burner. That’s where software can play a valuable role and make your teams’ jobs easier.
Tier 1: Smaller improvement initiatives
These are day-to-day operating questions, usually single-asset focused.
Did I ever run my batch like this? What happened?
How do certain parameters improve my batch quality when I compare good and bad runs?
They can be solved in a simple way by domain experts, either using Excel or self-service data analytics solutions.
Tier 2: Complex improvement initiatives
These cross multiple assets or stages, or involve soft sensors requiring complex modeling.
Based on the previous crystallization step, should I change my drying time to achieve the desired moisture content?
How much longer is it profitable for me to run my process with this catalyst before regenerating it?
How can I approximate an offline lab measurement in real time? What factors in my process influence this measurement, and how can I change them to get the desired result?
Sometimes these problems can be solved by a Six Sigma green or black belt. However, they are often left on the back burner due to a lack of time and resources.
Hiring a data scientist, data engineer, and chemical engineer with data skills is one solution to address these complex problems. However, it’s impractical and time-consuming to build custom code for every use case. Also, data scientists are not process experts—so it’s also time-consuming for them to collaborate with engineers and operators from the production facility.
Maintenance of custom models is also an issue—they need to be tested and retrained continually in order to ensure predictive accuracy remains high. And even with the best-trained models, process changes or variability in product mixes can mean the underlying assumptions are off. These are two of the main reasons why many industrial machine learning projects fail to realize ROI.
Tier 3: High-impact, high-complexity initiatives
These are very specific problems that can’t be solved using “out of the box” methods.
How can I optimize my reaction kinetics to improve yield in a tightly controlled environment?
How can I optimize the operation of my distillation column?
If you just think about these examples, various problems can arise:
- You could be missing crucial kinetic/thermodynamic parameters not measured in the lab
- If there is little variation in the data, purely data-driven methods will not produce good results
Therefore, you often need to opt for a purely specific physical model or combine a physical model with an individual data model to truly represent reality.
Tackling such problems typically requires deep expertise in either physical or data-driven modeling. In Six Sigma terms, those would be black belt projects, or even master black belt projects. Such projects take a lot of time, interactions with production personnel, and manpower. The project cost can quickly add up to millions of dollars, not to mention you must think about total cost of ownership as the solution has to be maintained. Usually organizations find themselves dealing with bottlenecks in availability of highly trained experts as well as production personnel.
How factory optimization software helps
Tier 1 initiatives would not benefit much from process optimization software. In these cases, simple analytics tools or statistics software like JMP or Minitab should be the first tools.
However, solving Tier 2 problems with software will have a dramatic effect on your process. In this case, your organization will benefit from putting an AI tool in the hands of domain experts. This will enable them to tackle those issues in relatively little time, freeing up data resources to be applied to more specialized projects. Data scientists no longer have to work on these problems, as process engineers can do that instead. And if they do end up working on them, they can solve them faster and more efficiently because the software helps with topics including data preparation, collaboration with engineers, and deployments.
When it comes to Tier 3, traditional analytics also can’t help—you need a solution that can work quickly and spot correlations that humans can’t. Investing in software to solve such a use case will turn out to be far more effective than hiring engineers. Moreover, you’ll transform your production management by enabling your personnel to be proactive rather than reactive, which will have long-term benefits far beyond this single use case.
Beyond the complexity of individual use cases, the sheer volume of use cases is another reason why adding headcount is an inefficient solution. A single optimization project can range from weeks to months; more complex optimization tasks take as long as a year. All of these require substantial work from an engineer and possibly data scientist, as well as many interactions with plant personnel. To optimize across all of a plant’s use cases, therefore, would require scaling your headcount massively.
Software, in contrast, requires a comparatively small investment. Armed with the typical machine learning software solution, your existing engineers can quickly optimize multiple use cases. If a solution works, it can quickly be deployed in other areas of the plant and even other plants belonging to the same manufacturer, creating a more efficient approach to production.