So you now have a use case. (If you don’t, you can find one using our previous article). You have good enough data. You’re excited to take the next step and start optimizing your process using a machine learning solution. The only question is…which one?
These days, there are many software vendors offering industrial machine learning technology. In addition, many manufacturers prefer to build their own models, or are at least considering this method. It can be difficult to determine which solution will provide the most value, even before embarking on this already complex journey.
In this post, we’ll break down the process of selecting a machine learning solution so you can be assured of choosing the optimal solution for your plant. Before making a decision, there are 4 main questions to consider:
1. What action do you want to take?
Are you looking to use machine learning for planning operational policy, or optimizing processes in real time?
For the former, you simply need a ML solution that allows you to work with historical data by uploading a file and conducting an analysis or simulation. This is the solution you want if your goal is answering questions about a process in general, e.g. whether permanently raising the temperature by a certain number of degrees will improve yield. This type of solution can complement your typical Six Sigma workflow and be quickly applied by engineers to new use cases without calling IT.
However, the reality of a factory is that planning can only take you so far. Process conditions are constantly changing. Today’s raw material amount will likely be marginally different from yesterday’s, and other factors will be different too.
If you’re looking for a solution that can optimize processes in real time based on constantly changing conditions, you’ll need the capability to do live optimization. Let’s say you want to know if your steel will fail to meet mechanical criteria before it’s too late to regrade it, or when your heat exchanger will become too fouled to sustain the production rate or fulfill its cooling duty and need to be serviced. Those questions can only be answered with real-time machine learning.
This type of solution inherently requires more setup as it must be connected to IT systems. However, it will yield higher ROI, as recommendations based on real-time sensor information will inevitably be more accurate than those based on historical data. And if you have the chance to consistently increase the output of your process by operating at optimal levels, you can turn that single return into recurring value.
2. How much context do you need?
Machine learning solutions currently on the market can be classified as either black-box or white-box. As the name suggests, black-box methods give you recommendations without context, like Netflix suggesting what movie you should watch based on some mysterious algorithm. Conversely, white-box solutions provide a wealth of information to guide decision-making. With a white-box approach, users can understand why the software is making its recommendations and predictions and verify that those insights align with their existing process knowledge.
The specific information provided depends on the solution, but it can range from confidence intervals—how confident is the software that the recommended action will actually produce the desired outcome? What might the results look like?—to root cause breakdowns pinpointing the key factors causing an issue. You can even investigate counterfactuals and hypotheticals with some solutions, tackling questions such as "What would have happened if I produced yesterday’s batch at a higher temperature?" or "What if I produced tomorrow's at a lower temperature?" If trust is a major obstacle to adoption of AI and ML at your organization, as it is in many plants, such context can go a long way towards ensuring broader uptake of optimization solutions.
A black-box model won’t give you any of that information. However, in some cases, you don’t need it. You can still enjoy a Netflix movie without knowing why the algorithm recommended it to you; similarly, there are places in manufacturing where white-box models aren’t necessary. Computer vision is one such area. If you’re looking to do some kind of computer vision work—for instance, replacing a human monitoring a production line for surface defects with a camera—a black-box model will be sufficient. On the other hand, if you’re looking to gain a deeper understanding of a particular process and optimize it for one or multiple variables, you may want to consider a white-box solution where more context is provided.
3. Who will be working with the solution?
Another point to consider is who will be using it. Even for the same problem, different users may have different skill sets that let them solve it in different ways. So your chosen ML solution should take this into account.
Operators, for example, will need a solution that lets them take in new suggestions and run them on the fly, such as set point recommendations. Engineers, meanwhile, will be focusing on planning and optimization and will require a solution that helps them find new ways to operate and simulate different scenarios. And data scientists will need a tool that lets them explore new possibilities, such as how models can transfer over to similar process technology or different chemical types.
4. Will the solution connect directly to the factory or include a human in the loop?
With a human involved in production decisions, any recommendation made by the machine learning model can be verified and further analyzed before proceeding. While trust is still important, this method will ensure for a backup check on accuracy.
In some plants, however, the aim of a machine learning project may be to involve the software by directly connecting it to the factory so it can have an immediate impact on production. In these cases, it’s even more critical to have a solution with confidence intervals. That way, it can be set to only commit to certain actions if a basic confidence interval threshold is met, avoiding any situations where the machine learning recommendations create drastic errors.
With answers to these questions in hand, you should be able to select the machine learning solution that’s right for you. (If you’re weighing the pros of building internally vs. buying an external solution, check out our upcoming post). Once you have selected your solution and use case, the next step to think about is how to structure a pilot...coming soon!
