AI is “new” and everywhere these days. The funny thing is that, at least for Industrial Applications, AI has been around for years–buried under words like “Optimization,” “Analytics,” or “Workflow.” The same characteristics applied to use case evaluation five years ago still apply and are relevant today.
Let’s examine each “must-have" and explain how it drives success when present and poses significant challenges when absent.
Always begin with the end. Why even start a project or use case unless it aligns with corporate objectives? A strong value proposition should demonstrate how an Industrial AI project will deliver tangible business benefits, such as cost savings, revenue growth, or efficiency improvements. Substantial, tangible economic benefit acts as a proxy for:
Every project needs a go-to person to solve intractable problems. Ideally, this person has the most to gain when the project succeeds, grants the use case visibility at the highest levels, and distributes the money necessary to support it.
Industrial AI projects require, at a minimum, IT resources, process engineers, and a project manager. Others, such as data scientists, work process automation engineers, etc., may also be involved. None of these people have extra cycles. So, to get them to focus on your use case, you need that senior-level executive sponsor and the knowledge that this is work worth doing.
Good use cases drive economic benefit at the outset–on the first line or piece of equipment. The best use cases identify similar processes and equipment for rapid, duplicable deployment across the company–driving benefits at scale.
Ask your team, “When this use case is solved, what will be different? Whose job will change? What processes change? Who has ownership over changes moving forward?” Change management is, and always has been, an important part of economic success. If there aren’t answers to these questions, take the time to find them. This gap is where proof of concepts dies.
This brings me to the next “must-have.”
Once you determine the economic benefit, you must tie it to industrial-level metrics and outcomes. To do so, you need to answer:
This information establishes a baseline and allows you to track progress toward your goals.
Like any analytical capability, AI is more accurate the more valuable data it can access when building its models. The more data (three months to over a year), the better it can account for factors such as seasonality, long-term recipe changes, capital equipment changeouts, etc.
Get data from as many sources as possible, even if you’re unsure it’s relevant. Well-functioning AI systems will keep what works and discard the rest. So when considering a use case, such as downtime, you will want to include not only process historian data but quality data, MES data, energy management system data, cost data, maintenance data, and more. The more data, the better.
The data should be of good quality. Don’t worry if it’s imperfect. Sound Industrial AI systems will have data contextualization, cleansing, and creation capabilities.
The data should have multiple instances at a high frequency. For example, multiple temperature readings a day.
You should insist on continuously streaming new data as it is created. Industrial AI is a learning system that will continue to provide value far into the future—so long as its data stays current.
When you have the data as described, you are allowing Industrial AI access to all possible factors contributing to your use case. When you don't, you risk missing the true root cause of your central problem or, worse, having the system make an invalid recommendation.
So now you have the data. Next is to ensure the project exhibits a few key characteristics.
Industrial AI systems need variation in order to identify contributing factors to anomalous performance. Without variation, examining different data elements and working the model is impossible. A lack of variation may indicate that the metric isn’t often measured enough.
For example, recently, I worked with a customer who wanted to reduce their dryer temperature. The dryer temperature was recorded only once per day despite multiple shifts and SKUs requiring dryer use during that same day. Does that mean the dryer temperature was constant? No. It meant it wasn’t being measured at a granular enough level.
Okay, now that you've developed your business case and examined the data for your pilot implementation of the use case. Let's move on to the next characteristic, which will deal with scale.
You want the biggest payback possible when implementing a use case using Industrial AI. So ask yourself:
With the right Industrial AI solution, you will find that the work involved in integration at the first site or line will be significantly reduced as you deploy across multiple similar processes or assets. Even better news? Your overall return on investment (ROI) will improve as you roll out.
Note: I do not recommend using cases that are a one-off. If you follow these guidelines, you are unlikely to get this far with one-offs. They typically don't have the economic feasibility to prove much value.
This part of the process comes last because change will likely happen once the other boxes are checked. Let’s be frank. You’re asking busy people to take on more, at least temporarily.
So, what can you do beyond what’s above to ensure success for your Industrial AI solution? Here are a few recommendations.
If you nail these, you’ll avoid pilot purgatory and deliver a scalable project with real economic benefit for your company.