Most industrial plants are already highly automated. Programmable controllers, sensors, computers have been common for decades and robots assemble products. But operative leaders naturally need to look ahead. We have automated our processes but what will the next step be? How to make our production line even more efficient? Could the data we collect be the answer?
Terms like artificial intelligence, machine learning and neural networks are hot buzzwords today. Many would probably go for this if asked what the next revolution in industry automation will be. But is that just riding the hype wave, or is it insightful? It could be both. AI is a large umbrella of different technologies that are both widely hyped and poorly understood. But there’s definitively a huge potential here. Computers can often outperform humans, especially when the problem involves large amounts of data in many dimensions. And that’s not uncommon in industrial environments.
So how to tell hype from opportunity in the industrial automation sector? One important fundamental that we at Etteplan want to uphold is to be problem-centric. To talk more about what you want to achieve and less about the tools. That’s why we prefer to talk about analytics, that naturally can be built using state-of-the-art artificial intelligence. It means utilizing the data you already collect to make smarter decisions, improve quality and avoid downtime. Or as Jan Müller, Director of Business Development at Etteplan, puts it: “Automation systems tells you what’s happening. Analytics systems tells you why, or what’s going to happen, which is invaluable when optimizing the process.”
Automation systems tells you what’s happening. Analytics systems tells you why something is happening, or what is going to happen.
What would analytics mean in practice?
No, it’s not a product you can purchase, drop into the system and expect immediate results. Effective analytics is a continuous process. It’s tightly integrated with your production process and uses process-specific data to solve process-specific problems. Jan Müller continues: “You need three things to get valuable analytics - understanding of the process itself, understanding of the suitable analytic methods, and seamless cooperation between those two competences.” So, should industrial companies start investing in machine learning competence? No, that’s the hard and slow way. Etteplan is happy to bring competence number two and three to the table. And even “Analytics as a Service”, when feasible.
Ok, how to implement analytics for our production line? First, we must ask the fundamental question: Why? There are several possible answers. “We have to use [insert favorite buzzword here] because all competitors are doing it and we fall behind otherwise.” Wrong. That’s not focused, and you are unlikely to achieve any valuable results. But what about: “Let’s dig deeper into that exiting pile of data we have, there must be something valuable in there!” Still wrong and unfocused, even if there probably is hidden value to be discovered.
What are the right questions?
“How can we predict breakdowns of this machine and replace parts as scheduled maintenance instead?” Yes, the ultimate goal for maintenance managers. Or: “These 13 parameters all affect product quality. How can we use them to detect situations leading to quality failures, and take preventive actions?” It’s easy to tell how these questions differ from the wrong ones. They focus on a concrete problem, not on tools or technology.
Additional things to consider
The process owner is of course in a key role when defining the problem, but the analytics expert can add a lot of value too. It’s very important to understand the capabilities and limitations of the analytical methods, to be able to set the bar high enough, but not too high. This is one of our key competences.
The next step is to select the suitable methods. And of course, do we have all the data we need? Sometimes more sensors or data dimensions may be needed to achieve the desired accuracy. Also think out of the box, data from outside of the factory may be useful. Could we bring in data from the administrative systems, or external open data?
Expectation management is an important issue to consider. There are three main points one must be aware of. First, systems of this kind require a learning cycle. The machine can’t do anything unless we teach it how to do it. This process is based on analyzing large data volumes and tuning the system’s internal decision-making parameters. It may be a quick task if you have a backlog of useful data, or a time-consuming phase if you need to collect the data first. The other thing slowing down deployment is that the analytics process often needs to be refined before the goals are met. Industrial processes are all different, and there’s usually no standard analytics practice that you just drop in. This makes the effort an iterative process.
Finally, it’s important to understand how inputs from different sources affect the result, and what can be expected with the available data. That’s individual for every case. A realistic evaluation of the achievable and required accuracy is of course essential.
The world is in the middle of a new revolution, often called Industry 4.0. Every company becomes part of a global digital web of companies. You not only tie business relationships to entities you cooperate with. Your computerized systems are also connected, enabling you to become a part of a global business machinery. Large amounts of data are available to us and fast decision-making based on this data is critical. Can one survive without modern computerized analytics in this kind of environment?
“We have access to a constantly increasing amount of data, both from our own automation and the business ecosystem. But this data is often underutilized. Deeper analysis is performed by humans, even in cases where modern computer algorithms could do a better job.”