To implement AI in the production process, a number of things must be met. The good news: this is largely consistent with what is needed for effective Business Intelligence. In practice, therefore, we see that a strong business intelligence environment is the most important stepping stone to get started with AI.
The BI environment as a stepping stone to AI
AI models are only as good as the data on which they are trained. Before a model is suitable for use in production, it must first be developed, tested and validated on a historical dataset. The richer and more complete this historical data set, the more information the model can learn about the part of the production process being focused on. Placing the data in the right context – in this case, the piece of the production process being focused on – is often a challenge in practice. It is essential that historical data be of good quality, accurately reflect reality and preferably easily linked to the right context.
AI development environment
The image below shows the components of a development environment for AI. These components are explained below the image.
Data analytisch platform
In the operational practices of most manufacturing companies, data already plays an important role, but this data is often in isolated systems. Consider ERP systems, the MES and the Historian. To get the maximum value from AI, the very relationships between all these data sources are important. Bringing these various data sources together in a well-structured way in a Data Analytic Platform is essential.
All the data together: Ingestion
The first priority in setting up the Data platform is to determine the relevant data sources and collect the data from the various data sources. In this step, all relevant systems are connected and their data are unlocked in raw form to the Data platform. To illustrate; time series data comes from the Historian. Order, Equipment, and batch data comes from MES and information about the quality of raw materials used comes from MES, LIMS, or ERP.
Aggregate and combine into insight
Once all the data is unlocked in raw form to the data platform, the next task is to get a grip on the meaning of the data. This requires domain experts to collaborate with Data engineers to work from a raw data source, in several steps, to cleaned and aggregated data, suitable for analysis. Traditionally, this is often done on an ad hoc basis in various BI applications and dashboards. This often leads to divergent definitions and inconsistent understandings. The data platform is the appropriate place to centralize and manage these definitions.
The step to AI: training a model
Although the steps named so far are not yet specifically related to AI, they create the preconditions of a reliable and analyzable dataset. With this, starting with AI is only a small step. A simple statistical correlation between measured values can often be the first step in starting to make predictions. Using the data platform, these correlations can be validated on historical data, and verified with domain experts.
”AI always sounds very complex, but when the data is in good order, and the business context is well mapped, most of the complexity is actually behind it. A very large part of the challenge is often on the front end’
Luuk Schagen – AI Engineer at Enjins
AI in Manufacturing
The above steps create the preconditions for implementing AI in the factory. A well-trained and validated model in a development environment can be packaged and transferred to production. The version of the AI model ported to production does not learn independently. The goal of the first version of the AI model is to have proven (manageable and) stable behavior. In the expert-in-the-loop situation (see Expert in the loop versus expert on the loop), among other things, the final choice of the expert will be used to make the next version of the AI model better.
Depending on the product being produced, a different AI model must be used or the AI model is fed different set points. Also, the AI model must be fed with values of current variables, think of; Analysis results, process conditions, environmental conditions and weather forecasts. The software can quickly and reliably exchange data with level 2 control systems (PLC, DCS, SCADA). See below for clarification.
Specific challenges for AI in manufacturing
Using AI is a challenge for most industries. Manufacturing has a few additional specific challenges.
Aggregate and combine into insight
This involves combining the data from different systems into a whole. A scheduled start time in a production order from ERP does not yet say anything about the actual start time of a production order. In fact, a production order may consist of several operations; first A must be created, A is a raw material of B, and B is dried on the drying tower. When the context of an operation is missing, combining and aggregating the data becomes a challenge.
Many manufacturing companies plan their production process based on 1 or 2 production orders, even though the production process often consists of many more processing steps. Of these steps, there is often no order context known in IT systems, but only on production forms.
Jens Hensen – Consultant at Greywise Consultancy
Expert on the loop
Having the AI autonomously perform part of the production process is sensitive. This could compromise safety. In addition, knowledge is lost because operators will eventually be unable to follow the decisions of the AI model anyway. Before starting a project, therefore, it should be considered whether an expert-on-the-loop situation will ever be accepted, because the expert-in-the-loop always causes a delay in the desired adjustment and therefore probably does not ensure the most optimal result
Reliability and latency
Short interruptions to the production process can be costly. It is therefore important to build in fallbacks so that production can continue when the AI model encounters problems. Thus, there must always be an option to disable the AI and there must be insight into the IT performance of the AI model, for example, it quickly becomes clear that and why the AI model is slow to respond (latency).
It is also important to implement sanity checks and business rules. These ensure that the system can deal with outliers in the data or faulty sensors. Taking these measures will ensure that the system functions reliably.
Well-functioning IT, checks and business rules, combined with high reliability of the AI model are prerequisites for moving to the Expert-on-the-loop situation.
Cybersecurity plays an important role in manufacturing companies. Over the years, manufacturing companies have become increasingly dependent on IT systems and, as a result, the importance of cyber security is increasing. Manufacturing companies prefer to minimize the number of lines to the outside world.
At the same time, we are seeing a big increase in the cloud. The cloud is also relevant to AI. As already mentioned, the AI model can be developed in the cloud. Using software in the cloud is flexible, cost-effective and reliable. A long-term risk may be that too much business-critical information will end up in the cloud, but starting small can negate this argument.
However, the AI software running the AI model will have to run locally, as it must be able to respond quickly and reliably. New versions of the AI model are transferred from the cloud to the factory AI software at low frequency.
However, the software running the AI model is not yet standardized. There is no clear best-practice known in the industry. Companies’ cybersecurity requirements make standard solutions difficult and allow for customization that varies by manufacturing company. Still, a large gap can be filled by unified, well-thought-out software for running the AI model locally.
More interesting articles on Artificial Intelligence (AI)
About Greywise Consultancy
Greywise helps manufacturing companies accelerate the transition to the factory of the future.
This starts at the strategic level and ends with successful implementation of manufacturing-IT solutions. Consider MES, WMS, MOMS, Scheduling and QC applications.
Translating the technical concepts into impact on operations in clear non-technical language and creating an overview that includes the customer in the change is where Greywise’s strength lies.
An independent party between the supplier and the customer is a proven strategy for success.
Enjins develops strategies and implements machine learning solutions at innovative companies.
They help companies leverage their data and AI for smarter decision-making and automation. In addition to manufacturing companies, they help companies in different industries and at different levels of data and AI maturity.